Repository navigation
feat: list the steering files each chat receives under Context breakdown - #33600
Conversation
GPT 6.1 Review — ✅ no blocking findings (all downgraded on adjudication)GPT 6.1 flagged blocking issues on This comment is updated in place on each push. Review detailsBLOCKING -- src/kiro_crew/dashboard/handlers/chat_steering.py:120 -- Body text can falsely mark an omitted guide as included
Adjudication (Opus 5.5) — is blocking on each finding proportionate?F1: LOW harm. The defect is a wrong label in the read-only Context breakdown. Nothing is lost, corrupted or exposed, and the model's actual context is not changed. It needs all of these: a non-member chat ( The real fix is to change False positive or not applicable? A repository writer can comment: |
15afad1 to
524ff55
Compare
UX Review (Opus 5.5) — ✅ PASS (a BLOCK with no Trigger and no Outcome does not block)UX-level review of UX-Verdict: PASS cannot evaluate: missing screenshots of the new "Steering in this chat" list. The description's screenshot table shows "[image removed]" in all four cells, so no image was admitted and the blind read was not performed. Evidence gaps
Notes
[UX-REVIEWED] 3433344 Description read[DESCRIPTION-READ] 9de70a92cea3e4b086ae16727d19df96788c89121603ad819f681af72d7a717b That sha256 names everything this verdict read: the PR description, plus the 3 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 |
Design Review (Opus 5.5) — ✅ PASS (a BLOCK with no Trigger and no Outcome does not block)Design-level review of Design-Verdict: PASS cannot evaluate: missing a screenshot of the "Steering in this chat" list under Context breakdown. The description's four screenshot cells say "[image removed]", and the evidence list shows 0 images or recordings downloaded or committed. BlockersNo rendered evidence for a new dashboard panel Notes
[DESIGN-REVIEWED] 3433344 Description read[DESCRIPTION-READ] 2d0edce1a29cbcd73eee19a4517c82da9690c4ffdc557df1aa2663cd5a76d678 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 One note: the new backend-to-label helper is a second copy of mapping logic that Notes
What this change shipsInventory (8 items) — 7 justifiedIntent: let a person see which steering files reach a chat, and why a file does not. ADDITION (P1 of
[FIRST-PRINCIPLES-REVIEWED] 3433344 Description read[DESCRIPTION-READ] 16fb64787a5c644e448389679c55c7c9ddf7bdd222e26188da9dba03f41248d2 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: |
524ff55 to
da2b6b5
Compare
|
Review round 1, fixed on da2b6b5:
Downgraded and left as is: F1 (folder cap) and F3 (shadowed project agent spec) are marked shortcuts in the code; F5 (stale list after a project change) clears on the next mount. |
da2b6b5 to
280de71
Compare
|
Review round 2, fixed on 280de71:
Sensitive change review
|
280de71 to
2b216fa
Compare
|
Rebased onto main for the conflict; no code change since the last note. Head 2b216fa. Sensitive change review
|
2b216fa to
4717164
Compare
|
Fixed two test reds on 4717164: the route now imports through the dashboard _shared facade, and the project agent spec path comes from config.paths.project_agents_dir. The new agent-spec read site is pinned in test_agent_spec_hardened_reads.py. The other reds on the previous head were GitHub rate limits and 504s. Sensitive change review
|
4717164 to
7c58ac6
Compare
|
Review round 3, on 7c58ac6. Two findings came back a second time, so the approach changed instead of patching another case:
Sensitive change review
|
7c58ac6 to
af98d83
Compare
|
On af98d83: the new folder-cap test stubs the folder collection, so it pins the section-cap verdicts and not the collector (it failed on the Windows stability job). The kirocrew-worktree-dev skill-size red in backend shard 3 also fails on main; #33795 fixes it. Sensitive change review
|
af98d83 to
c2c6f45
Compare
|
Review round 4, on c2c6f45:
Sensitive change review
|
A person cannot see which steering files reach the chat they are in; on several harnesses the global and project trees do not reach the model at all. The Context breakdown tab now lists every global, project and folder steering file with its source, when it applies, and whether it reaches this chat, and why not. The verdicts come from one plan that applies the gates context assembly already applies, and a parity test pins the Claude Code verdicts to what the loader sends. Read-only: nothing that reaches the model changes. Implements P1 of the accepted RFC (#33133). Part of #33112.
c2c6f45 to
3433344
Compare
|
Review round 5, on 3433344. The three upheld GPT findings are fixed; each new test fails on the previous head:
F2 and F5 were downgraded by the adjudicator and are left as they are. Sensitive change review
|
bolichen97
left a comment
There was a problem hiding this comment.
