feat(ai, ai-harness): add the harness stack with replay and adapter parity

[AlemTuzlak]

Review the combined harness and replay stack in this PR. The harness keeps one agent conversation open across turns, restarts, and model changes. This combined stack adds the harness, CLI, dashboard, MCP in both directions, durable sessions, host hooks, and compaction. It also keeps saved history usable across adapters and checks final tool input before execution or client dispatch.

🎯 Changes

Compatibility changes

Area	Caller impact
Reasoning	The owner stack replaces provider reasoning fields in modelOptions with chat({ reasoning }). The migration guide remains part of this PR.
Default execution and caching	Server tools run in parallel. Dependent tools need sequential: true or toolExecution: 'sequential'. Prompt caching is on by default. Use promptCache: 'none' to disable it.
Usage and saved history	Anthropic, Bedrock, and Claude Code report total input in promptTokens. Assistant block order can produce more AG-UI rows. Source metadata separates the requested model from the response model.
Host and compaction	Host joins stop at a refused or cancelled steer. Old fold checkpoints rebuild once. Compaction events add reason, and withCompaction adds compactNext. Durable turns keep LogRecordsCapability.
Model ids	The first-party factories take any model id string, so a misspelled or retired id is no longer a type error. It fails at the provider instead.
Tool validation and durable resume	Final validation follows middleware. Invalid input can reject a call that previously ran. Older same-run phases work when saved approval and tool bindings match. Cross-run phases need trusted saved approval context. Without it, resume is rejected and the client phase stays pending. Malformed context is rejected in both cases.

Saved subagent cards use a separate view of the transcript. Ordinary text and ambiguous hosts stay in that view. Both host creators preserve unrelated metadata. They choose an unused ID for missing IDs and explicit duplicate IDs, while keeping a unique explicit ID.

Replay changes provider input. It keeps the saved transcript unchanged. History without source metadata retains its legacy source treatment.

Existing harness stack

This remains the single combined PR for phases P0 to P14. It replaces #1551 and the closed phase PRs. Their descriptions remain the review record.

Phase	PR	Behavior
P0	#1513	Subagents return values and call activities.
P1	#1515	Sessions, typed agents, and plugins.
P2	#1518	Resume crashed turns and serve sessions to clients.
P3	#1519	Commands, settings, auth, and first-party plugins.
P4	#1520	Subagent tree limits and harness children.
P5	#1521	Build artifacts, worker mode, and remote harnessText.
P6	#1522	Dashboard and runnable example.
P7	#1523	MCP connectors with browser sign-in and runtime tools.
P8	#1524	Code mode with pluggable isolates.
P9	#1538	Delegate to coding agents.
P10	#1540	Continue work until a goal judge accepts the result.
P11	#1546	Live session state for any UI.
P12	#1549	CLI with a replaceable UI.
P13	#1550	Middleware and usage for every agent run.
P14	#1554	Harness MCP server, --mcp, and /mcp on --serve.

The owner changes also keep media, provider keys, durable tool steps, host hooks, turn leases, skills, prompt caching, routing, and per-turn overrides. Thinking, text, and tool calls keep their block order. Mid-conversation tool and prompt changes preserve the cached start where the model supports them.

MCP in both directions. The client keeps connectors and browser sign-in. createHarnessMcpServer exposes chat, approvals, agents, and commands to MCP clients. The CLI serves stdio MCP or authenticated HTTP MCP. MCP inputs keep parsed template variables and per-answer MIME types.

createMCPClient keeps toolName, request timeouts, and strict toolFilter checks. Missing or repeated filtered names raise MCPToolFilterError. Tool metadata keeps titles and frozen annotations. Final input validation runs before an accepted call reaches its tool.

Compaction. withCompaction({ countTokens: 'usage' }) uses reported usage. compactNext(threadId) requests compaction at the next model call. Durable compaction writes session-log records through LogRecordsCapability. The host view rebuilds the context from those records.

Compaction reports its reason, summary usage, and errors. Summary options keep turn cuts, updates to previous summaries, and bounded tool output. These opt-in features remain part of the owner scope.

