Repository navigation
Conversation
UX Review (Opus 5.5) — ⏭️ skippedRevision |
First Principles Review (Opus 5.5) — 🔴 BLOCKPremise-level review of First-Principles-Verdict: BLOCK Product-shape change without accepted RFC: every goal conductor now writes its own plan's worker count as Notes
What this change shipsInventory (5 items) — 2 justifiedIntent: a Bot runs no more workers at once than its approved plan said. FIX in name, but it reverses a default.
Blockersproduct-shape change without accepted RFC: the conductor caps its own concurrency by default — product-shape [FIRST-PRINCIPLES-REVIEWED] 7f9f171 Description read[DESCRIPTION-READ] c83670a9c59a21ae2446061744d6c7b364ae94a015a168ae3a2fc25ed280cc3b That sha256 names the PR description this verdict read. Recompute it with |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Nothing blocks. The change only edits prompt text and a reversal is a clean revert. It does go against RFC §7.2 ("a bound the user set is Notes
[DESIGN-REVIEWED] 7f9f171 Description read[DESCRIPTION-READ] 5f90a5a71b76f124fa500b457b0f5dfbe39ed9301351f33dad1bf81aa77dffa5 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 |
Opus 5.5 Review — 🔴 changes requested (blocking)Reviewed Blocking: this PR overturns the repository owner's accepted §7.2 decision on BLOCKING — src/kiro_crew/builtin_skills/goal-conductor/SKILL.md:616 [BLOCK-MERGE] 7f9f171 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: A person should decideWhy: the decider raised 1 possible problem and estimates a 72% chance that this change introduces a defect; a person should check it. This prompt-only change makes the conductor write its plan's worker count as the goal's concurrency_cap. That directly reverses the owner's 2026-11-13 decision in the implemented RFC rfc-work-ledger-concurrency-cap.md §7: no micro-managing, the cap is only for a bound the user sets, and resources decide the wave. No accepted RFC amendment or override covers the reversal, and it also changes the shared rule that Ultra Mode's conductor reads. Request changes until an RFC amendment lands or a maintainer records an override. Things to check (possible problems the decider raised; at this confidence many turn out to be fine):
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 |
|
span=ebe6efe49155
|
324fabb to
9fb1f89
Compare
|
span=ecd02edd22be
|
9fb1f89 to
d9a6988
Compare
|
Rebased: main moved in, and the RFC for #37361 (#37649) is merged. Its open decision 7 asks whether the plan's worker count is a bound the user set, which is the §7.2 question the First Principles lane raised here. Conflicts resolved by keeping both sides:
Local: 1681 related backend tests pass, plus black, flake8, comment-history, brand and line-cap gates. |
d9a6988 to
7f9f171
Compare
|
span=9ad0ef453319
|
|
span=1c12c52038a7
|
|
Closing per the owner's decision: an approved plan's worker count does not become the goal's concurrency cap. RFC #37649 section 7.2 stands as written. |
Pull request was closed
Part of #37361
Problem / Motivation
Goal: A Bot runs no more workers at once than its approved plan said, unless the user raises the number.
A goal conductor's plan names its concurrency ("two or three"), but nothing writes that number into the work ledger. The ledger's
concurrency_capstays at its default of 32, so the plan's number is prose that nothing enforces. In the report on #37361 the Bot ran far more workers at once than its plan said.Why it matters
Every worker the conductor starts beyond its plan costs model time, opens more pull requests and makes the goal harder to follow. The ledger already refuses a
createpastconcurrency_cap; it just never receives the plan's number.Not a goal
resource_statusor how a wave is sized under memory pressure.What changed (motivation → approach → change)
The plan's worker count never reached the ledger, because the rules told the conductor to write
concurrency_caponly for a bound the user set. The conductor now writes the number its approved plan named asconcurrency_cap, with the goal record, before its firstcreate. A number the user sets wins. A range like "two or three" writes the upper end. Only the user can raise it. The plan has to state concurrency as one number of items running at once. The goal-conductor skill (src/kiro_crew/builtin_skills/goal-conductor/SKILL.md), the two conductor prompt rules, theconcurrency_capparameter text ofwork_ledger_recordand the work-ledger doc now say the same thing.This changes a recorded owner decision.
rfc-work-ledger-concurrency-cap.md§7.2 says only a number the user sets isconcurrency_cap, andresource_statussizes each wave. This PR counts the number in a plan the owner approved as one the user set. The RFC for #37361 (#37649, merged as a draft) lists this as open decision 7, with the recommendation "yes". It is not decided yet, so this PR waits on the owner's decision before it merges.Backwards compatibility
Compatible: only instructions and a parameter description change; the ledger accepts the same
concurrency_capvalues (1 to 32) and a goal with none written keeps the default of 32.Removes nothing: no field, path, validator or record is removed or narrowed.
Tests
test/test_conductor_plan_concurrency_cap.pypins that the goal-conductor skill, the conductor'sresource_waveandspend_boundedrules, and theconcurrency_capparameter description all tell the conductor to write the plan's number as the cap before the first create, and to raise it only on the user's word. All five tests fail on the parent commit and pass on this head. Reverting only the SKILL.md change turns the two skill tests red (prove.py cannot revert Markdown, so that check was by hand).test/test_agent_refactor_spec_bytes.pyandtest/test_agent_refactor_contracts.pypin the bytes of the written conductor specs and of_CONDUCTOR_SYSTEM_PROMPT. Their stored digests forkirocrew-conductor.json,kirocrew-ledger-conductor.jsonand_CONDUCTOR_SYSTEM_PROMPTare refreshed to the new prompt text; no other digest moved.Proof: PROVEN test/test_agent_refactor_contracts.py::test_every_pre_split_name_resolves_with_its_pre_split_shape, test/test_agent_refactor_spec_bytes.py::test_every_written_spec_matches_the_pre_split_bytes[clean_over_customized], test/test_agent_refactor_spec_bytes.py::test_every_written_spec_matches_the_pre_split_bytes[customized], test/test_agent_refactor_spec_bytes.py::test_every_written_spec_matches_the_pre_split_bytes[forks], test/test_agent_refactor_spec_bytes.py::test_every_written_spec_matches_the_pre_split_bytes[fresh] (head 9fb1f89)
Determinism: 20/20 runs at
-n0, 10/10 at-n 4besidetest/test_mcp_work.pyandtest/test_mate_conductor_parity.py, and green in a shuffled order. The tests only read static text, so no clock, thread or shared state is involved.Manual verification
N/A — the change is prompt and skill text, and the tests read the exact text a conductor receives.
Related Issues
Part of #37361
Sibling sweep
Swept:
git grep -n "concurrency_cap\|a bound the user set" -- src/Sibling: src/kiro_crew/agent_parts/spec_prompts.py -- changed here
Sibling: src/kiro_crew/mcp_work.py -- changed here
Sibling: src/kiro_crew/docs/work-ledger.md -- changed here
Sibling: src/kiro_crew/work_ledger_writes.py -- not affected -- its refusal already says only the user raises the cap
Pattern harvest
Not generalizable: the defect is one rule that named only the user's number as the cap; it has no code shape a lint could match.
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)