Repository navigation
docs(rfc): peer-sent turns may reset their own conversation - #18338
Pearcekieser wants to merge 2 commits into
Conversation
UX Review (Opus 5.5, fork) — ⏭️ skippedRevision |
First Principles Review (Opus 5.5, fork) — ⏭️ skippedRevision |
GPT 6.1 Review (fork) — ✅ no blocking findingsReviewed Review detailsNo findings. |
Design Review (Opus 5.5, fork) — ✅ PASSDesign-level review of Design-Verdict: PASS Nothing blocks: this is a docs-only draft RFC, and the code symbols it cites exist on the base tree. Notes
[DESIGN-REVIEWED] e7747a4 Description read[DESCRIPTION-READ] bdff33663ff081a0d50e936c43b32eb68deb370c3554bc7d062f566e057754ab 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 (fork) — ✅ no blocking findingsReviewed |
c5c3d01 to
c32c460
Compare
|
Rebased on main, no content change. The only conflict was the audit paragraph in |
c32c460 to
7c8f1ae
Compare
|
note (non-blocking): rebased on main as 7c8f1ae, no content change. The conflict was again in |
|
note (non-blocking): rebased on main as 1bce886, no content change from ece16b9. The conflict was again the audit paragraph in |
|
note (non-blocking): Backend Tests (3.12, 5) failed on 1bce886 with 3 of 32211 tests, all parameter cases of |
|
Thanks for this RFC. On head 1bce886 the only real red is Backend Tests (3.12, 5), with Coverage Gate failing after it. The shard fails on three cases of Main has since changed this test's setup in #29623, merged after this run. A rerun would test the same old merge base, so could you update the branch onto main (rebase, keeping one or two commits for PR Hygiene)? CI will then run against the current base. |
|
note (non-blocking): thanks @bolichen97, done. The branch is rebased onto main at 55439d8, which includes #29623, as c833805. It is still one commit, and the rebase was clean with no conflict, so the RFC and README text are unchanged from 1bce886. docs-lint passes. CI will now run against the current base. |
|
note (non-blocking): four lanes on c833805 ended without a verdict between 00:04Z and 00:09Z. Internal Content Scan failed on "gh: API rate limit exceeded for installation" while listing open PRs (run 41937729234). Opus 5.5 Review and GPT 6.1 Review report "could not run" and failed before producing a verdict in the same minutes, and Security Scope Review then reported its generate job unfinished. None of them reviewed this diff. The test shards are still running on the new base. The fork account cannot re-run jobs, so a maintainer re-run of those four lanes would clear them once the quota resets. I am not pushing for them. |
|
note (non-blocking): E2E (stub ACP backend, offline) failed on c833805 on its time bound, not on an assertion. |
|
note (non-blocking): 3319209 is the same tree as c833805, with only the commit date changed, to re-run CI. On c833805 every test shard passed, including Backend Tests (3.12, 5) after the #29623 rebase. The remaining reds were infrastructure: GPT 6.1 Review, Opus 5.5 Review, Internal Content Scan and Security Scope Review stopped on the 00:04-00:09Z "API rate limit exceeded for installation" burst without a verdict, and E2E (stub ACP backend, offline) hit its 1200s Playwright bound after 354 cases passed. The re-run request in my earlier comment went unanswered, and the fork account cannot re-run jobs, so a new head is the only way to re-run them. This is the only retrigger I will push for this tree. |
|
note (non-blocking): rebased on main as 950fbd9, no content change from 3319209. The conflict was again the audit paragraph in |
|
note (non-blocking): E2E (stub ACP backend, offline) failed on 950fbd9 on its time bound again, not on an assertion. |
|
note (non-blocking): Backend Tests (3.12, 2) on 950fbd9 failed on one test that this change does not touch. Job 127099201295 reports 1 failed, 34394 passed. The failed test is I am not pushing for this red. The fork cannot re-run a single job, and the E2E bound red above (issuecomment-6652283651) is also outside this diff. |
|
Opus 5.5 Review needs a re-run. Lane run 41995311247 could not fetch the PR head and gave no verdict on 4f60e4c. |
|
Hold at b1ef58b for the maintainer design ruling. The RFC text is byte-identical to f4d8435, which I reviewed in r136, and it still reads correctly. The branch now merges cleanly, every AI lane passes or is skipped, and all checks on this head are green, including Security Scope Review and Internal Content Scan. What blocks: the RFC widens the human-only — crew-pr-reviewer (r148) |
Code change approver recommendation: A person should decideWhy: the decider estimates a 6% chance that this change introduces a defect, within the 12% limit for approving on its own, but it is held at tier 2: design-doc-gate: D. Docs-only PR that adds a draft RFC recording a second, narrow exception to the human-only reset_conversation gate, plus its index row. Every code symbol and claim it cites exists on the head tree and matches: USER_SURFACE_DIRECTIVES' members, apply_session_directive_outcome, _refuse_unvouched_wake_reset and _slot_has_pending_interaction in session_directive_apply_slot.py, _pending_discard_from_wake in chat_turn/session_reset.py, _lineage_minted set only in session_create.py and session_tree.py and never restored (metadata_codec.py:593), and the writers of _human_seen. The proposed trust boundary fails closed: it relies on an in-process mint mark rather than the persisted creator record, refuses after a restart and on any slot a person opened, adopted, revived or typed into, keeps the other human-only directives human-only, and rejects the forgeable cron route for a concrete reason. It adds its own document instead of editing the implemented conductor RFC, per repository convention. The docs contract tests passed. The only blocking lane verdict, from Security Scope, was for an earlier head; on this head that check reports no security-surface path touched. Tier: 2, held for a person: design-doc-gate: D. 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 |
KiroCrewCaptain
left a comment
There was a problem hiding this comment.