Latest owner update preserved

The eight owner commits through c9f8e5f901f82e5dcba3dbd505d7ee2ab9e110e3 stay in the stack. Shared threads run each chat input and command as the principal returned by authorize. Principal.tenantId keeps credential scopes apart. Reads use the user's key before a shared tenant key. Commands save only through their own sender's credentials. Different senders do not join a running turn by default. Client context stays untrusted; server context wins when objects merge.

The harness run route now supports useChat, transcript hydration, interrupted phases, and cursor-based rejoin after reload. Repeated run IDs keep their receipt. A conflicting sender or already-used run record returns a conflict. Invalid resolve input leaves the interrupt pending. A resolve sent as a turn finishes can start immediately after it.

Sign-in wait uses credentials.require(id, { wait: true }) inside a chat tool. The stopped sender's sign-in resumes that turn. lifetime: 'turn' names a plugin that lives for one turn; 'run' remains an alias. TurnInfo carries message, input context, principal, input ID, and overrides. A plugin may supply each turn's adapter, so HarnessConfig.adapter is optional in that case. harnessText accepts explicit inputModalities.

The view separates approvals, client tools, and sign-in requests. ClientToolCall, view.on('clientTool', ...), and its resolve/fail actions answer client tools. The CLI accepts a caller-owned host and principal, keeps that host open, and keeps its bearer authorization gate. fakeText keeps parent run identity, distinct tool IDs between adapter instances, and exact optional typing. All eight new owner changesets, shared-thread docs, sandbox-provider docs, and their navigation entries remain.

MCP sender isolation. The new per-sender connection cache now encodes tenant and user as a JSON tuple. Names that contain : cannot reuse another sender's authenticated client or tools. A new real-SDK/HTTP regression covers two names that collided before, distinct scoped credentials, separate tools, and same-sender reuse. This new regression was source-reviewed and was not run locally under the user's push instruction. The owner changeset already includes the @tanstack/ai-mcp patch.

Other owner fixes retained. The stack keeps MCP interrupt-kind handling, the Linear OAuth issuer, unique exported tool names, v2 connector credentials, background-agent lease recovery, routed handoff hooks and child resume, interrupted-session recovery, and MCP sessions needed for elicitation. Workspace approval, lazy code-mode tools, skills, Cloudflare example, model catalog, and live plugin commands remain part of the owner scope.

Usage, resumable agents, thread settings, fork, and reset

These five owner features come from the pi-durable comparison. Each one is opt-in or read-only for existing callers.

• Usage totals. session.usage() and snapshot().usage total each model call by model and by sender, with the cost that providers report. A durable host keeps one harness.usage record per call. A harness.usage event fires on each call.
• Resumable background agents. agents.start(agent, input, { resume: true }) needs a durable host. The host that takes over runs the agent again from its saved transcript. ctx.step.do(name, fn) keeps side effects from running twice. durability.maxAttempts caps the runs.
• Thread settings. defineHarness({ models }) and session.configure() store the model, reasoning, instructions, tools, plugins, and working folder of a thread. A client can change only the fields in expose.settings, and none by default.
• Fork and reset. host.fork() copies a thread up to a message into a new thread, with its settings and media. session.reset(note?) starts a fresh model context and keeps the full transcript, with a marker.
• Fixes. A resolve that a crash interrupted continues with its routed plan, sender, and context. On a host without a log, the transcript is saved before each model call, so a tool result stays when the next model call fails. The used-up interrupt rule from e677b7516 stays as it is.

