diff --git a/scripts/architecture-baseline.json b/scripts/architecture-baseline.json
index 3d1514d..a9325d5 100644
--- a/scripts/architecture-baseline.json
+++ b/scripts/architecture-baseline.json
@@ -2,10 +2,7 @@
  "version": 1,
  "violations": [
   "clone 003d2210c257 src/rsi/__tests__/controller-candidates.test.ts,src/rsi/__tests__/controller-fleet-results.test.ts",
-  "clone 00803486e758 src/engine/__tests__/loop-helpers-history.test.ts,src/engine/__tests__/loop-helpers-readonly.test.ts,src/engine/__tests__/parser.test.ts,src/engine/__tests__/summarize-tool-arg.test.ts,src/memory/__tests__/embed.test.ts,src/memory/__tests__/summarizer.test.ts",
   "clone 00815c1c8497 src/rsi/__tests__/controller-candidates.test.ts,src/rsi/__tests__/controller-integrity.test.ts",
-  "clone 0081a7d462ab src/orchestrator/__tests__/cost-history-cli.test.ts,src/orchestrator/__tests__/cost-history.test.ts",
-  "clone 0152e2318e4e src/engine/__tests__/loop-helpers-history.test.ts,src/engine/__tests__/loop-helpers-readonly.test.ts,src/engine/__tests__/new-task.test.ts,src/engine/__tests__/parser.test.ts,src/engine/__tests__/summarize-tool-arg.test.ts,src/engine/__tests__/switch-mode.test.ts",
   "clone 01a56a433e82 src/orchestrator/task-files.ts",
   "clone 01e5839d0892 src/permissions/commands.ts",
   "clone 0288db0ddefd src/orchestrator/cli.ts",
@@ -13,14 +10,12 @@
   "clone 0338cd3dd229 src/orchestrator/__tests__/cli.test.ts",
   "clone 03d1744fcd1c src/orchestrator/task-files.ts",
   "clone 04288103b140 src/dashboard/__tests__/server.test.ts,src/dashboard/__tests__/session-launch.test.ts",
-  "clone 044be972ef45 src/codesearch/__tests__/backend-selection.test.ts,src/codesearch/__tests__/index.test.ts,src/codesearch/__tests__/search.test.ts,src/config/__tests__/mode-models.test.ts,src/dashboard/__tests__/file-browser-permissions.test.ts,src/dashboard/__tests__/settings.test.ts,src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/model-resolution.test.ts,src/orchestrator/__tests__/resume.test.ts,src/orchestrator/__tests__/watch.test.ts,src/permissions/__tests__/protected-files.test.ts,src/tools/__tests__/codebase_search.test.ts",
   "clone 0502285714dd src/engine/__tests__/loop.test.ts",
   "clone 0547f20476cc src/rsi/__tests__/postgres-queue.test.ts",
   "clone 058c180b1b12 src/orchestrator/resume.ts",
   "clone 05d4573a34dd src/codesearch/embedder.ts,src/memory/embed.ts",
   "clone 05f9940a16d5 src/decision-proxy/cli.ts,src/orchestrator/cli.ts,src/orchestrator/resume.ts",
   "clone 0654c838248c src/llm/muse.ts,src/llm/openrouter.ts",
-  "clone 0665ce33abe0 src/engine/__tests__/loop-helpers-history.test.ts,src/engine/__tests__/loop-helpers-readonly.test.ts,src/engine/__tests__/parser.test.ts,src/engine/__tests__/summarize-tool-arg.test.ts",
   "clone 06b27f6a49e2 src/orchestrator/__tests__/analyze-cli.test.ts",
   "clone 06c2ac64ec1e src/engine/__tests__/condense.test.ts",
   "clone 070bf0935a24 src/engine/loop.ts",
@@ -32,11 +27,9 @@
   "clone 085a3dc0f084 src/orchestrator/__tests__/spawn-model.test.ts,src/orchestrator/__tests__/spawn-state-reset.test.ts",
   "clone 08693511f4f2 src/orchestrator/task-files.ts",
   "clone 08a6c01fdb2d src/orchestrator/__tests__/analyze-cli.test.ts",
-  "clone 08c495e8fc18 src/orchestrator/__tests__/cost-history-cli.test.ts,src/orchestrator/__tests__/cost-history.test.ts",
   "clone 093e5318134f src/codesearch/embedder.ts,src/memory/embed.ts",
   "clone 097bb86875ad src/llm/openrouter.ts",
   "clone 097dbebdafaf src/github/__tests__/installations.test.ts,src/github/__tests__/pr.test.ts",
-  "clone 0a607f8c651c src/orchestrator/__tests__/cost-history-cli.test.ts,src/orchestrator/__tests__/cost-history.test.ts",
   "clone 0a8cd67b23bf src/cloud/__tests__/openshell-provider.test.ts",
   "clone 0ab505c08776 src/engine/loop.ts",
   "clone 0aba39a8736e src/checkpoints/cli.ts,src/decision-proxy/cli.ts",
@@ -74,27 +67,20 @@
   "clone 12dd86c98dc3 src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/resume.test.ts",
   "clone 136eebd8cce3 src/dashboard/server.ts",
   "clone 1405529d1b27 src/engine/__tests__/condense.test.ts,src/engine/__tests__/events.test.ts",
-  "clone 1418c0249247 src/engine/__tests__/global-modes.test.ts,src/engine/__tests__/prompt.test.ts,src/engine/__tests__/stack-rules-content.test.ts,src/engine/__tests__/stacks.test.ts,src/llm/__tests__/muse.test.ts,src/llm/__tests__/openrouter.test.ts",
-  "clone 14ae7a9a5f1d src/codesearch/__tests__/backend-selection.test.ts,src/codesearch/__tests__/index.test.ts,src/codesearch/__tests__/search.test.ts,src/config/__tests__/mode-models.test.ts,src/dashboard/__tests__/file-browser-permissions.test.ts,src/dashboard/__tests__/settings.test.ts,src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/model-resolution.test.ts,src/orchestrator/__tests__/resume.test.ts,src/orchestrator/__tests__/spawn-model.test.ts,src/orchestrator/__tests__/watch.test.ts,src/permissions/__tests__/protected-files.test.ts,src/tools/__tests__/codebase_search.test.ts",
   "clone 14d7e81b5443 src/checkpoints/__tests__/cli.test.ts,src/codemap/__tests__/cli.test.ts,src/dashboard/__tests__/cli.test.ts",
   "clone 15440c983d57 src/checkpoints/cli.ts,src/decision-proxy/cli.ts",
-  "clone 157b83b06b3d src/engine/__tests__/loop-helpers-history.test.ts,src/engine/__tests__/loop-helpers-readonly.test.ts,src/engine/__tests__/parser.test.ts,src/engine/__tests__/summarize-tool-arg.test.ts,src/memory/__tests__/embed.test.ts,src/memory/__tests__/summarizer.test.ts",
   "clone 15d1e6928f43 src/dashboard/server.ts",
   "clone 16271d25ae1f src/llm/preflight.ts",
   "clone 162af6dd3851 src/decision-proxy/__tests__/proxy.test.ts,src/engine/__tests__/switch-mode.test.ts,src/tools/__tests__/codebase_search.test.ts,src/tools/__tests__/execute_command.test.ts",
   "clone 162cb5b72716 src/memory/__tests__/memory-wiring.test.ts,src/tools/__tests__/decision.test.ts",
-  "clone 168bc796174d src/codeintel/__tests__/registration.test.ts,src/tools/__tests__/batch.test.ts",
   "clone 16b63330e573 src/orchestrator/task-files.ts",
   "clone 16baec87b0bf src/orchestrator/task-files.ts",
-  "clone 17203871d0e8 src/orchestrator/__tests__/cost-history-cli.test.ts,src/orchestrator/__tests__/cost-history.test.ts",
   "clone 17924b477eb9 src/orchestrator/cli.ts",
   "clone 180e79a01e9e src/dashboard/__tests__/server.test.ts,src/dashboard/__tests__/session-launch.test.ts",
   "clone 18209827df5b src/llm/preflight.ts",
-  "clone 182bd02a4dda src/__tests__/cli-env.test.ts,src/codesearch/__tests__/embedder-types.test.ts,src/evidence/__tests__/verify-files.test.ts,src/orchestrator/__tests__/task-files.test.ts,src/orchestrator/__tests__/terminal-status.test.ts,src/tools/__tests__/batch-types.test.ts,src/tools/__tests__/paths.test.ts",
   "clone 1858558ef1ae src/engine/__tests__/new-task.test.ts,src/engine/__tests__/switch-mode.test.ts",
   "clone 1918e1837e10 src/orchestrator/task-files.ts",
   "clone 191b28b096c9 src/dashboard/server.ts",
-  "clone 19a8c1d20a2f src/codesearch/__tests__/backend-selection.test.ts,src/codesearch/__tests__/index.test.ts,src/codesearch/__tests__/search.test.ts,src/config/__tests__/mode-models.test.ts,src/dashboard/__tests__/file-browser-permissions.test.ts,src/dashboard/__tests__/settings.test.ts,src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/model-resolution.test.ts,src/orchestrator/__tests__/resume.test.ts,src/orchestrator/__tests__/watch.test.ts,src/permissions/__tests__/protected-files.test.ts,src/tools/__tests__/codebase_search.test.ts",
   "clone 19c4c40edbae src/engine/__tests__/loop.test.ts",
   "clone 19d774f4b47a src/engine/__tests__/new-task.test.ts,src/engine/__tests__/switch-mode.test.ts",
   "clone 19de83260e74 src/engine/__tests__/loop.test.ts",
@@ -107,7 +93,6 @@
   "clone 1b2b48c0199b src/orchestrator/cli.ts,src/watcher/cli.ts",
   "clone 1b404c3cbfae src/llm/__tests__/muse.test.ts,src/llm/__tests__/openrouter.test.ts",
   "clone 1b48355447df src/llm/__tests__/muse.test.ts,src/llm/__tests__/openrouter.test.ts",
-  "clone 1b759012c29a src/checkpoints/__tests__/cli.test.ts,src/codemap/__tests__/cli.test.ts,src/codemap/__tests__/units.test.ts,src/dashboard/__tests__/cli.test.ts,src/dashboard/__tests__/page-route.test.ts,src/dashboard/__tests__/self-improvement-metrics.test.ts,src/dashboard/__tests__/sigint-shutdown.test.ts,src/github/__tests__/installations.test.ts,src/orchestrator/__tests__/analyze-cli.test.ts,src/orchestrator/__tests__/cost-estimate.test.ts,src/orchestrator/__tests__/log-analysis.test.ts,src/orchestrator/__tests__/spawn-state-reset.test.ts,src/orchestrator/__tests__/status.test.ts,src/run-tests-script.test.ts,src/tools/__tests__/set-indentation.test.ts",
   "clone 1bac564acc3f src/engine/loop.ts",
   "clone 1c3546eef909 src/orchestrator/__tests__/cli.test.ts",
   "clone 1c4bbb9846bf src/orchestrator/task-files.ts",
@@ -115,10 +100,8 @@
   "clone 1c95522b17e5 src/dashboard/__tests__/projects-routes.test.ts,src/projects/__tests__/cli.test.ts",
   "clone 1cb8d4dbc25b src/codeintel/__tests__/call-graph.test.ts,src/codeintel/__tests__/outline.test.ts,src/codeintel/__tests__/registration.test.ts",
   "clone 1d93450ce46b src/dashboard/server.ts",
-  "clone 1daf4022de6f src/codeintel/__tests__/implementations.test.ts,src/orchestrator/__tests__/review-gates.test.ts",
   "clone 1e9f5c07bbb1 src/dashboard/__tests__/cost-history-routes.test.ts,src/orchestrator/__tests__/cost-history-cli.test.ts",
   "clone 1ea2ad860baf src/codesearch/files.ts,src/engine/stacks.ts",
-  "clone 1ec0943d340c src/engine/__tests__/global-modes.test.ts,src/engine/__tests__/stack-rules-content.test.ts",
   "clone 1ec65e3ed9d8 src/orchestrator/task-files.ts",
   "clone 1ed752dc9308 src/engine/local-explore.ts,src/llm/ollama.ts",
   "clone 1ee2355c21d5 src/llm/openrouter.ts",
@@ -135,24 +118,18 @@
   "clone 21fd3bfb944d src/dashboard/__tests__/cost-history-routes.test.ts,src/orchestrator/__tests__/cost-history-cli.test.ts",
   "clone 224a1f3d5365 src/orchestrator/resume.ts",
   "clone 22f7d06a0abd src/rsi/__tests__/controller-candidates.test.ts,src/rsi/__tests__/controller-fleet-results.test.ts",
-  "clone 230967798bac src/engine/__tests__/global-modes.test.ts,src/engine/__tests__/prompt.test.ts,src/engine/__tests__/stack-rules-content.test.ts,src/engine/__tests__/stacks.test.ts,src/llm/__tests__/muse.test.ts,src/llm/__tests__/openrouter.test.ts,src/stack-rules-content.test.ts",
   "clone 231bc945ff50 src/control/emergency-controller.ts",
   "clone 232f1b0a8918 src/engine/__tests__/loop.test.ts",
   "clone 2359b270d07f src/orchestrator/cli.ts",
-  "clone 23c11a917a3b src/codeintel/__tests__/implementations.test.ts,src/orchestrator/__tests__/review-gates.test.ts",
-  "clone 23ccd9c01538 src/__tests__/cli-env.test.ts,src/codesearch/__tests__/embedder-types.test.ts,src/evidence/__tests__/verify-files.test.ts,src/orchestrator/__tests__/task-files.test.ts,src/orchestrator/__tests__/terminal-status.test.ts,src/tools/__tests__/batch-types.test.ts,src/tools/__tests__/paths.test.ts",
   "clone 243c8df9399b src/engine/__tests__/switch-mode.test.ts",
   "clone 244ddb70203e src/dashboard/server.ts",
   "clone 245b47905e93 src/tools/__tests__/edit_file.test.ts,src/tools/__tests__/search_replace.test.ts",
   "clone 24948eb63643 src/checkpoints/__tests__/cli.test.ts,src/codemap/__tests__/cli.test.ts,src/dashboard/__tests__/cli.test.ts",
   "clone 24a65346f0d9 src/orchestrator/task-files.ts",
-  "clone 252a9b7735d0 src/codeintel/__tests__/implementations.test.ts,src/orchestrator/__tests__/review-gates.test.ts",
   "clone 25a39d706027 src/orchestrator/__tests__/cli.test.ts",
   "clone 25d503f4631a src/orchestrator/__tests__/reviewer.test.ts,src/qa/__tests__/qa.test.ts",
   "clone 2603e91a093c src/llm/preflight.ts",
   "clone 262737545dd9 src/cloud/docker-provider.ts,src/cloud/provider.ts",
-  "clone 2691a8885b90 src/orchestrator/__tests__/cost-history-cli.test.ts,src/orchestrator/__tests__/cost-history.test.ts",
-  "clone 271ce8cbe5bc src/codeintel/__tests__/registration.test.ts,src/tools/__tests__/batch.test.ts",
   "clone 27804cea573f src/codesearch/__tests__/airunner-embedder.test.ts,src/codesearch/__tests__/ollama-embedder.test.ts",
   "clone 27f1585b3e2a src/orchestrator/cli.ts",
   "clone 27f5663efaad src/orchestrator/task-files.ts",
@@ -175,11 +152,9 @@
   "clone 2c9f8493e38e src/llm/muse.ts,src/llm/openrouter.ts",
   "clone 2cb1493c835d src/orchestrator/__tests__/main-tree-tripwire.test.ts,src/orchestrator/__tests__/review-gates.test.ts",
   "clone 2e4f5c004508 src/init/__tests__/cli.test.ts,src/orchestrator/__tests__/cost-history-cli.test.ts",
-  "clone 2eefefd492bd src/checkpoints/__tests__/cli.test.ts,src/codemap/__tests__/cli.test.ts,src/codemap/__tests__/units.test.ts,src/dashboard/__tests__/cli.test.ts,src/dashboard/__tests__/page-route.test.ts,src/dashboard/__tests__/self-improvement-metrics.test.ts,src/dashboard/__tests__/sigint-shutdown.test.ts,src/github/__tests__/installations.test.ts,src/orchestrator/__tests__/analyze-cli.test.ts,src/orchestrator/__tests__/cost-estimate.test.ts,src/orchestrator/__tests__/log-analysis.test.ts,src/orchestrator/__tests__/spawn-state-reset.test.ts,src/orchestrator/__tests__/status.test.ts,src/run-tests-script.test.ts,src/tools/__tests__/set-indentation.test.ts",
   "clone 2ef48d0cbf0f src/engine/__tests__/new-task.test.ts,src/engine/__tests__/switch-mode.test.ts",
   "clone 2efe67471984 src/orchestrator/cli.ts",
   "clone 2f137ce3a1d0 src/llm/muse.ts,src/llm/openrouter.ts",
-  "clone 2f29adde55ae src/engine/__tests__/global-modes.test.ts,src/engine/__tests__/prompt.test.ts,src/engine/__tests__/stack-rules-content.test.ts,src/engine/__tests__/stacks.test.ts,src/llm/__tests__/muse.test.ts,src/llm/__tests__/openrouter.test.ts",
   "clone 2f4521043a87 src/engine/__tests__/condense.test.ts",
   "clone 2fc443c44382 src/orchestrator/cli.ts",
   "clone 30244c9090b3 src/checkpoints/cli.ts,src/cli.ts,src/codemap/cli.ts,src/codesearch/cli.ts,src/dashboard/cli.ts,src/dashboard/trend-cli.ts,src/decision-proxy/cli.ts,src/github/cli.ts,src/init/cli.ts,src/orchestrator/cleanup.ts,src/watcher/cli.ts",
@@ -192,7 +167,6 @@
   "clone 3319637fc066 src/engine/local-explore.ts,src/llm/ollama.ts",
   "clone 3341cac7ce47 src/orchestrator/task-files.ts",
   "clone 338cc9087bf2 src/dashboard/__tests__/cost-history-routes.test.ts,src/orchestrator/__tests__/cost-history-cli.test.ts,src/orchestrator/__tests__/cost-history.test.ts",
-  "clone 3390241b01d9 src/budget/__tests__/budget.test.ts,src/budget/__tests__/concurrency.test.ts,src/budget/__tests__/cost.test.ts,src/checkpoints/__tests__/service.test.ts,src/cloud/__tests__/provider.test.ts,src/codeintel/__tests__/call-graph.test.ts,src/codeintel/__tests__/call-hierarchy.test.ts,src/codeintel/__tests__/find-references.test.ts,src/codeintel/__tests__/go-to-definition.test.ts,src/codeintel/__tests__/hover.test.ts,src/codeintel/__tests__/import-graph.test.ts,src/codeintel/__tests__/outline.test.ts,src/codeintel/__tests__/rename-symbol.test.ts,src/codemap/__tests__/build.test.ts,src/codesearch/__tests__/airunner-embedder.test.ts,src/codesearch/__tests__/chunk.test.ts,src/codesearch/__tests__/files.test.ts,src/codesearch/__tests__/ollama-embedder.test.ts,src/dashboard/__tests__/aggregate.test.ts,src/dashboard/__tests__/chat-thread.test.ts,src/dashboard/__tests__/checkpoints-routes.test.ts,src/dashboard/__tests__/codemap-routes.test.ts,src/dashboard/__tests__/cost-history-routes.test.ts,src/dashboard/__tests__/events.test.ts,src/dashboard/__tests__/projects-routes.test.ts,src/dashboard/__tests__/server.test.ts,src/dashboard/__tests__/session-launch.test.ts,src/dashboard/__tests__/timeline.test.ts,src/deploy/__tests__/gate.test.ts,src/engine/__tests__/condense-async.test.ts,src/engine/__tests__/condense.test.ts,src/engine/__tests__/events.test.ts,src/engine/__tests__/inject-message.test.ts,src/engine/__tests__/local-explore.test.ts,src/engine/__tests__/loop.test.ts,src/engine/__tests__/new-task.test.ts,src/engine/__tests__/pause.test.ts,src/engine/__tests__/switch-mode.test.ts,src/github/__tests__/app-auth.test.ts,src/github/__tests__/cli.test.ts,src/github/__tests__/pr.test.ts,src/github/__tests__/provision.test.ts,src/github/__tests__/push.test.ts,src/index-util.test.ts,src/init/__tests__/cli.test.ts,src/init/__tests__/gitignore.test.ts,src/memory/__tests__/local.test.ts,src/memory/__tests__/memory-wiring.test.ts,src/memory/__tests__/uwuchat.test.ts,src/migrate/cli.test.ts,src/orchestrator/__tests__/arg-cursor.test.ts,src/orchestrator/__tests__/auto-split.test.ts,src/orchestrator/__tests__/cleanup.test.ts,src/orchestrator/__tests__/git-sync.test.ts,src/orchestrator/__tests__/merge-check.test.ts,src/orchestrator/__tests__/pipeline.test.ts,src/orchestrator/__tests__/reviewer.test.ts,src/orchestrator/__tests__/split.test.ts,src/orchestrator/__tests__/state.test.ts,src/orchestrator/__tests__/status-cli.test.ts,src/orchestrator/__tests__/stop-cli.test.ts,src/permissions/__tests__/commands.test.ts,src/permissions/__tests__/store-protection.test.ts,src/project-store.test.ts,src/projects/__tests__/cli.test.ts,src/qa/__tests__/qa.test.ts,src/tools/__tests__/apply_diff.test.ts,src/tools/__tests__/browser-action.test.ts,src/tools/__tests__/decision.test.ts,src/tools/__tests__/describe_image.test.ts,src/tools/__tests__/edit_file.test.ts,src/tools/__tests__/execute_command.test.ts,src/tools/__tests__/language-detect.test.ts,src/tools/__tests__/list-files.test.ts,src/tools/__tests__/local-summarization-gate.test.ts,src/tools/__tests__/output-summarizer.test.ts,src/tools/__tests__/read-cache.test.ts,src/tools/__tests__/run-tests.test.ts,src/tools/__tests__/search_replace.test.ts,src/tools/__tests__/update_todo_list.test.ts,src/tools/__tests__/write_to_file.test.ts,src/vision/__tests__/describe.test.ts,src/watcher/__tests__/github.test.ts,src/watcher/__tests__/state.test.ts,src/watcher/__tests__/watch.test.ts",
   "clone 33c630390d21 src/rsi/__tests__/controller-candidates.test.ts,src/rsi/__tests__/controller-integrity.test.ts",
   "clone 33e68b8dfd03 src/llm/muse.ts,src/llm/openrouter.ts",
   "clone 3437313b4dfa src/llm/__tests__/muse.test.ts,src/llm/__tests__/openrouter.test.ts",
@@ -229,11 +203,8 @@
   "clone 3ee06defdfda src/llm/preflight.ts",
   "clone 3f80ca564a0a src/orchestrator/cli.ts,src/orchestrator/resume.ts",
   "clone 3fb76f830d07 src/engine/__tests__/inject-message.test.ts,src/engine/__tests__/loop.test.ts,src/engine/__tests__/pause.test.ts",
-  "clone 3fc784cdd83c src/__tests__/cli-env.test.ts,src/codesearch/__tests__/embedder-types.test.ts,src/evidence/__tests__/verify-files.test.ts,src/orchestrator/__tests__/task-files.test.ts,src/orchestrator/__tests__/terminal-status.test.ts,src/tools/__tests__/batch-types.test.ts,src/tools/__tests__/paths.test.ts",
-  "clone 3fd94d4af296 src/codeintel/__tests__/registration.test.ts,src/tools/__tests__/batch.test.ts",
   "clone 3fe29973675e src/orchestrator/resume.ts",
   "clone 404168195169 src/engine/loop.ts",
-  "clone 409d3c6aa64e src/engine/__tests__/new-task.test.ts,src/engine/__tests__/switch-mode.test.ts",
   "clone 40d56ba05f73 src/engine/__tests__/loop.test.ts,src/tools/__tests__/execute_command.test.ts",
   "clone 40df3509282c src/engine/__tests__/new-task.test.ts,src/engine/__tests__/switch-mode.test.ts",
   "clone 40e273ce917f src/engine/local-explore.ts,src/llm/ollama.ts",
@@ -248,14 +219,12 @@
   "clone 42dd6d326705 src/llm/__tests__/muse.test.ts,src/llm/__tests__/openrouter.test.ts",
   "clone 4372b69d7f78 src/orchestrator/resume.ts",
   "clone 43b0c11b7347 src/engine/loop.ts",
-  "clone 44848edc370f src/engine/__tests__/new-task.test.ts,src/engine/__tests__/switch-mode.test.ts",
   "clone 4494882fe028 src/llm/preflight.ts",
   "clone 44999a278fcd src/orchestrator/resume.ts",
   "clone 44d299e7bc4c src/dashboard/server.ts",
   "clone 44f4c5f6148f src/cloud/__tests__/openshell-provider.test.ts",
   "clone 45181f328e82 src/orchestrator/task-files.ts",
   "clone 45edfde1d33f src/orchestrator/task-files.ts",
-  "clone 464303e3e594 src/codeintel/__tests__/implementations.test.ts,src/orchestrator/__tests__/main-tree-tripwire.test.ts,src/orchestrator/__tests__/review-gates.test.ts",
   "clone 465b90f368f7 src/dashboard/__tests__/server.test.ts,src/dashboard/__tests__/session-launch.test.ts",
   "clone 4666d04d2662 src/llm/muse.ts,src/llm/openrouter.ts",
   "clone 46ffdeff854a src/orchestrator/resume.ts",
@@ -279,7 +248,6 @@
   "clone 4c52d635b095 src/cloud/docker-provider.ts,src/cloud/provider.ts",
   "clone 4d4b7ced132f src/orchestrator/cli.ts",
   "clone 4d70161ff45e src/checkpoints/cli.ts,src/decision-proxy/cli.ts",
-  "clone 4dd292628b52 src/codeintel/__tests__/registration.test.ts,src/tools/__tests__/batch.test.ts",
   "clone 4dee6b8a0f46 src/llm/preflight.ts",
   "clone 4e55134d37dc src/orchestrator/cli.ts,src/orchestrator/resume.ts",
   "clone 4eaaac6576f7 src/dashboard/__tests__/cost-history-routes.test.ts,src/orchestrator/__tests__/cost-history-cli.test.ts",
@@ -345,24 +313,18 @@
   "clone 5e6943aa1599 src/dashboard/__tests__/session-launch.test.ts",
   "clone 5e9eb7a40cd3 src/engine/__tests__/condense.test.ts",
   "clone 5f09b011f3c9 src/engine/__tests__/condense.test.ts",
-  "clone 5f2c88e38fbe src/budget/__tests__/budget.test.ts,src/budget/__tests__/concurrency.test.ts,src/budget/__tests__/cost.test.ts,src/checkpoints/__tests__/service.test.ts,src/cloud/__tests__/provider.test.ts,src/codesearch/__tests__/airunner-embedder.test.ts,src/codesearch/__tests__/chunk.test.ts,src/codesearch/__tests__/files.test.ts,src/codesearch/__tests__/ollama-embedder.test.ts,src/dashboard/__tests__/aggregate.test.ts,src/dashboard/__tests__/chat-thread.test.ts,src/dashboard/__tests__/checkpoints-routes.test.ts,src/dashboard/__tests__/events.test.ts,src/dashboard/__tests__/projects-routes.test.ts,src/dashboard/__tests__/server.test.ts,src/dashboard/__tests__/session-launch.test.ts,src/dashboard/__tests__/timeline.test.ts,src/deploy/__tests__/gate.test.ts,src/engine/__tests__/condense-async.test.ts,src/engine/__tests__/condense.test.ts,src/engine/__tests__/events.test.ts,src/engine/__tests__/inject-message.test.ts,src/engine/__tests__/loop.test.ts,src/engine/__tests__/new-task.test.ts,src/engine/__tests__/pause.test.ts,src/engine/__tests__/switch-mode.test.ts,src/github/__tests__/app-auth.test.ts,src/github/__tests__/cli.test.ts,src/github/__tests__/pr.test.ts,src/github/__tests__/provision.test.ts,src/github/__tests__/push.test.ts,src/index-util.test.ts,src/init/__tests__/cli.test.ts,src/init/__tests__/gitignore.test.ts,src/memory/__tests__/local.test.ts,src/memory/__tests__/memory-wiring.test.ts,src/memory/__tests__/uwuchat.test.ts,src/migrate/cli.test.ts,src/orchestrator/__tests__/arg-cursor.test.ts,src/orchestrator/__tests__/auto-split.test.ts,src/orchestrator/__tests__/cleanup.test.ts,src/orchestrator/__tests__/git-sync.test.ts,src/orchestrator/__tests__/merge-check.test.ts,src/orchestrator/__tests__/pipeline.test.ts,src/orchestrator/__tests__/reviewer.test.ts,src/orchestrator/__tests__/split.test.ts,src/orchestrator/__tests__/state.test.ts,src/orchestrator/__tests__/status-cli.test.ts,src/orchestrator/__tests__/stop-cli.test.ts,src/permissions/__tests__/commands.test.ts,src/permissions/__tests__/store-protection.test.ts,src/project-store.test.ts,src/projects/__tests__/cli.test.ts,src/qa/__tests__/qa.test.ts,src/tools/__tests__/apply_diff.test.ts,src/tools/__tests__/decision.test.ts,src/tools/__tests__/describe_image.test.ts,src/tools/__tests__/edit_file.test.ts,src/tools/__tests__/execute_command.test.ts,src/tools/__tests__/language-detect.test.ts,src/tools/__tests__/list-files.test.ts,src/tools/__tests__/local-summarization-gate.test.ts,src/tools/__tests__/output-summarizer.test.ts,src/tools/__tests__/path-safety.test.ts,src/tools/__tests__/read-cache.test.ts,src/tools/__tests__/run-tests.test.ts,src/tools/__tests__/search_replace.test.ts,src/tools/__tests__/update_todo_list.test.ts,src/tools/__tests__/write_to_file.test.ts,src/vision/__tests__/describe.test.ts,src/watcher/__tests__/github.test.ts,src/watcher/__tests__/state.test.ts,src/watcher/__tests__/watch.test.ts",
   "clone 5f2ee4383718 src/github/pr.ts",
   "clone 5f66e587d856 src/engine/local-explore.ts,src/llm/ollama.ts",
   "clone 5f6b2a16f10d src/checkpoints/cli.ts,src/cli.ts,src/codemap/cli.ts,src/codesearch/cli.ts,src/dashboard/cli.ts,src/dashboard/trend-cli.ts,src/decision-proxy/cli.ts,src/github/cli.ts,src/init/cli.ts,src/orchestrator/cleanup.ts,src/watcher/cli.ts",
   "clone 5fff6e3ebed7 src/dashboard/__tests__/cost-history-routes.test.ts,src/orchestrator/__tests__/cost-history-cli.test.ts",
   "clone 60c4fdaf31bd src/codesearch/airunner-embedder.ts,src/codesearch/ollama-embedder.ts",
-  "clone 61046dbb6410 src/budget/__tests__/budget.test.ts,src/budget/__tests__/concurrency.test.ts,src/budget/__tests__/cost.test.ts,src/checkpoints/__tests__/service.test.ts,src/cloud/__tests__/provider.test.ts,src/codeintel/__tests__/call-graph.test.ts,src/codeintel/__tests__/call-hierarchy.test.ts,src/codeintel/__tests__/find-references.test.ts,src/codeintel/__tests__/go-to-definition.test.ts,src/codeintel/__tests__/hover.test.ts,src/codeintel/__tests__/import-graph.test.ts,src/codeintel/__tests__/outline.test.ts,src/codeintel/__tests__/rename-symbol.test.ts,src/codemap/__tests__/build.test.ts,src/codesearch/__tests__/airunner-embedder.test.ts,src/codesearch/__tests__/chunk.test.ts,src/codesearch/__tests__/files.test.ts,src/codesearch/__tests__/ollama-embedder.test.ts,src/dashboard/__tests__/aggregate.test.ts,src/dashboard/__tests__/chat-thread.test.ts,src/dashboard/__tests__/checkpoints-routes.test.ts,src/dashboard/__tests__/codemap-routes.test.ts,src/dashboard/__tests__/cost-history-routes.test.ts,src/dashboard/__tests__/events.test.ts,src/dashboard/__tests__/projects-routes.test.ts,src/dashboard/__tests__/server.test.ts,src/dashboard/__tests__/session-launch.test.ts,src/dashboard/__tests__/timeline.test.ts,src/deploy/__tests__/gate.test.ts,src/engine/__tests__/condense-async.test.ts,src/engine/__tests__/condense.test.ts,src/engine/__tests__/events.test.ts,src/engine/__tests__/inject-message.test.ts,src/engine/__tests__/local-explore.test.ts,src/engine/__tests__/loop.test.ts,src/engine/__tests__/new-task.test.ts,src/engine/__tests__/pause.test.ts,src/engine/__tests__/switch-mode.test.ts,src/github/__tests__/app-auth.test.ts,src/github/__tests__/cli.test.ts,src/github/__tests__/pr.test.ts,src/github/__tests__/provision.test.ts,src/github/__tests__/push.test.ts,src/index-util.test.ts,src/init/__tests__/cli.test.ts,src/init/__tests__/gitignore.test.ts,src/memory/__tests__/local.test.ts,src/memory/__tests__/memory-wiring.test.ts,src/memory/__tests__/uwuchat.test.ts,src/migrate/cli.test.ts,src/orchestrator/__tests__/arg-cursor.test.ts,src/orchestrator/__tests__/auto-split.test.ts,src/orchestrator/__tests__/cleanup.test.ts,src/orchestrator/__tests__/git-sync.test.ts,src/orchestrator/__tests__/merge-check.test.ts,src/orchestrator/__tests__/pipeline.test.ts,src/orchestrator/__tests__/reviewer.test.ts,src/orchestrator/__tests__/split.test.ts,src/orchestrator/__tests__/state.test.ts,src/orchestrator/__tests__/status-cli.test.ts,src/orchestrator/__tests__/stop-cli.test.ts,src/permissions/__tests__/commands.test.ts,src/permissions/__tests__/store-protection.test.ts,src/project-store.test.ts,src/projects/__tests__/cli.test.ts,src/qa/__tests__/qa.test.ts,src/tools/__tests__/apply_diff.test.ts,src/tools/__tests__/browser-action.test.ts,src/tools/__tests__/decision.test.ts,src/tools/__tests__/describe_image.test.ts,src/tools/__tests__/edit_file.test.ts,src/tools/__tests__/execute_command.test.ts,src/tools/__tests__/language-detect.test.ts,src/tools/__tests__/list-files.test.ts,src/tools/__tests__/local-summarization-gate.test.ts,src/tools/__tests__/output-summarizer.test.ts,src/tools/__tests__/path-safety.test.ts,src/tools/__tests__/read-cache.test.ts,src/tools/__tests__/run-tests.test.ts,src/tools/__tests__/search_replace.test.ts,src/tools/__tests__/update_todo_list.test.ts,src/tools/__tests__/write_to_file.test.ts,src/vision/__tests__/describe.test.ts,src/watcher/__tests__/github.test.ts,src/watcher/__tests__/state.test.ts,src/watcher/__tests__/watch.test.ts",
   "clone 61b9caaf613a src/rsi/__tests__/controller-candidates.test.ts,src/rsi/__tests__/controller-fleet-results.test.ts",
-  "clone 6255fbea917d src/orchestrator/__tests__/cost-history-cli.test.ts,src/orchestrator/__tests__/cost-history.test.ts",
-  "clone 6287c1413aa4 src/budget/__tests__/budget.test.ts,src/budget/__tests__/concurrency.test.ts,src/budget/__tests__/cost.test.ts,src/checkpoints/__tests__/service.test.ts,src/cloud/__tests__/provider.test.ts,src/codeintel/__tests__/call-graph.test.ts,src/codeintel/__tests__/call-hierarchy.test.ts,src/codeintel/__tests__/find-references.test.ts,src/codeintel/__tests__/go-to-definition.test.ts,src/codeintel/__tests__/hover.test.ts,src/codeintel/__tests__/import-graph.test.ts,src/codeintel/__tests__/outline.test.ts,src/codeintel/__tests__/rename-symbol.test.ts,src/codemap/__tests__/build.test.ts,src/codesearch/__tests__/airunner-embedder.test.ts,src/codesearch/__tests__/chunk.test.ts,src/codesearch/__tests__/files.test.ts,src/codesearch/__tests__/ollama-embedder.test.ts,src/dashboard/__tests__/aggregate.test.ts,src/dashboard/__tests__/chat-thread.test.ts,src/dashboard/__tests__/checkpoints-routes.test.ts,src/dashboard/__tests__/codemap-routes.test.ts,src/dashboard/__tests__/cost-history-routes.test.ts,src/dashboard/__tests__/events.test.ts,src/dashboard/__tests__/projects-routes.test.ts,src/dashboard/__tests__/server.test.ts,src/dashboard/__tests__/session-launch.test.ts,src/dashboard/__tests__/timeline.test.ts,src/deploy/__tests__/gate.test.ts,src/engine/__tests__/condense-async.test.ts,src/engine/__tests__/condense.test.ts,src/engine/__tests__/events.test.ts,src/engine/__tests__/inject-message.test.ts,src/engine/__tests__/local-explore.test.ts,src/engine/__tests__/loop-helpers-history.test.ts,src/engine/__tests__/loop-helpers-readonly.test.ts,src/engine/__tests__/loop.test.ts,src/engine/__tests__/new-task.test.ts,src/engine/__tests__/parser.test.ts,src/engine/__tests__/pause.test.ts,src/engine/__tests__/summarize-tool-arg.test.ts,src/engine/__tests__/switch-mode.test.ts,src/github/__tests__/app-auth.test.ts,src/github/__tests__/cli.test.ts,src/github/__tests__/pr.test.ts,src/github/__tests__/provision.test.ts,src/github/__tests__/push.test.ts,src/index-util.test.ts,src/init/__tests__/cli.test.ts,src/init/__tests__/gitignore.test.ts,src/memory/__tests__/embed.test.ts,src/memory/__tests__/local.test.ts,src/memory/__tests__/memory-wiring.test.ts,src/memory/__tests__/summarizer.test.ts,src/memory/__tests__/uwuchat.test.ts,src/migrate/cli.test.ts,src/orchestrator/__tests__/arg-cursor.test.ts,src/orchestrator/__tests__/auto-split.test.ts,src/orchestrator/__tests__/cleanup.test.ts,src/orchestrator/__tests__/git-sync.test.ts,src/orchestrator/__tests__/merge-check.test.ts,src/orchestrator/__tests__/pipeline.test.ts,src/orchestrator/__tests__/reviewer.test.ts,src/orchestrator/__tests__/split.test.ts,src/orchestrator/__tests__/state.test.ts,src/orchestrator/__tests__/status-cli.test.ts,src/orchestrator/__tests__/stop-cli.test.ts,src/permissions/__tests__/commands.test.ts,src/permissions/__tests__/store-protection.test.ts,src/project-store.test.ts,src/projects/__tests__/cli.test.ts,src/qa/__tests__/qa.test.ts,src/tools/__tests__/apply_diff.test.ts,src/tools/__tests__/browser-action.test.ts,src/tools/__tests__/decision.test.ts,src/tools/__tests__/describe_image.test.ts,src/tools/__tests__/edit_file.test.ts,src/tools/__tests__/execute_command.test.ts,src/tools/__tests__/language-detect.test.ts,src/tools/__tests__/list-files.test.ts,src/tools/__tests__/local-summarization-gate.test.ts,src/tools/__tests__/output-summarizer.test.ts,src/tools/__tests__/read-cache.test.ts,src/tools/__tests__/run-tests.test.ts,src/tools/__tests__/search_replace.test.ts,src/tools/__tests__/update_todo_list.test.ts,src/tools/__tests__/write_to_file.test.ts,src/vision/__tests__/describe.test.ts,src/watcher/__tests__/github.test.ts,src/watcher/__tests__/state.test.ts,src/watcher/__tests__/watch.test.ts",
   "clone 62d28c84ccfe src/orchestrator/cli.ts",
   "clone 62fb95ada976 src/rsi/controller.ts",
-  "clone 62ff58c84315 src/engine/__tests__/loop-helpers-history.test.ts,src/engine/__tests__/loop-helpers-readonly.test.ts,src/engine/__tests__/parser.test.ts,src/engine/__tests__/summarize-tool-arg.test.ts,src/memory/__tests__/embed.test.ts,src/memory/__tests__/summarizer.test.ts,src/stack-rules-content.test.ts",
   "clone 6376e4c9bfd3 src/github/__tests__/provision.test.ts",
   "clone 63fb58a6271f src/orchestrator/__tests__/cli.test.ts",
   "clone 6433777a2a28 src/orchestrator/__tests__/spawn-model.test.ts,src/orchestrator/__tests__/spawn-state-reset.test.ts",
   "clone 6466c80a5290 src/github/pr.ts",
-  "clone 6472f40613ee src/codesearch/__tests__/backend-selection.test.ts,src/codesearch/__tests__/index.test.ts,src/codesearch/__tests__/search.test.ts,src/config/__tests__/mode-models.test.ts,src/dashboard/__tests__/file-browser-permissions.test.ts,src/dashboard/__tests__/settings.test.ts,src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/model-resolution.test.ts,src/orchestrator/__tests__/resume.test.ts,src/orchestrator/__tests__/spawn-model.test.ts,src/orchestrator/__tests__/watch.test.ts,src/permissions/__tests__/protected-files.test.ts,src/tools/__tests__/codebase_search.test.ts",
   "clone 64820517d48e src/llm/__tests__/muse.test.ts,src/llm/__tests__/openrouter.test.ts",
   "clone 6489529d68a1 src/dashboard/server.ts",
   "clone 64a8e2de29d4 src/engine/__tests__/new-task.test.ts,src/engine/__tests__/switch-mode.test.ts",
@@ -370,7 +332,6 @@
   "clone 655648fae3f0 src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/resume.test.ts",
   "clone 659104753a4c src/checkpoints/cli.ts,src/cli.ts,src/codemap/cli.ts,src/codesearch/cli.ts,src/dashboard/cli.ts,src/dashboard/trend-cli.ts,src/decision-proxy/cli.ts,src/github/cli.ts,src/init/cli.ts,src/orchestrator/cleanup.ts,src/watcher/cli.ts",
   "clone 6592f5270188 src/orchestrator/__tests__/reviewer.test.ts,src/qa/__tests__/qa.test.ts",
-  "clone 65bf7deda970 src/engine/__tests__/stack-rules-content.test.ts,src/stack-rules-content.test.ts",
   "clone 65e7527aaf4e src/watcher/watch.ts",
   "clone 6620858d5a93 src/orchestrator/__tests__/cli.test.ts",
   "clone 663e0371f2b8 src/engine/__tests__/new-task.test.ts,src/engine/__tests__/switch-mode.test.ts",
@@ -380,27 +341,21 @@
   "clone 66ec447e84b0 src/engine/__tests__/switch-mode.test.ts",
   "clone 672782f20b04 src/orchestrator/__tests__/resume.test.ts",
   "clone 67833e34f406 src/rsi/controller.ts",
-  "clone 67cd395b3adb src/checkpoints/__tests__/cli.test.ts,src/codemap/__tests__/cli.test.ts,src/codemap/__tests__/units.test.ts,src/dashboard/__tests__/cli.test.ts,src/dashboard/__tests__/page-route.test.ts,src/dashboard/__tests__/self-improvement-metrics.test.ts,src/dashboard/__tests__/sigint-shutdown.test.ts,src/github/__tests__/installations.test.ts,src/orchestrator/__tests__/analyze-cli.test.ts,src/orchestrator/__tests__/cost-estimate.test.ts,src/orchestrator/__tests__/log-analysis.test.ts,src/orchestrator/__tests__/spawn-state-reset.test.ts,src/orchestrator/__tests__/status.test.ts,src/run-tests-script.test.ts,src/tools/__tests__/set-indentation.test.ts",
   "clone 6846b0b81f2a src/llm/__tests__/preflight.test.ts,src/orchestrator/__tests__/cli.test.ts",
   "clone 68612079ddc5 src/llm/preflight.ts",
-  "clone 68654f999273 src/engine/__tests__/stack-rules-content.test.ts,src/stack-rules-content.test.ts",
   "clone 687e46486dc5 src/engine/__tests__/stack-rules-content.test.ts,src/engine/__tests__/stacks.test.ts",
   "clone 6893410ec1b1 src/orchestrator/__tests__/main-tree-tripwire.test.ts,src/orchestrator/__tests__/review-gates.test.ts",
   "clone 68b1356447b2 src/orchestrator/cli.ts",
   "clone 68ceda9cfcd3 src/codesearch/files.ts,src/engine/stacks.ts",
+  "clone 6910fdd8603d src/codesearch/__tests__/backend-selection.test.ts,src/codesearch/__tests__/index.test.ts,src/codesearch/__tests__/search.test.ts,src/config/__tests__/mode-models.test.ts,src/dashboard/__tests__/file-browser-permissions.test.ts,src/dashboard/__tests__/settings.test.ts,src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/cost-history-cli.test.ts,src/orchestrator/__tests__/cost-history.test.ts,src/orchestrator/__tests__/model-resolution.test.ts,src/orchestrator/__tests__/resume.test.ts,src/orchestrator/__tests__/watch.test.ts,src/permissions/__tests__/protected-files.test.ts,src/tools/__tests__/codebase_search.test.ts",
   "clone 694e054dcbfb src/codesearch/files.ts,src/engine/stacks.ts",
   "clone 695d76cf4fae src/dashboard/__tests__/projects-routes.test.ts,src/projects/__tests__/cli.test.ts",
-  "clone 699b3b419af4 src/orchestrator/__tests__/cost-history-cli.test.ts,src/orchestrator/__tests__/cost-history.test.ts",
   "clone 69c5c3472368 src/dashboard/server.ts",
-  "clone 69d258407ea9 src/codesearch/__tests__/backend-selection.test.ts,src/codesearch/__tests__/index.test.ts,src/codesearch/__tests__/search.test.ts,src/config/__tests__/mode-models.test.ts,src/dashboard/__tests__/file-browser-permissions.test.ts,src/dashboard/__tests__/settings.test.ts,src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/model-resolution.test.ts,src/orchestrator/__tests__/resume.test.ts,src/orchestrator/__tests__/watch.test.ts,src/permissions/__tests__/protected-files.test.ts,src/tools/__tests__/codebase_search.test.ts",
   "clone 69ff81f6294e src/engine/__tests__/condense.test.ts,src/engine/__tests__/loop.test.ts",
   "clone 6ad43b2053fa src/engine/__tests__/loop.test.ts",
-  "clone 6b322a1d1b11 src/orchestrator/__tests__/cost-history-cli.test.ts,src/orchestrator/__tests__/cost-history.test.ts",
-  "clone 6b35c8a20e34 src/orchestrator/__tests__/cost-history-cli.test.ts,src/orchestrator/__tests__/cost-history.test.ts",
   "clone 6b5f2e4eff30 src/llm/muse.ts",
   "clone 6bf4008bd80c src/orchestrator/__tests__/cli.test.ts",
   "clone 6c4b7719d621 src/engine/__tests__/loop.test.ts",
-  "clone 6c650d02957b src/codeintel/__tests__/registration.test.ts,src/tools/__tests__/batch.test.ts",
   "clone 6cade99be57e src/rsi/__tests__/controller-candidates.test.ts,src/rsi/__tests__/controller-fleet-results.test.ts",
   "clone 6cc5e85990d1 src/orchestrator/__tests__/spawn-model.test.ts,src/orchestrator/__tests__/spawn-state-reset.test.ts",
   "clone 6cd34638ecf4 src/codesearch/airunner-embedder.ts,src/codesearch/ollama-embedder.ts",
@@ -410,7 +365,6 @@
   "clone 6de3bc55de51 src/engine/loop.ts",
   "clone 6e0f7e4d8bbd src/orchestrator/cli.ts,src/orchestrator/resume.ts",
   "clone 6e3afc8af712 src/codesearch/files.ts,src/engine/stacks.ts",
-  "clone 6e67c36eb0ac src/codeintel/__tests__/registration.test.ts,src/tools/__tests__/batch.test.ts",
   "clone 6ec154ac699f src/llm/preflight.ts",
   "clone 700ed9396d5d src/orchestrator/cli.ts",
   "clone 70f80df8706f src/orchestrator/cli.ts,src/watcher/cli.ts",
@@ -439,8 +393,6 @@
   "clone 76fd2afa5e53 src/dashboard/__tests__/projects-routes.test.ts,src/projects/__tests__/cli.test.ts",
   "clone 76fe651f8360 src/dashboard/server.ts",
   "clone 775035a6a231 src/dashboard/server.ts",
-  "clone 7771faf0bdbe src/budget/__tests__/budget.test.ts,src/budget/__tests__/concurrency.test.ts,src/budget/__tests__/cost.test.ts,src/checkpoints/__tests__/service.test.ts,src/cloud/__tests__/docker-provider.test.ts,src/cloud/__tests__/provider.test.ts,src/codeintel/__tests__/call-graph.test.ts,src/codeintel/__tests__/call-hierarchy.test.ts,src/codeintel/__tests__/find-references.test.ts,src/codeintel/__tests__/go-to-definition.test.ts,src/codeintel/__tests__/hover.test.ts,src/codeintel/__tests__/import-graph.test.ts,src/codeintel/__tests__/outline.test.ts,src/codeintel/__tests__/rename-symbol.test.ts,src/codemap/__tests__/build.test.ts,src/codesearch/__tests__/airunner-embedder.test.ts,src/codesearch/__tests__/backend-selection.test.ts,src/codesearch/__tests__/chunk.test.ts,src/codesearch/__tests__/files.test.ts,src/codesearch/__tests__/index.test.ts,src/codesearch/__tests__/ollama-embedder.test.ts,src/codesearch/__tests__/search.test.ts,src/config/__tests__/mode-models.test.ts,src/dashboard/__tests__/aggregate.test.ts,src/dashboard/__tests__/chat-thread.test.ts,src/dashboard/__tests__/checkpoints-routes.test.ts,src/dashboard/__tests__/codemap-routes.test.ts,src/dashboard/__tests__/cost-history-routes.test.ts,src/dashboard/__tests__/events.test.ts,src/dashboard/__tests__/file-browser-permissions.test.ts,src/dashboard/__tests__/projects-routes.test.ts,src/dashboard/__tests__/server.test.ts,src/dashboard/__tests__/session-launch.test.ts,src/dashboard/__tests__/settings.test.ts,src/dashboard/__tests__/timeline.test.ts,src/deploy/__tests__/gate.test.ts,src/engine/__tests__/condense-async.test.ts,src/engine/__tests__/condense.test.ts,src/engine/__tests__/events.test.ts,src/engine/__tests__/inject-message.test.ts,src/engine/__tests__/loop.test.ts,src/engine/__tests__/new-task.test.ts,src/engine/__tests__/pause.test.ts,src/engine/__tests__/switch-mode.test.ts,src/github/__tests__/app-auth.test.ts,src/github/__tests__/cli.test.ts,src/github/__tests__/pr.test.ts,src/github/__tests__/provision.test.ts,src/github/__tests__/push.test.ts,src/index-util.test.ts,src/init/__tests__/cli.test.ts,src/init/__tests__/gitignore.test.ts,src/memory/__tests__/local.test.ts,src/memory/__tests__/memory-wiring.test.ts,src/memory/__tests__/uwuchat.test.ts,src/migrate/cli.test.ts,src/orchestrator/__tests__/arg-cursor.test.ts,src/orchestrator/__tests__/auto-split.test.ts,src/orchestrator/__tests__/cleanup.test.ts,src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/git-sync.test.ts,src/orchestrator/__tests__/merge-check.test.ts,src/orchestrator/__tests__/model-resolution.test.ts,src/orchestrator/__tests__/pipeline.test.ts,src/orchestrator/__tests__/resume.test.ts,src/orchestrator/__tests__/reviewer.test.ts,src/orchestrator/__tests__/spawn-model.test.ts,src/orchestrator/__tests__/split.test.ts,src/orchestrator/__tests__/state.test.ts,src/orchestrator/__tests__/status-cli.test.ts,src/orchestrator/__tests__/stop-cli.test.ts,src/orchestrator/__tests__/watch.test.ts,src/permissions/__tests__/commands.test.ts,src/permissions/__tests__/protected-files.test.ts,src/permissions/__tests__/store-protection.test.ts,src/project-store.test.ts,src/projects/__tests__/cli.test.ts,src/qa/__tests__/qa.test.ts,src/tools/__tests__/apply_diff.test.ts,src/tools/__tests__/browser-action.test.ts,src/tools/__tests__/codebase_search.test.ts,src/tools/__tests__/decision.test.ts,src/tools/__tests__/describe_image.test.ts,src/tools/__tests__/edit_file.test.ts,src/tools/__tests__/execute_command.test.ts,src/tools/__tests__/language-detect.test.ts,src/tools/__tests__/list-files.test.ts,src/tools/__tests__/local-summarization-gate.test.ts,src/tools/__tests__/output-summarizer.test.ts,src/tools/__tests__/path-safety.test.ts,src/tools/__tests__/read-cache.test.ts,src/tools/__tests__/run-tests.test.ts,src/tools/__tests__/search_replace.test.ts,src/tools/__tests__/update_todo_list.test.ts,src/tools/__tests__/write_to_file.test.ts,src/vision/__tests__/describe.test.ts,src/watcher/__tests__/github.test.ts,src/watcher/__tests__/state.test.ts,src/watcher/__tests__/watch.test.ts",
-  "clone 77a1f2eb17cb src/codesearch/__tests__/backend-selection.test.ts,src/codesearch/__tests__/index.test.ts,src/codesearch/__tests__/search.test.ts,src/config/__tests__/mode-models.test.ts,src/dashboard/__tests__/file-browser-permissions.test.ts,src/dashboard/__tests__/settings.test.ts,src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/model-resolution.test.ts,src/orchestrator/__tests__/resume.test.ts,src/orchestrator/__tests__/spawn-model.test.ts,src/orchestrator/__tests__/watch.test.ts,src/permissions/__tests__/protected-files.test.ts,src/tools/__tests__/codebase_search.test.ts",
   "clone 77aa86621a85 src/decision-proxy/__tests__/proxy.test.ts,src/engine/__tests__/switch-mode.test.ts,src/tools/__tests__/codebase_search.test.ts,src/tools/__tests__/execute_command.test.ts",
   "clone 77db3aa3bdb6 src/dashboard/__tests__/projects-routes.test.ts,src/projects/__tests__/cli.test.ts",
   "clone 77f73cc7bae5 src/cli.ts,src/dashboard/cli.ts,src/dashboard/trend-cli.ts,src/orchestrator/cleanup.ts,src/watcher/cli.ts",
@@ -459,7 +411,7 @@