Approve on 3433344. I read the whole diff (2066 lines). This is P1 of the accepted RFC from #33133.
steering_resource_documentsnow holds the loader body, and_load_steering_resourcesjoins its output. The list and the Claude Code load therefore read one source.test_claude_code_verdicts_match_what_its_loader_sendsand the AGENTS.md and link tests pin that parity.- Folder rows come from the same
_folder_steering_delivery_snapshotskip and the samerender_folder_steeringatcaps.steeringthat the turn uses, sofolder_capreflects the real cut. declared_steering_targetsreads the spec through_read_agent_specand is registered intest_agent_spec_hardened_reads.py. For a pin or a project spec it returnsNone, and the row shows "cannot tell" instead of guessing.- The route is read-only. It has the same
slot_not_found/deny_app_slot_accesschecks asselection-capabilitiesand sits behind the same owner check on the chat routes. Paths go through_redact_meta, and no body leaves the route. The redaction sink is registered.test_an_app_cannot_read_a_slot_it_does_not_owncovers the app path. - The UI is lazy-loaded so the App chunk does not grow, and the strings are added to every locale.
- All 97 checks are green, and nobody else has an open change request. Auto-merge is already set by RohanK6.
Problem / Motivation
Goal: A person can see, on a chat's Context breakdown tab, which steering files reach that chat and why any one does not.
Implements P1 of the accepted RFC (#33133). Today the Context breakdown shows one steering token total and nothing more. A person cannot tell which of their global, project or folder steering files the agent actually got, and on several harnesses the global and project trees do not reach the model at all (#29940, #30158).
Why it matters
Steering is how a person tells the agent their rules. When a file silently does not reach a chat, the agent ignores those rules and nothing on screen says why. derreckFromSlalom (#29940), gwatsonk (#30158) and calebpollman (#31952) all hit this.
Not a goal
.kiro/steering andAGENTS.mdfrom the enclosing repository root when a session's folder is a subdirectory #31952's ask; the RFC keeps it out of scope).What changed (motivation → approach → change)
src/kiro_crew/steering_plan.py, applies the harness gates and, where Crew writes the text, reads the writer itself: the Claude Code load now has one reader,steering_resource_documents(context_assembly/agent_spec.py), whose list_load_steering_resourcesjoins, and folder rows come from the bounded folder section rendered at the turn's cap.GET /api/chat/slots/{slot}/steering(src/kiro_crew/dashboard/handlers/chat_steering.py) answers it, owner-gated like the other slot routes, and is registered as a redaction sink (security_posture_sinks_gateway.py): its display paths go through the Steering tab's_redact_meta. A cold chat's harness label comes from a newprovider_label_for_backendinagent_sdk/drivers/acp.py, so the route does not import the ACP layer.website/src/pages/ChatSteeringList.tsxrenders "Steering in this chat" under the Context breakdown tab: name, source, when it applies, and whether it reaches this chat with the reason when not. Strings in all 12 locales.docs/architecture/context-management.mdnames the plan;docs/feature-map/README.mdrow; the RFC moves topartialwith this PR inimplementation-prs, and itsreachesvocabulary gainsbriefing(a member chat's essentials path decides per guide).Backwards compatibility
Compatible: a new read-only route and a new list; no existing route, key or behaviour changes.
Removes nothing: the diff only adds a route, a module, a component and catalog keys.
Tests
test/test_chat_steering_plan.py(21): each gate's verdict, the kiro-cli agent-resources case, a project agent spec that kiro-cli serves instead, Codex global files, owner gating, a channel-linked chat listing its own project, a Claude Code project link the Steering tab does not read, a projectAGENTS.mdthe load reads, a folder document the section cap leaves out, and a parity test that pins the Claude Code verdicts to what_load_steering_resourcesreally sends. They fail on main: the route and module do not exist there.test/test_security_posture.py: the new route is a registered redaction sink, and_redact_metacounts as both scanners.website/src/test/ChatSteeringList.test.tsx(4): the list renders each verdict, the empty state, a labelled section with one list item per file, and a project change re-reads the list. Fails on main: the component does not exist there.Manual verification
Built the SPA from this branch and from main and opened a chat's Context tab against stubbed API answers (a two-turn context trace and three steering files), in en and de (screenshots below). Before: the breakdown alone. After: the same breakdown with "Steering in this chat" under it. zh-CN is checked by the i18n gates only: the capture host has no CJK font.
Screenshots / video
Related Issues
Part of #33112. Implements P1 of the accepted RFC (#33133). Credits #29940 (derreckFromSlalom), #30158 (gwatsonk), #31952 (calebpollman).
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)Contribution License Agreement
Contributed under the project's Apache-2.0 license.