Replay and adapter parity

Handoff item	Final source behavior
1. Foreign-model history	Record the provider, API, and requested model. Foreign replay removes incompatible signatures and redacted thinking. Readable thinking becomes text, and tool IDs stay paired.
2. Replay cleanup	Omit failed or aborted assistant batches and their results from provider input. Fill unanswered calls with No result provided errors. Keep system changes after tool results.
3. Claude thinking on Bedrock	Keep reasoning text, signatures, redacted bytes, indexes, and block order. Replay signed reasoning before the matching tool call.
4. Unknown finish reasons	Emit RUN_ERROR with the provider finish reason. Known content_filter behavior stays the same.
5. Images in tool results	Chat Completions sends tool text followed by user images. Mistral keeps images in the tool message. Bedrock uses native tool-result image blocks.
6. Bedrock error status	An error property sets status: 'error', including an empty error string.
7. Lone Unicode surrogates	Remove lone surrogates from outgoing text and decoded JSON strings. Keep valid pairs, saved raw arguments, signatures, URLs, and bytes.
8. Final argument validation	Validate after middleware and before execution or client dispatch. Keep Standard Schema transforms and full JSON Schema checks with coercion. Preserve own JSON keys and ignore inherited lookup entries across all ten runtime paths.
9. Response identity	Expose genuine response IDs and response models where the provider supplies them. Keep the requested source model separate. Do not invent a generation ID.
10. Chat Completions content	Null or missing text adds no text and still permits tool calls. Object or array text emits the specified error. Keep main's malformed-argument recovery and stream drain.
11. Anthropic Bearer and OAuth	Support Bearer credentials, environment precedence, and OAuth identity and betas through the existing SDK. Bearer auth omits x-api-key.
12. Azure OpenAI	Add Azure Responses support through the existing OpenAI SDK. Keep endpoint, version, credentials, and deployment-name precedence explicit. Deployment maps use own entries only.
13. Gateway thinking	allowEmptySignature permits unsigned same-source thinking on configured gateways. The default remains false.
14. Empty tools with tool history	Chat Completions sends tools: [] for tool history without active tools. Wrappers retain that empty list. The no-history request stays unchanged.

Final validation. Standard Schema runs first on the exact input. When a safe input schema permits a coercion retry, the authored schema checks the retry result. A successful transform keeps its value. Raw JSON Schema uses full schema checks. Pending or denied calls do not reach tool hooks, execution, or client dispatch.

The only new runtime dependency for this replay work is approved typebox at ^1.3.27, locked to 1.3.34. TypeBox uses its check engine when dynamic evaluation is unavailable. Source review confirms this fallback. Focused controls check validation behavior. They do not prove a deployed Worker.

Own-key preservation covers schema copies, nullable widening maps, Standard Schema root copies, Gemini parameter sanitation, and Azure deployment lookup. OpenRouter keeps raw arguments separate from normalized input across all three terminal paths. Scalar, array, null, and missing input keep their contracts.

The host selector still returns a message index or -1. Presence guards make that return type explicit. This repairs type inference; it does not change caller sentinels or claim a runtime undefined defect.

Integration. The candidate preserves the earlier 16 owner commits and the eight newer owner commits through c9f8e5f901f82e5dcba3dbd505d7ee2ab9e110e3, then retains the approved merge with pinned main 6d8e6485f92c98a6f2200471e2aff21ab50be013. Main's cancellation and malformed-JSON controls remain. The route registry includes the owner and main routes. Model metadata keeps main's catalog and the owner's modality and reasoning policy.

Main merge and data-loss fixes

fdaeabe8f merges main at 7fb4a5f7f. That brings in #1620, live durable streams and one hydrate GET in Strict Mode. These commits fix the data-loss issues and the CI failures that showed up after the merge:

Commit	Fix
581efd816	A host record from the last model call of a turn stays in the log. Before, rebase dropped it when the turn ended.
af7274d6f	An empty summary fails the compaction. Before, it replaced the history with nothing. On the overflow path the turn fails with the overflow error.
4c12b1e89, e677b7516	A queued resume turn keeps its parent run. An interrupt is not offered again after its tool ran, including when a later model call fails.
bf329d5e3, 809259930	A tool with no input or null arguments runs with {} again (#265). The final input check made this fail. Other scalars and malformed JSON are still rejected. 809259930 updates the five adapter replay tests that expected the rejection.
c1e9fdad8	The OTel middleware records no execute_tool span for a client tool. The server does not run it.
15868e693	Tests for two compaction functions that had none: part text in the summary prompt, and summary usage after the turn. Coverage saw ai-compaction functions drop from 100% to 97.43%.

