Repository navigation
docs(skills): rename prepare-pr, slim it and match current ci - #16740
Conversation
Design Review (Opus 5.5) — 🟡 CONCERNSDesign-level review of Design-Verdict: CONCERNS An agent can now override AI-review findings on its own PR, and the loop arms auto-merge by default. That adds a policy to a PR whose goal is a rename and a "describe-only" update. Watch
[DESIGN-REVIEWED] 608b4aa Description read[DESCRIPTION-READ] 6aa8164ea868be0b55da2e0a3997366d481408e6cdd1488da2327fe3b4b800da That sha256 names everything this verdict read: the PR description, plus the 1 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with |
First Principles Review (Opus 5.5) — ✅ human override accepted@bolichen97 overrode this lane for |
Opus 5.5 Review — ✅ no blocking findingsReviewed Review detailsNo findings block this merge. One advisory finding below. FINDING — src/kiro_crew/builtin_skills/kirocrew-dev/kirocrew-prepare-pr/SKILL.md:24 — The trigger list now includes "fix ci", "make it green" and "commit and push". With these, a bare "fix ci" falls through to the default "Full loop", and the Guardrails at line 293 say "A request to run this loop authorizes its own pushes, … auto-merge, and the red-main fix PR it opens". So the agent pushes, arms auto-merge and opens new PRs even though nobody asked it to push. That contradicts AGENTS.md ("Do NOT [OPUS-REVIEWED] 608b4aa Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
374b615 to
8214759
Compare
GPT 6.1 Review — ✅ no blocking findingsGPT 6.1 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: |
8214759 to
3828984
Compare
|
Frozen-goal edit, at the maintainer's request. @bolichen97 asked for the skill to be renamed to The rename keeps working for existing setups: |
Security Scope Review -- 🟡 CONCERNSWhich legitimate operations does Verdict basis: the reviewer's own judgement; no confirmed regression. Security scope review -- adjudicated candidates
No candidate flipped from allowed to refused. Scope-Verdict: CONCERNS TIGHTENING: built-in skill relocation/quarantine in ADJUDICATED: 18 candidates, verdicts pending from the harness (none were handed to me in this run, so I report no confirmed newly-refused rows) PLATFORM COVERAGE:
UNADJUDICATED:
[SCOPE-REVIEWED] 608b4aa |
ab1cc58 to
3088197
Compare
|
/ai-review override first-principles 3088197: agent: records maintainer @bolichen97's explicit decisions for #16731 - full loop arms auto-merge by default with a 24h watch, a writer's agent may override false-positive/over-engineering findings (never security-class) and must report it, red main gets its own T1 fix PR, the skill is renamed kirocrew-prepare-pr and scoped to this repo. |
Human judgment recorded@bolichen97 marked the first-principles AI finding as false positive, not applicable, or explicitly accepted for
This decision applies only to this commit. A new push requires a new judgment. |
3088197 to
8ba217e
Compare
|
/ai-review override first-principles 8ba217e: agent: records maintainer @bolichen97's explicit decisions for #16731 - full loop arms auto-merge by default with a 24h watch, a writer's agent may override false-positive/over-engineering findings (never security-class) and must report it, red main gets its own T1 fix PR, the skill is renamed kirocrew-prepare-pr and scoped to this repo. |
Human judgment recorded@bolichen97 marked the first-principles AI finding as false positive, not applicable, or explicitly accepted for
This decision applies only to this commit. A new push requires a new judgment. |
|
Human judgment recorded, but the following reviewer lane(s) could not be re-run automatically: First Principles Review. Re-run the lane's latest workflow run for |
|
|
|
8ba217e to
67e1a30
Compare
|
/ai-review override first-principles 67e1a30: agent: records maintainer @bolichen97's explicit decisions for #16731 - full loop arms auto-merge by default with a 24h watch, a writer's agent may override false-positive/over-engineering findings (never security-class) and must report it, red main gets its own T1 fix PR, the skill is renamed kirocrew-prepare-pr and scoped to this repo. |
92f9275 to
b52e945
Compare
|