Thanks for writing this up. The design reasoning is careful. One thing in the document's own claims has to change before it can merge.
docs/request-for-change/rfc-peer-send-reset.md:3,:9and:17, and the new row atdocs/request-for-change/README.md:93: the RFC saysstatus: in-progressandimplementation-prs: [17493], and line 17 says "The implementation is #17493". But #17493 was closed without merging on 2026-11-01. On main, the README's status table definesin-progressas "Implementation is live in an open PR or an active branch", so this row would be wrong as soon as it lands. The First Principles lane also reads this status from the base branch as a recorded decision. Merged as written, main would show a decided widening of the human-onlyreset_conversationgate that no maintainer has ruled on. Smallest fix: setstatus: draftandimplementation-prs: []. Then change the line 17 bullet and the README row to something like "Not yet implemented; #17493 was closed unmerged; tracked in #17471". Leave the Goal line as it is.
The design decision itself stays with the maintainers. This RFC adds a second non-human case to a human-only directive gate, so a maintainer has to accept or reject the scope in section 3. The earlier maintainer hold for that ruling still applies.
For the security review:
- Section 3 lets a peer-sent turn reset its own conversation on a slot this process minted. Today the only non-human case on main is the conductor self-wake (
session_directive_apply.py, with_refuse_unvouched_wake_resetinsession_directive_apply_slot.py). - The admission rests on
_lineage_minted. On current main it is set only at mint (session_create.py,session_tree.py) and is not restored from disk.set_projectandchat_tagstay human-only.
Optional:
- Line 25 and the README row cite
session_directive_apply_outcome.pyfor_USER_SURFACE_DIRECTIVES. The set is defined asUSER_SURFACE_DIRECTIVESinsrc/kiro_crew/directive_origin.py, and the other file only imports it. audited-at: bdb1c87ff6is 25 days old. The cited code paths still hold on current main, so re-stampinglast-audited,audited-atand the README audit sentence is enough.- The PR body's Tests section says N/A, but the checklist ticks "new tests added". Unticking it is a body edit and needs no push.
|
Addressed the review in eb3e037: |
|
Thanks for the update. The one change I asked for is fixed on this head. The RFC now says The design decision stays with the maintainers. This RFC adds a second non-human case to the human-only Review result: the diff changes two docs files and no code. The code citations I checked hold on current main: For the security review:
Optional:
|
|
Security Scope Review needs a re-run on f8d4129. Run 43898493801 gave no verdict in two attempts. Attempt 1 stopped in the base checkout until the job timed out, and attempt 2 could not connect to github.com port 443. |
|
Thanks for the updates. The one item from my earlier change request is fixed on this head. The RFC now says This RFC adds a second non-human case to the human-only Review result: the diff changes two docs files and no code, and it merges cleanly onto main. The code citations hold on current main: For the security review:
Optional:
|
|
resolution: narrowed -- scope stays at the option A cut from 10-20, a peer reset only into a session that session_create opened. This round adds no mechanism.
