Skip to content

[Chore] Remove the completion check - #3139

Merged
mrubens merged 3 commits into
developfrom
chore/remove-completion-gate
Sep 22, 2026
Merged

mrubens merged 3 commits into
developfrom
chore/remove-completion-gate

Conversation

@mrubens

@mrubens mrubens commented Sep 22, 2026

Copy link
Copy Markdown
Contributor

Why

The completion check (#3071, #3083, #3085, #3113) set out to replace the judge's minutes-long completion pass with a sub-second decision-model check. The judge part worked; the check did not. On nightly it flagged 49-60% of coding turns, mostly wrongly, because its hardest questions (did the tests pass on the shipped code?) depended on reconstructing validation evidence from the transcript: a shell-command window, file fingerprints, staleness rules. Each fix to that evidence exposed another gap. CI and pull request review already answer those questions from the source of truth.

What this removes

  • Turn-end check and reopen in the OpenCode harness, including the held-message and settle plumbing it needed.
  • The roomote-completion-gate.js OpenCode plugin and the commands.checkCompletionBeforeTool sandbox procedure (tool-time holds).
  • Command recording, diff collection, and code fingerprinting (completion-gate-runtime.ts, completion-gate-evidence.ts).
  • POST /api/mcp/tasks/runs/:runId/completion_check, the server gate, its types, and the ROOMOTE_COMPLETION_GATE flag set at dequeue.
  • Docs describing the check.

What stays

Rollback

No schema involved. A worker from the previous release calling the removed route gets a 404, which it already treats as a skipped check.

Validation

pnpm lint:fast, pnpm check-types:fast, pnpm knip; worker sandbox-server and run-task suites (remaining failures are the known local-only command-executor, tail-file-path, and OpenRouter variant tests), cloud-agents workflows, sdk task-runs, api tasks.

@roomote-community

roomote-community Bot commented Sep 22, 2026

Copy link
Copy Markdown
Contributor

No new code issues found. See task

  • The judge's shared prompt still requests a full implementation review, so visual-proof runs retain the completion pass this change intends to remove. apps/worker/src/run-task/agent-home.ts:1389
  • The proof-only judge no longer receives the final shipped diff, so it cannot detect source drift after capture begins. packages/cloud-agents/src/server/workflows/skills/standard/implement-changes/resources/default-workflow.md:37

Reviewed e118987

Comment thread apps/worker/src/run-task/agent-home.ts
@mrubens
mrubens marked this pull request as ready for review September 22, 2026 20:53
@mrubens
mrubens merged commit 1aabae6 into develop Sep 22, 2026
18 checks passed
@mrubens
mrubens deleted the chore/remove-completion-gate branch September 22, 2026 20:53
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant