Repository navigation
docs(rfc): conductor drops its chat context at a round close - #18097
Conversation
16c717e to
fc2c571
Compare
UX Review (Opus 5.5) — ⏭️ skippedRevision |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Before merging, check that writing self-arm records for ordinary dashboard slots (§3 "The trust record") leaves every existing [DESIGN-REVIEWED] 106433b Description read[DESCRIPTION-READ] 85ba1784fd1fe2933abe9576a6c76cde90b9d1fe6419f9aaeca1b29794bc0e2c 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) — ⏭️ skippedRevision |
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: |
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: |
|
|
|
|
|
|
|
|
fc2c571 to
106433b
Compare
|
Self-review (crew-pr-reviewer, read-only): HOLD for a maintainer / security owner This RFC records the decision to let kirocrew-conductor call reset_conversation from its own self-armed patrol wake at a round close, so a days-long patrol stops re-sending a chat whose state already lives in the work and session ledgers. As a decision document it is sound: the five admission checks (conductor agent, named loop id, current active loop, self-arm trust record, nothing pending) are well bounded, and section 5 now honestly says the trust record is forgeable from an in-sandbox shell (trust is a visible sandbox leaf) and rests the case on bounded harm instead. The earlier Design/GPT/Opus findings on fc2c571 are each answered and fixed in the current head 106433b, and every lane on that head is PASS or skipped, with base main and all checks green. I tried to break the Design lane's open question: on origin/main the only reader of is_recorded_self_arm is GatewayOrchestrator._dashboard_mode_admits in src/kiro_crew/slack/gateway.py, which returns early for any mode other than crew/member, so writing records for ordinary dashboard slots does not change any existing reader. The one leftover gap is a slot whose mode changes later under the same slot_key, which the implementation PR (#18113) should pin. HOLD only because this is an RFC that narrows a human-only gate (reset_conversation admitted from an unattended wake, plus new trust-record writes). A person needs to sign off on that permission change. Otherwise the document is ready. — crew-pr-reviewer |
buluoray
left a comment
There was a problem hiding this comment.
Approve at 106433b5cb043702e2fdc853a689627eda937af0 (docs-only RFC).
I checked the RFC's claims about current code against source: the directive apply path (src/kiro_crew/dashboard/session_directive_apply.py:74,299), the self-arm authorization gate (src/kiro_crew/autonudge_authz.py:1063-1075,1256-1259), the loop mutation path (src/kiro_crew/autonudge_service/mutations.py:1245-1276) and the sandbox reference (src/kiro_crew/sandbox.py:745-755) match what the RFC describes. scripts/docs-lint.sh passes.
Non-blocking follow-up for the implementation PR: the RFC says test/test_autonudge_member_self_arm.py already pins that a wake cannot call reset_conversation. That file only exercises set_project; reset_conversation appears once, in a docstring (:1430). Please add the actual test when this is implemented, and soften the sentence here so a reader does not assume it is already covered.
CI: 85 success, 13 skipped, 0 failing. No tests run (docs-only).
Problem / Motivation
Goal: Record the decision that a goal conductor may drop its own chat context at a round close, from its own patrol wake, so the implementation can land.
The implementation PR changes who may call
reset_conversation. Its First Principles lane reads the decision off the base branch, so the RFC lands first.Why it matters
A conductor that patrols for days re-sends its whole chat on every wake, even when nothing changed. Its state already lives in the work ledger and its session ledger, so the chat is cost with no use.
Not a goal
set_projectorchat_tag.What changed (motivation → approach → change)
Adds
docs/request-for-change/rfc-conductor-round-reset.md(statusin-progress): grantreset_conversationto the conductor agent only, admit it from a wake only when the wake is the slot's own live self-armed loop and nothing is pending, log it at WARNING, and reset only after the ledger write at a round close. Adds its row todocs/request-for-change/README.md. Implementation: #18113.Backwards compatibility
Compatible: documentation only.
Removes nothing: adds one document and one index row.
Tests
N/A — docs only;
scripts/docs-lint.shandtest/test_docs_lint_fact_checks.pypass.Manual verification
N/A — docs only.
Related Issues
Part of #18093
Checklist