Repository navigation
fix(goal-conductor): treat an owner's remark as a suggestion until the owner makes it a rule - #37651
Conversation
…e owner makes it a rule
UX Review (Opus 5.5) — ⏭️ skippedRevision |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS No design concerns: the change adds one paragraph of skill text, placed right after the "save a correction as a lesson" paragraph that caused the problem, and it removes nothing. Notes
[DESIGN-REVIEWED] 9ef5b17 Description read[DESCRIPTION-READ] bc8500dfb175aa172bbd86417a9422f8eef2d4a3b9fbb5e361474e7cd27e8f30 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) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Nothing blocks: the change adds one paragraph to the conductor skill, its defect is tracked in #37361, and it does not touch the explicit-correction path. What this change shipsInventory (4 items) — 4 justifiedIntent: stop a goal-conductor Bot from turning an owner's one-off "maybe" into a costly standing rule for every change. This is a FIX.
[FIRST-PRINCIPLES-REVIEWED] 9ef5b17 Description read[DESCRIPTION-READ] 61aa7c5bd61967d58f111b622ff310d0232cdd717677d7ec3f01980182f34662 That sha256 names the PR description this verdict read. Recompute it with |
Opus 5.5 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: |
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: |
Code change approver recommendation: ApproveWhy: the decider estimates a 6% chance that this change introduces a defect, within the 12% limit for approving on its own, and every check passes. The PR adds one paragraph to the Patrol section of goal-conductor/SKILL.md. It says an owner's passing remark is a suggestion for that case unless the owner says it is a rule. Before making it a rule for every change, the Bot asks once and names the cost, and the same check applies before saving the remark with learn_add. The existing correction-as-lesson path is unchanged. The new test reads the shipped skill and slices it between '### Patrol' (line 401) and '### Close the round' (line 739). The paragraph (lines 465–471) falls inside that range, so the test fails if the paragraph is removed. It imports _BUILTIN_SKILLS_DIR from kiro_crew.skills, which re-exports it and which seven other tests import the same way. The parity tests in test_mate_conductor_parity.py only check that anchor phrases are present, so added text cannot break them. I found no skill size or snapshot limit. No other skill uses the 'as a lesson' wording, and docs/decisions has no ruling on remarks or lesson saving. The change is wording only: no data, trust or lifecycle impact. An automated recommendation, not a review: a maintainer still decides. It approves on its own only when the estimated chance of a defect is at most 12% and the change has at most 400 changed lines. Decider v3c (217fe607) · commit |
|
This PR is part of #37361, and that issue is tracked as a design-level change to how goal-driven Bots run. This adds a new concept, so it needs a design sign-off from a maintainer before it can merge. A maintainer will post the decision here. Review result on this head: no blocking findings. The diff adds one paragraph to the Patrol section of Optional:
|
bolichen97
left a comment
There was a problem hiding this comment.
Approving 9ef5b1773a.
- The new Patrol paragraph is scoped well: a remark stays a suggestion for its own case, the Bot asks once and names the cost before widening it, and the same check sits in front of
learn_add. A real correction still becomes a lesson as before. - Ran
test/test_conductor_owner_remark_suggestion.pylocally: 3 passed. CI is green on this head.
AI-assisted, sent by Bolin Review bot
Part of #37361
Problem / Motivation
Goal: A Bot treats an owner's passing remark as a suggestion for that case, and asks once, naming the cost, before it turns the remark into a rule for every change.
An owner wrote "maybe even a separate security reviewer, for the credentials" once. The Bot made it a security review, with re-checks, for every change that touches credentials, roles, the API, install or deploy. It never asked what that rule costs.
Why it matters
A remark that silently becomes a rule for every change multiplies the work on every later change. The owner only finds out days later, when the Bot's work has grown past what it planned.
Not a goal
learn_add's code or to how saved lessons are injected.crewmate_prompts.json; it says nothing about lessons.What changed (motivation → approach → change)
The goal-conductor skill told the Bot to save a person's correction as a lesson, and said nothing about remarks. A remark therefore had no default reading, and the Bot read "maybe" as a rule.
The skill's Patrol section now says that an owner's remark applies to the case it was said about unless the owner says it is a rule. Before the Bot turns a remark into a rule for every change, it asks the owner once and names what the rule adds to each change. The same check comes before saving a remark with
learn_add: the Bot saves it as a rule only when the owner said so or agreed, and keeps the scope the owner gave.Backwards compatibility
Compatible: skill text only. A correction the owner gives still becomes a lesson as before; only an unconfirmed remark waits for one question.
Removes nothing: the diff adds one paragraph to
src/kiro_crew/builtin_skills/goal-conductor/SKILL.mdand one test file.Tests
test/test_conductor_owner_remark_suggestion.pypins the three rules in the Patrol section: a remark is a suggestion unless the owner says it is a rule; ask once and name the cost before a rule for every change; the same check beforelearn_add. Mutation-checked: removing the paragraph turns all three red, and removing each phrase turns its own test red. Determinism: the file passed 5 repeats and a reversed test order.Proof: PROVEN test/test_conductor_owner_remark_suggestion.py::test_owner_remark_is_a_suggestion_unless_the_owner_says_it_is_a_rule, test/test_conductor_owner_remark_suggestion.py::test_bot_asks_once_and_names_the_cost_before_a_rule_for_every_change, test/test_conductor_owner_remark_suggestion.py::test_learn_add_gets_the_same_check (head 9ef5b17)
Manual verification
N/A — the change is skill wording, and the test reads that wording from the shipped skill file.
Related Issues
Part of #37361
Sibling sweep
Swept: git grep -n -i "as a lesson|learn_add" -- src/kiro_crew/builtin_skills src/kiro_crew/crewmate_prompts.json
Sibling: src/kiro_crew/builtin_skills/goal-conductor/SKILL.md:461 -- changed here
Sibling: src/kiro_crew/builtin_skills/kirocrew-dev/dashboard-template/agent-spec.md:119 -- not affected -- that agent is told never to call learn_add
Pattern harvest
Not generalizable: the defect is one missing default in the conductor's instructions, not a code shape a lint rule can match.
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)