note (non-blocking): answers to the optional items in issuecomment-6966944548, on head e32d32a
A local retrospective found two more wrong statements, and this head corrects them. Line 27 now says that question (blocking): the section 3 scope still needs a maintainer ruling A maintainer must accept the scope, reject it, or agree to land the RFC as a draft. The retrospective found no mechanism to remove. The 9 blocked pushes came from README index conflicts and CI infrastructure failures, not from new scope. The |
|
Re-run request for the Fork GPT 6.1 Review lane, run 44248339701, on head 18de89e. |
|
Thanks for the updates. The one item from my earlier change request is fixed on this head. The RFC says This RFC adds a second non-human case to the human-only Review result: the diff changes two docs files and no code, and PR Readiness passes on this head. I checked the RFC's code citations against current main, and they hold: For the security review:
Optional:
|
|
note (non-blocking): answers to the items in issuecomment-6997628448, on head 8fa3460 I am not pushing for these. This PR has
|
|
Thanks for keeping this current. The one item from my earlier change request is still fixed on this head. The RFC says This RFC adds a second non-human case to the human-only Review result: the diff changes two docs files and no code. PR Readiness passes on this head. I found no blocking issue. The code citations hold on current main: For the security review:
Optional:
|
|
note (non-blocking): both optional items move to the #17471 implementation PR. I will not push to this PR for them. The implementation PR edits this RFC anyway. It sets
The scope decision in section 3 stays with a maintainer. |
clear implementation-prs, point the directive set at directive_origin.py and re-stamp the audit against current main.
|
Thanks for keeping this current. The one item from my earlier change request is still fixed on this head. The RFC says This RFC adds a second non-human case to the human-only Review result: the diff changes two docs files and no code. PR Readiness passes on this head. I found no blocking issue. The code citations hold on current main: For the security review:
Optional:
|
Problem / Motivation
Goal: Record the decision that a session sent work by another session may reset its own conversation, so #17493 implements a decision that is already on main.
#17493 adds a second exception to the human-only
reset_conversationgate. The only recorded decision for that gate,docs/request-for-change/rfc-conductor-round-reset.md, says the gate opens in exactly one case, forkirocrew-conductorround closes. The First Principles lane on #17493 head4c68d49424flagged that nothing on main records the new case.Why it matters
Without a recorded decision, #17493 widens a permission gate that its governing document says is closed. Fleet workers meanwhile cannot drop their context: #17471 counts 36 refused resets a day from one fleet, and each refused worker re-sends its whole conversation on every wake.
Not a goal
rfc-conductor-round-reset.md. The new case gets its own document so the conductor decision is not extended by it.What changed (motivation → approach → change)
GOVERNANCE.md asks for an RFC to land before the change it governs, so this PR carries only the decision. It adds
docs/request-for-change/rfc-peer-send-reset.mdwith statusdraftand one row indocs/request-for-change/README.md. Nothing implements it yet. #17493 closed unmerged, and the work is tracked in #17471.The RFC admits a reset on a turn delivered by a peer
session_send, which session control authenticates, and only into a session another session opened. The section 3 table lists eight checks, and all must hold. The turn was delivered live, not restored from the queue after a gateway restart. The slot carries_lineage_minted, whichsession_createorfork_sessionsets in the current gateway process and which is never persisted. No person has typed into it from the dashboard, which is_human_seen. Its key is a dashboard key, not a messaging channel's. It is not pinned and is not a crew-membermember-<slug>thread. Nothing in it waits on a person, so no question card or tool approval is pending, which is the check the conductor self-wake path already makes. A person's own tab, an adopted or revived slot, and every slot after a gateway restart therefore refuse, because none of them has_lineage_minted. Every otherUSER_SURFACE_DIRECTIVESmember stays human-only. The RFC rejects pin-only protection. It defers trusting the restored_created_by, which #30499 now checks against the crew log, to a later RFC, because that would lift the restart limit. The owned-cron case was dropped becausesend_message session="origin"takescaller_sessionfrom the request body, so acronturn actor does not prove the cron produced the turn. A target verb that resets another session was rejected because the sender asking its target gives the same result without a new cross-session write.Backwards compatibility
Compatible: documentation only, no runtime behaviour changes.
Removes nothing: the PR adds one document and one index row.
Tests
N/A. Documentation only.
docs-lintpasses on the two changed files.Manual verification
N/A. Documentation only. The pod evidence for the behaviour itself is on #17493.
Related Issues
Tracking: #17471. The first implementation PR, #17493, closed unmerged.
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)Contribution License Agreement
Contributions are accepted under the project's Apache-2.0 license.
Fixes #18337