The test commits 71d5e933c, f02f0ddcd, and f07b75e0a keep the MCP and E2E tests true under the new behavior. Five patch changesets cover the five fixes above. The tools-test adapter now counts only the tool results of the current turn. Replay gives an unanswered call from an earlier turn a No result provided result, so the old count ended the run after Stop too early.

Adapters from catalog records

A catalog model whose id is not in the adapter's own table got no reasoning field. That was 582 of the 709 catalog reasoning models on first-party wires. Now:

• Reasoning from the config. Each first-party text adapter takes reasoning?: ModelReasoning: Anthropic (incl. Vertex), OpenAI and Azure, Gemini (AI Studio and Vertex), Bedrock Converse, Mistral, and Cloudflare. It wins over the table for the request fields and for the levels chat({ reasoning }) takes. false sends nothing. The factories take any model id. @tanstack/ai-models adds modelReasoning(record).
• Prices. modelCost uses the models.dev cost.tiers (long context) and prices cacheWrite1h at 2 times the input price. Usage adds promptTokensDetails.cacheWrite1hTokens, filled by Anthropic and Bedrock Converse.
• Catalog. The google-vertex catalog leaves out the Claude models, which had the Gemini wire.
• Responses replay. Each answer item keeps its id and phase in metadata.tanstack.responseItems, and a same-model replay sends them back.

The remaining 66 models with no reasoning field are Bedrock Converse models that are not Claude. pi also sends Converse thinking fields only for Claude.

This batch also merges main two times, at 8ce7ad97c and a32782c3e. The merges keep the activity messages, video persistence, stream backpressure, and the Anthropic max_tokens usage, replay, and Sonnet 5.5 fixes from main. Activity chunks no longer get the assistant call metadata.

Anthropic thinking shape from the record

Flue compared the Anthropic thinking fields with pi's for every anthropic-messages reasoning record. 195 records were different. Gateway ids with dots (anthropic/claude-opus-4.7) got budget thinking, older Claude on OpenRouter got adaptive thinking, and most other models on the wire got adaptive thinking. Now:

• Adaptive or budget from the record. ModelReasoning.adaptive picks the shape. modelReasoning(record) sets it from compat.forceAdaptiveThinking, and the catalog flag now equals pi's on all 291 shared records. Adaptive thinking without a level map sends pi's default effort.
• Mid-conversation effort. ModelReasoning.midConversationEffort (from the new compat.supportsMidConvoEffort, pi's 5 models) puts the level into the messages, so a level change keeps the cached start. Each answer keeps its level in metadata.tanstack.reasoningEffort.
• One channel at a time. midConversationChannels takes { tools?, systemPrompts? }.
• Catalog data. Seven Fireworks records take Fireworks' own levels (pi), and Vercel google/gemma-4-31b-it reasons. @tanstack/ai-vercel-gateway gets the same Gemma fix (6656c1d71), so the reasoning drift test agrees. Vercel's own model list tags it reasoning.
• Coverage. 618d71121 adds an ai-isolate-e2b test for stderr without an out-of-memory report. That branch side was hit only when V8's first GC block came in its own stderr chunk, so Linux CI missed it on this PR.

The local copy of the Flue check finds only group E (top-level effort on Claude 4.6, accepted in the 2026-09-30 report). Before this batch, 53fab0187 brought the harness-cli-batman lockfile entry in line with harness-cli, and 280263df4 and ef76ab6b1 merged main.

Docs and changesets

The replay work updates 13 existing pages and docs/config.json. It keeps the owner's harness, MCP, compaction, migration, and example docs. Examples cover server and client code where required. OpenAI text examples use gpt-5.5 and do not use type-assertion casts.

The tool-approval page explains same-run legacy support and the trusted context required for cross-run resume. The adapter-switching page explains the separate card view and conservative host preservation. Only those content edits move their updatedAt dates to October 5. Other content dates and all addedAt dates stay unchanged.