   "clone 7a1b9f629554 src/engine/__tests__/new-task.test.ts,src/engine/__tests__/switch-mode.test.ts",
   "clone 7a53d4a3fcc7 src/codesearch/files.ts,src/engine/stacks.ts",
   "clone 7aa1ee87f445 src/orchestrator/resume.ts",
-  "clone 7aff3d23b7c0 src/engine/__tests__/global-modes.test.ts,src/engine/__tests__/stack-rules-content.test.ts",
+  "clone 7aac966a5ced src/codesearch/__tests__/backend-selection.test.ts,src/codesearch/__tests__/index.test.ts,src/codesearch/__tests__/search.test.ts,src/config/__tests__/mode-models.test.ts,src/dashboard/__tests__/file-browser-permissions.test.ts,src/dashboard/__tests__/settings.test.ts,src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/cost-history-cli.test.ts,src/orchestrator/__tests__/cost-history.test.ts,src/orchestrator/__tests__/model-resolution.test.ts,src/orchestrator/__tests__/resume.test.ts,src/orchestrator/__tests__/watch.test.ts,src/permissions/__tests__/protected-files.test.ts,src/tools/__tests__/codebase_search.test.ts",
   "clone 7b4511a4efb0 src/orchestrator/__tests__/spawn-model.test.ts,src/orchestrator/__tests__/spawn-state-reset.test.ts",
   "clone 7bab51178bf0 src/orchestrator/task-files.ts",
   "clone 7bae656083ca src/dashboard/__tests__/server.test.ts,src/dashboard/__tests__/session-launch.test.ts",
@@ -473,7 +425,6 @@
   "clone 7dd51a187649 src/rsi/__tests__/controller-candidates.test.ts,src/rsi/__tests__/controller-fleet-results.test.ts,src/rsi/__tests__/controller-integrity.test.ts",
   "clone 7ddc99484482 src/dashboard/server.ts",
   "clone 7dfc9cbdc2b8 src/orchestrator/task-files.ts",
-  "clone 7e1bcf45c7b2 src/engine/__tests__/loop-helpers-history.test.ts,src/engine/__tests__/loop-helpers-readonly.test.ts,src/engine/__tests__/parser.test.ts,src/engine/__tests__/summarize-tool-arg.test.ts,src/memory/__tests__/embed.test.ts,src/memory/__tests__/summarizer.test.ts",
   "clone 7e604775e29f src/orchestrator/cli.ts",
   "clone 7e9f9c4012ef src/orchestrator/__tests__/analyze-cli.test.ts",
   "clone 7f13e4c18807 src/orchestrator/__tests__/cli.test.ts",
@@ -505,7 +456,6 @@
   "clone 86154b195a9d src/engine/__tests__/loop.test.ts",
   "clone 862b3b509236 src/engine/__tests__/loop.test.ts,src/tools/__tests__/execute_command.test.ts",
   "clone 862e77a7bd8b src/engine/__tests__/loop.test.ts",
-  "clone 86477118fb95 src/__tests__/cli-env.test.ts,src/codesearch/__tests__/embedder-types.test.ts,src/evidence/__tests__/verify-files.test.ts,src/orchestrator/__tests__/task-files.test.ts,src/orchestrator/__tests__/terminal-status.test.ts,src/tools/__tests__/batch-types.test.ts,src/tools/__tests__/paths.test.ts",
   "clone 868374f44959 src/orchestrator/cli.ts,src/watcher/cli.ts",
   "clone 8737d3d18a94 src/dashboard/trend-cli.ts,src/orchestrator/cleanup.ts",
   "clone 8744bca39679 src/engine/__tests__/pause.test.ts",
@@ -516,7 +466,6 @@
   "clone 88f198a525f7 src/orchestrator/task-files.ts",
   "clone 896cd21c8bf7 src/orchestrator/resume.ts",
   "clone 898239c06549 src/engine/__tests__/condense.test.ts",
-  "clone 8a0887cd8c83 src/engine/__tests__/new-task.test.ts,src/engine/__tests__/switch-mode.test.ts",
   "clone 8a2e5c3b48db src/orchestrator/cli.ts,src/watcher/cli.ts",
   "clone 8a757993c7f9 src/dashboard/__tests__/projects-routes.test.ts,src/projects/__tests__/cli.test.ts",
   "clone 8aa676959e90 src/codesearch/embedder.ts,src/memory/embed.ts",
@@ -525,7 +474,6 @@
   "clone 8b4721c34bf4 src/engine/loop.ts",
   "clone 8b665567608f src/orchestrator/__tests__/spawn-model.test.ts,src/orchestrator/__tests__/spawn-state-reset.test.ts",
   "clone 8c15aef04331 src/orchestrator/task-files.ts",
-  "clone 8ca0903cd323 src/codesearch/__tests__/backend-selection.test.ts,src/codesearch/__tests__/index.test.ts,src/codesearch/__tests__/search.test.ts,src/config/__tests__/mode-models.test.ts,src/dashboard/__tests__/file-browser-permissions.test.ts,src/dashboard/__tests__/settings.test.ts,src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/cost-history-cli.test.ts,src/orchestrator/__tests__/cost-history.test.ts,src/orchestrator/__tests__/model-resolution.test.ts,src/orchestrator/__tests__/resume.test.ts,src/orchestrator/__tests__/watch.test.ts,src/permissions/__tests__/protected-files.test.ts,src/tools/__tests__/codebase_search.test.ts",
   "clone 8cbf4767d294 src/engine/local-explore.ts,src/llm/ollama.ts",
   "clone 8d24cff0cf64 src/engine/__tests__/loop.test.ts",
   "clone 8d49805d7bfd src/orchestrator/__tests__/status-cli.test.ts",
@@ -542,7 +490,6 @@
   "clone 8fbaf4437025 src/dashboard/__tests__/cost-history-routes.test.ts,src/orchestrator/__tests__/cost-history-cli.test.ts",
   "clone 8fe789c46bc4 src/orchestrator/__tests__/reviewer.test.ts,src/qa/__tests__/qa.test.ts",
   "clone 9003da86dcf3 src/codesearch/files.ts,src/engine/stacks.ts",
-  "clone 902284a95450 src/budget/__tests__/budget.test.ts,src/budget/__tests__/concurrency.test.ts,src/budget/__tests__/cost.test.ts,src/checkpoints/__tests__/service.test.ts,src/cloud/__tests__/provider.test.ts,src/codeintel/__tests__/call-graph.test.ts,src/codeintel/__tests__/call-hierarchy.test.ts,src/codeintel/__tests__/find-references.test.ts,src/codeintel/__tests__/go-to-definition.test.ts,src/codeintel/__tests__/hover.test.ts,src/codeintel/__tests__/import-graph.test.ts,src/codeintel/__tests__/outline.test.ts,src/codeintel/__tests__/rename-symbol.test.ts,src/codemap/__tests__/build.test.ts,src/codesearch/__tests__/airunner-embedder.test.ts,src/codesearch/__tests__/chunk.test.ts,src/codesearch/__tests__/files.test.ts,src/codesearch/__tests__/ollama-embedder.test.ts,src/dashboard/__tests__/aggregate.test.ts,src/dashboard/__tests__/chat-thread.test.ts,src/dashboard/__tests__/checkpoints-routes.test.ts,src/dashboard/__tests__/codemap-routes.test.ts,src/dashboard/__tests__/cost-history-routes.test.ts,src/dashboard/__tests__/events.test.ts,src/dashboard/__tests__/projects-routes.test.ts,src/dashboard/__tests__/server.test.ts,src/dashboard/__tests__/session-launch.test.ts,src/dashboard/__tests__/timeline.test.ts,src/deploy/__tests__/gate.test.ts,src/engine/__tests__/condense-async.test.ts,src/engine/__tests__/condense.test.ts,src/engine/__tests__/events.test.ts,src/engine/__tests__/inject-message.test.ts,src/engine/__tests__/local-explore.test.ts,src/engine/__tests__/loop-helpers-history.test.ts,src/engine/__tests__/loop-helpers-readonly.test.ts,src/engine/__tests__/loop.test.ts,src/engine/__tests__/new-task.test.ts,src/engine/__tests__/parser.test.ts,src/engine/__tests__/pause.test.ts,src/engine/__tests__/summarize-tool-arg.test.ts,src/engine/__tests__/switch-mode.test.ts,src/github/__tests__/app-auth.test.ts,src/github/__tests__/cli.test.ts,src/github/__tests__/pr.test.ts,src/github/__tests__/provision.test.ts,src/github/__tests__/push.test.ts,src/index-util.test.ts,src/init/__tests__/cli.test.ts,src/init/__tests__/gitignore.test.ts,src/memory/__tests__/embed.test.ts,src/memory/__tests__/local.test.ts,src/memory/__tests__/memory-wiring.test.ts,src/memory/__tests__/summarizer.test.ts,src/memory/__tests__/uwuchat.test.ts,src/migrate/cli.test.ts,src/orchestrator/__tests__/arg-cursor.test.ts,src/orchestrator/__tests__/auto-split.test.ts,src/orchestrator/__tests__/cleanup.test.ts,src/orchestrator/__tests__/git-sync.test.ts,src/orchestrator/__tests__/merge-check.test.ts,src/orchestrator/__tests__/pipeline.test.ts,src/orchestrator/__tests__/reviewer.test.ts,src/orchestrator/__tests__/split.test.ts,src/orchestrator/__tests__/state.test.ts,src/orchestrator/__tests__/status-cli.test.ts,src/orchestrator/__tests__/stop-cli.test.ts,src/permissions/__tests__/commands.test.ts,src/permissions/__tests__/store-protection.test.ts,src/project-store.test.ts,src/projects/__tests__/cli.test.ts,src/qa/__tests__/qa.test.ts,src/tools/__tests__/apply_diff.test.ts,src/tools/__tests__/browser-action.test.ts,src/tools/__tests__/decision.test.ts,src/tools/__tests__/describe_image.test.ts,src/tools/__tests__/edit_file.test.ts,src/tools/__tests__/execute_command.test.ts,src/tools/__tests__/language-detect.test.ts,src/tools/__tests__/list-files.test.ts,src/tools/__tests__/local-summarization-gate.test.ts,src/tools/__tests__/output-summarizer.test.ts,src/tools/__tests__/read-cache.test.ts,src/tools/__tests__/run-tests.test.ts,src/tools/__tests__/search_replace.test.ts,src/tools/__tests__/update_todo_list.test.ts,src/tools/__tests__/write_to_file.test.ts,src/vision/__tests__/describe.test.ts,src/watcher/__tests__/github.test.ts,src/watcher/__tests__/state.test.ts,src/watcher/__tests__/watch.test.ts",
   "clone 90309d08a06e src/engine/local-explore.ts,src/llm/ollama.ts",
   "clone 90919b289545 src/engine/__tests__/loop.test.ts,src/tools/__tests__/execute_command.test.ts",
   "clone 90ad7e210e7a src/engine/__tests__/condense.test.ts,src/engine/__tests__/loop.test.ts",
@@ -563,12 +510,10 @@
   "clone 95c346543423 src/orchestrator/task-files.ts",
   "clone 95e9585718ac src/engine/loop.ts",
   "clone 96160c6998c6 src/dashboard/__tests__/server.test.ts,src/dashboard/__tests__/session-launch.test.ts",
-  "clone 9624c77068ad src/engine/__tests__/loop-helpers-history.test.ts,src/engine/__tests__/loop-helpers-readonly.test.ts,src/engine/__tests__/parser.test.ts,src/engine/__tests__/summarize-tool-arg.test.ts,src/memory/__tests__/embed.test.ts,src/memory/__tests__/summarizer.test.ts",
   "clone 9653e98afda5 src/codesearch/__tests__/airunner-embedder.test.ts,src/codesearch/__tests__/ollama-embedder.test.ts",
   "clone 966fcd888ef7 src/orchestrator/cli.ts,src/orchestrator/resume.ts",
   "clone 9674b1ab1696 src/engine/loop.ts",
   "clone 9677b9319ac8 src/engine/__tests__/condense.test.ts",
-  "clone 96a19c419888 src/codeintel/__tests__/registration.test.ts,src/tools/__tests__/batch.test.ts",
   "clone 96d294f3a39d src/orchestrator/resume.ts",
   "clone 96d730011cef src/llm/openrouter.ts",
   "clone 9716d28923cf src/dashboard/trend-cli.ts,src/orchestrator/cleanup.ts",
@@ -580,14 +525,12 @@
   "clone 981c36f38d34 src/orchestrator/reviewer.ts,src/qa/qa.ts",
   "clone 9820cd6df71a src/codesearch/files.ts,src/engine/stacks.ts",
   "clone 986c6f8c6c8e src/checkpoints/__tests__/cli.test.ts,src/codemap/__tests__/cli.test.ts,src/dashboard/__tests__/cli.test.ts",
-  "clone 988e6970e5ab src/codesearch/__tests__/backend-selection.test.ts,src/codesearch/__tests__/index.test.ts,src/codesearch/__tests__/search.test.ts,src/config/__tests__/mode-models.test.ts,src/dashboard/__tests__/file-browser-permissions.test.ts,src/dashboard/__tests__/settings.test.ts,src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/cost-history-cli.test.ts,src/orchestrator/__tests__/cost-history.test.ts,src/orchestrator/__tests__/model-resolution.test.ts,src/orchestrator/__tests__/resume.test.ts,src/orchestrator/__tests__/spawn-model.test.ts,src/orchestrator/__tests__/watch.test.ts,src/permissions/__tests__/protected-files.test.ts,src/tools/__tests__/codebase_search.test.ts",
   "clone 98d735782817 src/llm/muse.ts",
   "clone 98e64d6f23d8 src/checkpoints/cli.ts,src/decision-proxy/cli.ts",
   "clone 993a1632940a src/llm/muse.ts,src/llm/openrouter.ts",
   "clone 998a739cc867 src/engine/__tests__/new-task.test.ts,src/engine/__tests__/switch-mode.test.ts",
   "clone 99b150e9e3dc src/codesearch/__tests__/airunner-embedder.test.ts,src/codesearch/__tests__/ollama-embedder.test.ts",
   "clone 99ba2224e54b src/llm/muse.ts,src/llm/openrouter.ts",
-  "clone 99c8753a4cab src/checkpoints/__tests__/cli.test.ts,src/codemap/__tests__/cli.test.ts,src/codemap/__tests__/units.test.ts,src/dashboard/__tests__/cli.test.ts,src/dashboard/__tests__/page-route.test.ts,src/dashboard/__tests__/self-improvement-metrics.test.ts,src/dashboard/__tests__/sigint-shutdown.test.ts,src/dashboard/__tests__/trend.test.ts,src/github/__tests__/installations.test.ts,src/orchestrator/__tests__/analyze-cli.test.ts,src/orchestrator/__tests__/cost-estimate.test.ts,src/orchestrator/__tests__/log-analysis.test.ts,src/orchestrator/__tests__/spawn-state-reset.test.ts,src/orchestrator/__tests__/status.test.ts,src/run-tests-script.test.ts,src/tools/__tests__/set-indentation.test.ts",
   "clone 99ebebf858ee src/codesearch/__tests__/airunner-embedder.test.ts,src/codesearch/__tests__/ollama-embedder.test.ts",
   "clone 9a2d9c8ece5a src/rsi/__tests__/controller-candidates.test.ts,src/rsi/__tests__/controller-integrity.test.ts",
   "clone 9a34727f0f71 src/engine/__tests__/condense.test.ts",
@@ -617,13 +560,9 @@
   "clone 9f5a1c3e6be5 src/orchestrator/resume.ts",
   "clone 9fd2184a9f18 src/rsi/__tests__/controller-candidates.test.ts,src/rsi/__tests__/controller-integrity.test.ts",
   "clone a003fde0f022 src/llm/muse.ts,src/llm/openrouter.ts",
-  "clone a023cb3be29b src/budget/__tests__/budget.test.ts,src/budget/__tests__/concurrency.test.ts,src/budget/__tests__/cost.test.ts,src/checkpoints/__tests__/cli.test.ts,src/checkpoints/__tests__/service.test.ts,src/cloud/__tests__/provider.test.ts,src/codemap/__tests__/cli.test.ts,src/codemap/__tests__/units.test.ts,src/codesearch/__tests__/airunner-embedder.test.ts,src/codesearch/__tests__/chunk.test.ts,src/codesearch/__tests__/files.test.ts,src/codesearch/__tests__/ollama-embedder.test.ts,src/dashboard/__tests__/aggregate.test.ts,src/dashboard/__tests__/chat-thread.test.ts,src/dashboard/__tests__/checkpoints-routes.test.ts,src/dashboard/__tests__/cli.test.ts,src/dashboard/__tests__/events.test.ts,src/dashboard/__tests__/page-route.test.ts,src/dashboard/__tests__/projects-routes.test.ts,src/dashboard/__tests__/self-improvement-metrics.test.ts,src/dashboard/__tests__/server.test.ts,src/dashboard/__tests__/session-launch.test.ts,src/dashboard/__tests__/sigint-shutdown.test.ts,src/dashboard/__tests__/timeline.test.ts,src/deploy/__tests__/gate.test.ts,src/engine/__tests__/condense-async.test.ts,src/engine/__tests__/condense.test.ts,src/engine/__tests__/events.test.ts,src/engine/__tests__/inject-message.test.ts,src/engine/__tests__/new-task.test.ts,src/engine/__tests__/pause.test.ts,src/engine/__tests__/prompt.test.ts,src/engine/__tests__/switch-mode.test.ts,src/github/__tests__/app-auth.test.ts,src/github/__tests__/cli.test.ts,src/github/__tests__/installations.test.ts,src/github/__tests__/pr.test.ts,src/github/__tests__/provision.test.ts,src/github/__tests__/push.test.ts,src/index-util.test.ts,src/init/__tests__/cli.test.ts,src/init/__tests__/gitignore.test.ts,src/llm/__tests__/muse.test.ts,src/llm/__tests__/openrouter.test.ts,src/memory/__tests__/local.test.ts,src/memory/__tests__/memory-wiring.test.ts,src/memory/__tests__/uwuchat.test.ts,src/migrate/cli.test.ts,src/orchestrator/__tests__/analyze-cli.test.ts,src/orchestrator/__tests__/arg-cursor.test.ts,src/orchestrator/__tests__/auto-split.test.ts,src/orchestrator/__tests__/cleanup.test.ts,src/orchestrator/__tests__/cost-estimate.test.ts,src/orchestrator/__tests__/git-sync.test.ts,src/orchestrator/__tests__/log-analysis.test.ts,src/orchestrator/__tests__/merge-check.test.ts,src/orchestrator/__tests__/pipeline.test.ts,src/orchestrator/__tests__/review-gates.test.ts,src/orchestrator/__tests__/reviewer.test.ts,src/orchestrator/__tests__/spawn-state-reset.test.ts,src/orchestrator/__tests__/split.test.ts,src/orchestrator/__tests__/state.test.ts,src/orchestrator/__tests__/status-cli.test.ts,src/orchestrator/__tests__/status.test.ts,src/orchestrator/__tests__/stop-cli.test.ts,src/permissions/__tests__/commands.test.ts,src/permissions/__tests__/store-protection.test.ts,src/project-store.test.ts,src/projects/__tests__/cli.test.ts,src/qa/__tests__/qa.test.ts,src/run-tests-script.test.ts,src/tools/__tests__/apply_diff.test.ts,src/tools/__tests__/decision.test.ts,src/tools/__tests__/describe_image.test.ts,src/tools/__tests__/edit_file.test.ts,src/tools/__tests__/execute_command.test.ts,src/tools/__tests__/language-detect.test.ts,src/tools/__tests__/list-files.test.ts,src/tools/__tests__/local-summarization-gate.test.ts,src/tools/__tests__/output-summarizer.test.ts,src/tools/__tests__/path-safety.test.ts,src/tools/__tests__/read-cache.test.ts,src/tools/__tests__/run-tests.test.ts,src/tools/__tests__/search_replace.test.ts,src/tools/__tests__/set-indentation.test.ts,src/tools/__tests__/update_todo_list.test.ts,src/tools/__tests__/write_to_file.test.ts,src/vision/__tests__/describe.test.ts,src/watcher/__tests__/github.test.ts,src/watcher/__tests__/state.test.ts,src/watcher/__tests__/watch.test.ts",
-  "clone a0e38675893c src/checkpoints/__tests__/cli.test.ts,src/codemap/__tests__/cli.test.ts,src/codemap/__tests__/units.test.ts,src/dashboard/__tests__/cli.test.ts,src/dashboard/__tests__/page-route.test.ts,src/dashboard/__tests__/self-improvement-metrics.test.ts,src/dashboard/__tests__/sigint-shutdown.test.ts,src/github/__tests__/installations.test.ts,src/orchestrator/__tests__/analyze-cli.test.ts,src/orchestrator/__tests__/cost-estimate.test.ts,src/orchestrator/__tests__/log-analysis.test.ts,src/orchestrator/__tests__/spawn-state-reset.test.ts,src/orchestrator/__tests__/status.test.ts,src/run-tests-script.test.ts,src/tools/__tests__/set-indentation.test.ts",
   "clone a100929d0996 src/codesearch/__tests__/airunner-embedder.test.ts,src/codesearch/__tests__/ollama-embedder.test.ts",
   "clone a1132d316866 src/checkpoints/cli.ts,src/cli.ts,src/codemap/cli.ts,src/codesearch/cli.ts,src/dashboard/cli.ts,src/dashboard/trend-cli.ts,src/decision-proxy/cli.ts,src/github/cli.ts,src/init/cli.ts,src/orchestrator/cleanup.ts,src/watcher/cli.ts",
-  "clone a11ed931f825 src/codeintel/__tests__/registration.test.ts,src/tools/__tests__/batch.test.ts",
   "clone a1a26fbf0a56 src/orchestrator/__tests__/status-cli.test.ts",
-  "clone a1c96da14be8 src/codesearch/__tests__/backend-selection.test.ts,src/codesearch/__tests__/index.test.ts,src/codesearch/__tests__/search.test.ts,src/config/__tests__/mode-models.test.ts,src/dashboard/__tests__/file-browser-permissions.test.ts,src/dashboard/__tests__/settings.test.ts,src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/model-resolution.test.ts,src/orchestrator/__tests__/resume.test.ts,src/orchestrator/__tests__/watch.test.ts,src/permissions/__tests__/protected-files.test.ts,src/tools/__tests__/codebase_search.test.ts",
   "clone a1cf343dc5c6 src/orchestrator/cli.ts",
   "clone a1f6b098886c src/rsi/__tests__/controller-candidates.test.ts,src/rsi/__tests__/controller-integrity.test.ts",
   "clone a1fd626e0e34 src/dashboard/__tests__/checkpoints-routes.test.ts,src/dashboard/__tests__/file-browser-permissions.test.ts",
@@ -633,7 +572,6 @@
   "clone a2856ade0b68 src/llm/muse.ts,src/llm/openrouter.ts",
   "clone a290c29733ab src/codeintel/__tests__/call-graph.test.ts,src/codeintel/__tests__/find-references.test.ts,src/codeintel/__tests__/hover.test.ts,src/codeintel/__tests__/import-graph.test.ts,src/codeintel/__tests__/outline.test.ts,src/codeintel/__tests__/registration.test.ts,src/codeintel/__tests__/rename-symbol.test.ts,src/tools/__tests__/run-tests.test.ts",
   "clone a331eb790bd5 src/orchestrator/resume.ts",
-  "clone a3534d1a9458 src/engine/__tests__/global-modes.test.ts,src/engine/__tests__/prompt.test.ts,src/engine/__tests__/stack-rules-content.test.ts,src/engine/__tests__/stacks.test.ts,src/llm/__tests__/muse.test.ts,src/llm/__tests__/openrouter.test.ts",
   "clone a3552e07fb57 src/dashboard/server.ts",
   "clone a42a6d879ca3 src/dashboard/__tests__/session-launch.test.ts,src/dashboard/__tests__/settings.test.ts",
   "clone a4983caf62b6 src/llm/preflight.ts",
@@ -646,11 +584,11 @@
   "clone a705b5057bf3 src/codesearch/embedder.ts,src/memory/embed.ts",
   "clone a70726307e4f src/orchestrator/cli.ts",
   "clone a7524c1d0b75 src/dashboard/__tests__/projects-routes.test.ts,src/projects/__tests__/cli.test.ts",
-  "clone a7ac9999b88e src/checkpoints/__tests__/cli.test.ts,src/codemap/__tests__/cli.test.ts,src/codemap/__tests__/units.test.ts,src/dashboard/__tests__/cli.test.ts,src/dashboard/__tests__/page-route.test.ts,src/dashboard/__tests__/self-improvement-metrics.test.ts,src/dashboard/__tests__/sigint-shutdown.test.ts,src/github/__tests__/installations.test.ts,src/orchestrator/__tests__/analyze-cli.test.ts,src/orchestrator/__tests__/cost-estimate.test.ts,src/orchestrator/__tests__/cost-history-cli.test.ts,src/orchestrator/__tests__/cost-history.test.ts,src/orchestrator/__tests__/log-analysis.test.ts,src/orchestrator/__tests__/spawn-state-reset.test.ts,src/orchestrator/__tests__/status.test.ts,src/run-tests-script.test.ts,src/tools/__tests__/set-indentation.test.ts",
   "clone a85870eb1a5b src/decision-proxy/__tests__/proxy.test.ts,src/engine/__tests__/switch-mode.test.ts",
   "clone a89b7f2f4a03 src/checkpoints/cli.ts,src/decision-proxy/cli.ts",
   "clone a8a18bc5e095 src/engine/local-explore.ts,src/llm/ollama.ts",
   "clone a8b8282008b3 src/rsi/__tests__/controller-candidates.test.ts,src/rsi/__tests__/controller-fleet-results.test.ts",
+  "clone a94c299950cc src/codesearch/__tests__/backend-selection.test.ts,src/codesearch/__tests__/index.test.ts,src/codesearch/__tests__/search.test.ts,src/config/__tests__/mode-models.test.ts,src/dashboard/__tests__/file-browser-permissions.test.ts,src/dashboard/__tests__/settings.test.ts,src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/cost-history-cli.test.ts,src/orchestrator/__tests__/cost-history.test.ts,src/orchestrator/__tests__/model-resolution.test.ts,src/orchestrator/__tests__/resume.test.ts,src/orchestrator/__tests__/watch.test.ts,src/permissions/__tests__/protected-files.test.ts,src/tools/__tests__/codebase_search.test.ts",
   "clone a990dfadf953 src/tools/__tests__/apply_diff.test.ts,src/tools/__tests__/edit_file.test.ts,src/tools/__tests__/search_replace.test.ts",
   "clone a99cd9437614 src/permissions/commands.ts",
   "clone aa35046643a4 src/rsi/controller-proof-archive.ts",
@@ -677,7 +615,6 @@
   "clone afbcfb5dcc72 src/rsi/controller.ts",
   "clone afec75fcac46 src/dashboard/server.ts",
   "clone b013f5230fd0 src/dashboard/server.ts",
-  "clone b0d2cf6ccae4 src/codeintel/__tests__/call-graph.test.ts,src/codeintel/__tests__/call-hierarchy.test.ts,src/codeintel/__tests__/find-references.test.ts,src/codeintel/__tests__/go-to-definition.test.ts,src/codeintel/__tests__/hover.test.ts,src/codeintel/__tests__/import-graph.test.ts,src/codeintel/__tests__/outline.test.ts,src/codeintel/__tests__/rename-symbol.test.ts,src/codemap/__tests__/build.test.ts,src/dashboard/__tests__/codemap-routes.test.ts,src/dashboard/__tests__/cost-history-routes.test.ts",
   "clone b1157b77dcf6 src/budget/__tests__/budget.test.ts",
   "clone b12348060bcb src/llm/muse.ts",
   "clone b1263c10d333 src/orchestrator/task-files.ts",
@@ -712,7 +649,6 @@
   "clone b94a982634f6 src/orchestrator/reviewer.ts,src/qa/qa.ts",
   "clone b9cced506295 src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/resume.test.ts",
   "clone ba0241fb0a20 src/engine/events.ts,src/engine/usage.ts",
-  "clone ba4c232793d7 src/budget/__tests__/budget.test.ts,src/budget/__tests__/concurrency.test.ts,src/budget/__tests__/cost.test.ts,src/checkpoints/__tests__/service.test.ts,src/cloud/__tests__/provider.test.ts,src/codeintel/__tests__/call-graph.test.ts,src/codeintel/__tests__/call-hierarchy.test.ts,src/codeintel/__tests__/find-references.test.ts,src/codeintel/__tests__/go-to-definition.test.ts,src/codeintel/__tests__/hover.test.ts,src/codeintel/__tests__/import-graph.test.ts,src/codeintel/__tests__/outline.test.ts,src/codeintel/__tests__/rename-symbol.test.ts,src/codemap/__tests__/build.test.ts,src/codesearch/__tests__/airunner-embedder.test.ts,src/codesearch/__tests__/chunk.test.ts,src/codesearch/__tests__/files.test.ts,src/codesearch/__tests__/ollama-embedder.test.ts,src/dashboard/__tests__/aggregate.test.ts,src/dashboard/__tests__/chat-thread.test.ts,src/dashboard/__tests__/checkpoints-routes.test.ts,src/dashboard/__tests__/codemap-routes.test.ts,src/dashboard/__tests__/cost-history-routes.test.ts,src/dashboard/__tests__/events.test.ts,src/dashboard/__tests__/projects-routes.test.ts,src/dashboard/__tests__/server.test.ts,src/dashboard/__tests__/session-launch.test.ts,src/dashboard/__tests__/timeline.test.ts,src/deploy/__tests__/gate.test.ts,src/engine/__tests__/condense-async.test.ts,src/engine/__tests__/condense.test.ts,src/engine/__tests__/events.test.ts,src/engine/__tests__/inject-message.test.ts,src/engine/__tests__/loop.test.ts,src/engine/__tests__/new-task.test.ts,src/engine/__tests__/pause.test.ts,src/engine/__tests__/switch-mode.test.ts,src/github/__tests__/app-auth.test.ts,src/github/__tests__/cli.test.ts,src/github/__tests__/pr.test.ts,src/github/__tests__/provision.test.ts,src/github/__tests__/push.test.ts,src/index-util.test.ts,src/init/__tests__/cli.test.ts,src/init/__tests__/gitignore.test.ts,src/memory/__tests__/local.test.ts,src/memory/__tests__/memory-wiring.test.ts,src/memory/__tests__/uwuchat.test.ts,src/migrate/cli.test.ts,src/orchestrator/__tests__/arg-cursor.test.ts,src/orchestrator/__tests__/auto-split.test.ts,src/orchestrator/__tests__/cleanup.test.ts,src/orchestrator/__tests__/git-sync.test.ts,src/orchestrator/__tests__/merge-check.test.ts,src/orchestrator/__tests__/pipeline.test.ts,src/orchestrator/__tests__/reviewer.test.ts,src/orchestrator/__tests__/split.test.ts,src/orchestrator/__tests__/state.test.ts,src/orchestrator/__tests__/status-cli.test.ts,src/orchestrator/__tests__/stop-cli.test.ts,src/permissions/__tests__/commands.test.ts,src/permissions/__tests__/store-protection.test.ts,src/project-store.test.ts,src/projects/__tests__/cli.test.ts,src/qa/__tests__/qa.test.ts,src/tools/__tests__/apply_diff.test.ts,src/tools/__tests__/browser-action.test.ts,src/tools/__tests__/decision.test.ts,src/tools/__tests__/describe_image.test.ts,src/tools/__tests__/edit_file.test.ts,src/tools/__tests__/execute_command.test.ts,src/tools/__tests__/language-detect.test.ts,src/tools/__tests__/list-files.test.ts,src/tools/__tests__/local-summarization-gate.test.ts,src/tools/__tests__/output-summarizer.test.ts,src/tools/__tests__/path-safety.test.ts,src/tools/__tests__/read-cache.test.ts,src/tools/__tests__/run-tests.test.ts,src/tools/__tests__/search_replace.test.ts,src/tools/__tests__/update_todo_list.test.ts,src/tools/__tests__/write_to_file.test.ts,src/vision/__tests__/describe.test.ts,src/watcher/__tests__/github.test.ts,src/watcher/__tests__/state.test.ts,src/watcher/__tests__/watch.test.ts",
   "clone baa20b9ca397 src/codesearch/__tests__/airunner-embedder.test.ts,src/codesearch/__tests__/ollama-embedder.test.ts",
   "clone bac150db44ec src/llm/preflight.ts",
   "clone bad5d67208cf src/orchestrator/task-files.ts",
@@ -732,29 +668,26 @@
   "clone bd99f1f211f3 src/engine/__tests__/inject-message.test.ts,src/engine/__tests__/loop.test.ts,src/engine/__tests__/pause.test.ts",
   "clone be4b5da32f15 src/orchestrator/cli.ts",
   "clone bf4577eed8ec src/engine/__tests__/loop.test.ts",
-  "clone bf5199ff5773 src/codeintel/__tests__/implementations.test.ts,src/orchestrator/__tests__/review-gates.test.ts",
   "clone bf96094425f1 src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/resume.test.ts",
-  "clone bfda8ae55db7 src/engine/__tests__/loop-helpers-history.test.ts,src/engine/__tests__/loop-helpers-readonly.test.ts,src/engine/__tests__/loop.test.ts,src/engine/__tests__/new-task.test.ts,src/engine/__tests__/parser.test.ts,src/engine/__tests__/summarize-tool-arg.test.ts,src/engine/__tests__/switch-mode.test.ts",
   "clone c1b7cee2cf47 src/codesearch/files.ts,src/engine/stacks.ts",
   "clone c1e1b41c0b87 src/orchestrator/__tests__/spawn-model.test.ts,src/orchestrator/__tests__/spawn-state-reset.test.ts",
   "clone c1f158e49080 src/orchestrator/reviewer.ts,src/qa/qa.ts",
   "clone c25de2e7a85a src/orchestrator/cli.ts",
   "clone c26e4fd9d44b src/rsi/__tests__/controller-candidates.test.ts,src/rsi/__tests__/controller-integrity.test.ts",
   "clone c28d9afd9221 src/engine/__tests__/loop.test.ts,src/tools/__tests__/execute_command.test.ts",
+  "clone c29c3c3368bb src/engine/__tests__/global-modes.test.ts,src/engine/__tests__/stack-rules-content.test.ts",
   "clone c3486aeec199 src/codesearch/__tests__/airunner-embedder.test.ts,src/codesearch/__tests__/ollama-embedder.test.ts",
   "clone c3571181b0ad src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/resume.test.ts",
   "clone c3716be39566 src/orchestrator/cli.ts",
   "clone c383dd436114 src/engine/__tests__/switch-mode.test.ts",
   "clone c3e667db531c src/github/cli.ts",
   "clone c46a996d1f67 src/dashboard/server.ts",
-  "clone c4a733213681 src/codesearch/__tests__/backend-selection.test.ts,src/codesearch/__tests__/index.test.ts,src/codesearch/__tests__/search.test.ts,src/config/__tests__/mode-models.test.ts,src/dashboard/__tests__/file-browser-permissions.test.ts,src/dashboard/__tests__/settings.test.ts,src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/model-resolution.test.ts,src/orchestrator/__tests__/resume.test.ts,src/orchestrator/__tests__/watch.test.ts,src/permissions/__tests__/protected-files.test.ts,src/tools/__tests__/codebase_search.test.ts",
   "clone c4fb71be60ac src/codesearch/embedder.ts,src/memory/embed.ts",
   "clone c4ffe9d533ad src/orchestrator/cli.ts",
   "clone c514b8a70ac1 src/engine/__tests__/condense.test.ts",
   "clone c5206c729975 src/dashboard/__tests__/cost-history-routes.test.ts,src/orchestrator/__tests__/cost-history-cli.test.ts",
   "clone c5f99dad818d src/engine/__tests__/new-task.test.ts,src/engine/__tests__/switch-mode.test.ts",
   "clone c5fae546f069 src/watcher/watch.ts",
-  "clone c620cb9d14ba src/engine/__tests__/loop-helpers-history.test.ts,src/engine/__tests__/loop-helpers-readonly.test.ts,src/engine/__tests__/parser.test.ts,src/engine/__tests__/summarize-tool-arg.test.ts,src/memory/__tests__/embed.test.ts,src/memory/__tests__/summarizer.test.ts",
   "clone c68a90c711b3 src/codesearch/files.ts,src/engine/stacks.ts",
   "clone c6af3683a9b2 src/rsi/__tests__/worker.test.ts",
   "clone c6b95dabb537 src/engine/local-explore.ts,src/llm/ollama.ts",
@@ -766,7 +699,6 @@
   "clone c7d2cfba43b6 src/engine/__tests__/new-task.test.ts,src/engine/__tests__/switch-mode.test.ts",
   "clone c9648e69e45e src/orchestrator/__tests__/cli.test.ts",
   "clone c9726dcfaf3b src/engine/local-explore.ts,src/llm/ollama.ts",
-  "clone c9fe2d9273f4 src/codeintel/__tests__/implementations.test.ts,src/orchestrator/__tests__/main-tree-tripwire.test.ts,src/orchestrator/__tests__/review-gates.test.ts",
   "clone ca31b6bb0f50 src/budget/__tests__/concurrency.test.ts,src/watcher/__tests__/watch.test.ts",
   "clone cb208607f464 src/dashboard/server.ts",
   "clone cb54ce2ca134 src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/resume.test.ts",
@@ -783,10 +715,7 @@
   "clone cdff5c34ea22 src/orchestrator/task-files.ts",
   "clone ce27f097c90c src/rsi/__tests__/controller-candidates.test.ts,src/rsi/__tests__/controller-fleet-results.test.ts",
   "clone ce4455f3e99f src/orchestrator/task-files.ts",
-  "clone ce8ca6626552 src/engine/__tests__/global-modes.test.ts,src/engine/__tests__/prompt.test.ts,src/engine/__tests__/stack-rules-content.test.ts,src/engine/__tests__/stacks.test.ts,src/llm/__tests__/muse.test.ts,src/llm/__tests__/openrouter.test.ts,src/stack-rules-content.test.ts",
-  "clone cebb36184bf4 src/codesearch/__tests__/backend-selection.test.ts,src/codesearch/__tests__/index.test.ts,src/codesearch/__tests__/search.test.ts,src/config/__tests__/mode-models.test.ts,src/dashboard/__tests__/file-browser-permissions.test.ts,src/dashboard/__tests__/settings.test.ts,src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/model-resolution.test.ts,src/orchestrator/__tests__/resume.test.ts,src/orchestrator/__tests__/spawn-model.test.ts,src/orchestrator/__tests__/watch.test.ts,src/permissions/__tests__/protected-files.test.ts,src/tools/__tests__/codebase_search.test.ts",
   "clone cee648fd9d03 src/engine/__tests__/new-task.test.ts,src/engine/__tests__/switch-mode.test.ts",
-  "clone cf91fc906ce8 src/engine/__tests__/global-modes.test.ts,src/engine/__tests__/prompt.test.ts,src/engine/__tests__/stack-rules-content.test.ts,src/engine/__tests__/stacks.test.ts,src/llm/__tests__/muse.test.ts,src/llm/__tests__/openrouter.test.ts",
   "clone cfbc5b2083c1 src/engine/local-explore.ts,src/llm/ollama.ts",
   "clone d07fbc6b767f src/tools/__tests__/edit_file.test.ts,src/tools/__tests__/search_replace.test.ts",
   "clone d08cb9f5e40b src/dashboard/__tests__/projects-routes.test.ts,src/projects/__tests__/cli.test.ts",
@@ -828,7 +757,6 @@
   "clone d8dfa6d1a846 src/orchestrator/cli.ts",
   "clone d8e97665e9ec src/checkpoints/__tests__/cli.test.ts,src/codemap/__tests__/cli.test.ts,src/dashboard/__tests__/cli.test.ts",
   "clone d9003b766f88 src/dashboard/__tests__/server.test.ts,src/dashboard/__tests__/session-launch.test.ts",
-  "clone d95677f0fc38 src/codeintel/__tests__/implementations.test.ts,src/orchestrator/__tests__/main-tree-tripwire.test.ts,src/orchestrator/__tests__/review-gates.test.ts",
   "clone d96d4c124839 src/dashboard/trend-cli.ts,src/orchestrator/cleanup.ts",
   "clone d9ab3b61fdbc src/llm/preflight.ts",
   "clone d9cbef8f79d1 src/engine/local-explore.ts,src/llm/ollama.ts",
@@ -842,7 +770,6 @@
   "clone dc983745440d src/llm/muse.ts,src/llm/openrouter.ts",
   "clone dcf2b502a3dd src/github/__tests__/pr.test.ts,src/orchestrator/__tests__/merge-check.test.ts",
   "clone dd0c493b152f src/tools/__tests__/execute_command.test.ts",
-  "clone dda5b65e734c src/engine/__tests__/prompt.test.ts,src/llm/__tests__/muse.test.ts,src/llm/__tests__/openrouter.test.ts",
   "clone dde98d818073 src/cloud/provider.ts,src/orchestrator/cli.ts",
   "clone ddf5d3691b74 src/checkpoints/cli.ts,src/cli.ts,src/codemap/cli.ts,src/codesearch/cli.ts,src/dashboard/cli.ts,src/dashboard/trend-cli.ts,src/decision-proxy/cli.ts,src/github/cli.ts,src/init/cli.ts,src/orchestrator/cleanup.ts,src/watcher/cli.ts",
   "clone de0450878713 src/llm/muse.ts,src/llm/openrouter.ts",
@@ -853,7 +780,6 @@
   "clone df2e214d885a src/engine/__tests__/loop.test.ts",
   "clone dfe022ad1548 src/orchestrator/cli.ts",
   "clone e010fc53c9f4 src/engine/loop.ts",
-  "clone e023eaa49b35 src/tools/__tests__/apply_diff.test.ts,src/tools/__tests__/edit_file.test.ts,src/tools/__tests__/search_replace.test.ts",
   "clone e0baba4315c3 src/orchestrator/cli.ts,src/orchestrator/resume.ts,src/watcher/cli.ts",
   "clone e0f1e0a708cc src/orchestrator/__tests__/spawn-model.test.ts,src/orchestrator/__tests__/spawn-state-reset.test.ts",
   "clone e1591587b296 src/engine/local-explore.ts,src/llm/ollama.ts",
@@ -862,7 +788,6 @@
   "clone e217ea15e165 src/orchestrator/task-files.ts",
   "clone e24148d569bb src/github/__tests__/installations.test.ts,src/orchestrator/__tests__/merge-check.test.ts",
   "clone e244829eb955 src/orchestrator/cli.ts,src/orchestrator/resume.ts",
-  "clone e29c9dd71a32 src/codeintel/__tests__/call-hierarchy.test.ts,src/codeintel/__tests__/find-references.test.ts,src/codeintel/__tests__/import-graph.test.ts",
   "clone e2f7c4138acf src/llm/muse.ts,src/llm/openrouter.ts",
   "clone e2f849efd7f2 src/cloud/docker-provider.ts,src/cloud/provider.ts",
   "clone e34ac968803e src/engine/__tests__/inject-message.test.ts,src/engine/__tests__/pause.test.ts",
@@ -875,16 +800,13 @@
   "clone e416c30b4843 src/llm/preflight.ts",
   "clone e498eb38bb29 src/orchestrator/resume.ts",
   "clone e4c0c632d2e8 src/engine/__tests__/condense.test.ts",
-  "clone e4cd97daa489 src/__tests__/cli-env.test.ts,src/codesearch/__tests__/embedder-types.test.ts,src/evidence/__tests__/verify-files.test.ts,src/orchestrator/__tests__/task-files.test.ts,src/orchestrator/__tests__/terminal-status.test.ts,src/tools/__tests__/batch-types.test.ts,src/tools/__tests__/paths.test.ts",
   "clone e4f3cf1fa942 src/rsi/__tests__/controller-candidates.test.ts,src/rsi/__tests__/controller-fleet-results.test.ts",
   "clone e5b435a1ec6d src/llm/openrouter.ts",
   "clone e5c94c55dadd src/orchestrator/__tests__/spawn-model.test.ts,src/orchestrator/__tests__/spawn-state-reset.test.ts",
   "clone e5e1bcaff8b1 src/orchestrator/__tests__/spawn-model.test.ts,src/orchestrator/__tests__/spawn-state-reset.test.ts",
   "clone e627e238091f src/engine/__tests__/condense.test.ts",
   "clone e670cea95cf5 src/llm/muse.ts",
-  "clone e69e899590f4 src/checkpoints/__tests__/cli.test.ts,src/codemap/__tests__/cli.test.ts,src/codemap/__tests__/units.test.ts,src/dashboard/__tests__/cli.test.ts,src/dashboard/__tests__/page-route.test.ts,src/dashboard/__tests__/self-improvement-metrics.test.ts,src/dashboard/__tests__/sigint-shutdown.test.ts,src/dashboard/__tests__/trend.test.ts,src/github/__tests__/installations.test.ts,src/orchestrator/__tests__/analyze-cli.test.ts,src/orchestrator/__tests__/cost-estimate.test.ts,src/orchestrator/__tests__/log-analysis.test.ts,src/orchestrator/__tests__/spawn-state-reset.test.ts,src/orchestrator/__tests__/status.test.ts,src/run-tests-script.test.ts,src/tools/__tests__/set-indentation.test.ts",
   "clone e6c233c3ca88 src/llm/muse.ts,src/llm/openrouter.ts",
-  "clone e76f1d15c039 src/tools/__tests__/apply_diff.test.ts,src/tools/__tests__/edit_file.test.ts,src/tools/__tests__/search_replace.test.ts",
   "clone e7c40e42987c src/orchestrator/task-files.ts",
   "clone e8c2bd5a02ed src/engine/__tests__/new-task.test.ts,src/engine/__tests__/switch-mode.test.ts",
   "clone e8c3d99ff835 src/engine/__tests__/inject-message.test.ts,src/engine/__tests__/pause.test.ts",
@@ -913,12 +835,10 @@
   "clone ef26fdd9523d src/llm/preflight.ts",
   "clone ef55d3fea271 src/dashboard/__tests__/cost-history-routes.test.ts,src/orchestrator/__tests__/cost-history-cli.test.ts",
   "clone ef70ff0a9386 src/engine/local-explore.ts,src/llm/ollama.ts",
-  "clone ef71bbe16ec5 src/codeintel/__tests__/registration.test.ts,src/tools/__tests__/batch.test.ts",
   "clone f01f54f65b50 src/codesearch/files.ts,src/engine/stacks.ts",
   "clone f075d1e0c08a src/orchestrator/resume.ts",
   "clone f0ce557b48e0 src/watcher/watch.ts",
   "clone f0f979cf4060 src/llm/muse.ts,src/llm/openrouter.ts",
-  "clone f0fbb3826fed src/codeintel/__tests__/registration.test.ts,src/tools/__tests__/batch.test.ts",
   "clone f14b65dc96ac src/engine/__tests__/loop.test.ts,src/tools/__tests__/execute_command.test.ts",
   "clone f22d61a091e1 src/engine/__tests__/condense.test.ts",
   "clone f2684b8a8b7a src/permissions/commands.ts",
@@ -936,10 +856,10 @@
   "clone f602bc3ce671 src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/resume.test.ts",
   "clone f687d6368f68 src/llm/muse.ts",
   "clone f69d991fe89e src/orchestrator/__tests__/cli.test.ts",
-  "clone f6a98fae1855 src/engine/__tests__/new-task.test.ts,src/engine/__tests__/switch-mode.test.ts",
   "clone f6c0de337c8b src/orchestrator/resume.ts",
   "clone f6ed5475b786 src/orchestrator/task-files.ts",
   "clone f730e676e214 src/orchestrator/task-files.ts",
+  "clone f7e01a037e17 src/codesearch/__tests__/backend-selection.test.ts,src/codesearch/__tests__/index.test.ts,src/codesearch/__tests__/search.test.ts,src/config/__tests__/mode-models.test.ts,src/dashboard/__tests__/file-browser-permissions.test.ts,src/dashboard/__tests__/settings.test.ts,src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/cost-history-cli.test.ts,src/orchestrator/__tests__/cost-history.test.ts,src/orchestrator/__tests__/model-resolution.test.ts,src/orchestrator/__tests__/resume.test.ts,src/orchestrator/__tests__/spawn-model.test.ts,src/orchestrator/__tests__/watch.test.ts,src/permissions/__tests__/protected-files.test.ts,src/tools/__tests__/codebase_search.test.ts",
   "clone f83833b6f3b1 src/engine/loop.ts",
   "clone f876220170d5 src/orchestrator/__tests__/analyze-cli.test.ts",
   "clone f8a191f41b2a src/orchestrator/__tests__/cli.test.ts,src/orchestrator/__tests__/resume.test.ts",
@@ -953,7 +873,6 @@
   "clone fa817379580e src/checkpoints/__tests__/cli.test.ts,src/codemap/__tests__/cli.test.ts,src/dashboard/__tests__/cli.test.ts",
   "clone facb47dc5d4e src/llm/muse.ts,src/llm/openrouter.ts",
   "clone faf040a5922d src/orchestrator/task-files.ts",
-  "clone fbc6a19ebf9a src/__tests__/cli-env.test.ts,src/codesearch/__tests__/embedder-types.test.ts,src/evidence/__tests__/verify-files.test.ts,src/orchestrator/__tests__/task-files.test.ts,src/orchestrator/__tests__/terminal-status.test.ts,src/tools/__tests__/batch-types.test.ts,src/tools/__tests__/paths.test.ts",
   "clone fc018e6a7371 src/orchestrator/__tests__/cli.test.ts",
   "clone fc35fa2848fb src/codesearch/files.ts,src/engine/stacks.ts",
   "clone fc4aa9a9577b src/dashboard/__tests__/projects-routes.test.ts,src/projects/__tests__/cli.test.ts",
@@ -1025,9 +944,6 @@
   "src/codemap/extract.ts console 1e0b59d485f1",
   "src/codemap/extract.ts console bece561cd9ec",
   "src/codemap/flows.ts complexity 8005c8ba0de4",
-  "src/codesearch/__tests__/backend-selection.test.ts nesting 5a1db74d0b54",
-  "src/codesearch/__tests__/index.test.ts nesting 5a1db74d0b54",
-  "src/codesearch/__tests__/search.test.ts nesting 5a1db74d0b54",
   "src/codesearch/airunner-embedder.ts complexity 0fdf477e9e4b",
   "src/codesearch/chunk.ts nesting 0521a2813b3b",
   "src/codesearch/cli.ts complexity 0f0bb1f16669",
@@ -1040,7 +956,6 @@
   "src/codesearch/index.ts complexity f2fd182e4c58",
   "src/codesearch/lexical.ts complexity 7f089b589f5d",
   "src/codesearch/ollama-embedder.ts complexity 0fdf477e9e4b",
-  "src/config/__tests__/mode-models.test.ts nesting 5a1db74d0b54",
   "src/config/mode-models.ts complexity 54346a7c8ba3",
   "src/config/mode-models.ts complexity 81355004b185",
   "src/config/mode-models.ts nesting 54346a7c8ba3",
@@ -1050,8 +965,6 @@
   "src/control/emergency-controller.ts nesting 71308ea5154c",
   "src/control/session-runtime.ts complexity d35d302346c1",
   "src/control/session-runtime.ts nesting d35d302346c1",
-  "src/dashboard/__tests__/file-browser-permissions.test.ts nesting 5a1db74d0b54",
-  "src/dashboard/__tests__/settings.test.ts nesting 5a1db74d0b54",
   "src/dashboard/chat-thread.ts complexity 7a1720a83e64",
   "src/dashboard/checkpoints.ts complexity 80236917ad55",
   "src/dashboard/cli.ts complexity b22fbbebd876",
@@ -1185,14 +1098,7 @@
   "src/memory/__tests__/uwuchat.test.ts nesting 2d2e45536b71",
   "src/memory/summarizer.ts nesting 7cee04a9645b",
   "src/monitoring/features.ts complexity 24d38f3a6c14",
-  "src/orchestrator/__tests__/cli.test.ts nesting 5a1db74d0b54",
-  "src/orchestrator/__tests__/cost-history-cli.test.ts nesting 5a1db74d0b54",
-  "src/orchestrator/__tests__/cost-history.test.ts nesting 5a1db74d0b54",
-  "src/orchestrator/__tests__/model-resolution.test.ts nesting 5a1db74d0b54",
   "src/orchestrator/__tests__/muse-executor.test.ts complexity 8e2ad4b546b7",
-  "src/orchestrator/__tests__/resume.test.ts nesting 5a1db74d0b54",
-  "src/orchestrator/__tests__/spawn-model.test.ts nesting 5a1db74d0b54",
-  "src/orchestrator/__tests__/watch.test.ts nesting 5a1db74d0b54",
   "src/orchestrator/analyze-cli.ts complexity aa16ab05bcec",
   "src/orchestrator/analyze-cli.ts nesting aa16ab05bcec",
   "src/orchestrator/cleanup.ts complexity 192b651b4c9a",
@@ -1249,7 +1155,6 @@
   "src/orchestrator/watch.ts complexity 4b74c496e627",
   "src/orchestrator/watch.ts complexity 5ef3a7678de8",
   "src/orchestrator/watch.ts complexity 73c5129596ac",
-  "src/permissions/__tests__/protected-files.test.ts nesting 5a1db74d0b54",
   "src/permissions/commands.ts complexity 603320ccaacb",
   "src/permissions/commands.ts complexity 88bfec36eee2",
   "src/permissions/commands.ts complexity a59b0df8679a",
@@ -1503,8 +1408,14 @@
   "src/terminal/chat.ts nesting a0d2a773413a",
   "src/terminal/input.ts complexity 12599998b43a",
   "src/terminal/renderer.ts complexity 7f5ba673c15f",
+  "src/test-runner.ts console 820c4507ed73",
+  "src/test-runner.ts console 8d8544d9af06",
+  "src/test-runner.ts console a8f563293e4f",
+  "src/test-runner.ts console c770bfde2dce",
+  "src/test-runner.ts console e17350daf90f",
+  "src/test-runner.ts console e450d263e401",
+  "src/test-runner.ts process-exit cbde29b17391",
   "src/tools/__tests__/browser-close-backstop.test.ts complexity d96a5159e95f",
-  "src/tools/__tests__/codebase_search.test.ts nesting 5a1db74d0b54",
   "src/tools/browser/handler.ts complexity 564b954e5e1c",
   "src/tools/browser/handler.ts nesting 564b954e5e1c",
   "src/tools/browser/service.ts complexity 79f179fd0b89",
diff --git a/src/__tests__/cli-env.test.ts b/src/__tests__/cli-env.test.ts
index 552d216..59eed41 100644
--- a/src/__tests__/cli-env.test.ts
+++ b/src/__tests__/cli-env.test.ts
@@ -19,6 +19,7 @@ import {
 	envBooleanValue as cliEnvBooleanValue,
 	resolvePerModeEnv as cliResolvePerModeEnv,
 } from "../cli.js"
+import { runMain } from "../test-runner.js"
 