/ai-review override first-principles b52e945: agent: records maintainer @bolichen97's explicit decisions for #16731 - full loop arms auto-merge by default with a 24h watch, a writer's agent may override false-positive/over-engineering findings (never security-class) and must report it, red main gets its own T1 fix PR, the skill is renamed kirocrew-prepare-pr and scoped to this repo. |
Human judgment recorded@bolichen97 marked the first-principles AI finding as false positive, not applicable, or explicitly accepted for
This decision applies only to this commit. A new push requires a new judgment. |
|
Human judgment recorded, but the following reviewer lane(s) could not be re-run automatically: First Principles Review. Re-run the lane's latest workflow run for |
The skill every Kiro Crew author's agent runs to ship a PR had drifted from CI and sat at its 68000-byte budget. It now states the CI rules for inside and outside authors (issue, triage, tiers, every blocking lane, PR Hygiene's Pattern harvest and Reader lines, overrides, the merge queue) and adds three loop rules: a red main gets its own T1 issue and fix PR while the original waits; the full loop arms auto-merge and watches up to 24 hours, rebasing conflicts; and a writer's agent may override a false positive or over-engineering finding and must report it, while security-class findings always go to a person. It is renamed to kirocrew-prepare-pr everywhere it is referenced, and carries `repo_scope: src/kiro_crew`, so it loads in every Kiro Crew session and never in another repository's; the CRUX `cr` triggers are gone. _RELOCATED_SKILLS quarantines an installed kirocrew-dev/prepare-pr copy on upgrade, and a saved `$prepare-pr` still resolves to it. The text shrinks from 67911 to about 55000 bytes and the ceiling test ratchets to 56000. The CI doc's deferral paragraph now names the tier-based path. Closes #16731
b52e945 to
608b4aa
Compare
|
|
/ai-review override first-principles 608b4aa: agent: records maintainer @bolichen97's explicit decisions for #16731 - full loop arms auto-merge by default with a 24h watch, a writer's agent may override false-positive/over-engineering findings (never security-class) and must report it, red main gets its own T1 fix PR, the skill is renamed kirocrew-prepare-pr and scoped to this repo. |
Human judgment recorded@bolichen97 marked the first-principles AI finding as false positive, not applicable, or explicitly accepted for
This decision applies only to this commit. A new push requires a new judgment. |
|
Human judgment recorded, but the following reviewer lane(s) could not be re-run automatically: First Principles Review. Re-run the lane's latest workflow run for |
|
|
|
PR Readiness reads main's lane list, which waits on an Issue Gate check. Issue Gate is a pull_request workflow, so GitHub reads it from the PR's merge ref, and release/0.8.0 has no issue-gate.yml. Every PR into the release branch therefore waits on a check that never starts. Copy .github/workflows/issue-gate.yml from main unchanged. Its grammar script is read from a checkout of the default branch at run time, so the workflow file is the only piece the release branch needs. Source: origin/main 2359397 (file last changed there in #16740; introduced by #16241, #16634, #16697). Fixes #16868
Problem / Motivation
Goal: Every Kiro Crew author's agent ships a PR with a
prepare-prskill that matches today's CI and states the repository's process, in a file short enough to read on every load.The skill had drifted from the CI it drives. It never named
Security Scope RevieworInternal Content Scan, listed one of seven override targets, missed PR Hygiene's## Pattern harvestandReader:rules, and still sent deferrals throughneeds-triage, whichIssue Gatestopped reading in #16697. It told an outside author nothing about issues, triage or tiers. It sat at its 68000-byte budget, and its description did not say "Kiro Crew PR", so the agent did not always load it.Why it matters
Each missing rule costs an author a CI round to learn, and an outside contributor cannot learn the issue and tier process from the loop at all. Main stays red until someone happens to fix it, and green PRs decay unmerged.
