Repository navigation
feat(prepare-pr): freeze a PR's goal sections at open - #14181
Conversation
Design Review (Fable 5) — ✅ PASSDesign-level review of Design-Verdict: PASS Verify the fork lane post-merge: Suggestions
[DESIGN-REVIEWED] 6a77b5b |
GPT 5.6 Review — ✅ no blocking findingsGPT 5.6 completed its review of This comment is updated in place on each push. Review detailsNo findings. False positive or not applicable? A repository writer can comment: |
First Principles Review (Fable 5) — 🟡 CONCERNSPremise-level review of First-Principles-Verdict: CONCERNS A memory-index owner-only change (memory.py, vector_memory.py, windows-install.md, a new test file) rides along undeclared — and base tip 4f35e56 (#14204) already ships it. Not justified as shipped
Every product-shape item here (the same-repo blocking gate, the reversed "Fork-only by design" rule, the prepare-pr revert verdict, the cost imposed on open PRs) is covered by What this change shipsInventory (10 items) — 9 justifiedIntent: a PR's goal is written once at open and every later review round is judged against it — an ADDITION (new gate plus process change).
[FIRST-PRINCIPLES-REVIEWED] 6a77b5b |
Opus 5 Review — ✅ no blocking findingsReviewed Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
c4651b0 to
2770bbe
Compare
|
self-added: no
|
|
self-added: no
|
|
self-added: no
|
|
self-added: no
|
|
self-added: no
|
|
self-added: no
|
|
self-added: no
|
|
self-added: no
|
2770bbe to
c1283ed
Compare
|
self-added: no
|
|
self-added: no
|
kirodotdev#14181 left the skill at 68015 bytes, so test_slim_skills_do_not_regrow_duplicate_guidance_or_wire_model_ids fails on main. Drop the duplicate 'never edit the goal' clause; the frozen-goal paragraph already says only the user may edit it.
kirodotdev#14181 left the skill at 68015 bytes, so test_slim_skills_do_not_regrow_duplicate_guidance_or_wire_model_ids fails on main. Drop the duplicate 'never edit the goal' clause; the frozen-goal paragraph already says only the user may edit it.
…on (#14049) * feat(session-control): session_create takes a model for the new session spawn_run and spawn_continue already take a per-run model, but session_create had no way to pick one, so a conductor opening worker sessions could not put them on a different model than the default. Add an optional model argument: held to the spawn_run model-id charset by the schema, refused with model_rejected by the same guard the dashboard model picker runs, pinned on the slot with the pick-generation bump a human pick gets, and recorded in the persist-at-birth metadata so an idle child keeps it across a restart. * fix(prepare-pr): trim SKILL.md back under its 68000-byte ceiling #14181 left the skill at 68015 bytes, so test_slim_skills_do_not_regrow_duplicate_guidance_or_wire_model_ids fails on main. Drop the duplicate 'never edit the goal' clause; the frozen-goal paragraph already says only the user may edit it.
Problem / Motivation
Goal: A PR's stated goal stays fixed from the moment it opens, so the review loop judges every later change against that goal instead of against its own growth.
The prepare-pr loop rewrites the whole PR body every round to match the diff. Its intent lived in a separate
<!-- prepare-pr-intent -->comment that many PRs never got. So each round the body's "why" drifted to describe whatever the loop had just added. Reviewers and the every-third-round retrospective then read that drifted text as the intent, and absorbed the loop's own mechanisms into it. PRs grew round after round with nothing left to measure the growth against.Why it matters
Agent-driven PRs on this repo routinely run ten or more review rounds. Without a fixed goal, each round can widen the PR and still look "in scope", so maintainers get large diffs that no longer match what was asked for, and review cost keeps climbing.
Not a goal
/intent approvecommand that detects edits to the frozen sections. That lands in a later PR.first-principles-review.yml,fork-first-principles-review.yml).What changed (motivation → approach → change)
Every PR body now opens with a frozen goal: a one-sentence
**Goal:**line,## Why it mattersand## Not a goal. Agents never rewrite those three, the review loop measures the whole diff against them, and PR Hygiene blocks a PR that lacks them.Design and decisions: docs/request-for-change/rfc-pr-goal-freeze.md (status
in-progress, merged in #14265)..github/PULL_REQUEST_TEMPLATE.md**Goal:**line under Problem / Motivation, new## Not a goal, and a top note: the first three sections are frozen; everything from What changed down tracks the diff.AGENTS.md.github/scripts/pr-description-check.sh## Not a goal, plus a filled**Goal:**line under Problem / Motivation. The failure names the missing part and links the template on main.code-review.yml) andfork-pr-description.ymldocs/ci/ci-and-reviews.mdlists it as PR Hygiene's fourth check.origin/<base>...HEADdiff against them and may rule "revert to<head sha>and redo via a smaller path".pr_findings.py --roundsprints the Goal line.No same-repo required-headings check existed before. The fork check said so and left same-repo PRs unchecked. PR Hygiene was the closest same-repo check, so the rule joined it there rather than as a new workflow. The rules moved into a shared script so the two checks cannot drift. The fork workflow reads it through a sparse checkout of the base branch, never the fork.
Backwards compatibility
Breaking: PR Hygiene now fails any PR whose body lacks
## Not a goalor a filled**Goal:**line, and the fork check reports the same. Writer sweep on origin/main @ 67f79c3: the only in-repo PR-body writer is prepare-pr, which scaffolds from the template and is updated here. Every open PR body predates the change, so each fails on its next push or body edit. The fix is a body edit with no push, since PR Hygiene runs onedited. Dependabot PRs are skipped. The off-repo auto-approval cron still uses the old list, so a fork's runs may be approved while its description check is red; PR Hygiene still blocks it.Tests
test/test_fork_pr_description_workflow.pyruns the shared script for real: missing Goal line, empty scaffold, Goal in the wrong section, fenced or indented Goal, a subheading, and the shipped template failing until its Goal is filled. It also pins the base-only sparse checkout, thecontents: readpermission, and PR Hygiene running the same script for every PR.test/test_prepare_pr_rounds_view.pychecks--roundsprints the Goal line redacted, reads an empty or absent one as missing, and that the skill no longer names the intent comment.Manual verification
N/A — unit coverage sufficient: the check logic is shell executed by the tests, and the skill change is prose pinned by its tests.
Related Issues
no linked issue: requested directly by the maintainer.
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)