 async function testPerModeOverrideWins(): Promise<void> {
 	const env = { HEADLESSCODE_FOO: "global", HEADLESSCODE_FOO__CODE: "per-mode" }
@@ -60,40 +61,13 @@ async function testCliReexportsTheLeaf(): Promise<void> {
 	assert.equal(cliEnvBoolean, envBoolean)
 }
 
-async function runOne(name: string, fn: () => Promise<void>): Promise<boolean> {
-	try {
-		await fn()
-		console.log(`  ok   ${name}`)
-		return true
-	} catch (error) {
-		console.error(`  FAIL ${name}`)
-		if (error instanceof Error) console.error(error.stack ?? error)
-		else console.error(String(error))
-		return false
-	}
-}
-
-async function main(): Promise<void> {
-	const tests: Array<[string, () => Promise<void>]> = [
-		["per-mode override wins", testPerModeOverrideWins],
-		["global falls back", testGlobalFallsBack],
-		["mode slug normalization", testModeSlugNormalization],
-		["missing is undefined", testMissingIsUndefined],
-		["env boolean values", testEnvBooleanValues],
-		["cli re-exports the leaf", testCliReexportsTheLeaf],
-	]
-	let failed = 0
-	for (const [name, fn] of tests) {
-		if (!(await runOne(name, fn))) failed++
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} cli-env tests passed`)
-}
+const tests: Array<[string, () => Promise<void>]> = [
+	["per-mode override wins", testPerModeOverrideWins],
+	["global falls back", testGlobalFallsBack],
+	["mode slug normalization", testModeSlugNormalization],
+	["missing is undefined", testMissingIsUndefined],
+	["env boolean values", testEnvBooleanValues],
+	["cli re-exports the leaf", testCliReexportsTheLeaf],
+]
 
-main().catch((error) => {
-	console.error("Test runner crashed:", error)
-	process.exit(1)
-})
+runMain("cli-env", tests)
diff --git a/src/budget/__tests__/budget.test.ts b/src/budget/__tests__/budget.test.ts
index cdbf6ab..044d3e2 100644
--- a/src/budget/__tests__/budget.test.ts
+++ b/src/budget/__tests__/budget.test.ts
@@ -19,6 +19,7 @@ import {
 	type SessionBudget,
 } from "../budget.js"
 import type { PricingTable } from "../cost.js"
+import { runMain } from "../../test-runner.js"
 
 // Cheap model for deterministic math: $1/1M in, $2/1M out → 1000 in + 1000 out
 // = (1000*1 + 1000*2)/1e6 = $0.003 per call.
@@ -346,26 +347,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["session: iterations cap trips before the next LLM call", testSessionBudgetIterationsCap],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} budget tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("budget", tests)
diff --git a/src/budget/__tests__/concurrency.test.ts b/src/budget/__tests__/concurrency.test.ts
index f340746..9d65778 100644
--- a/src/budget/__tests__/concurrency.test.ts
+++ b/src/budget/__tests__/concurrency.test.ts
@@ -24,6 +24,7 @@ import { saveWatcherState, type WatcherState } from "../../watcher/state.js"
 import { watchIssues, type SpawnBatch, type WatcherConfig } from "../../watcher/watch.js"
 import { loadWatcherState } from "../../watcher/state.js"
 import type { GhClient, GitHubIssue } from "../../watcher/github.js"
+import { runMain } from "../../test-runner.js"
 
 const SILENT = new Logger({ level: "silent" })
 
@@ -260,26 +261,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["watcher: pre-existing running fleet at cap -> all new issues deferred", testWatcherGlobalCapRespectsActiveFleet],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} concurrency tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("concurrency", tests)
diff --git a/src/budget/__tests__/cost.test.ts b/src/budget/__tests__/cost.test.ts
index 69578b2..11101ef 100644
--- a/src/budget/__tests__/cost.test.ts
+++ b/src/budget/__tests__/cost.test.ts
@@ -21,6 +21,7 @@ import {
 	type EndpointPricingEntry,
 	type PricingTable,
 } from "../cost.js"
+import { runMain } from "../../test-runner.js"
 
 /**
  * Regression test for a real incident (2026-08-01): `deepseek/deepseek-v4-flash`
@@ -356,26 +357,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["mergeLivePrice overrides the hardcoded entry, feeds estimateCost", testMergeLivePriceOverridesDefaults],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} cost tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("cost", tests)
diff --git a/src/checkpoints/__tests__/cli.test.ts b/src/checkpoints/__tests__/cli.test.ts
index 249f50c..9ae7759 100644
--- a/src/checkpoints/__tests__/cli.test.ts
+++ b/src/checkpoints/__tests__/cli.test.ts
@@ -16,6 +16,7 @@ import * as path from "node:path"
 
 import { checkpointsMain, parseCheckpointsArgs } from "../cli.js"
 import { createCheckpointService } from "../service.js"
+import { runMain } from "../../test-runner.js"
 
 async function tmpRepo(prefix = "hc-ckpt-cli-"): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -223,26 +224,4 @@ const tests: Array<[string, () => void | Promise<void>]> = [
 	["checkpointsMain: diff against the working tree exits 0", testMainDiffAgainstWorkingTree],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? (err.stack ?? err.message) : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} checkpoints CLI tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("checkpoints CLI", tests)
diff --git a/src/checkpoints/__tests__/service.test.ts b/src/checkpoints/__tests__/service.test.ts
index e171dc3..5e4b0ea 100644
--- a/src/checkpoints/__tests__/service.test.ts
+++ b/src/checkpoints/__tests__/service.test.ts
@@ -15,6 +15,7 @@ import * as os from "node:os"
 import * as path from "node:path"
 
 import { createCheckpointService } from "../service.js"
+import { runMain } from "../../test-runner.js"
 
 async function mkTmpWorkspace(prefix: string): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -198,26 +199,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["diff: reports before/after content for a changed file", testDiffReportsChangedFile],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} checkpoints tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("checkpoints", tests)
diff --git a/src/cloud/__tests__/docker-provider.test.ts b/src/cloud/__tests__/docker-provider.test.ts
index 0c96b98..bb340eb 100644
--- a/src/cloud/__tests__/docker-provider.test.ts
+++ b/src/cloud/__tests__/docker-provider.test.ts
@@ -32,6 +32,7 @@ import * as path from "node:path"
 
 import { DockerSessionProvider } from "../docker-provider.js"
 import type { CloudSessionRequest, CommandResult, SessionHandle } from "../provider.js"
+import { runSuites } from "../../test-runner.js"
 
 // ─── Docker availability probe ───────────────────────────────────────────────
 
@@ -302,18 +303,8 @@ if (probe.ok) {
 
 async function main(): Promise<void> {
 	console.log(`  note: ${dockerNote}`)
-	let failed = 0
 	let skipped = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
+	const failed = await runSuites("docker-provider", tests)
 	if (!probe.ok) {
 		skipped = 2
 		console.log(`  SKIP ${skipped} real-daemon test(s) — ${probe.reason}`)
@@ -322,7 +313,7 @@ async function main(): Promise<void> {
 		console.error(`\n${failed} test(s) failed`)
 		process.exit(1)
 	}
-	console.log(`\nAll ${tests.length} docker-provider tests passed (${skipped} skipped: no Docker daemon)`)
+	console.log(`  note: ${skipped} skipped (no Docker daemon)`)
 }
 
 main().catch((err) => {
diff --git a/src/cloud/__tests__/provider.test.ts b/src/cloud/__tests__/provider.test.ts
index 4746bee..822b9a0 100644
--- a/src/cloud/__tests__/provider.test.ts
+++ b/src/cloud/__tests__/provider.test.ts
@@ -26,6 +26,7 @@ import {
 	type CommandResult,
 	type SessionHandle,
 } from "../provider.js"
+import { runMain } from "../../test-runner.js"
 
 async function tmpRepo(): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), "headlesscode-provider-repo-"))
@@ -188,26 +189,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["HetznerDockerProvider is a documented sketch (not implemented)", testHetznerSketchIsDocumentedNotImplemented],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} provider tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("provider", tests)
diff --git a/src/codeintel/__tests__/call-graph.test.ts b/src/codeintel/__tests__/call-graph.test.ts
index 8df7eed..e9626ba 100644
--- a/src/codeintel/__tests__/call-graph.test.ts
+++ b/src/codeintel/__tests__/call-graph.test.ts
@@ -18,6 +18,7 @@ import * as path from "node:path"
 
 import { getCallGraph } from "../call-graph.js"
 import { getCodeIntelCache, resetCodeIntelCaches } from "../program.js"
+import { runMain } from "../../test-runner.js"
 
 const FIXTURE_A = `import { greet, Greeter } from "./b.js"
 
@@ -101,30 +102,8 @@ async function testCallTargetsResolveToDefiningFile(): Promise<void> {
 	}
 }
 
-async function main(): Promise<void> {
-	const tests: Array<[string, () => Promise<void>]> = [
-		["call targets resolve to their defining file via the TS checker", testCallTargetsResolveToDefiningFile],
-	]
-
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} call-graph tests passed`)
-}
+const tests: Array<[string, () => Promise<void>]> = [
+	["call targets resolve to their defining file via the TS checker", testCallTargetsResolveToDefiningFile],
+]
 
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("call-graph", tests)
diff --git a/src/codeintel/__tests__/call-hierarchy.test.ts b/src/codeintel/__tests__/call-hierarchy.test.ts
index f8a98c1..26f48bf 100644
--- a/src/codeintel/__tests__/call-hierarchy.test.ts
+++ b/src/codeintel/__tests__/call-hierarchy.test.ts
@@ -18,6 +18,7 @@ import * as path from "node:path"
 import type { ToolContext } from "../../engine/types.js"
 import { resetCodeIntelCaches } from "../program.js"
 import { callHierarchyHandler } from "../call-hierarchy.js"
+import { runMain } from "../../test-runner.js"
 
 // Line numbers below are 1-based and VERIFIED against the fixture text:
 // leaf=1, mid=5 (calls leaf on 6 and 7), top=10 (calls mid on 11),
@@ -202,37 +203,15 @@ async function testNonTsFile(): Promise<void> {
 	}
 }
 
-async function main(): Promise<void> {
-	const tests: Array<[string, () => Promise<void>]> = [
-		["direct caller found from a definition", testDirectCallerFound],
-		["caller lookup works from a call site", testCallerViaCallSite],
-		["direct callees found", testDirectCalleesFound],
-		["callee lookup from a call site starts at the selected callee", testCalleesFromCallSiteStartAtSelectedCallee],
-		["default direction returns both sides", testBothDirectionsDefault],
-		["max_depth=1 excludes depth-2 edges", testMaxDepthPrunesDeeperEdges],
-		["any-typed call is tolerated", testAnyTypedCallTolerated],
-		["non-TS file returns the not-supported result", testNonTsFile],
-	]
-
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} call_hierarchy tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+const tests: Array<[string, () => Promise<void>]> = [
+	["direct caller found from a definition", testDirectCallerFound],
+	["caller lookup works from a call site", testCallerViaCallSite],
+	["direct callees found", testDirectCalleesFound],
+	["callee lookup from a call site starts at the selected callee", testCalleesFromCallSiteStartAtSelectedCallee],
+	["default direction returns both sides", testBothDirectionsDefault],
+	["max_depth=1 excludes depth-2 edges", testMaxDepthPrunesDeeperEdges],
+	["any-typed call is tolerated", testAnyTypedCallTolerated],
+	["non-TS file returns the not-supported result", testNonTsFile],
+]
+
+runMain("call_hierarchy", tests)
diff --git a/src/codeintel/__tests__/find-references.test.ts b/src/codeintel/__tests__/find-references.test.ts
index 0c194b0..4695d79 100644
--- a/src/codeintel/__tests__/find-references.test.ts
+++ b/src/codeintel/__tests__/find-references.test.ts
@@ -23,6 +23,7 @@ import * as path from "node:path"
 import { createHeadlessExecutor } from "../../tools/executor.js"
 import { resetCodeIntelCaches } from "../program.js"
 import { MAX_REFERENCES } from "../find-references.js"
+import { runMain } from "../../test-runner.js"
 
 const FIXTURE_UTIL = `export function util(): number {
 	return 1
@@ -139,33 +140,11 @@ async function testNonTsFile(): Promise<void> {
 	}
 }
 
-async function main(): Promise<void> {
-	const tests: Array<[string, () => Promise<void>]> = [
-		["returns exact reference sites for a symbol with few usages", testExactReferenceSites],
-		["caps output on a symbol with many usages and announces the overflow", testCapOnManyUsages],
-		["no-resolvable-references gets a clear message", testNoReferences],
-		["non-TS file returns the not-supported result", testNonTsFile],
-	]
-
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} find_references tests passed`)
-}
+const tests: Array<[string, () => Promise<void>]> = [
+	["returns exact reference sites for a symbol with few usages", testExactReferenceSites],
+	["caps output on a symbol with many usages and announces the overflow", testCapOnManyUsages],
+	["no-resolvable-references gets a clear message", testNoReferences],
+	["non-TS file returns the not-supported result", testNonTsFile],
+]
 
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("find_references", tests)
diff --git a/src/codeintel/__tests__/go-to-definition.test.ts b/src/codeintel/__tests__/go-to-definition.test.ts
index 9f426ad..8e73c90 100644
--- a/src/codeintel/__tests__/go-to-definition.test.ts
+++ b/src/codeintel/__tests__/go-to-definition.test.ts
@@ -22,6 +22,7 @@ import * as path from "node:path"
 
 import { createHeadlessExecutor } from "../../tools/executor.js"
 import { resetCodeIntelCaches } from "../program.js"
+import { runMain } from "../../test-runner.js"
 
 // Line numbers below are 1-based and VERIFIED against the fixture text:
 // add=6, PI=10, first overload=18, second overload=19, loose=24,
@@ -216,36 +217,14 @@ async function testMissingFile(): Promise<void> {
 	}
 }
 
-async function main(): Promise<void> {
-	const tests: Array<[string, () => Promise<void>]> = [
-		["plain function resolves to its declaration", testPlainFunctionDefinition],
-		["character disambiguates symbols on one line", testCharacterDisambiguation],
-		["definition in a .d.ts via tsconfig paths alias resolves", testDefinitionInDtsThroughPathsAlias],
-		["overloaded function returns ALL definitions", testOverloadReturnsAllDefinitions],
-		["unresolvable (any-typed) symbol gets a clear message", testUnresolvableSymbol],
-		["non-TS file returns the not-supported result", testNonTsFile],
-		["missing file errors clearly", testMissingFile],
-	]
-
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} go_to_definition tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+const tests: Array<[string, () => Promise<void>]> = [
+	["plain function resolves to its declaration", testPlainFunctionDefinition],
+	["character disambiguates symbols on one line", testCharacterDisambiguation],
+	["definition in a .d.ts via tsconfig paths alias resolves", testDefinitionInDtsThroughPathsAlias],
+	["overloaded function returns ALL definitions", testOverloadReturnsAllDefinitions],
+	["unresolvable (any-typed) symbol gets a clear message", testUnresolvableSymbol],
+	["non-TS file returns the not-supported result", testNonTsFile],
+	["missing file errors clearly", testMissingFile],
+]
+
+runMain("go_to_definition", tests)
diff --git a/src/codeintel/__tests__/hover.test.ts b/src/codeintel/__tests__/hover.test.ts
index 1a44c7d..3f062e3 100644
--- a/src/codeintel/__tests__/hover.test.ts
+++ b/src/codeintel/__tests__/hover.test.ts
@@ -15,6 +15,7 @@ import * as path from "node:path"
 import type { ToolContext } from "../../engine/types.js"
 import { typeAtPositionHandler } from "../hover.js"
 import { resetCodeIntelCaches } from "../program.js"
+import { runMain } from "../../test-runner.js"
 
 // Line numbers below are 1-based and VERIFIED against the fixture text:
 // lib.ts: doc=1-6, add=7, greeting=11. main.ts: import=1, sum=3, shout=5.
@@ -151,37 +152,15 @@ async function testNoIdentifierGraceful(): Promise<void> {
 	}
 }
 
-async function main(): Promise<void> {
-	const tests: Array<[string, () => Promise<void>]> = [
-		["inferred-type const shows its type text", testInferredConstType],
-		["function use shows the full signature", testFunctionSignature],
-		[
-			"import specifier resolves through the import",
-			testImportSpecifierResolves,
-		],
-		["docs and JSDoc tags follow the signature", testDocsAndTags],
-		["blank spot yields the graceful message", testNoIdentifierGraceful],
-	]
-
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} type_at_position tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+const tests: Array<[string, () => Promise<void>]> = [
+	["inferred-type const shows its type text", testInferredConstType],
+	["function use shows the full signature", testFunctionSignature],
+	[
+		"import specifier resolves through the import",
+		testImportSpecifierResolves,
+	],
+	["docs and JSDoc tags follow the signature", testDocsAndTags],
+	["blank spot yields the graceful message", testNoIdentifierGraceful],
+]
+
+runMain("type_at_position", tests)
diff --git a/src/codeintel/__tests__/implementations.test.ts b/src/codeintel/__tests__/implementations.test.ts
index 878e7ec..6c337eb 100644
--- a/src/codeintel/__tests__/implementations.test.ts
+++ b/src/codeintel/__tests__/implementations.test.ts
@@ -19,6 +19,7 @@ import * as path from "node:path"
 import type { ToolContext } from "../../engine/types.js"
 import { resetCodeIntelCaches } from "../program.js"
 import { findImplementationsHandler } from "../implementations.js"
+import { runMain } from "../../test-runner.js"
 
 // Line numbers below are 1-based and VERIFIED against the fixture text:
 // Shape=1, area=2, Circle=5, Circle.area=6, Square=11, Square.area=12,
@@ -198,35 +199,12 @@ async function testNonTsFile(): Promise<void> {
 	}
 }
 
-async function main(): Promise<void> {
-	const tests: Array<[string, () => Promise<void>]> = [
-		["interface method finds class methods", testInterfaceMethodImplementations],
-		["base method finds the subclass override", testSubclassOverride],
-		["empty spot gets the graceful message", testNoImplementationsMessage],
-		["symbol arg disambiguates one line", testSymbolArgDisambiguation],
-		["non-TS file is refused, not guessed", testNonTsFile],
-	]
-
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			const detail = err instanceof Error ? err.stack ?? err.message : String(err)
-			console.error(detail)
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} find_implementations tests passed`)
-}
+const tests: Array<[string, () => Promise<void>]> = [
+	["interface method finds class methods", testInterfaceMethodImplementations],
+	["base method finds the subclass override", testSubclassOverride],
+	["empty spot gets the graceful message", testNoImplementationsMessage],
+	["symbol arg disambiguates one line", testSymbolArgDisambiguation],
+	["non-TS file is refused, not guessed", testNonTsFile],
+]
 
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("find_implementations", tests)
diff --git a/src/codeintel/__tests__/import-graph.test.ts b/src/codeintel/__tests__/import-graph.test.ts
index 1709ff5..04c5d4c 100644
--- a/src/codeintel/__tests__/import-graph.test.ts
+++ b/src/codeintel/__tests__/import-graph.test.ts
@@ -19,6 +19,7 @@ import * as path from "node:path"
 
 import { createHeadlessExecutor } from "../../tools/executor.js"
 import { getCodeIntelCache, resetCodeIntelCaches } from "../program.js"
+import { runMain } from "../../test-runner.js"
 
 const FIXTURE_A = `import { b1 } from "./b.js"
 import type { c1 } from "./c.js"
@@ -146,33 +147,11 @@ async function testNonTsFile(): Promise<void> {
 	}
 }
 
-async function main(): Promise<void> {
-	const tests: Array<[string, () => Promise<void>]> = [
-		["imports resolve to real file paths (incl. type-only and unresolved)", testImportsResolvedToRealPaths],
-		["reverse lookup sees type imports and re-exports", testReverseLookupIncludingReExportAndTypeImport],
-		["program cache is reused across calls and invalidated on file change", testProgramCacheReusedAcrossCalls],
-		["non-TS file returns the not-supported result", testNonTsFile],
-	]
-
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} import_graph tests passed`)
-}
+const tests: Array<[string, () => Promise<void>]> = [
+	["imports resolve to real file paths (incl. type-only and unresolved)", testImportsResolvedToRealPaths],
+	["reverse lookup sees type imports and re-exports", testReverseLookupIncludingReExportAndTypeImport],
+	["program cache is reused across calls and invalidated on file change", testProgramCacheReusedAcrossCalls],
+	["non-TS file returns the not-supported result", testNonTsFile],
+]
 
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("import_graph", tests)
diff --git a/src/codeintel/__tests__/outline.test.ts b/src/codeintel/__tests__/outline.test.ts
index 6d378d3..808852f 100644
--- a/src/codeintel/__tests__/outline.test.ts
+++ b/src/codeintel/__tests__/outline.test.ts
@@ -19,6 +19,7 @@ import * as path from "node:path"
 
 import { createHeadlessExecutor } from "../../tools/executor.js"
 import { resetCodeIntelCaches } from "../program.js"
+import { runMain } from "../../test-runner.js"
 
 const FIXTURE_MATH = `export interface Point {
 	x: number
@@ -155,34 +156,12 @@ async function testOutlinePathTraversalRejected(): Promise<void> {
 	}
 }
 
-async function main(): Promise<void> {
-	const tests: Array<[string, () => Promise<void>]> = [
-		["outline returns the exact fixture structure", testOutlineExactStructure],
-		["outline signatures exclude bodies and duplicate keywords", testOutlineSignatureHasNoBody],
-		["outline of a missing file errors", testOutlineMissingFile],
-		["outline of a non-TS file returns the not-supported error", testOutlineUnsupportedFileType],
-		["outline rejects paths escaping the workspace", testOutlinePathTraversalRejected],
-	]
-
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} outline tests passed`)
-}
+const tests: Array<[string, () => Promise<void>]> = [
+	["outline returns the exact fixture structure", testOutlineExactStructure],
+	["outline signatures exclude bodies and duplicate keywords", testOutlineSignatureHasNoBody],
+	["outline of a missing file errors", testOutlineMissingFile],
+	["outline of a non-TS file returns the not-supported error", testOutlineUnsupportedFileType],
+	["outline rejects paths escaping the workspace", testOutlinePathTraversalRejected],
+]
 
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("outline", tests)
diff --git a/src/codeintel/__tests__/registration.test.ts b/src/codeintel/__tests__/registration.test.ts
index 2b94874..afc3aa5 100644
--- a/src/codeintel/__tests__/registration.test.ts
+++ b/src/codeintel/__tests__/registration.test.ts
@@ -19,6 +19,7 @@ import * as path from "node:path"
 import { createHeadlessExecutor } from "../../tools/executor.js"
 import { resetCodeIntelCaches } from "../program.js"
 import { CODE_INTEL_TOOL_NAMES } from "../tools.js"
+import { runMain } from "../../test-runner.js"
 
 // Line numbers below are 1-based and VERIFIED against the fixture text:
 // Shape=1, area=2, Circle=5, Circle.area=6, double=11, quadruple=15
@@ -185,45 +186,20 @@ async function testAdvertisedInToolList(): Promise<void> {
 	}
 }
 
-async function main(): Promise<void> {
-	const tests: Array<[string, () => Promise<void>]> = [
-		["all five tools are registered", testAllRegistered],
-		["get_diagnostics runs via executor", testDiagnosticsThroughExecutor],
-		["type_at_position runs via executor", testHoverThroughExecutor],
-		[
-			"workspace_symbol runs via executor",
-			testWorkspaceSymbolThroughExecutor,
-		],
-		[
-			"find_implementations runs via executor",
-			testImplementationsThroughExecutor,
-		],
-		["call_hierarchy runs via executor", testCallHierarchyThroughExecutor],
-		["all five are in CODE_INTEL_TOOLS", testAdvertisedInToolList],
-	]
-
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			const detail = err instanceof Error
-				? (err.stack ?? err.message)
-				: String(err)
-			console.error(detail)
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} registration tests passed`)
-}
+const tests: Array<[string, () => Promise<void>]> = [
+	["all five tools are registered", testAllRegistered],
+	["get_diagnostics runs via executor", testDiagnosticsThroughExecutor],
+	["type_at_position runs via executor", testHoverThroughExecutor],
+	[
+		"workspace_symbol runs via executor",
+		testWorkspaceSymbolThroughExecutor,
+	],
+	[
+		"find_implementations runs via executor",
+		testImplementationsThroughExecutor,
+	],
+	["call_hierarchy runs via executor", testCallHierarchyThroughExecutor],
+	["all five are in CODE_INTEL_TOOLS", testAdvertisedInToolList],
+]
 
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("registration", tests)
diff --git a/src/codeintel/__tests__/rename-symbol.test.ts b/src/codeintel/__tests__/rename-symbol.test.ts
index 90e0e9e..a507615 100644
--- a/src/codeintel/__tests__/rename-symbol.test.ts
+++ b/src/codeintel/__tests__/rename-symbol.test.ts
@@ -21,6 +21,7 @@ import * as path from "node:path"
 import { createHeadlessExecutor } from "../../tools/executor.js"
 import { getCodeIntelCache, resetCodeIntelCaches } from "../program.js"
 import { renameSymbol } from "../rename-symbol.js"
+import { runMain } from "../../test-runner.js"
 
 const FIXTURE_UTIL = `export function util(): number {
 	return 1
@@ -189,33 +190,11 @@ async function testMidWriteFailureRollsBackWrittenFiles(): Promise<void> {
 	}
 }
 
-async function main(): Promise<void> {
-	const tests: Array<[string, () => Promise<void>]> = [
-		["renames every real site across files and leaves lookalike identifiers untouched", testRenameAcrossAllSites],
-		["invalid new name is refused with nothing changed", testInvalidNewNameLeavesEverythingUntouched],
-		["stale snapshot aborts before any write (no partial edits)", testStaleSnapshotAbortsWithNoPartialWrites],
-		["mid-write failure rolls back the files already written", testMidWriteFailureRollsBackWrittenFiles],
-	]
+const tests: Array<[string, () => Promise<void>]> = [
+	["renames every real site across files and leaves lookalike identifiers untouched", testRenameAcrossAllSites],
+	["invalid new name is refused with nothing changed", testInvalidNewNameLeavesEverythingUntouched],
+	["stale snapshot aborts before any write (no partial edits)", testStaleSnapshotAbortsWithNoPartialWrites],
+	["mid-write failure rolls back the files already written", testMidWriteFailureRollsBackWrittenFiles],
+]
 
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} rename_symbol tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("rename_symbol", tests)
diff --git a/src/codemap/__tests__/build.test.ts b/src/codemap/__tests__/build.test.ts
index 7e62245..8528bca 100644
--- a/src/codemap/__tests__/build.test.ts
+++ b/src/codemap/__tests__/build.test.ts
@@ -30,6 +30,7 @@ import { execFileSync } from "node:child_process"
 import { buildCodemap } from "../build.js"
 import { codemapDir, codemapJsonPath, codemapLockPath, codemapHtmlPath, loadLock, loadCodemap } from "../lock.js"
 import { resetCodeIntelCaches } from "../../codeintel/program.js"
+import { runMain } from "../../test-runner.js"
 
 const TSCONFIG = JSON.stringify(
 	{
@@ -273,35 +274,13 @@ async function testHtmlIsSelfContained(): Promise<void> {
 	}
 }
 
-async function main(): Promise<void> {
-	const tests: Array<[string, () => Promise<void>]> = [
-		["TS/JS modules get resolved import edges + correct roles", testTypeScriptEdgesAndRoles],
-		["Python modules get AST-subprocess import edges", testPythonEdges],
-		["C/C++ modules get regex include edges", testCppIncludeEdges],
-		["TS/JS call edges resolve cross-module calls + entrypoint flows", testCallEdgesAndEntrypointFlows],
-		["regeneration is fingerprint-aware (no writes when unchanged)", testFingerprintAwareRegeneration],
-		["html visualizer is self-contained and embeds the map", testHtmlIsSelfContained],
-	]
+const tests: Array<[string, () => Promise<void>]> = [
+	["TS/JS modules get resolved import edges + correct roles", testTypeScriptEdgesAndRoles],
+	["Python modules get AST-subprocess import edges", testPythonEdges],
+	["C/C++ modules get regex include edges", testCppIncludeEdges],
+	["TS/JS call edges resolve cross-module calls + entrypoint flows", testCallEdgesAndEntrypointFlows],
+	["regeneration is fingerprint-aware (no writes when unchanged)", testFingerprintAwareRegeneration],
+	["html visualizer is self-contained and embeds the map", testHtmlIsSelfContained],
+]
 
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} codemap build tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("codemap build", tests)
diff --git a/src/codemap/__tests__/cli.test.ts b/src/codemap/__tests__/cli.test.ts
index af64c93..74861cf 100644
--- a/src/codemap/__tests__/cli.test.ts
+++ b/src/codemap/__tests__/cli.test.ts
@@ -23,6 +23,7 @@ import { execFileSync } from "node:child_process"
 import { codemapMain, parseCodemapArgs } from "../cli.js"
 import { codemapJsonPath } from "../lock.js"
 import { resetCodeIntelCaches } from "../../codeintel/program.js"
+import { runMain } from "../../test-runner.js"
 
 function gitInit(root: string): void {
 	execFileSync("git", ["init", "-q"], { cwd: root, stdio: "ignore" })
@@ -157,26 +158,4 @@ const tests: Array<[string, () => void | Promise<void>]> = [
 	["codemapMain: --force always regenerates", testMainForceAlwaysRegenerates],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? (err.stack ?? err.message) : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} codemap CLI tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("codemap CLI", tests)
diff --git a/src/codemap/__tests__/units.test.ts b/src/codemap/__tests__/units.test.ts
index 3f65a5f..4d5d2bb 100644
--- a/src/codemap/__tests__/units.test.ts
+++ b/src/codemap/__tests__/units.test.ts
@@ -21,6 +21,7 @@ import { codemapFingerprint, contentHash, fingerprintMap, lockMatches } from "..
 import { computeEntrypointFlows } from "../flows.js"
 import { classifyRole, languageForExtension, statLines } from "../files.js"
 import type { Codemap, CodemapLock, ModuleEntry } from "../types.js"
+import { runMain } from "../../test-runner.js"
 
 // ─── fingerprint.ts ──────────────────────────────────────────────────────────
 
@@ -207,26 +208,4 @@ const tests: Array<[string, () => void | Promise<void>]> = [
 	["files: statLines counts bytes/lines and degrades to zeros on a missing file", testStatLinesCountsLinesAndBytes],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? (err.stack ?? err.message) : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} codemap unit tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("codemap unit", tests)
diff --git a/src/codesearch/__tests__/airunner-embedder.test.ts b/src/codesearch/__tests__/airunner-embedder.test.ts
index 333e420..da2df18 100644
--- a/src/codesearch/__tests__/airunner-embedder.test.ts
+++ b/src/codesearch/__tests__/airunner-embedder.test.ts
@@ -27,6 +27,7 @@ import {
 	resolveAirunnerEmbeddingModel,
 	resolveAirunnerUrl,
 } from "../airunner-embedder.js"
+import { runMain } from "../../test-runner.js"
 
 /** Deterministic fake vector for a text (dim 8), so tests can assert content. */
 function fakeVector(text: string): number[] {
@@ -239,26 +240,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["AIRunner URL/model env vars resolve with sensible defaults", testEnvResolution],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} airunner-embedder tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("airunner-embedder", tests)
diff --git a/src/codesearch/__tests__/backend-selection.test.ts b/src/codesearch/__tests__/backend-selection.test.ts
index 520847b..be5a83e 100644
--- a/src/codesearch/__tests__/backend-selection.test.ts
+++ b/src/codesearch/__tests__/backend-selection.test.ts
@@ -33,6 +33,7 @@ import { AirunnerEmbedder } from "../airunner-embedder.js"
 import { OllamaEmbedder } from "../ollama-embedder.js"
 import { OpenRouterEmbedder } from "../embedder.js"
 import { createHeadlessExecutor } from "../../tools/executor.js"
+import { runSuites } from "../../test-runner.js"
 
 async function mkTmpWorkspace(prefix: string): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -250,16 +251,7 @@ async function main(): Promise<void> {
 	process.env.HEADLESSCODE_DATA_DIR = storeTmp
 	let failed = 0
 	try {
-		for (const [name, fn] of tests) {
-			try {
-				await fn()
-				console.log(`  ok   ${name}`)
-			} catch (err) {
-				failed++
-				console.error(`  FAIL ${name}`)
-				console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-			}
-		}
+		failed = await runSuites("backend-selection", tests)
 	} finally {
 		delete process.env.HEADLESSCODE_DATA_DIR
 		await fs.rm(storeTmp, { recursive: true, force: true })
@@ -268,7 +260,6 @@ async function main(): Promise<void> {
 		console.error(`\n${failed} test(s) failed`)
 		process.exit(1)
 	}
-	console.log(`\nAll ${tests.length} backend-selection tests passed`)
 }
 
 main().catch((err) => {
diff --git a/src/codesearch/__tests__/chunk.test.ts b/src/codesearch/__tests__/chunk.test.ts
index 047f293..98ccf62 100644
--- a/src/codesearch/__tests__/chunk.test.ts
+++ b/src/codesearch/__tests__/chunk.test.ts
@@ -12,6 +12,7 @@ import * as path from "node:path"
 
 import { MAX_CHUNK_CHARS } from "../embedder.js"
 import { chunkFile } from "../chunk.js"
+import { runMain } from "../../test-runner.js"
 
 async function mkTmpWorkspace(files: Record<string, string>): Promise<string> {
 	const dir = await fs.mkdtemp(path.join(os.tmpdir(), "hc-chunk-"))
@@ -322,26 +323,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["gitignored + binary files are skipped (walker integration)", testSkipsBinaryAndGitignoredFiles],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} chunk tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("chunk", tests)
diff --git a/src/codesearch/__tests__/embedder-types.test.ts b/src/codesearch/__tests__/embedder-types.test.ts
index 149696a..89a5780 100644
--- a/src/codesearch/__tests__/embedder-types.test.ts
+++ b/src/codesearch/__tests__/embedder-types.test.ts
@@ -17,6 +17,7 @@ import { AirunnerEmbedder } from "../airunner-embedder.js"
 import { createEmbedder, OpenRouterEmbedder } from "../embedder.js"
 import type { Embedder, EmbedResult } from "../embedder-types.js"
 import { OllamaEmbedder } from "../ollama-embedder.js"
+import { runMain } from "../../test-runner.js"
 
 async function testBackendsSatisfyTheSharedInterface(): Promise<void> {
 	const url = "http://localhost:1"
@@ -50,37 +51,10 @@ async function testBackendsRejectEmptyInput(): Promise<void> {
 	await assert.rejects(() => airunner.embedBatch([""]), /empty/)
 }
 
-async function runOne(name: string, fn: () => Promise<void>): Promise<boolean> {
-	try {
-		await fn()
-		console.log(`  ok   ${name}`)
-		return true
-	} catch (error) {
-		console.error(`  FAIL ${name}`)
-		if (error instanceof Error) console.error(error.stack ?? error)
-		else console.error(String(error))
-		return false
-	}
-}
-
-async function main(): Promise<void> {
-	const tests: Array<[string, () => Promise<void>]> = [
-		["backends satisfy interface", testBackendsSatisfyTheSharedInterface],
-		["factory selects backends", testFactorySelectsBackends],
-		["backends reject empty input", testBackendsRejectEmptyInput],
-	]
-	let failed = 0
-	for (const [name, fn] of tests) {
-		if (!(await runOne(name, fn))) failed++
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} embedder-types tests passed`)
-}
+const tests: Array<[string, () => Promise<void>]> = [
+	["backends satisfy interface", testBackendsSatisfyTheSharedInterface],
+	["factory selects backends", testFactorySelectsBackends],
+	["backends reject empty input", testBackendsRejectEmptyInput],
+]
 
-main().catch((error) => {
-	console.error("Test runner crashed:", error)
-	process.exit(1)
-})
+runMain("embedder-types", tests)
diff --git a/src/codesearch/__tests__/files.test.ts b/src/codesearch/__tests__/files.test.ts
index 2af5f21..cb9a59e 100644
--- a/src/codesearch/__tests__/files.test.ts
+++ b/src/codesearch/__tests__/files.test.ts
@@ -13,6 +13,7 @@ import * as path from "node:path"
 import { promisify } from "node:util"
 
 import { walkSourceFiles } from "../files.js"
+import { runMain } from "../../test-runner.js"
 
 const execFileP = promisify(execFile)
 
@@ -90,26 +91,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["exclusion matches exact path segments only, not partial directory names", testDoesNotOverMatchPartialNames],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} files tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("files", tests)
diff --git a/src/codesearch/__tests__/index.test.ts b/src/codesearch/__tests__/index.test.ts
index e11f27b..1a061ba 100644
--- a/src/codesearch/__tests__/index.test.ts
+++ b/src/codesearch/__tests__/index.test.ts
@@ -19,6 +19,7 @@ import { buildIndex, chunkHash, indexFilePath, loadIndex, readJsonlEntries, writ
 import type { Embedder } from "../embedder.js"
 import type { IndexEntry } from "../types.js"
 import { walkSourceFiles } from "../files.js"
+import { runSuites } from "../../test-runner.js"
 
 /** Deterministic fake embedder with an embed-call counter. */
 class FakeEmbedder implements Embedder {
@@ -346,16 +347,7 @@ async function main(): Promise<void> {
 	process.env.HEADLESSCODE_DATA_DIR = storeTmp
 	let failed = 0
 	try {
-		for (const [name, fn] of tests) {
-			try {
-				await fn()
-				console.log(`  ok   ${name}`)
-			} catch (err) {
-				failed++
-				console.error(`  FAIL ${name}`)
-				console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-			}
-		}
+		failed = await runSuites("index", tests)
 	} finally {
 		delete process.env.HEADLESSCODE_DATA_DIR
 		await fs.rm(storeTmp, { recursive: true, force: true })
@@ -364,7 +356,6 @@ async function main(): Promise<void> {
 		console.error(`\n${failed} test(s) failed`)
 		process.exit(1)
 	}
-	console.log(`\nAll ${tests.length} index tests passed`)
 }
 
 main().catch((err) => {
diff --git a/src/codesearch/__tests__/ollama-embedder.test.ts b/src/codesearch/__tests__/ollama-embedder.test.ts
index e539c59..541d20b 100644
--- a/src/codesearch/__tests__/ollama-embedder.test.ts
+++ b/src/codesearch/__tests__/ollama-embedder.test.ts
@@ -27,6 +27,7 @@ import {
 	resolveOllamaEmbeddingModel,
 	resolveOllamaUrl,
 } from "../ollama-embedder.js"
+import { runMain } from "../../test-runner.js"
 
 /** Deterministic fake vector for a text (dim 8), so tests can assert content. */
 function fakeVector(text: string): number[] {
@@ -220,26 +221,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["Ollama URL/model env vars resolve with sensible defaults", testEnvResolution],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} ollama-embedder tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("ollama-embedder", tests)
diff --git a/src/codesearch/__tests__/search.test.ts b/src/codesearch/__tests__/search.test.ts
index cb13505..bb77040 100644
--- a/src/codesearch/__tests__/search.test.ts
+++ b/src/codesearch/__tests__/search.test.ts
@@ -14,6 +14,7 @@ import * as path from "node:path"
 import { buildIndex } from "../index.js"
 import { formatSearchResults, searchIndex } from "../search.js"
 import type { Embedder } from "../embedder.js"
+import { runSuites } from "../../test-runner.js"
 
 /** Deterministic fake embedder: vector[i] = count of the i-th letter. */
 class FakeEmbedder implements Embedder {
@@ -172,16 +173,7 @@ async function main(): Promise<void> {
 	process.env.HEADLESSCODE_DATA_DIR = storeTmp
 	let failed = 0
 	try {
-		for (const [name, fn] of tests) {
-			try {
-				await fn()
-				console.log(`  ok   ${name}`)
-			} catch (err) {
-				failed++
-				console.error(`  FAIL ${name}`)
-				console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-			}
-		}
+		failed = await runSuites("search", tests)
 	} finally {
 		delete process.env.HEADLESSCODE_DATA_DIR
 		await fs.rm(storeTmp, { recursive: true, force: true })
@@ -190,7 +182,6 @@ async function main(): Promise<void> {
 		console.error(`\n${failed} test(s) failed`)
 		process.exit(1)
 	}
-	console.log(`\nAll ${tests.length} search tests passed`)
 }
 
 main().catch((err) => {
diff --git a/src/config/__tests__/mode-models.test.ts b/src/config/__tests__/mode-models.test.ts
index d605337..c134254 100644
--- a/src/config/__tests__/mode-models.test.ts
+++ b/src/config/__tests__/mode-models.test.ts
@@ -20,6 +20,7 @@ import {
 	validateModeModelsBody,
 	DEFAULT_KEY,
 } from "../mode-models.js"
+import { runSuites } from "../../test-runner.js"
 
 let storeTmp: string
 
@@ -452,16 +453,7 @@ async function main(): Promise<void> {
 	process.env.HEADLESSCODE_DATA_DIR = storeTmp
 	let failed = 0
 	try {
-		for (const [name, fn] of tests) {
-			try {
-				await fn()
-				console.log(`  ok   ${name}`)
-			} catch (err) {
-				failed++
-				console.error(`  FAIL ${name}`)
-				console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-			}
-		}
+		failed = await runSuites("mode-models", tests)
 	} finally {
 		delete process.env.HEADLESSCODE_DATA_DIR
 		await fs.rm(storeTmp, { recursive: true, force: true })
@@ -470,7 +462,6 @@ async function main(): Promise<void> {
 		console.error(`\n${failed} test(s) failed`)
 		process.exit(1)
 	}
-	console.log(`\nAll ${tests.length} mode-models tests passed`)
 }
 
 main().catch((err) => {
diff --git a/src/dashboard/__tests__/aggregate.test.ts b/src/dashboard/__tests__/aggregate.test.ts
index f035e25..5eaa9e3 100644
--- a/src/dashboard/__tests__/aggregate.test.ts
+++ b/src/dashboard/__tests__/aggregate.test.ts
@@ -17,6 +17,7 @@ import * as path from "node:path"
 import { buildSummary, sumSessions } from "../aggregate.js"
 import type { LiveUsageRecord, UsageRecord } from "../../engine/usage.js"
 import { saveState, type OrchestratorState } from "../../orchestrator/state.js"
+import { runMain } from "../../test-runner.js"
 
 async function tmpRepo(prefix = "hc-dashboard-"): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -388,26 +389,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["sumSessions: empty input -> zeroed totals", testSumSessionsEmpty],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} dashboard aggregate tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("dashboard aggregate", tests)
diff --git a/src/dashboard/__tests__/chat-thread.test.ts b/src/dashboard/__tests__/chat-thread.test.ts
index 699a82f..c0ecb5a 100644
--- a/src/dashboard/__tests__/chat-thread.test.ts
+++ b/src/dashboard/__tests__/chat-thread.test.ts
@@ -13,6 +13,7 @@ import assert from "node:assert/strict"
 
 import { groupEventsIntoChat, summarizeArg } from "../chat-thread.js"
 import type { EventRecord } from "../../engine/events.js"
+import { runMain } from "../../test-runner.js"
 
 function event(type: string, extra: Record<string, unknown> = {}): EventRecord {
 	return { ts: "2026-08-01T00:00:00.000Z", sessionId: "session-1", type, ...extra }
@@ -278,26 +279,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["chat-thread: summarizeArg shortens long args, keeps short ones", testSummarizeArg],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} dashboard chat-thread tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("dashboard chat-thread", tests)
diff --git a/src/dashboard/__tests__/checkpoints-routes.test.ts b/src/dashboard/__tests__/checkpoints-routes.test.ts
index b4d2414..6ecc4c9 100644
--- a/src/dashboard/__tests__/checkpoints-routes.test.ts
+++ b/src/dashboard/__tests__/checkpoints-routes.test.ts
@@ -27,6 +27,7 @@ import * as path from "node:path"
 import { startDashboardServer } from "../server.js"
 import type { Server } from "node:http"
 import { createCheckpointService } from "../../checkpoints/service.js"
+import { runMain } from "../../test-runner.js"
 
 async function tmpRepo(prefix = "hc-ckpt-routes-"): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -341,26 +342,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["checkpoints: POST restore validates body + method (400/405/500, never silent)", testRestoreValidatesBodyAndMethod],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} dashboard checkpoints-routes tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("dashboard checkpoints-routes", tests)
diff --git a/src/dashboard/__tests__/cli.test.ts b/src/dashboard/__tests__/cli.test.ts
index 46c5868..edf4298 100644
--- a/src/dashboard/__tests__/cli.test.ts
+++ b/src/dashboard/__tests__/cli.test.ts
@@ -19,6 +19,7 @@ import assert from "node:assert/strict"
 
 import { dashboardMain, parseDashboardArgs } from "../cli.js"
 import { parseTrendArgs, trendMain } from "../trend-cli.js"
+import { runMain } from "../../test-runner.js"
 
 /** Capture stdout+stderr written during `fn()`. */
 async function captureOutput(fn: () => Promise<number>): Promise<{ exit: number; stdout: string; stderr: string }> {
@@ -134,26 +135,4 @@ const tests: Array<[string, () => void | Promise<void>]> = [
 	["trend: trendMain with a bad arg exits 2", testTrendMainBadArgExitsTwo],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? (err.stack ?? err.message) : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} dashboard/trend CLI tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("dashboard/trend CLI", tests)
diff --git a/src/dashboard/__tests__/codemap-routes.test.ts b/src/dashboard/__tests__/codemap-routes.test.ts
index eddc737..a80c601 100644
--- a/src/dashboard/__tests__/codemap-routes.test.ts
+++ b/src/dashboard/__tests__/codemap-routes.test.ts
@@ -17,6 +17,7 @@ import type { Server } from "node:http"
 import { startDashboardServer } from "../server.js"
 import { buildCodemap } from "../../codemap/build.js"
 import { resetCodeIntelCaches } from "../../codeintel/program.js"
+import { runMain } from "../../test-runner.js"
 
 async function tmpRepo(prefix = "hc-codemap-route-"): Promise<string> {
 	const ws = await fsp.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -119,33 +120,11 @@ async function testNoGeneratedMapIsClear404(): Promise<void> {
 	}
 }
 
-async function main(): Promise<void> {
-	const tests: Array<[string, () => Promise<void>]> = [
-		["GET /api/codemap?repo= serves the stored codemap.json", testServesCodemapJson],
-		["GET /api/codemap/html?repo= serves the visualizer", testServesCodemapHtml],
-		["missing ?repo= is a 400", testMissingRepoParamIs400],
-		["a repo with no generated map is a clear 404", testNoGeneratedMapIsClear404],
-	]
+const tests: Array<[string, () => Promise<void>]> = [
+	["GET /api/codemap?repo= serves the stored codemap.json", testServesCodemapJson],
+	["GET /api/codemap/html?repo= serves the visualizer", testServesCodemapHtml],
+	["missing ?repo= is a 400", testMissingRepoParamIs400],
+	["a repo with no generated map is a clear 404", testNoGeneratedMapIsClear404],
+]
 
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} codemap dashboard-route tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("codemap dashboard-route", tests)
diff --git a/src/dashboard/__tests__/cost-history-routes.test.ts b/src/dashboard/__tests__/cost-history-routes.test.ts
index 31a6a3a..76e3d69 100644
--- a/src/dashboard/__tests__/cost-history-routes.test.ts
+++ b/src/dashboard/__tests__/cost-history-routes.test.ts
@@ -23,6 +23,7 @@ import {
 	type CostHistoryRecord,
 	type SessionCostRecord,
 } from "../../orchestrator/cost-history.js"
+import { runMain } from "../../test-runner.js"
 
 async function tmpRepo(prefix = "hc-cost-history-route-"): Promise<string> {
 	return fsp.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -237,39 +238,17 @@ async function testNoHistoryIsEmpty200(): Promise<void> {
 	}
 }
 
-async function main(): Promise<void> {
-	const tests: Array<[string, () => Promise<void>]> = [
-		["GET /api/cost-history?repo= serves group + session records (newest first)", testServesGroupAndSessionRecords],
-		["?since=<iso> windows by recordedAt", testSinceWindowsByRecordedAt],
-		[
-			"?since= boundary: millisecond-precision recordedAt vs second-precision since includes the equal instant",
-			testSinceBoundaryMillisecondPrecisionIncludesEqualInstant,
-		],
-		["?limit=N keeps the most recent N", testLimitKeepsMostRecent],
-		["missing ?repo= is a 400", testMissingRepoIs400],
-		["invalid ?since= / ?limit= are 400s", testInvalidSinceAndLimitAre400],
-		["a repo with no history is an empty 200", testNoHistoryIsEmpty200],
-	]
-
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} cost-history dashboard-route tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+const tests: Array<[string, () => Promise<void>]> = [
+	["GET /api/cost-history?repo= serves group + session records (newest first)", testServesGroupAndSessionRecords],
+	["?since=<iso> windows by recordedAt", testSinceWindowsByRecordedAt],
+	[
+		"?since= boundary: millisecond-precision recordedAt vs second-precision since includes the equal instant",
+		testSinceBoundaryMillisecondPrecisionIncludesEqualInstant,
+	],
+	["?limit=N keeps the most recent N", testLimitKeepsMostRecent],
+	["missing ?repo= is a 400", testMissingRepoIs400],
+	["invalid ?since= / ?limit= are 400s", testInvalidSinceAndLimitAre400],
+	["a repo with no history is an empty 200", testNoHistoryIsEmpty200],
+]
+
+runMain("cost-history dashboard-route", tests)
diff --git a/src/dashboard/__tests__/events.test.ts b/src/dashboard/__tests__/events.test.ts
index c786537..b60ab22 100644
--- a/src/dashboard/__tests__/events.test.ts
+++ b/src/dashboard/__tests__/events.test.ts
@@ -15,6 +15,7 @@ import * as path from "node:path"
 
 import { readSessionEvents, findSessionEventsFile } from "../aggregate.js"
 import { eventsFilePath } from "../../engine/events.js"
+import { runMain } from "../../test-runner.js"
 
 async function tmpRepo(prefix = "hc-dash-events-"): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -190,26 +191,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["readSessionEvents: offset beyond EOF clamps to the file size", testOffsetBeyondEOFClampsToFileEnd],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} dashboard events tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("dashboard events", tests)
diff --git a/src/dashboard/__tests__/file-browser-permissions.test.ts b/src/dashboard/__tests__/file-browser-permissions.test.ts
index a5169d4..76f5dbf 100644
--- a/src/dashboard/__tests__/file-browser-permissions.test.ts
+++ b/src/dashboard/__tests__/file-browser-permissions.test.ts
@@ -29,6 +29,7 @@ import { startDashboardServer } from "../server.js"
 import type { Server } from "node:http"
 import { permissionsFilePath, resolvePermissions } from "../../permissions/config.js"
 import { DEFAULT_PROTECTED_FILES } from "../../permissions/protected-files.js"
+import { runSuites } from "../../test-runner.js"
 
 async function tmpRepo(prefix = "hc-fileperm-"): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -552,16 +553,10 @@ async function main(): Promise<void> {
 	process.env.HEADLESSCODE_DATA_DIR = storeTmp
 	let failed = 0
 	try {
-		for (const [name, fn] of tests) {
-			try {
-				await fn()
-				console.log(`  ok   ${name}`)
-			} catch (err) {
-				failed++
-				console.error(`  FAIL ${name}`)
-				console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-			}
-		}
+		failed = await runSuites(
+			"dashboard file-browser + permissions settings",
+			tests,
+		)
 	} finally {
 		delete process.env.HEADLESSCODE_DATA_DIR
 		await fs.rm(storeTmp, { recursive: true, force: true })
@@ -570,7 +565,6 @@ async function main(): Promise<void> {
 		console.error(`\n${failed} test(s) failed`)
 		process.exit(1)
 	}
-	console.log(`\nAll ${tests.length} dashboard file-browser + permissions settings tests passed`)
 }
 
 main().catch((err) => {
diff --git a/src/dashboard/__tests__/page-route.test.ts b/src/dashboard/__tests__/page-route.test.ts
index 5559b6b..a339f8b 100644
--- a/src/dashboard/__tests__/page-route.test.ts
+++ b/src/dashboard/__tests__/page-route.test.ts
@@ -24,6 +24,7 @@ import * as path from "node:path"
 
 import { startDashboardServer } from "../server.js"
 import type { Server } from "node:http"
+import { runMain } from "../../test-runner.js"
 
 async function tmpRepo(prefix = "hc-page-route-"): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -139,26 +140,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["GET / without --repo still renders, with an empty file-browser default", testGetRootWithoutRepoStillRendersWithEmptyDefault],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? (err.stack ?? err.message) : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} dashboard page-route tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("dashboard page-route", tests)
diff --git a/src/dashboard/__tests__/projects-routes.test.ts b/src/dashboard/__tests__/projects-routes.test.ts
index d0bdbef..84dcb1b 100644
--- a/src/dashboard/__tests__/projects-routes.test.ts
+++ b/src/dashboard/__tests__/projects-routes.test.ts
@@ -19,6 +19,7 @@ import * as path from "node:path"
 import type { Server } from "node:http"
 
 import { startDashboardServer } from "../server.js"
+import { runMain } from "../../test-runner.js"
 
 async function mkTmp(prefix: string): Promise<string> {
 	return fsp.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -202,26 +203,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["GET /api/projects: gitOwnerName derived from origin remote", testGitOwnerNameDerivedFromOrigin],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} projects-route tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("projects-route", tests)
diff --git a/src/dashboard/__tests__/self-improvement-metrics.test.ts b/src/dashboard/__tests__/self-improvement-metrics.test.ts
index dfa0f3e..3283522 100644
--- a/src/dashboard/__tests__/self-improvement-metrics.test.ts
+++ b/src/dashboard/__tests__/self-improvement-metrics.test.ts
@@ -30,6 +30,7 @@ import {
 	extractIssueRefs,
 } from "../self-improvement-metrics.js"
 import type { SessionRow } from "../aggregate.js"
+import { runMain } from "../../test-runner.js"
 
 const execFileP = promisify(execFile)
 
@@ -199,26 +200,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["computeIssueLifecycle: a repo with no GitHub remote degrades to undefined, never throws", testIssueLifecycleDegradesGracefullyWithoutGhRemote],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? (err.stack ?? err.message) : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} self-improvement-metrics tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("self-improvement-metrics", tests)
diff --git a/src/dashboard/__tests__/server.test.ts b/src/dashboard/__tests__/server.test.ts
index 8a7eab3..e59e2e5 100644
--- a/src/dashboard/__tests__/server.test.ts
+++ b/src/dashboard/__tests__/server.test.ts
@@ -17,6 +17,7 @@ import { setTimeout as sleep } from "node:timers/promises"
 import { startDashboardServer } from "../server.js"
 import type { Server } from "node:http"
 import { eventsFilePath } from "../../engine/events.js"
+import { runMain } from "../../test-runner.js"
 
 async function tmpRepo(prefix = "hc-server-"): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -359,26 +360,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["server: non-GET/non-POST methods to unrelated routes are still rejected (405)", testNonGetPostMethodsStillRejected],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} dashboard server tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("dashboard server", tests)
diff --git a/src/dashboard/__tests__/session-launch.test.ts b/src/dashboard/__tests__/session-launch.test.ts
index d600c13..66f0e4e 100644
--- a/src/dashboard/__tests__/session-launch.test.ts
+++ b/src/dashboard/__tests__/session-launch.test.ts
@@ -28,6 +28,7 @@ import type { Server } from "node:http"
 import { createHeadlessExecutor } from "../../tools/executor.js"
 import { eventsFilePath } from "../../engine/events.js"
 import { DECISION_MARKERS, validateSessionStartBody } from "../session-launch.js"
+import { runMain } from "../../test-runner.js"
 
 async function tmpRepo(prefix = "hc-launch-"): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -431,26 +432,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["session-launch: optional bearer token gates control-plane POSTs, off by default", testTokenGatesControlPlanePosts],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} session-launch tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("session-launch", tests)
diff --git a/src/dashboard/__tests__/settings.test.ts b/src/dashboard/__tests__/settings.test.ts
index 1ac27ec..509a923 100644
--- a/src/dashboard/__tests__/settings.test.ts
+++ b/src/dashboard/__tests__/settings.test.ts
@@ -15,6 +15,7 @@ import * as path from "node:path"
 import { startDashboardServer } from "../server.js"
 import type { Server } from "node:http"
 import { modeModelsFilePath } from "../../config/mode-models.js"
+import { runSuites } from "../../test-runner.js"
 
 async function tmpRepo(prefix = "hc-settings-"): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -240,16 +241,7 @@ async function main(): Promise<void> {
 	process.env.HEADLESSCODE_DATA_DIR = storeTmp
 	let failed = 0
 	try {
-		for (const [name, fn] of tests) {
-			try {
-				await fn()
-				console.log(`  ok   ${name}`)
-			} catch (err) {
-				failed++
-				console.error(`  FAIL ${name}`)
-				console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-			}
-		}
+		failed = await runSuites("dashboard settings", tests)
 	} finally {
 		delete process.env.HEADLESSCODE_DATA_DIR
 		await fs.rm(storeTmp, { recursive: true, force: true })
@@ -258,7 +250,6 @@ async function main(): Promise<void> {
 		console.error(`\n${failed} test(s) failed`)
 		process.exit(1)
 	}
-	console.log(`\nAll ${tests.length} dashboard settings tests passed`)
 }
 
 main().catch((err) => {
diff --git a/src/dashboard/__tests__/sigint-shutdown.test.ts b/src/dashboard/__tests__/sigint-shutdown.test.ts
index e19c7ae..4583aa3 100644
--- a/src/dashboard/__tests__/sigint-shutdown.test.ts
+++ b/src/dashboard/__tests__/sigint-shutdown.test.ts
@@ -35,6 +35,7 @@ import * as net from "node:net"
 import * as os from "node:os"
 import * as path from "node:path"
 import { fileURLToPath } from "node:url"
+import { runMain } from "../../test-runner.js"
 
 const REPO_ROOT = path.resolve(fileURLToPath(new URL("../../../", import.meta.url)))
 
@@ -207,26 +208,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["dashboard CLI: real SIGINT closes the HTTP server and releases the port", testDashboardCliSigintClosesServer],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? (err.stack ?? err.message) : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} sigint-shutdown tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("sigint-shutdown", tests)
diff --git a/src/dashboard/__tests__/timeline.test.ts b/src/dashboard/__tests__/timeline.test.ts
index c1d0edd..1b06d13 100644
--- a/src/dashboard/__tests__/timeline.test.ts
+++ b/src/dashboard/__tests__/timeline.test.ts
@@ -18,6 +18,7 @@ import { fileURLToPath } from "node:url"
 
 import { buildTimeline, classifyTool } from "../timeline.js"
 import type { EventRecord } from "../../engine/events.js"
+import { runMain } from "../../test-runner.js"
 
 const __dirname = path.dirname(fileURLToPath(import.meta.url))
 
@@ -325,26 +326,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["timeline: page.ts's TL_READ/WRITE/EXEC/SEARCH sets stay in sync with timeline.ts's", testPageTsWriteToolsStayInSyncWithTimelineTs],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} dashboard timeline tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("dashboard timeline", tests)
diff --git a/src/dashboard/__tests__/tool-exec.test.ts b/src/dashboard/__tests__/tool-exec.test.ts
index da74e55..52a066f 100644
--- a/src/dashboard/__tests__/tool-exec.test.ts
+++ b/src/dashboard/__tests__/tool-exec.test.ts
@@ -17,6 +17,7 @@ import * as path from "node:path"
 import { executeTool } from "../tool-exec.js"
 import { startDashboardServer } from "../server.js"
 import type { Server } from "node:http"
+import { runMain } from "../../test-runner.js"
 
 async function mkTmpWorkspace(prefix: string): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -194,23 +195,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["HTTP POST /api/tool/execute rejects bad body", testHttpEndpointRejectsBadBody],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`ok - ${name}`)
-		} catch (error) {
-			failed++
-			console.error(`FAIL - ${name}`)
-			console.error(error)
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed}/${tests.length} tool-exec tests failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} tool-exec tests passed`)
-}
-
-main()
+runMain("tool-exec", tests)
diff --git a/src/deploy/__tests__/gate.test.ts b/src/deploy/__tests__/gate.test.ts
index 78ef445..8862c81 100644
--- a/src/deploy/__tests__/gate.test.ts
+++ b/src/deploy/__tests__/gate.test.ts
@@ -7,6 +7,7 @@
 import assert from "node:assert/strict"
 
 import { decideApproval, type ApprovalInputs } from "../gate.js"
+import { runMain } from "../../test-runner.js"
 
 function inputs(over: Partial<ApprovalInputs>): ApprovalInputs {
 	return { interactive: false, approvalFileExists: false, ...over }
@@ -108,26 +109,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["all inputs empty denies", testNonInteractiveEverythingEmptyDenies],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} deploy-gate decision tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("deploy-gate decision", tests)
diff --git a/src/engine/__tests__/claims.test.ts b/src/engine/__tests__/claims.test.ts
index 61dbecf..3d129de 100644
--- a/src/engine/__tests__/claims.test.ts
+++ b/src/engine/__tests__/claims.test.ts
@@ -13,6 +13,7 @@ import {
 	type ClaimVerification,
 } from "../claims.js"
 import { resolvePermissions } from "../../permissions/config.js"
+import { runMain } from "../../test-runner.js"
 
 /** Deterministic tmp workspace per test. */
 async function makeTmpWorkspace(files: Record<string, string> = {}): Promise<string> {
@@ -337,51 +338,73 @@ function testAllClaimsVerifiedHelpers(): void {
 
 // --- runner ----------------------------------------------------------------
 
-const tests: Array<{ name: string; fn: () => void | Promise<void> }> = [
-	{ name: "extract: file-existence claims", fn: testExtractsFileExistenceClaims },
-	{ name: "extract: file claim behind a preposition (verb … to <path>)", fn: testExtractsFileClaimBehindPreposition },
-	{ name: "extract: prepositional phrase without a change verb is not a claim", fn: testPrepositionPatternStillNeedsAVerb },
-	{ name: "extract: bare path mentions without verbs ignored", fn: testIgnoresBarePathMentionsWithoutVerbs },
-	{ name: "extract: command-passed claims", fn: testExtractsCommandPassedClaims },
-	{ name: "extract: serial markers only when serial mentioned", fn: testExtractsSerialMarkersOnlyWhenSerialMentioned },
-	{ name: "extract: PR URL claims", fn: testExtractsPrUrlClaims },
-	{ name: "extract: dedupe identical claims", fn: testDeduplicatesClaims },
-	{ name: "verify: file exists against disk", fn: testFileExistsVerifiedAgainstDisk },
-	{ name: "verify: empty file fails", fn: testEmptyFileClaimFails },
-	{ name: "verify: command re-run passes", fn: testCommandClaimRerunPasses },
-	{ name: "verify: command re-run fails on non-zero exit", fn: testCommandClaimRerunFails },
-	{ name: "extract: quoted token in pass phrase becomes expectedMarker", fn: testExtractsExpectedMarkerFromQuotedToken },
-	{ name: "extract: all-caps word in pass phrase becomes expectedMarker", fn: testExtractsExpectedMarkerFromAllCapsWord },
-	{ name: "extract: lowercase adjectives are not output markers", fn: testLowercasePassPhraseDoesNotBecomeMarker },
-	{ name: "verify: exit 0 but missing claimed marker FAILS", fn: testCommandClaimExitZeroButMissingMarkerFails },
-	{ name: "verify: exit 0 with the claimed marker passes", fn: testCommandClaimExitZeroWithMarkerPasses },
-	{ name: "resolveClaimPath: ../ traversal and absolute paths rejected", fn: testResolveClaimPathRejectsEscapes },
-	{ name: "verify: denied command refused, never re-run", fn: testCommandClaimRefusedByPermissionGate },
-	{ name: "verify: nonexistent command fails", fn: testCommandClaimNonexistentCommandFails },
-	{ name: "verify: serial markers against newest serial log", fn: testSerialMarkerVerifiedAgainstLog },
-	{ name: "verify: no serial log fails closed", fn: testSerialMarkerNoLogFails },
-	{ name: "verify: PR number against real git history", fn: testPrUrlVerifiedAgainstGitHistory },
-	{ name: "helpers: allClaimsVerified / firstUnverifiedDetail", fn: testAllClaimsVerifiedHelpers },
+const tests: Array<[string, () => void | Promise<void>]> = [
+	["extract: file-existence claims", testExtractsFileExistenceClaims],
+	[
+		"extract: file claim behind a preposition (verb … to <path>)",
+		testExtractsFileClaimBehindPreposition,
+	],
+	[
+		"extract: prepositional phrase without a change verb is not a claim",
+		testPrepositionPatternStillNeedsAVerb,
+	],
+	[
+		"extract: bare path mentions without verbs ignored",
+		testIgnoresBarePathMentionsWithoutVerbs,
+	],
+	["extract: command-passed claims", testExtractsCommandPassedClaims],
+	[
+		"extract: serial markers only when serial mentioned",
+		testExtractsSerialMarkersOnlyWhenSerialMentioned,
+	],
+	["extract: PR URL claims", testExtractsPrUrlClaims],
+	["extract: dedupe identical claims", testDeduplicatesClaims],
+	["verify: file exists against disk", testFileExistsVerifiedAgainstDisk],
+	["verify: empty file fails", testEmptyFileClaimFails],
+	["verify: command re-run passes", testCommandClaimRerunPasses],
+	["verify: command re-run fails on non-zero exit", testCommandClaimRerunFails],
+	[
+		"extract: quoted token in pass phrase becomes expectedMarker",
+		testExtractsExpectedMarkerFromQuotedToken,
+	],
+	[
+		"extract: all-caps word in pass phrase becomes expectedMarker",
+		testExtractsExpectedMarkerFromAllCapsWord,
+	],
+	[
+		"extract: lowercase adjectives are not output markers",
+		testLowercasePassPhraseDoesNotBecomeMarker,
+	],
+	[
+		"verify: exit 0 but missing claimed marker FAILS",
+		testCommandClaimExitZeroButMissingMarkerFails,
+	],
+	[
+		"verify: exit 0 with the claimed marker passes",
+		testCommandClaimExitZeroWithMarkerPasses,
+	],
+	[
+		"resolveClaimPath: ../ traversal and absolute paths rejected",
+		testResolveClaimPathRejectsEscapes,
+	],
+	[
+		"verify: denied command refused, never re-run",
+		testCommandClaimRefusedByPermissionGate,
+	],
+	["verify: nonexistent command fails", testCommandClaimNonexistentCommandFails],
+	[
+		"verify: serial markers against newest serial log",
+		testSerialMarkerVerifiedAgainstLog,
+	],
+	["verify: no serial log fails closed", testSerialMarkerNoLogFails],
+	[
+		"verify: PR number against real git history",
+		testPrUrlVerifiedAgainstGitHistory,
+	],
+	[
+		"helpers: allClaimsVerified / firstUnverifiedDetail",
+		testAllClaimsVerifiedHelpers,
+	],
 ]
 
-async function run(): Promise<void> {
-	let failed = 0
-	for (const t of tests) {
-		try {
-			await t.fn()
-			console.log(`PASS ${t.name}`)
-		} catch (err) {
-			failed++
-			console.error(`FAIL ${t.name}`)
-			console.error(err instanceof Error ? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		process.exitCode = 1
-		console.error(`\n${failed}/${tests.length} tests failed`)
-	} else {
-		console.log(`\nAll ${tests.length} claims tests passed`)
-	}
-}
-
-void run()
+runMain("claims", tests)
diff --git a/src/engine/__tests__/condense-async.test.ts b/src/engine/__tests__/condense-async.test.ts
index f10c15a..e92106c 100644
--- a/src/engine/__tests__/condense-async.test.ts
+++ b/src/engine/__tests__/condense-async.test.ts
@@ -23,6 +23,7 @@ import { HeadlessSession } from "../loop.js"
 import { DEFAULT_WINDOW_SIZE } from "../loop.js"
 import type { ChatMessage, ChatToolCall, LlmClient, LlmRequest, LlmResponse } from "../types.js"
 import { BudgetTracker } from "../../budget/budget.js"
+import { runMain } from "../../test-runner.js"
 