Not a goal
Issue Gateor the merge queue.What changed (motivation → approach → change)
The skill now carries the rules an author needs, and the loop finishes the job instead of stopping at review-ready.
kirocrew-prepare-pr(directory, frontmatter, and every reference in CI, docs, skills and tests);repo_scope: src/kiro_crewloads it in every Kiro Crew session and hides it everywhere else; the CRUXcrtriggers are gone. An install that holds the oldkirocrew-dev/prepare-prcopy has it quarantined on upgrade, andskill_runtime/search.pyresolves a saved$prepare-pr(a hook, an AGENTS.md line) to the new nametier:T1..tier:T4and which tiers wait fortriaged; every blocking lane and what to do when it is red; PR Hygiene; the merge queuetier:T1issue and its own fix PR; the original PR waits, then rebasesmax_cycles=280)agent:, and must report it; security, data-loss, corruption, crash and removed-guard findings always go to a person.docs/ci/ci-and-reviews.md§ Human override says the sameThe rename rewrites only the name wherever it appears:
.github/workflows and scripts,CONTRIBUTING.md,.gitignorecomments,src/kiro_crew/agent.py,src/kiro_crew/frontmatter.py,src/kiro_crew/security_posture.py,src/kiro_crew/docs/agents.md, other skills, specs and tests;src/kiro_crew/skills.pymaps both old names to the new key.Repeated text moved out to
--help, the PR template, thebabysitskill andreferences/rationale.md, which also records why each new rule exists.SKILL.mddrops from 67911 to about 55000 bytes and its ceiling test ratchets to 56000.docs/ci/ci-and-reviews.mdnow names the tier-based deferral path. RFCs keep the old name: they are history.Backwards compatibility
Breaking: an agent spec that maps this skill by path, written by the crew page as
skill://~/.kiro/crew/skills/kirocrew-dev/prepare-pr/SKILL.md, loads nothing after upgrade, because the oldSKILL.mdis quarantined so two copies never load. Re-map the skill from the crew page to fix it. Writer sweep on origin/main @ 5bb27bd: no shipped agent template or spec maps it by path;src/kiro_crew/dashboard/handlers/_shared.py:_skill_uri_for_pathwrites the URI only when an operator maps a skill. A typed or hook-saved$prepare-prkeeps working through the alias, andpr-readiness.ymlandIssue Gateread the evaluator from the default branch, so each side always finds its own path.Reader: src/kiro_crew/agent_discovery.py:expand_skill_uri -- chat, subagent -- a saved old-path URI resolves to the quarantined file; the operator re-maps the skill
Reader: src/kiro_crew/skill_runtime/search.py:resolve_dollar_skills -- chat, cron --
$prepare-prresolves through_RELOCATED_SKILLS;test_renamed_builtin_answers_to_its_old_nameReader: test/test_review_repair_routing_skill.py:test_comment_aware_recipes_are_finite_ungated_and_keep_repair_routing -- chat -- pins the new
(280, 86400)recipeReader: test/test_pr_quality_gates.py:TestScreenshotEvidence -- chat -- the pinned screenshot bullets are kept verbatim
Tests
test_renamed_nested_copy_quarantined_when_new_name_presentpins the upgrade path for the old nested copy;test_renamed_builtin_answers_to_its_old_nameandtest_installed_skill_keeps_its_name_over_a_relocation_aliaspin the$prepare-pralias and that a user's own skill of that name still wins. Otherwise no new tests: every rule the skill text carries is already pinned by about 150 assertions across 18 test files, and all of them pass on this head. The ceiling and recipe pins are updated to the new values, and so are the security-conductor spec digests intest_agent_refactor_spec_bytes.py/test_agent_refactor_contracts.py, whose prompt names the skill.Manual verification
prepare-pr/SKILL.mdor its references: green.scripts/docs-lint.sh: green.gptandopusreviewers; the one Opus blocker (agent self-override on security findings) is fixed and re-verified.Related Issues
Closes #16731
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)