The OpenAI page now describes API-specific modelOptions and reasoning support accurately. That factual correction does not change its date.

The replay changeset covers 14 runtime packages: three minor releases and eleven patch releases. It includes the @tanstack/ai-utils patch for the actual own-key fix. The owner phase, media, reasoning, compaction, MCP, and latest shared-thread/auth/run/CLI/view changesets remain. The owner input-sender changeset also releases the MCP cache-key correction.

This batch adds docs/harness/usage.md, fork-and-reset.md, and thread-settings.md, and a resumable-agents section in subagents.md. It links them from five neighbor pages. Five new changesets cover usage, resumable agents (with a @tanstack/ai minor for ctx.step), thread settings and fork, reset, and the recovered-resolve fix.

The thinking shape batch updates docs/adapters/anthropic.md, docs/chat/reasoning.md, and docs/models/catalog.md, and adds the anthropic-thinking-shape changeset (@tanstack/ai, @tanstack/ai-anthropic, @tanstack/ai-models minor).

✅ Checklist

• I have followed the steps in the Contributing guide.
• I have tested code changes locally with pnpm run test:pr, or these tests do not apply to this pull request.
• I fully understand the code in this pull request, including any code generated with AI assistance.
• Docs: I updated docs/ for this change, or this change is not user-facing.
• Changeset: I added a changeset (pnpm changeset), or this PR does not change a published package.

Docs, changesets, and final source review are complete. The full local-test checkbox stays unchecked because the canonical command failed. The user explicitly authorized push despite the outstanding checks.

🚀 Release Impact

• This change affects published code, and I have generated a changeset.
• This change is docs/CI/dev-only (no release).

Published runtime packages change. Reviewed changesets are present. Branch CI remains pending for the pushed SHA.

Testing

Commands run. The frozen pnpm@11.9.0 install passed after owner/main integration. Focused regressions reached production code and failed before the fixes. The final focused run passed 32 selected cases across eight native commands: 22 JSON cases, five OpenRouter compatibility controls, and five host cases.

The first eight-package type check failed only in persistence. Indexed selector reads inferred number | undefined. The selector repair preserves its existing index-or--1 contract. No caller sentinel changed.

Current focused evidence	Result
Persistence's full host-card and reconstruction files	30 of 30 passed. This includes the positive no-host transfer control.
Strengthened OpenRouter replay controls	12 selected cases passed. Exact serialized nested-null checks cover each required terminal path.
pnpm exec nx run-many --target=test:types --projects=@tanstack/ai-persistence,@tanstack/ai-openrouter --parallel=1 --outputStyle=stream --verbose	Native exit 0 after the selector repair. Seven packages passed in the earlier eight-package run.

The follow-up cases overlap the earlier 32 selected cases. Their counts are not additive. Focused results do not replace the full repository gate.

Required final evidence	State
Nine existing paths reproduced on clean pinned main, with ordinary-input controls	Complete: 25 intended failures and nine ordinary passing controls on pinned main. All protected runtime bytes stayed unchanged. The first Base fixture error is excluded; its corrected run supplies valid proof.
Same-bytes original repro on the integrated candidate	Complete: exact5525 bytes passed3/3, native exit0; archived exact hash and feature source removed.
Exact fresh-main SDK/HTTP fixture bytes on the feature candidate	Complete: Base2 + Mistral2 + OpenRouter12 passed. Hash-guard removal verified for all three temporary copies.
Full literal pnpm test:pr	Known failures; manually stopped with actual root exit1 and captured-only closure verified. Includes live provider tests. No full canonical green or unfinished-target pass is claimed.
Focused local normal checks through existing scripts/Nx	Cancelled by explicit user push override. Source-reviewed command packet was not run. Two lint repairs were not retested; Solid config is unchanged.
Fresh final test:docs and test:kiira output	Further local checks cancelled by explicit user push override; branch CI pending.
E2E branch CI and the dedicated replay/validation browser sources	Local E2E waived at the user's request. Review branch CI for the final pushed head; no current browser pass is claimed.
Final combined owner-c9/MCP regression and full harness checks	No further local run at the user's request. Branch CI pending for the final pushed SHA.

