In thorough, "Request changes" at the proposal review sends nothing back: the answer is recorded and apply goes ahead #176

Closed
opened 2026-10-16 16:50:43 -04:00 by cmoriarty · 1 comment
Owner

Found while planning run 66, the combined verification run of 2026-10-16.

thorough's proposal gate, human.review-proposal, offers three answers:

  • Approve: "the change goes to implementation as it stands";
  • Approve with edits: "approved, but say below what you want changed as it is built";
  • Request changes: "send it back; say what is wrong".

Nothing in thorough sends it back:

  • After the gate, script.record-review-decision only classes the answer rejected in .osf/review/outcomes.json (osf.pipeline.review.classify_answer).
  • Then script.dependencies → openspec.apply[…] build the proposal unchanged.
  • The operator's words reach .osf/decisions.md, where apply may or may not act on them. In practice that is "Approve with edits" under another name, recorded as a rejection.

solo (#151) does what the option says. After "Request changes", openspec.revise.solo revises the change in the proposing session and commits it, and the gate opens again as "Review the proposal (round 2)", up to round 3. The condition, proposal_changes_requested, reads the latest answered "Review the proposal…" gate in decisions.md, and thorough's gate has the same header.

Wanted: in thorough and the pipelines derived from it, "Request changes" revises the change with the operator's words and asks again, as solo does, up to round 3. The evidence of what each round asked is kept. If that is too much, the option should at least stop promising "send it back".

Done when: a simulated run answers Request changes at the gate and sees a revise and a round-2 gate before apply, and the guide says what each answer does.

Found while planning run 66, the combined verification run of 2026-10-16. `thorough`'s proposal gate, `human.review-proposal`, offers three answers: - Approve: "the change goes to implementation as it stands"; - Approve with edits: "approved, but say below what you want changed as it is built"; - **Request changes: "send it back; say what is wrong"**. Nothing in `thorough` sends it back: - After the gate, `script.record-review-decision` only classes the answer `rejected` in `.osf/review/outcomes.json` (`osf.pipeline.review.classify_answer`). - Then `script.dependencies` → `openspec.apply[…]` build the proposal unchanged. - The operator's words reach `.osf/decisions.md`, where apply may or may not act on them. In practice that is "Approve with edits" under another name, recorded as a rejection. `solo` (#151) does what the option says. After "Request changes", `openspec.revise.solo` revises the change in the proposing session and commits it, and the gate opens again as "Review the proposal (round 2)", up to round 3. The condition, `proposal_changes_requested`, reads the latest answered "Review the proposal…" gate in `decisions.md`, and `thorough`'s gate has the same header. **Wanted:** in `thorough` and the pipelines derived from it, "Request changes" revises the change with the operator's words and asks again, as `solo` does, up to round 3. The evidence of what each round asked is kept. If that is too much, the option should at least stop promising "send it back". **Done when:** a simulated run answers Request changes at the gate and sees a revise and a round-2 gate before apply, and the guide says what each answer does.
Author
Owner

Shipped in bb5d87a (change operator-request-changes-revises, commit 2e8a70a), live since 00:07Z on 2026-10-17, and archived in 2f7d842.

  • In thorough and git-flow, Request changes at the proposal gate now runs agent.revise.operator with the operator's words (#182). It revises the change, validates it strictly, and writes .osf/review/operator-revision-<n>.md, saying what it changed for each point.
  • The gate then opens again as "Review the proposal (round 2)". A second request leads to round 3, the last, which offers only Approve and Approve with edits.
  • script.record-review-decision records every answered round.
  • The console names the new steps, and draws a round only once it began.
  • Checked on production in the verification run (run_01M53R7621CJW6PCZG7S6R94HS, thorough on cmoriarty/braid-combined-test, vLLM). Request changes ("After a delete, the Links page should say 'Deleted .' at the top, once…") ran the revise. It wrote its note and added the requirement, two scenarios and the tasks. The revision was committed, and round 2's gate opened. Approve there skipped round 3 and went on to apply. outcomes.json recorded rejected, then approved_unchanged.
  • Tests: graph, condition, brief and record tests, vitest for the names and the spine, a browser spec on a fixture run, and the simulated run thorough-request-changes. git-flow's simulated run now also checks the loop is skipped on an approval.

Closing.

Shipped in bb5d87a (change `operator-request-changes-revises`, commit 2e8a70a), live since 00:07Z on 2026-10-17, and archived in 2f7d842. - In `thorough` and `git-flow`, Request changes at the proposal gate now runs `agent.revise.operator` with the operator's words (#182). It revises the change, validates it strictly, and writes `.osf/review/operator-revision-<n>.md`, saying what it changed for each point. - The gate then opens again as "Review the proposal (round 2)". A second request leads to round 3, the last, which offers only Approve and Approve with edits. - `script.record-review-decision` records every answered round. - The console names the new steps, and draws a round only once it began. - Checked on production in the verification run (`run_01M53R7621CJW6PCZG7S6R94HS`, thorough on cmoriarty/braid-combined-test, vLLM). Request changes ("After a delete, the Links page should say 'Deleted <code>.' at the top, once…") ran the revise. It wrote its note and added the requirement, two scenarios and the tasks. The revision was committed, and round 2's gate opened. Approve there skipped round 3 and went on to apply. `outcomes.json` recorded `rejected`, then `approved_unchanged`. - Tests: graph, condition, brief and record tests, vitest for the names and the spine, a browser spec on a fixture run, and the simulated run `thorough-request-changes`. `git-flow`'s simulated run now also checks the loop is skipped on an approval. Closing.
Sign in to join this conversation.
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
cmoriarty/braid#176
No description provided.