 // ─── Fakes ───────────────────────────────────────────────────────────────────
 
@@ -379,26 +380,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["async condense: session-end before resolution discards the result but keeps its usage in the BudgetTracker", testSessionEndBeforeBackgroundResolveDiscardsButAccounts],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} condense-async tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("condense-async", tests)
diff --git a/src/engine/__tests__/condense.test.ts b/src/engine/__tests__/condense.test.ts
index 31ee911..c06a797 100644
--- a/src/engine/__tests__/condense.test.ts
+++ b/src/engine/__tests__/condense.test.ts
@@ -45,6 +45,7 @@ import {
 } from "../condense.js"
 import type { ChatMessage, ChatToolCall, LlmClient, LlmRequest, LlmResponse } from "../types.js"
 import { BudgetTracker } from "../../budget/budget.js"
+import { runMain } from "../../test-runner.js"
 
 // ─── Fakes ───────────────────────────────────────────────────────────────────
 
@@ -1338,26 +1339,4 @@ const tests: Array<[string, () => Promise<void> | void]> = [
 	["session: condensation call times out instead of hanging forever (non-fatal)", testSessionCondenseCallTimesOutNonFatally],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} condense tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("condense", tests)
diff --git a/src/engine/__tests__/events.test.ts b/src/engine/__tests__/events.test.ts
index 02c9e1e..b07d189 100644
--- a/src/engine/__tests__/events.test.ts
+++ b/src/engine/__tests__/events.test.ts
@@ -21,6 +21,7 @@ import { HeadlessSession } from "../loop.js"
 import { eventsDir, eventsFilePath, readEventsFile, EVENT_TRUNCATE_CHARS, truncateField } from "../events.js"
 import { readLimitFromEnv } from "../../tools/executor.js"
 import type { ChatMessage, ChatToolCall, LlmClient, LlmRequest, LlmResponse } from "../types.js"
+import { runMain } from "../../test-runner.js"
 
 // ─── Fake LLM client (mirrors src/engine/__tests__/loop.test.ts) ────────────
 
@@ -761,26 +762,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["events: streamed response usage feeds the budget tracker exactly once (no double-count)", testStreamedResponseUsageFeedsBudgetOnce],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} events tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("events", tests)
diff --git a/src/engine/__tests__/global-modes.test.ts b/src/engine/__tests__/global-modes.test.ts
index 3612733..9c43b0e 100644
--- a/src/engine/__tests__/global-modes.test.ts
+++ b/src/engine/__tests__/global-modes.test.ts
@@ -29,6 +29,7 @@ const ROOT_SANDBOX = path.join(os.tmpdir(), `hc-global-modes-root-${process.pid}
 import assert from "node:assert/strict"
 
 import { buildSystemPrompt, loadCustomModes } from "../prompt.js"
+import { runSuites } from "../../test-runner.js"
 
 // ─── Per-test HOME sandboxing ───────────────────────────────────────────────
 
@@ -232,22 +233,11 @@ async function main(): Promise<void> {
 	)
 	assert.equal(process.env.HOME, path.join(ROOT_SANDBOX, "home-0"))
 
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack : err)
-		}
-	}
+	const failed = await runSuites("global-modes", tests)
 	if (failed > 0) {
 		console.error(`\n${failed} test(s) failed`)
 		process.exit(1)
 	}
-	console.log(`\nAll ${tests.length} global-modes tests passed`)
 }
 
 main().finally(async () => {
diff --git a/src/engine/__tests__/inject-message.test.ts b/src/engine/__tests__/inject-message.test.ts
index ed48c1c..8111668 100644
--- a/src/engine/__tests__/inject-message.test.ts
+++ b/src/engine/__tests__/inject-message.test.ts
@@ -21,6 +21,7 @@ import * as path from "node:path"
 import { HeadlessSession } from "../loop.js"
 import { eventsDir, eventsFilePath, readEventsFile } from "../events.js"
 import type { ChatMessage, ChatToolCall, LlmClient, LlmRequest, LlmResponse } from "../types.js"
+import { runMain } from "../../test-runner.js"
 
 const INJECT_MESSAGE = ".harness.inject-message"
 const INJECTED_TEXT = "Wait — switch to the other approach first, then continue."
@@ -310,26 +311,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["inject: malformed marker is removed and ignored, never wedges the loop", testMalformedMarkerIsRemovedAndIgnored],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} inject-message tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("inject-message", tests)
diff --git a/src/engine/__tests__/lazy-tools.test.ts b/src/engine/__tests__/lazy-tools.test.ts
index 65ede1b..12768d9 100644
--- a/src/engine/__tests__/lazy-tools.test.ts
+++ b/src/engine/__tests__/lazy-tools.test.ts
@@ -18,6 +18,7 @@ import {
 	splitCoreAndLazyTools,
 } from "../lazy-tools.js"
 import type { ChatTool } from "../types.js"
+import { runMain } from "../../test-runner.js"
 
 function fnTool(name: string, description = `${name} description`): ChatTool {
 	return { type: "function", function: { name, description, parameters: { type: "object", properties: {} } } }
@@ -133,22 +134,4 @@ const tests: Array<[string, () => void]> = [
 	["renderToolIndex: empty case", testRenderToolIndexEmptyCase],
 ]
 
-function main(): void {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			fn()
-			console.log(`  ok   ${name}`)
-		} catch (error) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(error)
-		}
-	}
-	console.log(`\n${failed === 0 ? "All" : `${tests.length - failed}/${tests.length}`} ${tests.length} tests passed`)
-	if (failed > 0) {
-		process.exit(1)
-	}
-}
-
-main()
+runMain("", tests)
diff --git a/src/engine/__tests__/local-explore.test.ts b/src/engine/__tests__/local-explore.test.ts
index acba3e3..a35bdd4 100644
--- a/src/engine/__tests__/local-explore.test.ts
+++ b/src/engine/__tests__/local-explore.test.ts
@@ -55,6 +55,7 @@ import {
 } from "../local-explore.js"
 import type { LocalChatClient, LocalChatResponse, LocalExploreOptions, LocalExploreRequest } from "../local-explore.js"
 import type { ChatMessage, ChatToolCall, LlmClient, LlmRequest, LlmResponse } from "../types.js"
+import { runMain } from "../../test-runner.js"
 
 // ─── Fakes ───────────────────────────────────────────────────────────────────
 
@@ -178,10 +179,10 @@ async function makeSession(options: {
 
 // ─── Tests ───────────────────────────────────────────────────────────────────
 
-const tests: Array<{ name: string; fn: () => Promise<void> | void }> = []
+const tests: Array<[string, () => Promise<void> | void]> = []
 
 function test(name: string, fn: () => Promise<void> | void) {
-	tests.push({ name, fn })
+	tests.push([name, fn])
 }
 
 // 1. Iteration cap enforcement: a local model that never stops must be cut
@@ -631,26 +632,4 @@ test("Ollama client wraps HTTP and network failures in LocalExploreError", async
 
 // ─── Runner ──────────────────────────────────────────────────────────────────
 
-async function main() {
-	let failed = 0
-	for (const { name, fn } of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} local-explore tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("local-explore", tests)
diff --git a/src/engine/__tests__/loop-helpers-history.test.ts b/src/engine/__tests__/loop-helpers-history.test.ts
index 5006527..7fc61b1 100644
--- a/src/engine/__tests__/loop-helpers-history.test.ts
+++ b/src/engine/__tests__/loop-helpers-history.test.ts
@@ -14,6 +14,7 @@ import {
 	truncateHistory,
 } from "../loop.js"
 import type { ChatMessage, ChatTool } from "../types.js"
+import { runMain } from "../../test-runner.js"
 
 function msg(role: ChatMessage["role"], content: string | null): ChatMessage {
 	return { role, content }
@@ -94,23 +95,4 @@ const tests: Array<[string, () => void]> = [
 	["messagesPrefixEquals empty snapshot", testPrefixEqualsEmptySnapshot],
 ]
 
-function main(): void {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} tests passed`)
-}
-
-main()
+runMain("", tests)
diff --git a/src/engine/__tests__/loop-helpers-readonly.test.ts b/src/engine/__tests__/loop-helpers-readonly.test.ts
index 9ac65bd..4829dd9 100644
--- a/src/engine/__tests__/loop-helpers-readonly.test.ts
+++ b/src/engine/__tests__/loop-helpers-readonly.test.ts
@@ -13,6 +13,7 @@ import {
 	isReadOnlyDirectCall,
 	isReadOnlyTurnCall,
 } from "../loop.js"
+import { runMain } from "../../test-runner.js"
 
 function testCombineAbortedWhenInputAborted(): void {
 	const c = new AbortController()
@@ -147,23 +148,4 @@ const tests: Array<[string, () => void]> = [
 	["artifact nudge names streak+path", testArtifactNudgeNamesStreakAndPath],
 ]
 
-function main(): void {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} tests passed`)
-}
-
-main()
+runMain("", tests)
diff --git a/src/engine/__tests__/loop.test.ts b/src/engine/__tests__/loop.test.ts
index 1d25350..d87f50d 100644
--- a/src/engine/__tests__/loop.test.ts
+++ b/src/engine/__tests__/loop.test.ts
@@ -38,6 +38,7 @@ import type { EvidenceInput, EvidenceJournal, EvidenceReceipt } from "../../evid
 import type { ToolExecutor } from "../../tools/executor.js"
 import type { SessionBudget } from "../../budget/budget.js"
 import type { ChatMessage, ChatToolCall, LlmClient, LlmRequest, LlmResponse } from "../types.js"
+import { runSuites } from "../../test-runner.js"
 
 // ─── Fake LLM client ─────────────────────────────────────────────────────────
 
@@ -4375,22 +4376,11 @@ process.on("unhandledRejection", (reason) => {
 })
 
 async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
+	const failed = await runSuites("", tests)
 	if (failed > 0) {
 		console.error(`\n${failed} test(s) failed`)
 		process.exit(1)
 	}
-	console.log(`\nAll ${tests.length} tests passed`)
 	testsFinished = true
 	process.exitCode = 0
 }
diff --git a/src/engine/__tests__/new-task.test.ts b/src/engine/__tests__/new-task.test.ts
index 58fac88..f927d33 100644
--- a/src/engine/__tests__/new-task.test.ts
+++ b/src/engine/__tests__/new-task.test.ts
@@ -25,6 +25,7 @@ import { eventsDir, readEventsFile, type EventRecord } from "../events.js"
 import { createCheckpointService } from "../../checkpoints/service.js"
 import type { SessionBudget } from "../../budget/budget.js"
 import type { ChatMessage, ChatToolCall, LlmClient, LlmRequest, LlmResponse } from "../types.js"
+import { runMain } from "../../test-runner.js"
 
 // ─── Scripted fake LLM client (parent + child share it) ──────────────────────
 
@@ -431,26 +432,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["new_task: malformed child completion fails closed", testMalformedChildCompletionFailsClosed],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("", tests)
diff --git a/src/engine/__tests__/parser.test.ts b/src/engine/__tests__/parser.test.ts
index 802e2b6..a07cc7b 100644
--- a/src/engine/__tests__/parser.test.ts
+++ b/src/engine/__tests__/parser.test.ts
@@ -14,6 +14,7 @@ import assert from "node:assert/strict"
 
 import { bestEffortPartialJson, extractEmbeddedToolCall, parseToolCall, parseToolCalls } from "../parser.js"
 import type { ChatMessage, ChatToolCall } from "../types.js"
+import { runMain } from "../../test-runner.js"
 
 function call(overrides: Partial<ChatToolCall> = {}, id = "call_1"): ChatToolCall {
 	return {
@@ -256,23 +257,4 @@ const tests: Array<[string, () => void]> = [
 	],
 ]
 
-function main(): void {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} tests passed`)
-}
-
-main()
+runMain("", tests)
diff --git a/src/engine/__tests__/pause.test.ts b/src/engine/__tests__/pause.test.ts
index e771794..6156cb1 100644
--- a/src/engine/__tests__/pause.test.ts
+++ b/src/engine/__tests__/pause.test.ts
@@ -20,6 +20,7 @@ import { HeadlessSession } from "../loop.js"
 import { BudgetTracker } from "../../budget/budget.js"
 import { eventsDir, eventsFilePath, readEventsFile } from "../events.js"
 import type { ChatMessage, ChatToolCall, LlmClient, LlmRequest, LlmResponse } from "../types.js"
+import { runMain } from "../../test-runner.js"
 
 const PAUSE_REQUESTED = ".harness.pause-requested"
 const PAUSED = ".harness.paused"
@@ -299,26 +300,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["pause: no marker -> zero behavior change", testNoMarkerMeansNoBlocking],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} pause tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("pause", tests)
diff --git a/src/engine/__tests__/prompt.test.ts b/src/engine/__tests__/prompt.test.ts
index 41f418b..c67274b 100644
--- a/src/engine/__tests__/prompt.test.ts
+++ b/src/engine/__tests__/prompt.test.ts
@@ -31,6 +31,7 @@ import {
 	selectToolsForMode,
 } from "../prompt.js"
 import type { ChatTool } from "../types.js"
+import { runMain } from "../../test-runner.js"
 
 function findStrictModeViolations(tools: ChatTool[]): string[] {
 	const violations: string[] = []
@@ -143,23 +144,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["headless conventions are not shown to read-only modes", testHeadlessConventionsNotShownToReadOnlyModes],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack : err)
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} prompt/tool-schema tests passed`)
-}
-
-main()
+runMain("prompt/tool-schema", tests)
diff --git a/src/engine/__tests__/stack-rules-content.test.ts b/src/engine/__tests__/stack-rules-content.test.ts
index 9cd4f30..86107c2 100644
--- a/src/engine/__tests__/stack-rules-content.test.ts
+++ b/src/engine/__tests__/stack-rules-content.test.ts
@@ -26,6 +26,7 @@ const ROOT_SANDBOX = path.join(os.tmpdir(), `hc-stack-content-${process.pid}-${D
 
 import { buildSystemPrompt } from "../prompt.js"
 import { STACK_RULES_HEADER, getCentralStackRulesDir, loadStackRules } from "../stacks.js"
+import { runSuites } from "../../test-runner.js"
 
 /** Repo root = 4 levels up from src/engine/__tests__/<this file>. */
 const REPO_ROOT = path.resolve(fileURLToPath(import.meta.url), "..", "..", "..", "..")
@@ -157,22 +158,11 @@ async function main(): Promise<void> {
 	)
 	assert.equal(process.env.HOME, path.join(ROOT_SANDBOX, "home-0"))
 
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack : err)
-		}
-	}
+	const failed = await runSuites("stack-rules-content", tests)
 	if (failed > 0) {
 		console.error(`\n${failed} test(s) failed`)
 		process.exit(1)
 	}
-	console.log(`\nAll ${tests.length} stack-rules-content tests passed`)
 }
 
 main().finally(async () => {
diff --git a/src/engine/__tests__/stacks.test.ts b/src/engine/__tests__/stacks.test.ts
index b450f3d..dc3b237 100644
--- a/src/engine/__tests__/stacks.test.ts
+++ b/src/engine/__tests__/stacks.test.ts
@@ -43,6 +43,7 @@ import {
 	getCentralStackRulesDir,
 	loadStackRules,
 } from "../stacks.js"
+import { runSuites } from "../../test-runner.js"
 
 // ─── Per-test HOME sandboxing (mirrors global-modes.test.ts) ────────────────
 
@@ -580,22 +581,11 @@ async function main(): Promise<void> {
 		"getCentralStackRulesDir() must resolve under the per-test data dir",
 	)
 
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack : err)
-		}
-	}
+	const failed = await runSuites("stack-rules", tests)
 	if (failed > 0) {
 		console.error(`\n${failed} test(s) failed`)
 		process.exit(1)
 	}
-	console.log(`\nAll ${tests.length} stack-rules tests passed`)
 }
 
 main().finally(async () => {
diff --git a/src/engine/__tests__/summarize-tool-arg.test.ts b/src/engine/__tests__/summarize-tool-arg.test.ts
index e9e3161..4ae22c8 100644
--- a/src/engine/__tests__/summarize-tool-arg.test.ts
+++ b/src/engine/__tests__/summarize-tool-arg.test.ts
@@ -14,6 +14,7 @@ import assert from "node:assert/strict"
 
 import { readLimitFromEnv } from "../../tools/executor.js"
 import { summarizeToolArg } from "../loop.js"
+import { runMain } from "../../test-runner.js"
 
 function testSliceModeWithExplicitOffsetLimit(): void {
 	assert.equal(
@@ -104,23 +105,4 @@ const tests: Array<[string, () => void]> = [
 	["edit_file/search_replace use file_path", testEditFileAndSearchReplaceUseFilePath],
 ]
 
-function main(): void {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} tests passed`)
-}
-
-main()
+runMain("", tests)
diff --git a/src/engine/__tests__/switch-mode.test.ts b/src/engine/__tests__/switch-mode.test.ts
index 02c8254..df3dc5c 100644
--- a/src/engine/__tests__/switch-mode.test.ts
+++ b/src/engine/__tests__/switch-mode.test.ts
@@ -29,6 +29,7 @@ import { appendExtensionTools } from "../../extensions/tools.js"
 import type { SkillRegistry } from "../../skills/registry.js"
 import { createReadOnlyHeadlessExecutor } from "../../tools/executor.js"
 import type { ChatMessage, ChatTool, ChatToolCall, LlmClient, LlmRequest, LlmResponse } from "../types.js"
+import { runMain } from "../../test-runner.js"
 
 const NEEDS_DECISION = ".harness.needs-decision"
 const DECISION_ANSWER = ".harness.decision-answer"
@@ -690,26 +691,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["switch_mode: differing target tool set rebuilds the prompt (messages[0] replaced)", testToolSetChangeRebuildsPrompt],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("", tests)
diff --git a/src/evidence/__tests__/verify-files.test.ts b/src/evidence/__tests__/verify-files.test.ts
index 08e59b4..d14a3fc 100644
--- a/src/evidence/__tests__/verify-files.test.ts
+++ b/src/evidence/__tests__/verify-files.test.ts
@@ -20,6 +20,7 @@ import * as path from "node:path"
 import { FileEvidenceCollector, UnixEvidenceJournal } from "../collector.js"
 import { verifyEvidenceFiles } from "../verify.js"
 import { verifyEvidenceFiles as indexVerifyEvidenceFiles } from "../index.js"
+import { runMain } from "../../test-runner.js"
 
 interface SignedBundlePaths {
 	bundlePath: string
@@ -103,37 +104,10 @@ async function testIndexReexportsTheLeaf(): Promise<void> {
 	assert.equal(indexVerifyEvidenceFiles, verifyEvidenceFiles)
 }
 
-async function runOne(name: string, fn: () => Promise<void>): Promise<boolean> {
-	try {
-		await fn()
-		console.log(`  ok   ${name}`)
-		return true
-	} catch (error) {
-		console.error(`  FAIL ${name}`)
-		if (error instanceof Error) console.error(error.stack ?? error)
-		else console.error(String(error))
-		return false
-	}
-}
-
-async function main(): Promise<void> {
-	const tests: Array<[string, () => Promise<void>]> = [
-		["accepts a signed bundle", testVerifyFilesAcceptsSignedBundle],
-		["rejects a tampered bundle", testVerifyFilesRejectsTamperedBundle],
-		["index re-exports the leaf", testIndexReexportsTheLeaf],
-	]
-	let failed = 0
-	for (const [name, fn] of tests) {
-		if (!(await runOne(name, fn))) failed++
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} verify-files tests passed`)
-}
+const tests: Array<[string, () => Promise<void>]> = [
+	["accepts a signed bundle", testVerifyFilesAcceptsSignedBundle],
+	["rejects a tampered bundle", testVerifyFilesRejectsTamperedBundle],
+	["index re-exports the leaf", testIndexReexportsTheLeaf],
+]
 
-main().catch((error) => {
-	console.error("Test runner crashed:", error)
-	process.exit(1)
-})
+runMain("verify-files", tests)
diff --git a/src/github/__tests__/app-auth.test.ts b/src/github/__tests__/app-auth.test.ts
index 9370015..6b6cb12 100644
--- a/src/github/__tests__/app-auth.test.ts
+++ b/src/github/__tests__/app-auth.test.ts
@@ -18,6 +18,7 @@ import assert from "node:assert/strict"
 import * as crypto from "node:crypto"
 
 import { createAppAuthClient, type TokenCache } from "../app-auth.js"
+import { runMain } from "../../test-runner.js"
 
 /** A real RSA key (PKCS#1 PEM, like GitHub provides) so JWT signing works. */
 function fakePrivateKey(): string {
@@ -217,26 +218,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["exchange failure surfaces as typed AppAuthError", testExchangeFailureThrowsTypedError],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} app-auth tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("app-auth", tests)
diff --git a/src/github/__tests__/cli.test.ts b/src/github/__tests__/cli.test.ts
index 16eb368..1fb543c 100644
--- a/src/github/__tests__/cli.test.ts
+++ b/src/github/__tests__/cli.test.ts
@@ -11,6 +11,7 @@ import assert from "node:assert/strict"
 
 import type { createAppAuthClient } from "../app-auth.js"
 import { parsePushPrArgs, pushPrMain, type PushPrDeps } from "../cli.js"
+import { runMain } from "../../test-runner.js"
 
 /** The token used for API calls in these tests (fake — never a real one). */
 const FAKE_TOKEN = "ghs_TESTTOKEN1234567890"
@@ -227,26 +228,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["src/cli.ts main() dispatches push-pr to pushPrMain", testCliMainDispatchesPushPr],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} push-pr CLI tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("push-pr CLI", tests)
diff --git a/src/github/__tests__/installations.test.ts b/src/github/__tests__/installations.test.ts
index e70eec4..04365fa 100644
--- a/src/github/__tests__/installations.test.ts
+++ b/src/github/__tests__/installations.test.ts
@@ -10,6 +10,7 @@
 import assert from "node:assert/strict"
 
 import { listInstallationRepos, toInstallationRepo } from "../installations.js"
+import { runMain } from "../../test-runner.js"
 
 const FAKE_TOKEN = "ghs_TESTTOKEN1234567890"
 const API_BASE = "https://api.github.com"
@@ -197,26 +198,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["listInstallationRepos: a missing repositories field degrades to []", testListInstallationReposHandlesEmptyRepositoriesField],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? (err.stack ?? err.message) : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} github/installations tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("github/installations", tests)
diff --git a/src/github/__tests__/pr.test.ts b/src/github/__tests__/pr.test.ts
index 2acf015..f7c4c47 100644
--- a/src/github/__tests__/pr.test.ts
+++ b/src/github/__tests__/pr.test.ts
@@ -9,6 +9,7 @@
 import assert from "node:assert/strict"
 
 import { NoDiffError, OpenPrError, getPullRequestStatus, getRepoDefaultBranch, openPullRequest } from "../pr.js"
+import { runMain } from "../../test-runner.js"
 
 /** The token used for API calls in these tests (fake — never a real one). */
 const FAKE_TOKEN = "ghs_TESTTOKEN1234567890"
@@ -384,26 +385,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["getPullRequestStatus rejects a malformed 2xx body", testGetPullRequestStatusMalformedBodyThrows],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} PR tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("PR", tests)
diff --git a/src/github/__tests__/provision.test.ts b/src/github/__tests__/provision.test.ts
index 8719995..a38b990 100644
--- a/src/github/__tests__/provision.test.ts
+++ b/src/github/__tests__/provision.test.ts
@@ -17,6 +17,7 @@ import * as os from "node:os"
 import * as path from "node:path"
 
 import { provisionRepo, readOriginUrl } from "../provision.js"
+import { runMain } from "../../test-runner.js"
 
 /** The token used for clones in these tests (fake — never a real one). */
 const FAKE_TOKEN = "ghs_TESTTOKEN1234567890"
@@ -175,26 +176,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["refuses when target exists as a file", testRefusesExistingFileTarget],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} provision tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("provision", tests)
diff --git a/src/github/__tests__/push.test.ts b/src/github/__tests__/push.test.ts
index 1a7207c..f2f1594 100644
--- a/src/github/__tests__/push.test.ts
+++ b/src/github/__tests__/push.test.ts
@@ -27,6 +27,7 @@ import * as os from "node:os"
 import * as path from "node:path"
 
 import { PushError, pushBranch } from "../push.js"
+import { runMain } from "../../test-runner.js"
 
 /** The token used for pushes in these tests (fake — never a real one). */
 const FAKE_TOKEN = "ghs_TESTTOKEN1234567890"
@@ -207,26 +208,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["refuses when the token supplier returns nothing", testRefusesMissingToken],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} push tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("push", tests)
diff --git a/src/index-util.test.ts b/src/index-util.test.ts
index a03114a..26adc4d 100644
--- a/src/index-util.test.ts
+++ b/src/index-util.test.ts
@@ -18,6 +18,7 @@ import * as path from "node:path"
 
 import { hasCodebaseIndex, indexFilePathFor } from "./index-util.js"
 import { migrateLegacyProjectData, projectStoreRoot, resolveProjectDataDir, resolveProjectIdentity } from "./project-store.js"
+import { runMain } from "./test-runner.js"
 
 async function mkTmp(prefix: string): Promise<string> {
 	return fsp.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -122,26 +123,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["legacy workspace .headlesscode index migrates on first resolution", testLegacyWorkspaceIndexIsMigratedOnFirstResolution],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} index-util smoke tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("index-util smoke", tests)
diff --git a/src/init/__tests__/cli.test.ts b/src/init/__tests__/cli.test.ts
index 762a3eb..5fe2061 100644
--- a/src/init/__tests__/cli.test.ts
+++ b/src/init/__tests__/cli.test.ts
@@ -19,6 +19,7 @@ import * as path from "node:path"
 
 import { initMain, parseInitArgs } from "../cli.js"
 import { projectKeyFor, resolveProjectIdentity } from "../../project-store.js"
+import { runMain } from "../../test-runner.js"
 
 async function mkWs(files: Record<string, string> = {}): Promise<string> {
 	const ws = await fs.mkdtemp(path.join(os.tmpdir(), "hc-init-cli-"))
@@ -226,26 +227,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["existing .gitignore coverage reported as unchanged, file byte-identical", testExistingGitignoreCoverageReportedAsUnchanged],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} init cli tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("init cli", tests)
diff --git a/src/init/__tests__/gitignore.test.ts b/src/init/__tests__/gitignore.test.ts
index ac35edd..229c601 100644
--- a/src/init/__tests__/gitignore.test.ts
+++ b/src/init/__tests__/gitignore.test.ts
@@ -12,6 +12,7 @@ import * as os from "node:os"
 import * as path from "node:path"
 
 import { ensureWorkspaceGitignore } from "../gitignore.js"
+import { runMain } from "../../test-runner.js"
 
 const EXPECTED_ENTRY = "# headlesscode session artifacts (events, reports, usage, scratch, memory)\n/.headlesscode/\n"
 
@@ -141,26 +142,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["missing / non-directory workspace is a silent no-op (no .gitignore created)", testNonDirectoryWorkspaceIsSkipped],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} gitignore tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("gitignore", tests)
diff --git a/src/llm/__tests__/muse.test.ts b/src/llm/__tests__/muse.test.ts
index 5113497..ec0be43 100644
--- a/src/llm/__tests__/muse.test.ts
+++ b/src/llm/__tests__/muse.test.ts
@@ -20,6 +20,7 @@ import {
 	resolveLlmProviderOrReport,
 } from "../muse.js"
 import type { LlmRequest } from "../../engine/types.js"
+import { runMain } from "../../test-runner.js"
 
 const baseRequest: LlmRequest = {
 	model: "muse-spark-1.3",
@@ -452,23 +453,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["MUSE_BASE_URL is the documented Meta endpoint", testBaseUrlDefault],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack : err)
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} muse client tests passed`)
-}
-
-main()
+runMain("muse client", tests)
diff --git a/src/llm/__tests__/openrouter.test.ts b/src/llm/__tests__/openrouter.test.ts
index e10aca2..75f8a79 100644
--- a/src/llm/__tests__/openrouter.test.ts
+++ b/src/llm/__tests__/openrouter.test.ts
@@ -13,6 +13,7 @@ import { estimateCost } from "../../budget/cost.js"
 import { buildRequestBody, isRetryableOpenRouterError, OpenRouterClient, OpenRouterError, parseReasoningEffort } from "../openrouter.js"
 import { PromptByteBudgetError, TOOL_RESULT_BUDGET_NOTICE } from "../request-budget.js"
 import type { LlmRequest } from "../../engine/types.js"
+import { runMain } from "../../test-runner.js"
 
 const baseRequest: LlmRequest = {
 	model: "irrelevant-here",
@@ -1109,23 +1110,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["fetchModelInfo: one fetch resolves window AND price together", testFetchModelInfoSharesOneFetch],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack : err)
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} openrouter request-building tests passed`)
-}
-
-main()
+runMain("openrouter request-building", tests)
diff --git a/src/memory/__tests__/embed.test.ts b/src/memory/__tests__/embed.test.ts
index 36e19c7..59d8539 100644
--- a/src/memory/__tests__/embed.test.ts
+++ b/src/memory/__tests__/embed.test.ts
@@ -7,6 +7,7 @@
 import assert from "node:assert/strict"
 
 import { cosine, createLocalEmbedder, similarity } from "../embed.js"
+import { runMain } from "../../test-runner.js"
 
 // ─── Tests ───────────────────────────────────────────────────────────────────
 
@@ -72,23 +73,4 @@ const tests: Array<[string, () => void]> = [
 	["createLocalEmbedder validates dims", testDimsValidation],
 ]
 
-function main(): void {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} embed tests passed`)
-}
-
-main()
+runMain("embed", tests)
diff --git a/src/memory/__tests__/local.test.ts b/src/memory/__tests__/local.test.ts
index 3bd0523..3f4b2b3 100644
--- a/src/memory/__tests__/local.test.ts
+++ b/src/memory/__tests__/local.test.ts
@@ -11,6 +11,7 @@ import * as path from "node:path"
 
 import { LocalMemoryStore, sanitizeProject } from "../local.js"
 import type { SessionSummary } from "../types.js"
+import { runMain } from "../../test-runner.js"
 
 async function tmpDir(prefix: string): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -198,26 +199,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["sanitizeProject produces filesystem-safe slugs", testSanitizeProject],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} local store tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("local store", tests)
diff --git a/src/memory/__tests__/memory-wiring.test.ts b/src/memory/__tests__/memory-wiring.test.ts
index 0a561cb..ab0391e 100644
--- a/src/memory/__tests__/memory-wiring.test.ts
+++ b/src/memory/__tests__/memory-wiring.test.ts
@@ -21,6 +21,7 @@ import { HeadlessSession } from "../../engine/loop.js"
 import type { ChatMessage, LlmClient, LlmRequest, LlmResponse } from "../../engine/types.js"
 import { LocalMemoryStore } from "../local.js"
 import type { MemoryStore } from "../types.js"
+import { runMain } from "../../test-runner.js"
 
 const PROJECT = "wiring-project"
 
@@ -204,26 +205,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["memory failures are non-fatal to the session", testMemoryFailureDoesNotFailSession],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} memory-wiring tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("memory-wiring", tests)
diff --git a/src/memory/__tests__/summarizer.test.ts b/src/memory/__tests__/summarizer.test.ts
index 18ffff9..cf1a7f5 100644
--- a/src/memory/__tests__/summarizer.test.ts
+++ b/src/memory/__tests__/summarizer.test.ts
@@ -9,6 +9,7 @@ import assert from "node:assert/strict"
 import { buildRollingSummary, extractFacts, extractSessionSummary } from "../summarizer.js"
 import type { ChatMessage, SessionResult } from "../../engine/types.js"
 import type { SessionSummary } from "../types.js"
+import { runMain } from "../../test-runner.js"
 
 function toolMessage(name: string, content: string): ChatMessage {
 	return { role: "tool", tool_call_id: `c_${Math.random()}`, name, content }
@@ -150,23 +151,4 @@ const tests: Array<[string, () => void]> = [
 	["rolling summary renders key facts", testRollingSummaryIncludesKeyFacts],
 ]
 
-function main(): void {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} summarizer tests passed`)
-}
-
-main()
+runMain("summarizer", tests)
diff --git a/src/memory/__tests__/uwuchat.test.ts b/src/memory/__tests__/uwuchat.test.ts
index 683a27a..87401cf 100644
--- a/src/memory/__tests__/uwuchat.test.ts
+++ b/src/memory/__tests__/uwuchat.test.ts
@@ -8,6 +8,7 @@
 import assert from "node:assert/strict"
 
 import { API_PREFIX, UwUChatMemoryError, UwUChatMemoryStore } from "../uwuchat.js"
+import { runMain } from "../../test-runner.js"
 
 const BASE = "https://uwu.example"
 
@@ -163,26 +164,4 @@ const tests: Array<[string, () => Promise<void> | void]> = [
 	["typed UwUChatMemoryError on non-2xx", testTypedErrorOnNon2xx],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} uwuchat tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("uwuchat", tests)
diff --git a/src/migrate/cli.test.ts b/src/migrate/cli.test.ts
index ee9f7db..708a4f8 100644
--- a/src/migrate/cli.test.ts
+++ b/src/migrate/cli.test.ts
@@ -12,6 +12,7 @@ import * as path from "node:path"
 
 import { migrateMain } from "./cli.js"
 import { projectStoreRoot, resolveProjectDataDir } from "../project-store.js"
+import { runMain } from "../test-runner.js"
 
 async function mkTmp(prefix: string): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -86,26 +87,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["migrate respects the store-root override", testMigrateUsesStoreRoot],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} migrate CLI tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("migrate CLI", tests)
diff --git a/src/orchestrator/__tests__/analyze-cli.test.ts b/src/orchestrator/__tests__/analyze-cli.test.ts
index f8a133a..f1955a3 100644
--- a/src/orchestrator/__tests__/analyze-cli.test.ts
+++ b/src/orchestrator/__tests__/analyze-cli.test.ts
@@ -11,6 +11,7 @@ import * as path from "node:path"
 
 import { appendEvent, type EventRecord } from "../../engine/events.js"
 import { analyzeCliMain, parseAnalyzeArgs } from "../analyze-cli.js"
+import { runMain } from "../../test-runner.js"
 