Usage, settings, fork, reset, and resumable agents. These checks ran on the merged batch, before the rebase onto e677b7516:

Check	Result
@tanstack/ai: build, vitest run, test:types, test:oxlint, test:build	2404 of 2404 tests passed. Every other step exited 0.
@tanstack/ai-harness vitest run (one worker)	830 of 831 passed. The one failure was a build-stub timeout under load, and that file passed alone (8 of 8).
thread-settings.test.ts after the expose.settings gate	15 of 15 passed. tsc --noEmit for the harness exited 0.
kiira check on the changed docs pages	No errors.

After the rebase onto e677b7516, no local check ran, at the user's request. The E2E specs are written, and branch CI runs them.

Main merge and data-loss fixes. These checks ran on f02f0ddcd, after the rebase onto 6f3ea4ad6:

Check	Result
test:lib for ai, ai-harness, ai-compaction, ai-mcp	2409, 835, 105, and 413 tests passed. No failures.
test:types for those four packages and testing/e2e	All exited 0.
test:oxlint for those four packages	Exit 0. Only the warnings that were already there.
testing/e2e test:types and oxfmt on f07b75e0a	Exit 0.

On f07b75e0a, branch CI passed E2E, Bun, and Preview. Test and Coverage failed in five adapter suites that still expected '' and null to be rejected. 809259930 fixes them. After that fix, test:lib, test:types, and test:oxlint passed locally for openai-base, ai-ollama, ai-gemini, ai-mistral, and ai-openrouter (327, 70, 416, 102, and 303 tests).

On 809259930, Test passed and Coverage failed: ai-compaction functions dropped from 100% to 97.43% (76 of 78). 15868e693 adds a test for each of the 2 functions. Each test fails when its function is broken. ai-compaction passes 107 of 107 tests, with types, lint, and format clean.

Local E2E did not run, at the user's request. Branch CI runs it.

Adapters from catalog records. These checks ran on c0d7bfe21:

Check	Result
test:lib for the changed packages	ai 2499, ai-anthropic 264, ai-openai 434, ai-gemini 423, ai-bedrock 166, ai-mistral 107, ai-cloudflare 42, ai-models 33, openai-base 330. No failures.
test:lib for every dependent of @tanstack/ai and openai-base	Pass, except local environment cases: live Docker sandbox tests, Windows-only Solid and harness build tests, a duplicate @google/genai install, and an unbuilt example dependency.
test:types, test:oxlint for the changed packages and testing/e2e	Exit 0.
test:docs, kiira check	No broken links. 1806 snippets passed.
Catalog wire check (every reasoning record through the thinking builder of its adapter)	66 of 709 send no field, all Bedrock models that are not Claude. Before: 582.

Local E2E did not run. The new spec adapter-config-reasoning-wire.spec.ts runs in branch CI.

Anthropic thinking shape. These checks ran on 57ea69fed:

Check	Result
vitest run for ai, ai-anthropic, ai-models	2499, 273, and 36 tests passed. No failures.
tsc --noEmit and oxlint --type-aware for those packages	Exit 0. Only the warnings that were already there.
test:docs, kiira check, sherif, knip	No broken links. 1807 snippets passed. No new findings.
Local copy of the Flue shape check (1746 record and level pairs)	Only group E is different (24 pairs on 4 records).
Catalog flags against the pi snapshot	forceAdaptiveThinking equal on 291 of 291 records. supportsMidConvoEffort on the same 5 models.

Local E2E did not run. adapter-config-reasoning-wire.spec.ts now checks four record configs, and branch CI runs it.

On 6656c1d71, branch CI passed Test, E2E (after one rerun of generation-persistence-resume.spec.ts, a reload race from main), Bun, and Preview. Coverage failed only on ai-isolate-e2b. The new e2b test passed 5 of 5 local runs (29 of 29 tests), with tsc and oxlint clean.