 // ─── Helpers ─────────────────────────────────────────────────────────────────
 
@@ -148,26 +149,4 @@ const tests: [string, () => Promise<void>][] = [
 	["main --json prints raw analysis", testMainJsonOutput],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? (err.stack ?? err.message) : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} analyze-cli tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("analyze-cli", tests)
diff --git a/src/orchestrator/__tests__/arg-cursor.test.ts b/src/orchestrator/__tests__/arg-cursor.test.ts
index 57d4f5b..422993e 100644
--- a/src/orchestrator/__tests__/arg-cursor.test.ts
+++ b/src/orchestrator/__tests__/arg-cursor.test.ts
@@ -9,6 +9,7 @@
 import assert from "node:assert/strict"
 
 import { createArgCursor } from "../arg-cursor.js"
+import { runMain } from "../../test-runner.js"
 
 function cursor(argv: string[], i: number): {
 	flag: string
@@ -83,26 +84,4 @@ const tests: Array<[string, () => void | Promise<void>]> = [
 	["nextRaw accepts dashed values", testNextRawAcceptsDashedValue],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} arg-cursor tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("arg-cursor", tests)
diff --git a/src/orchestrator/__tests__/auto-split.test.ts b/src/orchestrator/__tests__/auto-split.test.ts
index 5f8a6d7..2bfb74e 100644
--- a/src/orchestrator/__tests__/auto-split.test.ts
+++ b/src/orchestrator/__tests__/auto-split.test.ts
@@ -12,6 +12,7 @@ import assert from "node:assert/strict"
 import { autoSplitOversizedIssues, proposeSemanticSplit } from "../auto-split.js"
 import type { IssueSizeWarning, SplitIssue } from "../split.js"
 import type { LlmClient, LlmRequest, LlmResponse } from "../../engine/types.js"
+import { runMain } from "../../test-runner.js"
 
 function issue(number: number, title: string, body?: string): SplitIssue {
 	return { number, title, body }
@@ -241,26 +242,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["autoSplitOversizedIssues: preserves the relative order of unflagged issues", testAutoSplitPreservesOrderOfUnflaggedIssues],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} auto-split tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("auto-split", tests)
diff --git a/src/orchestrator/__tests__/cleanup.test.ts b/src/orchestrator/__tests__/cleanup.test.ts
index 2658f0f..a72664e 100644
--- a/src/orchestrator/__tests__/cleanup.test.ts
+++ b/src/orchestrator/__tests__/cleanup.test.ts
@@ -35,6 +35,7 @@ import {
 	type OrchestratorGroup,
 	type OrchestratorState,
 } from "../state.js"
+import { runMain } from "../../test-runner.js"
 
 const ownedFixtureDirectories = new Set<string>()
 const FIXTURE_REMOVE_ATTEMPTS = 5
@@ -771,26 +772,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["orchestrate status shows the cleanup column (eligible) for a merged terminal group", testStatusShowsCleanupColumn],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} cleanup tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("cleanup", tests)
diff --git a/src/orchestrator/__tests__/cli.test.ts b/src/orchestrator/__tests__/cli.test.ts
index 0567844..d431a3f 100644
--- a/src/orchestrator/__tests__/cli.test.ts
+++ b/src/orchestrator/__tests__/cli.test.ts
@@ -34,6 +34,7 @@ import { watchGroups } from "../watch.js"
 import { defaultState, loadStateSync, saveStateSync, updateGroup, type OrchestratorGroup, type OrchestratorState } from "../state.js"
 import type { ReviewResult } from "../reviewer.js"
 import type { QaResult } from "../../qa/qa.js"
+import { runSuites } from "../../test-runner.js"
 
 async function tmpRepo(): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), "headlesscode-cli-rework-"))
@@ -1839,16 +1840,7 @@ async function main(): Promise<void> {
 	process.env.HEADLESSCODE_DATA_DIR = storeTmp
 	let failed = 0
 	try {
-		for (const [name, fn] of tests) {
-			try {
-				await fn()
-				console.log(`  ok   ${name}`)
-			} catch (err) {
-				failed++
-				console.error(`  FAIL ${name}`)
-				console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-			}
-		}
+		failed = await runSuites("cli", tests)
 	} finally {
 		delete process.env.HEADLESSCODE_DATA_DIR
 		await fs.rm(storeTmp, { recursive: true, force: true })
@@ -1857,7 +1849,6 @@ async function main(): Promise<void> {
 		console.error(`\n${failed} test(s) failed`)
 		process.exit(1)
 	}
-	console.log(`\nAll ${tests.length} cli tests passed`)
 }
 
 main().catch((err) => {
diff --git a/src/orchestrator/__tests__/cost-estimate.test.ts b/src/orchestrator/__tests__/cost-estimate.test.ts
index 0701b63..776933f 100644
--- a/src/orchestrator/__tests__/cost-estimate.test.ts
+++ b/src/orchestrator/__tests__/cost-estimate.test.ts
@@ -23,6 +23,7 @@ import {
 	VERIFICATION_MULTIPLIER_FULL_SAMPLES,
 	verificationMultiplier,
 } from "../cost-estimate.js"
+import { runMain } from "../../test-runner.js"
 
 // ─── Helpers ─────────────────────────────────────────────────────────────────
 
@@ -538,26 +539,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["LOW-CONFIDENCE flagged on an under-sampled direct match (#126)", testLowConfidenceFlaggedOnUnderSampledDirectMatch],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? (err.stack ?? err.message) : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} cost-estimate tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("cost-estimate", tests)
diff --git a/src/orchestrator/__tests__/cost-history-cli.test.ts b/src/orchestrator/__tests__/cost-history-cli.test.ts
index b1ac56c..3a92be7 100644
--- a/src/orchestrator/__tests__/cost-history-cli.test.ts
+++ b/src/orchestrator/__tests__/cost-history-cli.test.ts
@@ -12,6 +12,7 @@ import { execFileSync } from "node:child_process"
 
 import { appendCostHistoryRecord, appendSessionCostRecord, type CostHistoryRecord, type SessionCostRecord } from "../cost-history.js"
 import { costHistoryCliMain, parseCostHistoryArgs } from "../cost-history-cli.js"
+import { runSuites } from "../../test-runner.js"
 
 // ─── Helpers ─────────────────────────────────────────────────────────────────
 
@@ -327,16 +328,7 @@ async function main(): Promise<void> {
 	process.env.HEADLESSCODE_DATA_DIR = storeTmp
 	let failed = 0
 	try {
-		for (const [name, fn] of tests) {
-			try {
-				await fn()
-				console.log(`  ok   ${name}`)
-			} catch (err) {
-				failed++
-				console.error(`  FAIL ${name}`)
-				console.error(err instanceof Error ? (err.stack ?? err.message) : String(err))
-			}
-		}
+		failed = await runSuites("cost-history-cli", tests)
 	} finally {
 		delete process.env.HEADLESSCODE_DATA_DIR
 		await fs.rm(storeTmp, { recursive: true, force: true })
@@ -345,7 +337,6 @@ async function main(): Promise<void> {
 		console.error(`\n${failed} test(s) failed`)
 		process.exit(1)
 	}
-	console.log(`\nAll ${tests.length} cost-history-cli tests passed`)
 }
 
 main().catch((err) => {
diff --git a/src/orchestrator/__tests__/cost-history.test.ts b/src/orchestrator/__tests__/cost-history.test.ts
index da58658..a789354 100644
--- a/src/orchestrator/__tests__/cost-history.test.ts
+++ b/src/orchestrator/__tests__/cost-history.test.ts
@@ -21,6 +21,7 @@ import {
 	type CostHistoryRecord,
 } from "../cost-history.js"
 import type { OrchestratorGroup } from "../state.js"
+import { runSuites } from "../../test-runner.js"
 
 // ─── Helpers ─────────────────────────────────────────────────────────────────
 
@@ -393,16 +394,7 @@ async function main(): Promise<void> {
 	process.env.HEADLESSCODE_DATA_DIR = storeTmp
 	let failed = 0
 	try {
-		for (const [name, fn] of tests) {
-			try {
-				await fn()
-				console.log(`  ok   ${name}`)
-			} catch (err) {
-				failed++
-				console.error(`  FAIL ${name}`)
-				console.error(err instanceof Error ? (err.stack ?? err.message) : String(err))
-			}
-		}
+		failed = await runSuites("cost-history", tests)
 	} finally {
 		delete process.env.HEADLESSCODE_DATA_DIR
 		await fs.rm(storeTmp, { recursive: true, force: true })
@@ -411,7 +403,6 @@ async function main(): Promise<void> {
 		console.error(`\n${failed} test(s) failed`)
 		process.exit(1)
 	}
-	console.log(`\nAll ${tests.length} cost-history tests passed`)
 }
 
 main().catch((err) => {
diff --git a/src/orchestrator/__tests__/git-sync.test.ts b/src/orchestrator/__tests__/git-sync.test.ts
index 0537b76..f34479c 100644
--- a/src/orchestrator/__tests__/git-sync.test.ts
+++ b/src/orchestrator/__tests__/git-sync.test.ts
@@ -21,6 +21,7 @@ import {
 	syncWarningLines,
 	TRIVIAL_DRIFT_AHEAD,
 } from "../git-sync.js"
+import { runMain } from "../../test-runner.js"
 
 function git(repo: string, args: string[]): string {
 	return execFileSync("git", ["-C", repo, ...args], { encoding: "utf-8", stdio: ["ignore", "pipe", "ignore"] }).trim()
@@ -274,26 +275,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["TRIVIAL_DRIFT_AHEAD threshold is exported", testTrivialDriftThresholdIsExported],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} git-sync tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("git-sync", tests)
diff --git a/src/orchestrator/__tests__/log-analysis.test.ts b/src/orchestrator/__tests__/log-analysis.test.ts
index 0f2e4d8..66e0948 100644
--- a/src/orchestrator/__tests__/log-analysis.test.ts
+++ b/src/orchestrator/__tests__/log-analysis.test.ts
@@ -12,6 +12,7 @@ import * as path from "node:path"
 
 import { appendEvent, type EventRecord } from "../../engine/events.js"
 import { analyzeWorktreeSessions, formatAnalysisReport, listSessionIds } from "../log-analysis.js"
+import { runMain } from "../../test-runner.js"
 
 // ─── Helpers ─────────────────────────────────────────────────────────────────
 
@@ -175,26 +176,4 @@ const tests: [string, () => Promise<void>][] = [
 	["formatAnalysisReport renders readable text", testFormatAnalysisReportIsReadable],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? (err.stack ?? err.message) : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} log-analysis tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("log-analysis", tests)
diff --git a/src/orchestrator/__tests__/main-tree-tripwire.test.ts b/src/orchestrator/__tests__/main-tree-tripwire.test.ts
index a9aad06..82c4499 100644
--- a/src/orchestrator/__tests__/main-tree-tripwire.test.ts
+++ b/src/orchestrator/__tests__/main-tree-tripwire.test.ts
@@ -9,6 +9,7 @@ import { fileURLToPath } from "node:url"
 import { removeHarnessArtifacts } from "../cleanup.js"
 import { buildStatusSummary, formatStatusText } from "../status.js"
 import { reconcileGroups } from "../status.js"
+import { runMain } from "../../test-runner.js"
 import {
 	DEFAULT_STALL_TIMEOUT_MS,
 	inspectGroup,
@@ -202,14 +203,14 @@ async function cleanupRemovesMarkers(dir: string): Promise<void> {
 	}
 }
 
-const tests: Array<[string, (fx: Fixture) => Promise<void>]> = [
-	["clean run leaves no marker", cleanRunLeavesNoMarker],
-	["escaped write is reported", escapedWriteIsReported],
-	["orchestrator dirs excluded", orchestratorDirsExcluded],
-	["missing baseline reports all", missingBaselineReportsAll],
-	["main checkout refuses loudly", mainCheckoutRefusesLoudly],
-	["qa role uses qa markers", qaRoleUsesQaMarkers],
-	["script and reader agree", scriptAndReaderAgree],
+const tests: Array<[string, () => Promise<void>]> = [
+	["clean run leaves no marker", () => withFixture(cleanRunLeavesNoMarker)],
+	["escaped write is reported", () => withFixture(escapedWriteIsReported)],
+	["orchestrator dirs excluded", () => withFixture(orchestratorDirsExcluded)],
+	["missing baseline reports all", () => withFixture(missingBaselineReportsAll)],
+	["main checkout refuses loudly", () => withFixture(mainCheckoutRefusesLoudly)],
+	["qa role uses qa markers", () => withFixture(qaRoleUsesQaMarkers)],
+	["script and reader agree", () => withFixture(scriptAndReaderAgree)],
 	["reader reads marker lines", () => withTmpDir(readerReadsMarkerLines)],
 	["inspect persists dirty", () => withTmpDir(inspectPersistsDirty)],
 	["reconcile persists dirty", () => withTmpDir(reconcilePersistsDirty)],
@@ -217,27 +218,4 @@ const tests: Array<[string, (fx: Fixture) => Promise<void>]> = [
 	["cleanup removes markers", () => withTmpDir(cleanupRemovesMarkers)],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await withFixture(fn)
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			const detail = err instanceof Error ? err.stack ?? err.message : String(err)
-			console.error(detail)
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} main-tree-tripwire tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("main-tree-tripwire", tests)
diff --git a/src/orchestrator/__tests__/merge-check.test.ts b/src/orchestrator/__tests__/merge-check.test.ts
index 66fa679..898c1be 100644
--- a/src/orchestrator/__tests__/merge-check.test.ts
+++ b/src/orchestrator/__tests__/merge-check.test.ts
@@ -18,6 +18,7 @@ import * as os from "node:os"
 import * as path from "node:path"
 
 import { checkMergedByAncestor, checkMergedByGitHubPr } from "../merge-check.js"
+import { runMain } from "../../test-runner.js"
 
 function git(repo: string, args: string[]): string {
 	return execFileSync("git", ["-C", repo, ...args], { encoding: "utf-8", stdio: ["ignore", "pipe", "ignore"] }).trim()
@@ -246,26 +247,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["github-pr: network error fails closed as unknown", testGitHubNetworkErrorFailsClosed],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} merge-check tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("merge-check", tests)
diff --git a/src/orchestrator/__tests__/model-resolution.test.ts b/src/orchestrator/__tests__/model-resolution.test.ts
index 22d0549..0bacf9b 100644
--- a/src/orchestrator/__tests__/model-resolution.test.ts
+++ b/src/orchestrator/__tests__/model-resolution.test.ts
@@ -21,6 +21,7 @@ import { resolveModelForMode } from "../../config/mode-models.js"
 import { handleReviewVerdict, parseOrchestrateArgs } from "../cli.js"
 import type { OrchestratorGroup } from "../state.js"
 import type { ReviewResult } from "../reviewer.js"
+import { runSuites } from "../../test-runner.js"
 
 async function tmpRepo(): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), "headlesscode-model-res-"))
@@ -157,16 +158,7 @@ async function main(): Promise<void> {
 	process.env.HEADLESSCODE_DATA_DIR = storeTmp
 	let failed = 0
 	try {
-		for (const [name, fn] of tests) {
-			try {
-				await fn()
-				console.log(`  ok   ${name}`)
-			} catch (err) {
-				failed++
-				console.error(`  FAIL ${name}`)
-				console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-			}
-		}
+		failed = await runSuites("model-resolution", tests)
 	} finally {
 		delete process.env.HEADLESSCODE_DATA_DIR
 		await fs.rm(storeTmp, { recursive: true, force: true })
@@ -175,7 +167,6 @@ async function main(): Promise<void> {
 		console.error(`\n${failed} test(s) failed`)
 		process.exit(1)
 	}
-	console.log(`\nAll ${tests.length} model-resolution tests passed`)
 }
 
 main().catch((err) => {
diff --git a/src/orchestrator/__tests__/pipeline.test.ts b/src/orchestrator/__tests__/pipeline.test.ts
index 30211ec..14dc1ef 100644
--- a/src/orchestrator/__tests__/pipeline.test.ts
+++ b/src/orchestrator/__tests__/pipeline.test.ts
@@ -25,6 +25,7 @@ import {
 	defaultResearchTaskText,
 	parseFilingResult,
 } from "../pipeline.js"
+import { runMain } from "../../test-runner.js"
 
 function testParseFilingResultStructuredLine(): void {
 	// The required `ISSUES:` line is authoritative.
@@ -92,26 +93,4 @@ const tests: Array<[string, () => void | Promise<void>]> = [
 	["mode constants match .roomodes slugs", testModeConstants],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} pipeline tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("pipeline", tests)
diff --git a/src/orchestrator/__tests__/resume.test.ts b/src/orchestrator/__tests__/resume.test.ts
index c882bac..6574479 100644
--- a/src/orchestrator/__tests__/resume.test.ts
+++ b/src/orchestrator/__tests__/resume.test.ts
@@ -39,6 +39,7 @@ import {
 } from "../resume.js"
 import { defaultState, loadStateSync, saveStateSync, type OrchestratorGroup, type OrchestratorState } from "../state.js"
 import { assessVerificationActivity, commandIdentity, parseRequiredCommandIdentities } from "../verification-gate.js"
+import { runSuites } from "../../test-runner.js"
 
 async function tmpRepo(): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), "headlesscode-resume-"))
@@ -1355,16 +1356,7 @@ async function main(): Promise<void> {
 	process.env.HEADLESSCODE_DATA_DIR = storeTmp
 	let failed = 0
 	try {
-		for (const [name, fn] of tests) {
-			try {
-				await fn()
-				console.log(`  ok   ${name}`)
-			} catch (err) {
-				failed++
-				console.error(`  FAIL ${name}`)
-				console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-			}
-		}
+		failed = await runSuites("resume", tests)
 	} finally {
 		delete process.env.HEADLESSCODE_DATA_DIR
 		await fs.rm(storeTmp, { recursive: true, force: true })
@@ -1373,7 +1365,6 @@ async function main(): Promise<void> {
 		console.error(`\n${failed} test(s) failed`)
 		process.exit(1)
 	}
-	console.log(`\nAll ${tests.length} resume tests passed`)
 }
 
 main().catch((err) => {
diff --git a/src/orchestrator/__tests__/review-gates.test.ts b/src/orchestrator/__tests__/review-gates.test.ts
index a1faddf..27f0efe 100644
--- a/src/orchestrator/__tests__/review-gates.test.ts
+++ b/src/orchestrator/__tests__/review-gates.test.ts
@@ -8,6 +8,7 @@ import type { CodeIntel } from "../../codeintel/program.js"
 import type { GateRunner } from "../review-gates.js"
 import { runReviewGates } from "../review-gates.js"
 import { hasRealWorktreeChanges } from "../verification-gate.js"
+import { runMain } from "../../test-runner.js"
 
 function git(repo: string, args: string[]): string {
 	return execFileSync("git", ["-C", repo, ...args], {
@@ -162,27 +163,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["uncovered sources flagged", () => withFixture(uncoveredSourceFlagged)],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			const detail = err instanceof Error ? err.stack ?? err.message : String(err)
-			console.error(detail)
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} review-gates tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("review-gates", tests)
diff --git a/src/orchestrator/__tests__/reviewer.test.ts b/src/orchestrator/__tests__/reviewer.test.ts
index ff9155c..96d2437 100644
--- a/src/orchestrator/__tests__/reviewer.test.ts
+++ b/src/orchestrator/__tests__/reviewer.test.ts
@@ -10,6 +10,7 @@ import * as path from "node:path"
 
 import { parseReviewResult, runReview, runReviewWithRetries } from "../reviewer.js"
 import type { ChatMessage, ChatToolCall, LlmClient, LlmRequest, LlmResponse } from "../../engine/types.js"
+import { runMain } from "../../test-runner.js"
 
 // ─── Fake LLM client (same pattern as engine/__tests__/loop.test.ts) ─────────
 
@@ -446,26 +447,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["runReviewWithRetries: exhausts retries and returns the final error verdict", testRunReviewWithRetriesExhaustsRetries],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} reviewer tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("reviewer", tests)
diff --git a/src/orchestrator/__tests__/spawn-model.test.ts b/src/orchestrator/__tests__/spawn-model.test.ts
index 10b8a71..762b019 100644
--- a/src/orchestrator/__tests__/spawn-model.test.ts
+++ b/src/orchestrator/__tests__/spawn-model.test.ts
@@ -23,6 +23,7 @@ import * as fsPromises from "node:fs/promises"
 import * as os from "node:os"
 import * as path from "node:path"
 import { fileURLToPath } from "node:url"
+import { runSuites } from "../../test-runner.js"
 
 /** This repo root (parent of src/orchestrator/__tests__/). */
 const REPO_ROOT = fileURLToPath(new URL("../../..", import.meta.url))
@@ -256,16 +257,7 @@ async function main(): Promise<void> {
 	process.env.HEADLESSCODE_DATA_DIR = storeTmp
 	let failed = 0
 	try {
-		for (const [name, fn] of tests) {
-			try {
-				await fn()
-				console.log(`  ok   ${name}`)
-			} catch (err) {
-				failed++
-				console.error(`  FAIL ${name}`)
-				console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-			}
-		}
+		failed = await runSuites("spawn-model", tests)
 	} finally {
 		delete process.env.HEADLESSCODE_DATA_DIR
 		await fsPromises.rm(storeTmp, { recursive: true, force: true })
@@ -274,7 +266,6 @@ async function main(): Promise<void> {
 		console.error(`\n${failed} test(s) failed`)
 		process.exit(1)
 	}
-	console.log(`\nAll ${tests.length} spawn-model tests passed`)
 }
 
 main().catch((err) => {
diff --git a/src/orchestrator/__tests__/spawn-state-reset.test.ts b/src/orchestrator/__tests__/spawn-state-reset.test.ts
index 4f351b5..e669b85 100644
--- a/src/orchestrator/__tests__/spawn-state-reset.test.ts
+++ b/src/orchestrator/__tests__/spawn-state-reset.test.ts
@@ -23,6 +23,7 @@ import * as fsPromises from "node:fs/promises"
 import * as os from "node:os"
 import * as path from "node:path"
 import { fileURLToPath } from "node:url"
+import { runMain } from "../../test-runner.js"
 
 const REPO_ROOT = fileURLToPath(new URL("../../..", import.meta.url))
 const SPAWN_SCRIPT = path.join(REPO_ROOT, "scripts", "spawn-parallel-worktrees.sh")
@@ -164,26 +165,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["reused worktree slot clears stale review_verdict/qa/cost_recorded (issue #28)", testReusedSlotClearsStaleReviewQaCostFields],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? (err.stack ?? err.message) : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} spawn-state-reset tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("spawn-state-reset", tests)
diff --git a/src/orchestrator/__tests__/split.test.ts b/src/orchestrator/__tests__/split.test.ts
index b9d3c8b..032d8e0 100644
--- a/src/orchestrator/__tests__/split.test.ts
+++ b/src/orchestrator/__tests__/split.test.ts
@@ -20,6 +20,7 @@ import {
 	ISSUE_SIZE_WARN_THRESHOLD,
 } from "../split.js"
 import type { SplitIssue, WorktreeSpec } from "../split.js"
+import { runMain } from "../../test-runner.js"
 
 function issue(number: number, title: string, body?: string): SplitIssue {
 	return { number, title, body }
@@ -278,26 +279,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["issueSizeWarnings flags only 3+-section issues (issue #53)", testIssueSizeWarningsOnlyFlagsThreePlusSections],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} split tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("split", tests)
diff --git a/src/orchestrator/__tests__/state.test.ts b/src/orchestrator/__tests__/state.test.ts
index 562c1fe..fc07d92 100644
--- a/src/orchestrator/__tests__/state.test.ts
+++ b/src/orchestrator/__tests__/state.test.ts
@@ -21,6 +21,7 @@ import {
 	updateGroup,
 	type OrchestratorState,
 } from "../state.js"
+import { runMain } from "../../test-runner.js"
 
 async function tmpStatePath(): Promise<string> {
 	const dir = await fs.mkdtemp(path.join(os.tmpdir(), "headlesscode-state-"))
@@ -209,26 +210,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["patchGroup survives two concurrent OS processes racing the same state file (issue #43)", testPatchGroupSurvivesTwoConcurrentProcesses],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} state tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("state", tests)
diff --git a/src/orchestrator/__tests__/status-cli.test.ts b/src/orchestrator/__tests__/status-cli.test.ts
index 89817b4..76278fd 100644
--- a/src/orchestrator/__tests__/status-cli.test.ts
+++ b/src/orchestrator/__tests__/status-cli.test.ts
@@ -22,6 +22,7 @@ import {
 	type OrchestratorGroup,
 	type OrchestratorState,
 } from "../state.js"
+import { runMain } from "../../test-runner.js"
 
 async function tmpRepo(): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), "headlesscode-status-"))
@@ -1037,26 +1038,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["wait reports the reconciled group", testWaitReportsReconciled],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} status-cli tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("status-cli", tests)
diff --git a/src/orchestrator/__tests__/status.test.ts b/src/orchestrator/__tests__/status.test.ts
index d498c95..b1c3fc9 100644
--- a/src/orchestrator/__tests__/status.test.ts
+++ b/src/orchestrator/__tests__/status.test.ts
@@ -15,6 +15,7 @@ import assert from "node:assert/strict"
 
 import { buildStatusSummary, formatStatusText } from "../status.js"
 import type { OrchestratorGroup, OrchestratorState } from "../state.js"
+import { runMain } from "../../test-runner.js"
 
 // ─── Helpers ─────────────────────────────────────────────────────────────────
 
@@ -149,26 +150,4 @@ const tests: [string, () => Promise<void>][] = [
 	["formatStatusText omits batch usage when absent", testFormattedTextOmitsBatchUsageWhenAbsent],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? (err.stack ?? err.message) : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} status tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("status", tests)
diff --git a/src/orchestrator/__tests__/stop-cli.test.ts b/src/orchestrator/__tests__/stop-cli.test.ts
index 22c923d..055d226 100644
--- a/src/orchestrator/__tests__/stop-cli.test.ts
+++ b/src/orchestrator/__tests__/stop-cli.test.ts
@@ -17,6 +17,7 @@ import * as path from "node:path"
 
 import { parseStopArgs, stopMain } from "../cli.js"
 import { defaultState, saveStateSync, updateGroup, type OrchestratorGroup, type OrchestratorState } from "../state.js"
+import { runMain } from "../../test-runner.js"
 
 async function tmpRepo(): Promise<string> {
 	return fsp.mkdtemp(path.join(os.tmpdir(), "headlesscode-stop-"))
@@ -235,26 +236,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["stop: usage errors exit 2, --help exits 0", testStopUsageErrors],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} stop-cli tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("stop-cli", tests)
diff --git a/src/orchestrator/__tests__/task-files.test.ts b/src/orchestrator/__tests__/task-files.test.ts
index ed8fedc..a93831a 100644
--- a/src/orchestrator/__tests__/task-files.test.ts
+++ b/src/orchestrator/__tests__/task-files.test.ts
@@ -29,6 +29,7 @@ import {
 	buildTaskFileContent as cliBuildTaskFileContent,
 	writeTaskFiles as cliWriteTaskFiles,
 } from "../cli.js"
+import { runMain } from "../../test-runner.js"
 
 function renderSameNumberPair(): [string, string] {
 	const slices = [
@@ -121,41 +122,14 @@ async function testOrchestratorCliReexportsTheLeaf(): Promise<void> {
 	assert.equal(cliWriteTaskFiles, writeTaskFiles)
 }
 
-async function runOne(name: string, fn: () => Promise<void>): Promise<boolean> {
-	try {
-		await fn()
-		console.log(`  ok   ${name}`)
-		return true
-	} catch (error) {
-		console.error(`  FAIL ${name}`)
-		if (error instanceof Error) console.error(error.stack ?? error)
-		else console.error(String(error))
-		return false
-	}
-}
-
-async function main(): Promise<void> {
-	const tests: Array<[string, () => Promise<void>]> = [
-		["inlines bodies", testBuildTaskFileContentInlineBodies],
-		["falls back", testBuildTaskFileContentFallback],
-		["plan-first mentions planning", testPlanFirstMentionsPlanning],
-		["rework embeds findings", testReworkEmbedsFindings],
-		["qa-rework and continuation", testQaReworkAndContinuation],
-		["writeTaskFiles writes both", testWriteTaskFilesWritesBoth],
-		["cli re-exports the leaf", testOrchestratorCliReexportsTheLeaf],
-	]
-	let failed = 0
-	for (const [name, fn] of tests) {
-		if (!(await runOne(name, fn))) failed++
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} task-files tests passed`)
-}
+const tests: Array<[string, () => Promise<void>]> = [
+	["inlines bodies", testBuildTaskFileContentInlineBodies],
+	["falls back", testBuildTaskFileContentFallback],
+	["plan-first mentions planning", testPlanFirstMentionsPlanning],
+	["rework embeds findings", testReworkEmbedsFindings],
+	["qa-rework and continuation", testQaReworkAndContinuation],
+	["writeTaskFiles writes both", testWriteTaskFilesWritesBoth],
+	["cli re-exports the leaf", testOrchestratorCliReexportsTheLeaf],
+]
 
-main().catch((error) => {
-	console.error("Test runner crashed:", error)
-	process.exit(1)
-})
+runMain("task-files", tests)
diff --git a/src/orchestrator/__tests__/terminal-status.test.ts b/src/orchestrator/__tests__/terminal-status.test.ts
index 936d250..3a1a6f8 100644
--- a/src/orchestrator/__tests__/terminal-status.test.ts
+++ b/src/orchestrator/__tests__/terminal-status.test.ts
@@ -18,6 +18,7 @@ import {
 	TERMINAL_STATUSES as STATUS_LIST,
 } from "../status.js"
 import { isTerminalStatus, TERMINAL_STATUSES } from "../terminal-status.js"
+import { runMain } from "../../test-runner.js"
 
 async function testTerminalStatusesAreTerminal(): Promise<void> {
 	for (const status of ["done", "failed", "needs-human", "orphaned"]) {
@@ -43,38 +44,11 @@ async function testStatusReexportsTheLeaf(): Promise<void> {
 	assert.equal(statusIsTerminal("running"), isTerminalStatus("running"))
 }
 
-async function runOne(name: string, fn: () => Promise<void>): Promise<boolean> {
-	try {
-		await fn()
-		console.log(`  ok   ${name}`)
-		return true
-	} catch (error) {
-		console.error(`  FAIL ${name}`)
-		if (error instanceof Error) console.error(error.stack ?? error)
-		else console.error(String(error))
-		return false
-	}
-}
-
-async function main(): Promise<void> {
-	const tests: Array<[string, () => Promise<void>]> = [
-		["terminal statuses are terminal", testTerminalStatusesAreTerminal],
-		["live statuses are not terminal", testLiveStatusesAreNotTerminal],
-		["blocked stays non-terminal", testBlockedStaysNonTerminal],
-		["status re-exports the leaf", testStatusReexportsTheLeaf],
-	]
-	let failed = 0
-	for (const [name, fn] of tests) {
-		if (!(await runOne(name, fn))) failed++
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} terminal-status tests passed`)
-}
+const tests: Array<[string, () => Promise<void>]> = [
+	["terminal statuses are terminal", testTerminalStatusesAreTerminal],
+	["live statuses are not terminal", testLiveStatusesAreNotTerminal],
+	["blocked stays non-terminal", testBlockedStaysNonTerminal],
+	["status re-exports the leaf", testStatusReexportsTheLeaf],
+]
 
-main().catch((error) => {
-	console.error("Test runner crashed:", error)
-	process.exit(1)
-})
+runMain("terminal-status", tests)
diff --git a/src/orchestrator/__tests__/watch.test.ts b/src/orchestrator/__tests__/watch.test.ts
index cf33ae7..ee265a0 100644
--- a/src/orchestrator/__tests__/watch.test.ts
+++ b/src/orchestrator/__tests__/watch.test.ts
@@ -32,6 +32,7 @@ import {
 	type OrchestratorGroup,
 	type OrchestratorState,
 } from "../state.js"
+import { runSuites } from "../../test-runner.js"
 
 async function tmpRepo(): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), "headlesscode-watch-blocked-"))
@@ -1045,16 +1046,7 @@ async function main(): Promise<void> {
 	process.env.HEADLESSCODE_DATA_DIR = storeTmp
 	let failed = 0
 	try {
-		for (const [name, fn] of tests) {
-			try {
-				await fn()
-				console.log(`  ok   ${name}`)
-			} catch (err) {
-				failed++
-				console.error(`  FAIL ${name}`)
-				console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-			}
-		}
+		failed = await runSuites("watch (blocked-status)", tests)
 	} finally {
 		delete process.env.HEADLESSCODE_DATA_DIR
 		await fs.rm(storeTmp, { recursive: true, force: true })
@@ -1063,7 +1055,6 @@ async function main(): Promise<void> {
 		console.error(`\n${failed} test(s) failed`)
 		process.exit(1)
 	}
-	console.log(`\nAll ${tests.length} watch (blocked-status) tests passed`)
 }
 
 main().catch((err) => {
diff --git a/src/permissions/__tests__/commands.test.ts b/src/permissions/__tests__/commands.test.ts
index 1335c74..f7aee57 100644
--- a/src/permissions/__tests__/commands.test.ts
+++ b/src/permissions/__tests__/commands.test.ts
@@ -31,6 +31,7 @@ import {
 	redirectTargets,
 	resolveRedirectTarget,
 } from "../commands.js"
+import { runMain } from "../../test-runner.js"
 
 // ─── (a) dangerous substitutions are always refused, regardless of config ───
 
@@ -570,26 +571,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["cd refusal precedes redirect; describes", testCdRefusalPrecedesRedirect],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} command-permissions tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("command-permissions", tests)
diff --git a/src/permissions/__tests__/protected-files.test.ts b/src/permissions/__tests__/protected-files.test.ts
index 35be5cf..8d7aecd 100644
--- a/src/permissions/__tests__/protected-files.test.ts
+++ b/src/permissions/__tests__/protected-files.test.ts
@@ -22,6 +22,7 @@ import {
 	resolvePermissions,
 	type PermissionsOverrides,
 } from "../config.js"
+import { runSuites } from "../../test-runner.js"
 
 async function mkTmpWorkspace(prefix: string): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -218,16 +219,7 @@ async function main(): Promise<void> {
 	process.env.HEADLESSCODE_DATA_DIR = storeTmp
 	let failed = 0
 	try {
-		for (const [name, fn] of tests) {
-			try {
-				await fn()
-				console.log(`  ok   ${name}`)
-			} catch (err) {
-				failed++
-				console.error(`  FAIL ${name}`)
-				console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-			}
-		}
+		failed = await runSuites("protected-file + config", tests)
 	} finally {
 		delete process.env.HEADLESSCODE_DATA_DIR
 		await fs.rm(storeTmp, { recursive: true, force: true })
@@ -236,7 +228,6 @@ async function main(): Promise<void> {
 		console.error(`\n${failed} test(s) failed`)
 		process.exit(1)
 	}
-	console.log(`\nAll ${tests.length} protected-file + config tests passed`)
 }
 
 main().catch((err) => {
diff --git a/src/permissions/__tests__/store-protection.test.ts b/src/permissions/__tests__/store-protection.test.ts
index 24ae19b..1454e99 100644
--- a/src/permissions/__tests__/store-protection.test.ts
+++ b/src/permissions/__tests__/store-protection.test.ts
@@ -39,6 +39,7 @@ import {
 	resolveProjectIdentity,
 	writeProjectMetadata,
 } from "../../project-store.js"
+import { runMain } from "../../test-runner.js"
 
 async function mkTmp(prefix: string): Promise<string> {
 	return fsp.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -373,26 +374,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["word splitting / tilde + env expansion / containment edges", testWordSplittingAndPathResolution],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} store-protection tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("store-protection", tests)
diff --git a/src/project-store.test.ts b/src/project-store.test.ts
index d460346..502fcbc 100644
--- a/src/project-store.test.ts
+++ b/src/project-store.test.ts
@@ -34,6 +34,7 @@ import {
 	resolveProjectIdentity,
 	writeProjectMetadata,
 } from "./project-store.js"
+import { runMain } from "./test-runner.js"
 
 async function mkTmp(prefix: string): Promise<string> {
 	return fsp.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -420,26 +421,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["OpenShell identity override applies only to TARGET_REPO", testOpenShellIdentityOverrideIsWorkspaceScoped],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} project-store tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("project-store", tests)
diff --git a/src/projects/__tests__/cli.test.ts b/src/projects/__tests__/cli.test.ts
index 18c8796..9590a60 100644
--- a/src/projects/__tests__/cli.test.ts
+++ b/src/projects/__tests__/cli.test.ts
@@ -18,6 +18,7 @@ import * as os from "node:os"
 import * as path from "node:path"
 
 import { projectsMain } from "../cli.js"
+import { runMain } from "../../test-runner.js"
 
 async function mkTmp(prefix: string): Promise<string> {
 	return fsp.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -330,26 +331,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["projects: --help and argument errors", testHelpAndArgErrors],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} projects CLI tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("projects CLI", tests)
diff --git a/src/qa/__tests__/qa.test.ts b/src/qa/__tests__/qa.test.ts
index 4aa01d8..14c1471 100644
--- a/src/qa/__tests__/qa.test.ts
+++ b/src/qa/__tests__/qa.test.ts
@@ -10,6 +10,7 @@ import * as path from "node:path"
 
 import { parseQaResult, runQa, runQaWithRetries } from "../qa.js"
 import type { ChatMessage, ChatToolCall, LlmClient, LlmRequest, LlmResponse } from "../../engine/types.js"
+import { runMain } from "../../test-runner.js"
 