Manual test.

1. Run pnpm --filter @tanstack/ai-persistence run test:lib --run --maxWorkers=1 tests/subagent-persistence-cards.test.ts tests/subagent-reconstruct.test.ts --no-file-parallelism.
2. Run pnpm --filter @tanstack/ai-openrouter run test:lib --run --maxWorkers=1 tests/replay-parity.test.ts --no-file-parallelism.
3. Review the final branch's E2E CI. Local E2E is not run at the user's request.
4. Read the adapter-switching and tool-approval docs for saved-history and approval compatibility.
5. Run pnpm --filter @tanstack/ai-harness exec vitest run tests/usage.test.ts tests/agent-resume.test.ts tests/thread-settings.test.ts tests/fork.test.ts tests/reset.test.ts. They cover the five new features.
6. Run pnpm --filter @tanstack/ai-anthropic exec vitest run tests/reasoning-request.test.ts tests/mid-conversation-effort.test.ts tests/mid-conversation-changes.test.ts. They cover the thinking shape, the two-turn effort prefix, and the channel object.

How this PR makes testing easy. Public tool-boundary tests cover final validation and approval. Host tests cover ordinary text, ambiguous markers, metadata, and occupied IDs. Adapter suites cover request and stream boundaries. Browser fixtures cover saved harness replay, client phases, streamed errors, and local cancellation. The owner's durable compaction and MCP option fixtures remain registered. The new batch adds usage.test.ts, agent-resume.test.ts, thread-settings.test.ts, fork.test.ts, and reset.test.ts in ai-harness, agent-step.test.ts in ai, and the E2E spec harness-thread-controls.spec.ts.

The full canonical run was manually stopped after known failures. Actual root exit1 and captured-only closure are verified. Failures were five lint findings, a Solid test collection error, and a baseline SBX pre-kill detector failure. Their current failure evidence stays separate from the parity controls. The user explicitly authorizes commit/push despite outstanding local checks. All further local checks and the Solid config repair are cancelled. The two lint repairs were source-reviewed but not rerun. The eight unchanged live Docker/SBX files and all browser E2E checks are left for branch CI. Its results remain pending until the new pushed head exists.

Limits. Provider HTTP mocks prove request and response handling. They do not prove live-provider credentials or service behavior. Browser fixtures define the UI checks for branch CI. No current browser pass is claimed. Source review of TypeBox's no-eval route does not prove a Worker deployment. Local verification uses the installed native Node 24.3 executable, not the repository's Node 24.8 pin. Coverage remains CI-only.

Risk / rollback

A reasoning config wins over the data of an adapter, so a wrong record sends wrong thinking fields. Without the config, nothing changes. Long-context tiers raise modelCost for large inputs, as the providers bill them. A record config now also picks adaptive or budget thinking. With midConversationEffort, every request thinks adaptively with a fixed output_config.effort: 'high' and sends no temperature.

Validation can reject input that previously reached a tool. Edited approval input is checked again before dispatch. Missing trusted context rejects cross-run legacy phases and keeps the client phase pending.

Replay changes provider input and preserves the saved transcript. Genuine response identity depends on what the provider reports. Adapters without a generation ID omit it.

Owner limitations remain visible. The stdio MCP loop lacks an automated end-to-end test. The owner report names three compaction limits. This PR now fixes the empty summary and the late host record. Reporting before the durable write is still open.

The new batch is opt-in, with two exceptions. First, a host without a log now saves the transcript before each model call, which is one more write per call. Second, snapshot() has a new usage field. A client configure input is refused unless expose.settings lists its fields.

Revert the final merge commit to undo the combined change. Opt-in host, gateway, and compaction settings can also be disabled.

Public API change

The owner stack adds the harness, CLI, dashboard, model catalog, and MCP server interfaces described above. Existing adapter factories remain usable. The replay work adds these approved caller surfaces:

Surface	Added contract
Source and response identity	Exported MessageSource has provider, api, and model. Optional message and saved-run fields carry source, response identity, and failed or aborted status. TextAdapter exposes optional readonly api and provider. RunFinishedEvent adds optional responseId. StructuredOutputResult adds optional responseId and model.
Anthropic configuration	AnthropicClientConfig extends the SDK options with optional apiKey; the SDK supplies authToken. AnthropicTextConfig adds oauth, allowEmptySignature, and provider. Existing factory and injected-client forms remain.
Azure Responses	Root exports add azureOpenaiText, AzureOpenAITextAdapter, and AzureOpenAITextConfig. Configuration covers apiKey, baseURL, resourceName, apiVersion, deploymentName, and deploymentNameMap.
Usage totals	session.usage(), snapshot().usage (SessionUsage, UsageCounts), and the harness.usage event.
Resumable agents	AgentStartOptions.resume, ctx.step.do(name, fn) in agent code, and optional SubagentBinding.step in @tanstack/ai.
Thread settings	defineHarness({ models, expose: { settings } }), session.configure(), session.settings(), client.configure(), and the configure input op.
Fork and reset	host.fork(harness, { threadId, newThreadId, at?, principal? }), session.reset(note?), client.reset(), and the reset input op.
Reasoning config	reasoning?: ModelReasoning on the first-party text adapter configs, ConfigReasoning in @tanstack/ai, and the *ModelId and *TextAdapterFor types of each adapter package.
Catalog and cost	modelReasoning(record), ModelReasoning, ModelCostRates.tiers, ModelCostTier, and TokenCounts.cacheWrite1h in @tanstack/ai-models.
Usage and replay	PromptTokensDetails.cacheWrite1hTokens, and responseItems on TanStackMessageMetadata and TanStackRunMetadata.
Thinking shape	ModelReasoning.adaptive and ModelReasoning.midConversationEffort (in @tanstack/ai and @tanstack/ai-models), ModelCompat.supportsMidConvoEffort, and reasoningEffort on TanStackMessageMetadata and TanStackRunMetadata.
Anthropic channels	midConversationChannels on AnthropicTextConfig also takes { tools?: boolean; systemPrompts?: boolean }.
Adapter utilities	@tanstack/ai/adapter-internals exports ReplayMessages, ReplayToolIdRule, transformMessagesForReplay, hashToolCallId, sanitizeUnicode, and sanitizeJsonArguments. These support adapter conversion; they add no application callback.

Before

import { chat } from '@tanstack/ai'
import { openaiText } from '@tanstack/ai-openai'

const stream = chat({
  adapter: openaiText('gpt-5.5'),
  messages: [{ role: 'user', content: 'Hello!' }],
})

for await (const chunk of stream) {
  if (chunk.type === 'TEXT_MESSAGE_CONTENT') console.log(chunk.delta)
}

After: Azure deployment

import { chat } from '@tanstack/ai'
import { azureOpenaiText } from '@tanstack/ai-openai'

const stream = chat({
  adapter: azureOpenaiText('gpt-5.5', {
    resourceName: 'my-resource',
    apiKey: process.env.AZURE_OPENAI_API_KEY,
    apiVersion: 'v1',
    deploymentNameMap: { 'gpt-5.5': 'production-chat' },
  }),
  messages: [{ role: 'user', content: 'Hello!' }],
})

for await (const chunk of stream) {
  if (chunk.type === 'TEXT_MESSAGE_CONTENT') console.log(chunk.delta)
}

Other caller examples remain in the adapter, tool, harness, and migration docs. No migration API or private-key editing procedure is added.

The latest owner P0–P14, shared-thread, MCP, compaction, auth, CLI, and run-route scope stays in this description. Historical owner checks are not presented as tests of this final merged candidate.

Pushed head: 618d71121. It adds the Anthropic thinking shape batch (a0eb57968, 57ea69fed), the Gemma drift fix (6656c1d71), and the e2b coverage test (618d71121) on top of ef76ab6b1, with plain pushes, no force. Branch CI passed every check on this head, Coverage included.