 // ─── Fake LLM client (same pattern as engine/__tests__/loop.test.ts) ─────────
 
@@ -349,26 +350,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["runQaWithRetries: exhausts retries and returns the final error verdict", testRunQaWithRetriesExhaustsRetries],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} QA parsing tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("QA parsing", tests)
diff --git a/src/run-tests-script.test.ts b/src/run-tests-script.test.ts
index ecf8250..80c6f60 100644
--- a/src/run-tests-script.test.ts
+++ b/src/run-tests-script.test.ts
@@ -14,6 +14,7 @@ import * as fs from "node:fs"
 import * as os from "node:os"
 import * as path from "node:path"
 import { fileURLToPath } from "node:url"
+import { runMain } from "./test-runner.js"
 
 const REPO_ROOT = fileURLToPath(new URL("..", import.meta.url))
 const SCRIPT = path.join(REPO_ROOT, "scripts", "run-tests.mjs")
@@ -135,26 +136,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["--help / -h print usage and exit 0 without running the suite (BUG-14)", testHelpFlagPrintsUsageAndExitsWithoutRunningTests],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? (err.stack ?? err.message) : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} run-tests-script tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("run-tests-script", tests)
diff --git a/src/stack-rules-content.test.ts b/src/stack-rules-content.test.ts
index bef6124..3d66ff4 100644
--- a/src/stack-rules-content.test.ts
+++ b/src/stack-rules-content.test.ts
@@ -20,6 +20,7 @@ import * as path from "node:path"
 import { fileURLToPath } from "node:url"
 
 import { STACK_NAMES } from "./engine/stacks.js"
+import { runMain } from "./test-runner.js"
 
 const repoRoot = path.resolve(path.dirname(fileURLToPath(import.meta.url)), "..")
 const SHIPPED_DIR = path.join(repoRoot, "shared", "stacks")
@@ -65,23 +66,4 @@ const tests: Array<[string, () => void]> = [
 	["javascript + python content shipped", testRoundStacksShipped],
 ]
 
-function main(): void {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack : err)
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} stack-rules-content tests passed`)
-}
-
-main()
+runMain("stack-rules-content", tests)
diff --git a/src/test-runner.test.ts b/src/test-runner.test.ts
new file mode 100644
index 0000000..615558b
--- /dev/null
+++ b/src/test-runner.test.ts
@@ -0,0 +1,111 @@
+import assert from "node:assert/strict"
+
+import { runMain, runSuites } from "./test-runner.js"
+
+function capture(stream: "log" | "error"): {
+	lines: string[]
+	restore: () => void
+} {
+	const lines: string[] = []
+	const orig = console[stream]
+	console[stream] = (...args: unknown[]): void => {
+		lines.push(args.map(String).join(" "))
+	}
+	return { lines, restore: () => void (console[stream] = orig) }
+}
+
+function stubExit(): { codes: number[]; restore: () => void } {
+	const codes: number[] = []
+	const orig = process.exit
+	process.exit = ((code?: number) => {
+		codes.push(code ?? 0)
+	}) as never
+	return { codes, restore: () => void (process.exit = orig) }
+}
+
+async function tick(): Promise<void> {
+	await new Promise<void>((resolve) => setImmediate(() => resolve()))
+}
+
+async function boom(): Promise<void> {
+	throw new Error("boom-detail")
+}
+
+async function passingSuiteLogsOkSummaryAndZero(): Promise<void> {
+	const out = capture("log")
+	try {
+		const failed = await runSuites("demo", [
+			["first", async () => {}],
+			["second", () => {}],
+		])
+		assert.equal(failed, 0)
+		const want = ["  ok   first", "  ok   second", "\nAll 2 demo tests passed"]
+		assert.deepEqual(out.lines, want)
+	} finally {
+		out.restore()
+	}
+}
+
+async function failingSuiteLogsFailStackAndCount(): Promise<void> {
+	const errOut = capture("error")
+	const out = capture("log")
+	try {
+		const failed = await runSuites("demo", [["bad", boom]])
+		assert.equal(failed, 1)
+	} finally {
+		errOut.restore()
+		out.restore()
+	}
+	assert.ok(errOut.lines.some((l) => l.includes("FAIL bad")))
+	assert.ok(errOut.lines.some((l) => l.includes("boom-detail")))
+	assert.ok(errOut.lines.some((l) => l.includes("1 test(s) failed")))
+}
+
+async function emptyLabelOmitsLabel(): Promise<void> {
+	const out = capture("log")
+	try {
+		await runSuites("", [["only", async () => {}]])
+		assert.deepEqual(out.lines, ["  ok   only", "\nAll 1 tests passed"])
+	} finally {
+		out.restore()
+	}
+}
+
+async function mainExitsOneOnFailure(): Promise<void> {
+	const errOut = capture("error")
+	const out = capture("log")
+	const exited = stubExit()
+	try {
+		runMain("demo", [["bad", boom]])
+		await tick()
+	} finally {
+		errOut.restore()
+		out.restore()
+		exited.restore()
+	}
+	assert.deepEqual(exited.codes, [1])
+}
+
+async function crashHandlerExitsOne(): Promise<void> {
+	const errOut = capture("error")
+	const exited = stubExit()
+	try {
+		runMain("demo", null as unknown as [])
+		await tick()
+	} finally {
+		errOut.restore()
+		exited.restore()
+	}
+	assert.deepEqual(exited.codes, [1])
+	assert.ok(errOut.lines.some((l) => l.includes("Test runner crashed:")))
+}
+
+const tests: Array<[string, () => Promise<void>]> = [
+	["passing suite logs ok summary and zero", passingSuiteLogsOkSummaryAndZero],
+	["failing suite logs fail stack and count", failingSuiteLogsFailStackAndCount],
+	["empty label omits label", emptyLabelOmitsLabel],
+	["main exits one on failure", mainExitsOneOnFailure],
+	["crash handler exits one", crashHandlerExitsOne],
+]
+
+runMain("test-runner", tests)
diff --git a/src/test-runner.ts b/src/test-runner.ts
new file mode 100644
index 0000000..3389bbc
--- /dev/null
+++ b/src/test-runner.ts
@@ -0,0 +1,78 @@
+/**
+ * Shared assert-style test runner (issue #138). Every `*.test.ts` file
+ * used to inline the same ~20-line main() loop; the copies were flagged
+ * as clone groups by the architecture gate. New and migrated files end
+ * with a single line instead:
+ *
+ *   import { runMain } from "../../test-runner.js"
+ *   runMain("my-module", tests)
+ *
+ * Semantics are the established house pattern: per-test try/catch,
+ * `  ok   <name>` / `  FAIL <name>` + stack lines, and an
+ * `All N <label> tests passed` summary (the label segment is omitted
+ * when empty, matching the labelless files). runSuites returns the
+ * failure count WITHOUT exiting so files with teardown can run it
+ * first (`const failed = await runSuites(...)`) — runMain exits 1.
+ */
+
+/** One test case: sync or async, either is awaited. */
+export type SuiteFn = () => void | Promise<void>
+
+/** One [name, fn] row; readonly so `as const` tables pass unchanged. */
+export type SuiteTest = readonly [string, SuiteFn]
+
+/**
+ * Run every row, logging per-test results plus the closing summary or
+ * failure count. Returns the number of failed tests; never exits, so
+ * callers with teardown can sequence it before their own exit.
+ */
+async function runOne(name: string, fn: SuiteFn): Promise<boolean> {
+	try {
+		await fn()
+		console.log(`  ok   ${name}`)
+		return true
+	} catch (err) {
+		console.error(`  FAIL ${name}`)
+		const detail = err instanceof Error ? (err.stack ?? err.message) : String(err)
+		console.error(detail)
+		return false
+	}
+}
+
+export async function runSuites(
+	label: string,
+	tests: readonly SuiteTest[],
+): Promise<number> {
+	let failed = 0
+	for (const [name, fn] of tests) {
+		if (!(await runOne(name, fn))) {
+			failed++
+		}
+	}
+	if (failed > 0) {
+		console.error(`\n${failed} test(s) failed`)
+		return failed
+	}
+	const scope = label === "" ? "" : ` ${label}`
+	console.log(`\nAll ${tests.length}${scope} tests passed`)
+	return 0
+}
+
+/**
+ * File entry point: run the suite, exiting 1 on any failure, and
+ * converting an unexpected crash (outside any single test) into a
+ * loud message + exit 1.
+ */
+export function runMain(label: string, tests: readonly SuiteTest[]): void {
+	runSuites(label, tests).then(
+		(failed) => {
+			if (failed > 0) {
+				process.exit(1)
+			}
+		},
+		(err: unknown) => {
+			console.error("Test runner crashed:", err)
+			process.exit(1)
+		},
+	)
+}
diff --git a/src/tools/__tests__/apply_diff.test.ts b/src/tools/__tests__/apply_diff.test.ts
index 77305de..4f01805 100644
--- a/src/tools/__tests__/apply_diff.test.ts
+++ b/src/tools/__tests__/apply_diff.test.ts
@@ -17,6 +17,7 @@ import * as path from "node:path"
 
 import { createHeadlessExecutor } from "../executor.js"
 import type { PermissionsConfig } from "../../permissions/config.js"
+import { runMain } from "../../test-runner.js"
 
 async function mkTmpWorkspace(prefix: string): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -239,26 +240,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["protected file: --allow-protected-writes escape hatch permits the edit", testProtectedFileEscapeHatch],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} apply_diff tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("apply_diff", tests)
diff --git a/src/tools/__tests__/batch-types.test.ts b/src/tools/__tests__/batch-types.test.ts
index 6a6a69c..83d2df7 100644
--- a/src/tools/__tests__/batch-types.test.ts
+++ b/src/tools/__tests__/batch-types.test.ts
@@ -21,6 +21,7 @@ import type { BatchExecutor, BatchSection } from "../batch-types.js"
 import { assembleSections, formatSummary } from "../batch-format.js"
 import { createHeadlessExecutor } from "../executor.js"
 import type { ToolContext } from "../../engine/types.js"
+import { runMain } from "../../test-runner.js"
 
 interface ExecResult {
 	content: string
@@ -83,38 +84,11 @@ async function testToolExecutorSatisfiesTheInterface(): Promise<void> {
 	assert.equal(asInterface.has("no-such-tool"), false)
 }
 
-async function runOne(name: string, fn: () => Promise<void>): Promise<boolean> {
-	try {
-		await fn()
-		console.log(`  ok   ${name}`)
-		return true
-	} catch (error) {
-		console.error(`  FAIL ${name}`)
-		if (error instanceof Error) console.error(error.stack ?? error)
-		else console.error(String(error))
-		return false
-	}
-}
-
-async function main(): Promise<void> {
-	const tests: Array<[string, () => Promise<void>]> = [
-		["runs against the interface", testBatchHandlerRunsAgainstTheInterface],
-		["unknown tool is a section error", testUnknownToolIsASectionError],
-		["sections flow through", testLeafSectionsFlowThroughTheFormatter],
-		["executor satisfies interface", testToolExecutorSatisfiesTheInterface],
-	]
-	let failed = 0
-	for (const [name, fn] of tests) {
-		if (!(await runOne(name, fn))) failed++
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} batch-types tests passed`)
-}
+const tests: Array<[string, () => Promise<void>]> = [
+	["runs against the interface", testBatchHandlerRunsAgainstTheInterface],
+	["unknown tool is a section error", testUnknownToolIsASectionError],
+	["sections flow through", testLeafSectionsFlowThroughTheFormatter],
+	["executor satisfies interface", testToolExecutorSatisfiesTheInterface],
+]
 
-main().catch((error) => {
-	console.error("Test runner crashed:", error)
-	process.exit(1)
-})
+runMain("batch-types", tests)
diff --git a/src/tools/__tests__/batch.test.ts b/src/tools/__tests__/batch.test.ts
index f042520..8fdbaef 100644
--- a/src/tools/__tests__/batch.test.ts
+++ b/src/tools/__tests__/batch.test.ts
@@ -16,6 +16,7 @@ import * as path from "node:path"
 
 import { createHeadlessExecutor } from "../executor.js"
 import type { ToolExecutor } from "../executor.js"
+import { runMain } from "../../test-runner.js"
 
 async function mkFixtureWorkspace(): Promise<string> {
 	const ws = await fs.mkdtemp(path.join(os.tmpdir(), "hc-batch-"))
@@ -208,46 +209,21 @@ async function exists(p: string): Promise<boolean> {
 	}
 }
 
-async function main(): Promise<void> {
-	const tests: Array<[string, () => Promise<void>]> = [
-		["all-reads batch returns all sections", testAllReadsBatch],
-		["edit-then-read sees new content", testEditThenReadSeesNewContent],
-		[
-			"refused tools are per-section errors",
-			testRefusedToolsArePerSectionErrors,
-		],
-		["unknown tool is a per-section error", testUnknownToolIsAPerSectionError],
-		["11 sub-calls is a batch-level error", testOverCapIsABatchLevelError],
-		["empty calls is a batch-level error", testEmptyCallsIsABatchLevelError],
-		[
-			"stop_on_error skips the serial remainder",
-			testStopOnErrorSkipsSerialRemainder,
-		],
-		["oversized section is truncated", testSectionTruncation],
-	]
-
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			const detail = err instanceof Error
-				? (err.stack ?? err.message)
-				: String(err)
-			console.error(detail)
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} batch tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+const tests: Array<[string, () => Promise<void>]> = [
+	["all-reads batch returns all sections", testAllReadsBatch],
+	["edit-then-read sees new content", testEditThenReadSeesNewContent],
+	[
+		"refused tools are per-section errors",
+		testRefusedToolsArePerSectionErrors,
+	],
+	["unknown tool is a per-section error", testUnknownToolIsAPerSectionError],
+	["11 sub-calls is a batch-level error", testOverCapIsABatchLevelError],
+	["empty calls is a batch-level error", testEmptyCallsIsABatchLevelError],
+	[
+		"stop_on_error skips the serial remainder",
+		testStopOnErrorSkipsSerialRemainder,
+	],
+	["oversized section is truncated", testSectionTruncation],
+]
+
+runMain("batch", tests)
diff --git a/src/tools/__tests__/browser-action.test.ts b/src/tools/__tests__/browser-action.test.ts
index 0d922f9..232e32c 100644
--- a/src/tools/__tests__/browser-action.test.ts
+++ b/src/tools/__tests__/browser-action.test.ts
@@ -31,6 +31,7 @@ import { createHeadlessExecutor } from "../executor.js"
 import { describeImage } from "../../vision/describe.js"
 import { setVisionCaptionerForTest } from "../browser/handler.js"
 import type { AuxLlmUsage } from "../../engine/types.js"
+import { runSuites } from "../../test-runner.js"
 
 /**
  * COV-6: on a fresh contributor machine, Playwright's `chromium` package is
@@ -480,17 +481,7 @@ async function main(): Promise<void> {
 		process.exit(0)
 	}
 
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
+	const failed = await runSuites("browser_action", tests)
 	if (failed > 0) {
 		console.error(`\n${failed} test(s) failed`)
 		// Explicit exit: Playwright's connection sockets can keep the event
@@ -498,7 +489,6 @@ async function main(): Promise<void> {
 		// finished suite.
 		process.exit(1)
 	}
-	console.log(`\nAll ${tests.length} browser_action tests passed`)
 	// See above — a closed Playwright browser can still hold a handle.
 	process.exit(0)
 }
diff --git a/src/tools/__tests__/codebase_search.test.ts b/src/tools/__tests__/codebase_search.test.ts
index 752eef9..75945b4 100644
--- a/src/tools/__tests__/codebase_search.test.ts
+++ b/src/tools/__tests__/codebase_search.test.ts
@@ -22,6 +22,7 @@ import * as path from "node:path"
 
 import { createHeadlessExecutor } from "../executor.js"
 import { chunkHash } from "../../codesearch/index.js"
+import { runSuites } from "../../test-runner.js"
 
 async function mkTmpWorkspace(prefix: string): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -130,16 +131,7 @@ async function main(): Promise<void> {
 	process.env.HEADLESSCODE_DATA_DIR = storeTmp
 	let failed = 0
 	try {
-		for (const [name, fn] of tests) {
-			try {
-				await fn()
-				console.log(`  ok   ${name}`)
-			} catch (err) {
-				failed++
-				console.error(`  FAIL ${name}`)
-				console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-			}
-		}
+		failed = await runSuites("codebase_search handler", tests)
 	} finally {
 		delete process.env.HEADLESSCODE_DATA_DIR
 		await fs.rm(storeTmp, { recursive: true, force: true })
@@ -148,7 +140,6 @@ async function main(): Promise<void> {
 		console.error(`\n${failed} test(s) failed`)
 		process.exit(1)
 	}
-	console.log(`\nAll ${tests.length} codebase_search handler tests passed`)
 }
 
 main().catch((err) => {
diff --git a/src/tools/__tests__/decision.test.ts b/src/tools/__tests__/decision.test.ts
index 7375556..ef2fcee 100644
--- a/src/tools/__tests__/decision.test.ts
+++ b/src/tools/__tests__/decision.test.ts
@@ -20,6 +20,7 @@ import { BudgetTracker } from "../../budget/budget.js"
 import { HeadlessSession } from "../../engine/loop.js"
 import type { ChatMessage, ChatToolCall, LlmClient, LlmRequest, LlmResponse } from "../../engine/types.js"
 import type { PermissionsConfig } from "../../permissions/config.js"
+import { runMain } from "../../test-runner.js"
 
 async function mkTmpWorkspace(prefix: string): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -380,26 +381,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["permissions: reviewer/QA read-only executor gates execute_command identically", testReadOnlyExecutorAlsoGatesCommands],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} executor tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("executor", tests)
diff --git a/src/tools/__tests__/describe_image.test.ts b/src/tools/__tests__/describe_image.test.ts
index f772349..f557e66 100644
--- a/src/tools/__tests__/describe_image.test.ts
+++ b/src/tools/__tests__/describe_image.test.ts
@@ -22,6 +22,7 @@ import { AddressInfo } from "node:net"
 
 import { createHeadlessExecutor } from "../executor.js"
 import type { AuxLlmUsage } from "../../engine/types.js"
+import { runMain } from "../../test-runner.js"
 
 const SCRIPTED_CONTENT =
 	"A dashboard titled 'Q3 Revenue Dashboard' shows $42,318.50 (+12.4%), Active users 1,284 (-3.1%), Error rate 0.07%."
@@ -197,27 +198,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["describe_image: bad path argument is a clear error", testBadArgumentRejected],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} describe_image tests passed`)
-	process.exit(0)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("describe_image", tests)
diff --git a/src/tools/__tests__/edit_file.test.ts b/src/tools/__tests__/edit_file.test.ts
index 6fa5b1d..07395d7 100644
--- a/src/tools/__tests__/edit_file.test.ts
+++ b/src/tools/__tests__/edit_file.test.ts
@@ -17,6 +17,7 @@ import * as path from "node:path"
 
 import { createHeadlessExecutor } from "../executor.js"
 import type { PermissionsConfig } from "../../permissions/config.js"
+import { runMain } from "../../test-runner.js"
 
 async function mkTmpWorkspace(prefix: string): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -353,26 +354,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["protected file: --allow-protected-writes escape hatch permits the edit", testProtectedFileEscapeHatch],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} edit_file tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("edit_file", tests)
diff --git a/src/tools/__tests__/execute_command.test.ts b/src/tools/__tests__/execute_command.test.ts
index ec40f9c..8f3946c 100644
--- a/src/tools/__tests__/execute_command.test.ts
+++ b/src/tools/__tests__/execute_command.test.ts
@@ -25,6 +25,7 @@ import {
 	collectSpawnDiagnostics,
 	formatSpawnDiagnostics,
 } from "../executor.js"
+import { runMain } from "../../test-runner.js"
 
 async function mkTmpWorkspace(prefix: string): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -657,26 +658,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["execute_command: normal command close does NOT emit the spawn-diagnostics line", testNoDiagnosticsOnNormalClose],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} execute_command timeout tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("execute_command timeout", tests)
diff --git a/src/tools/__tests__/language-detect.test.ts b/src/tools/__tests__/language-detect.test.ts
index 76881a2..8dc4ecb 100644
--- a/src/tools/__tests__/language-detect.test.ts
+++ b/src/tools/__tests__/language-detect.test.ts
@@ -12,6 +12,7 @@ import * as path from "node:path"
 
 import { detectWorkspaceLanguages, isTypeScriptWorkspace } from "../language-detect.js"
 import { createHeadlessExecutor, createReadOnlyHeadlessExecutor } from "../executor.js"
+import { runMain } from "../../test-runner.js"
 
 async function mkTmp(prefix: string): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -155,26 +156,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["executor registers TS-only tools iff the workspace is TS/JS", testExecutorGatesTsToolsByWorkspaceLanguage],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} language-detect tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("language-detect", tests)
diff --git a/src/tools/__tests__/list-files.test.ts b/src/tools/__tests__/list-files.test.ts
index 6180d77..edfa91c 100644
--- a/src/tools/__tests__/list-files.test.ts
+++ b/src/tools/__tests__/list-files.test.ts
@@ -12,6 +12,7 @@ import * as os from "node:os"
 import * as path from "node:path"
 
 import { createHeadlessExecutor, MAX_LIST_FILES } from "../executor.js"
+import { runMain } from "../../test-runner.js"
 
 async function mkTmpWorkspace(prefix: string): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -178,26 +179,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["disableReadFileCache also suppresses list_files' cache-hit notice (2026-09-02)", testDisableReadFileCacheAlsoCoversListFiles],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} list-files tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("list-files", tests)
diff --git a/src/tools/__tests__/local-summarization-gate.test.ts b/src/tools/__tests__/local-summarization-gate.test.ts
index 1e389c2..edf33ec 100644
--- a/src/tools/__tests__/local-summarization-gate.test.ts
+++ b/src/tools/__tests__/local-summarization-gate.test.ts
@@ -30,6 +30,7 @@ import * as path from "node:path"
 import { createHeadlessExecutor } from "../executor.js"
 import { MAX_RESULT_CHARS } from "../executor.js"
 import { LOCAL_SUMMARIZATION_DEFAULT_ENABLED } from "../output-summarizer.js"
+import { runMain } from "../../test-runner.js"
 
 const SUMMARIZATION_ENV = "HEADLESSCODE_LOCAL_SUMMARIZATION"
 const OLLAMA_URL_ENV = "HEADLESSCODE_OLLAMA_URL"
@@ -260,26 +261,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["ON + oversized result: summarized with transparency header", testOnOversizedResultIsSummarized],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} local-summarization-gate tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("local-summarization-gate", tests)
diff --git a/src/tools/__tests__/output-summarizer.test.ts b/src/tools/__tests__/output-summarizer.test.ts
index 57edc52..0c358a5 100644
--- a/src/tools/__tests__/output-summarizer.test.ts
+++ b/src/tools/__tests__/output-summarizer.test.ts
@@ -27,6 +27,7 @@ import {
 	LOCAL_SUMMARIZATION_DEFAULT_ENABLED,
 	LOCAL_SUMMARIZATION_ENV,
 } from "../output-summarizer.js"
+import { runMain } from "../../test-runner.js"
 
 // ─── Fake Ollama server ───────────────────────────────────────────────────────
 
@@ -243,26 +244,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["summarizer: oversized summary (> MAX_SUMMARY_CHARS) rejected", testOversizedSummaryRejected],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} output-summarizer tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("output-summarizer", tests)
diff --git a/src/tools/__tests__/path-safety.test.ts b/src/tools/__tests__/path-safety.test.ts
index 40366a1..9a21c74 100644
--- a/src/tools/__tests__/path-safety.test.ts
+++ b/src/tools/__tests__/path-safety.test.ts
@@ -27,6 +27,7 @@ import {
 	SymlinkEscapeError,
 	PathTraversalError,
 } from "../executor.js"
+import { runMain } from "../../test-runner.js"
 
 async function mkTmpWorkspace(prefix: string): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -307,23 +308,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["unit: SymlinkEscapeError vs PathTraversalError", testUnitErrorTypes],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`path-safety: ${failed} of ${tests.length} tests failed`)
-		process.exit(1)
-	}
-	console.log(`All ${tests.length} tests passed`)
-}
-
-await main()
+runMain("", tests)
diff --git a/src/tools/__tests__/paths.test.ts b/src/tools/__tests__/paths.test.ts
index 315bd85..50f1912 100644
--- a/src/tools/__tests__/paths.test.ts
+++ b/src/tools/__tests__/paths.test.ts
@@ -23,6 +23,7 @@ import {
 	PathTraversalError as ExecutorTraversalError,
 	resolveWithinWorkspace as executorResolve,
 } from "../executor.js"
+import { runMain } from "../../test-runner.js"
 
 async function mkWorkspace(prefix: string): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -76,40 +77,13 @@ async function testExecutorReexportsTheSameGuard(): Promise<void> {
 	assert.throws(() => executorResolve(ws, "../escape.txt"), PathTraversalError)
 }
 
-async function runOne(name: string, fn: () => Promise<void>): Promise<boolean> {
-	try {
-		await fn()
-		console.log(`  ok   ${name}`)
-		return true
-	} catch (error) {
-		console.error(`  FAIL ${name}`)
-		if (error instanceof Error) console.error(error.stack ?? error)
-		else console.error(String(error))
-		return false
-	}
-}
-
-async function main(): Promise<void> {
-	const tests: Array<[string, () => Promise<void>]> = [
-		["inside path resolves", testInsidePathResolves],
-		["dot path resolves to root", testDotPathResolvesToRoot],
-		["traversal refused", testTraversalRefused],
-		["symlink escape refused", testSymlinkEscapeRefused],
-		["escape is a traversal error", testSymlinkEscapeIsATraversalError],
-		["executor re-exports guard", testExecutorReexportsTheSameGuard],
-	]
-	let failed = 0
-	for (const [name, fn] of tests) {
-		if (!(await runOne(name, fn))) failed++
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} paths tests passed`)
-}
+const tests: Array<[string, () => Promise<void>]> = [
+	["inside path resolves", testInsidePathResolves],
+	["dot path resolves to root", testDotPathResolvesToRoot],
+	["traversal refused", testTraversalRefused],
+	["symlink escape refused", testSymlinkEscapeRefused],
+	["escape is a traversal error", testSymlinkEscapeIsATraversalError],
+	["executor re-exports guard", testExecutorReexportsTheSameGuard],
+]
 
-main().catch((error) => {
-	console.error("Test runner crashed:", error)
-	process.exit(1)
-})
+runMain("paths", tests)
diff --git a/src/tools/__tests__/read-cache.test.ts b/src/tools/__tests__/read-cache.test.ts
index 047aa8c..67a4973 100644
--- a/src/tools/__tests__/read-cache.test.ts
+++ b/src/tools/__tests__/read-cache.test.ts
@@ -16,6 +16,7 @@ import * as os from "node:os"
 import * as path from "node:path"
 
 import { createHeadlessExecutor, createReadOnlyHeadlessExecutor } from "../executor.js"
+import { runMain } from "../../test-runner.js"
 
 async function mkTmpWorkspace(prefix: string): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -384,26 +385,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["disableReadFileCache absent/false: default cache behavior unaffected", testCacheStillWorksWhenOptionIsAbsentOrFalse],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} read-cache tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("read-cache", tests)
diff --git a/src/tools/__tests__/run-tests.test.ts b/src/tools/__tests__/run-tests.test.ts
index 169a6f3..7c283bc 100644
--- a/src/tools/__tests__/run-tests.test.ts
+++ b/src/tools/__tests__/run-tests.test.ts
@@ -24,6 +24,7 @@ import { createHeadlessExecutor } from "../executor.js"
 import { getCodeIntelCache, resetCodeIntelCaches } from "../../codeintel/program.js"
 import { changedFilesFromGit, runOneTestFile } from "../run-tests.js"
 import { directTestMatch, isTestFile, selectTestsForChangedFiles } from "../test-selection.js"
+import { runMain } from "../../test-runner.js"
 
 const FIXTURE_TSCONFIG = JSON.stringify(
 	{
@@ -302,26 +303,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["end-to-end: selects and runs a real matched test via tsx", testEndToEndRunsMatchedTest],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} run_tests tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("run_tests", tests)
diff --git a/src/tools/__tests__/search_replace.test.ts b/src/tools/__tests__/search_replace.test.ts
index 4b51e15..1d66061 100644
--- a/src/tools/__tests__/search_replace.test.ts
+++ b/src/tools/__tests__/search_replace.test.ts
@@ -15,6 +15,7 @@ import * as path from "node:path"
 
 import { createHeadlessExecutor } from "../executor.js"
 import type { PermissionsConfig } from "../../permissions/config.js"
+import { runMain } from "../../test-runner.js"
 
 async function mkTmpWorkspace(prefix: string): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -226,26 +227,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["protected file: --allow-protected-writes escape hatch permits the edit", testProtectedFileEscapeHatch],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} search_replace tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("search_replace", tests)
diff --git a/src/tools/__tests__/set-indentation.test.ts b/src/tools/__tests__/set-indentation.test.ts
index 79532a7..52e3c41 100644
--- a/src/tools/__tests__/set-indentation.test.ts
+++ b/src/tools/__tests__/set-indentation.test.ts
@@ -16,6 +16,7 @@ import * as os from "node:os"
 import * as path from "node:path"
 
 import { createHeadlessExecutor, createReadOnlyHeadlessExecutor } from "../executor.js"
+import { runMain } from "../../test-runner.js"
 
 async function mkTmpWorkspace(prefix: string): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -160,26 +161,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["not registered on the read-only reviewer/QA executor", testNotAvailableOnReadOnlyExecutor],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? (err.stack ?? err.message) : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} set_indentation tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("set_indentation", tests)
diff --git a/src/tools/__tests__/update_todo_list.test.ts b/src/tools/__tests__/update_todo_list.test.ts
index 0364efe..0d61953 100644
--- a/src/tools/__tests__/update_todo_list.test.ts
+++ b/src/tools/__tests__/update_todo_list.test.ts
@@ -33,6 +33,7 @@ import { EventFeed, eventsDir, eventsFilePath, readEventsFile } from "../../engi
 import type { ChatMessage, ChatToolCall, LlmClient, LlmRequest, LlmResponse } from "../../engine/types.js"
 import { reviewTools } from "../../orchestrator/reviewer.js"
 import { qaTools } from "../../qa/qa.js"
+import { runMain } from "../../test-runner.js"
 
 async function mkTmpWorkspace(prefix: string): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -353,26 +354,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["summarizeToolArg previews update_todo_list args", testSummarizeToolArgForUpdateTodoList],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} update_todo_list tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("update_todo_list", tests)
diff --git a/src/tools/__tests__/write_to_file.test.ts b/src/tools/__tests__/write_to_file.test.ts
index b56dcee..6b29d57 100644
--- a/src/tools/__tests__/write_to_file.test.ts
+++ b/src/tools/__tests__/write_to_file.test.ts
@@ -19,6 +19,7 @@ import * as os from "node:os"
 import * as path from "node:path"
 
 import { createHeadlessExecutor } from "../executor.js"
+import { runMain } from "../../test-runner.js"
 
 async function mkTmpWorkspace(prefix: string): Promise<string> {
 	return fs.mkdtemp(path.join(os.tmpdir(), prefix))
@@ -118,26 +119,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["existing trivial file: unaffected even when the gate is on", testTrivialExistingFileUnaffected],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} write_to_file tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("write_to_file", tests)
diff --git a/src/vision/__tests__/describe.test.ts b/src/vision/__tests__/describe.test.ts
index c6cc77c..60f538e 100644
--- a/src/vision/__tests__/describe.test.ts
+++ b/src/vision/__tests__/describe.test.ts
@@ -24,6 +24,7 @@ import { AddressInfo } from "node:net"
 
 import { describeImage, VisionError } from "../describe.js"
 import { BudgetTracker } from "../../budget/budget.js"
+import { runMain } from "../../test-runner.js"
 
 /** Description the fake server returns, quoted in the tests below. */
 const SCRIPTED_CONTENT =
@@ -277,27 +278,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["describeImage: missing API key is a clear error", testMissingApiKeyError],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} vision describe tests passed`)
-	process.exit(0)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("vision describe", tests)
diff --git a/src/watcher/__tests__/github.test.ts b/src/watcher/__tests__/github.test.ts
index 2e01e50..0e0b277 100644
--- a/src/watcher/__tests__/github.test.ts
+++ b/src/watcher/__tests__/github.test.ts
@@ -16,6 +16,7 @@ import {
 	listIssues,
 	toIssue,
 } from "../github.js"
+import { runMain } from "../../test-runner.js"
 
 interface FakeInit {
 	method?: string
@@ -278,26 +279,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["headlesscode-job lifecycle operations use the expected GitHub API", testHeadlessCodeJobLifecycleApi],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} github client tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("github client", tests)
diff --git a/src/watcher/__tests__/state.test.ts b/src/watcher/__tests__/state.test.ts
index 694501a..bd80432 100644
--- a/src/watcher/__tests__/state.test.ts
+++ b/src/watcher/__tests__/state.test.ts
@@ -21,6 +21,7 @@ import {
 	type WatcherIssueEntry,
 	type WatcherState,
 } from "../state.js"
+import { runMain } from "../../test-runner.js"
 
 async function tmpStatePath(): Promise<string> {
 	const dir = await fs.mkdtemp(path.join(os.tmpdir(), "headlesscode-watcher-state-"))
@@ -203,26 +204,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["sync load/save variants", testSyncVariants],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} watcher state tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("watcher state", tests)
diff --git a/src/watcher/__tests__/watch.test.ts b/src/watcher/__tests__/watch.test.ts
index 3ee7983..6fdccef 100644
--- a/src/watcher/__tests__/watch.test.ts
+++ b/src/watcher/__tests__/watch.test.ts
@@ -20,6 +20,7 @@ import { Logger } from "../../engine/logger.js"
 import { watchIssues, type SpawnBatch, type SpawnResult, type WatcherConfig } from "../watch.js"
 import { loadWatcherState } from "../state.js"
 import type { GhClient, GitHubIssue } from "../github.js"
+import { runMain } from "../../test-runner.js"
 
 const SILENT = new Logger({ level: "silent" })
 
@@ -260,26 +261,4 @@ const tests: Array<[string, () => Promise<void>]> = [
 	["write-ahead: spawned entry durable before the spawn step", testWriteAheadOrdering],
 ]
 
-async function main(): Promise<void> {
-	let failed = 0
-	for (const [name, fn] of tests) {
-		try {
-			await fn()
-			console.log(`  ok   ${name}`)
-		} catch (err) {
-			failed++
-			console.error(`  FAIL ${name}`)
-			console.error(err instanceof Error ? err.stack ?? err.message : String(err))
-		}
-	}
-	if (failed > 0) {
-		console.error(`\n${failed} test(s) failed`)
-		process.exit(1)
-	}
-	console.log(`\nAll ${tests.length} watcher tests passed`)
-}
-
-main().catch((err) => {
-	console.error("Test runner crashed:", err)
-	process.exit(1)
-})
+runMain("watcher", tests)