Repository navigation
feat(chat): render every warm welcome fact in the gateway - #33032
iamwhatever wants to merge 1 commit into
Conversation
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 screenshot or recording of the version 4 BlockersThis revision has no rendered evidence for the new chat and Board surfaces Notes
[DESIGN-REVIEWED] fc1db03 Description read[DESCRIPTION-READ] 27cb51213fbab1fb24314587bd006eda1ffb18ab516b289158a1ebb1137db65e 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) — 🔴 BLOCKPremise-level review of First-Principles-Verdict: BLOCK The warm welcome's default shape changes for every owner: the model call is removed and the gateway renders the welcome itself. The only record of that decision is RFC section 3.11, and this PR adds that section itself. What this change shipsInventory (10 items) — 9 justifiedIntent: give an owner a welcome made only of facts the gateway checked, so they can act on it without asking what is true. ADDITION.
Blockersproduct-shape change without accepted RFC: the gateway-rendered welcome (item 1) [FIRST-PRINCIPLES-REVIEWED] e89ddc6 Description read[DESCRIPTION-READ] a5e34acfe3e9727a3738b56bb26a0590fea59c0e2cb21546df83684eb0112356 That sha256 names the PR description this verdict read. Recompute it with |
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 and a recording. The PR changes the warm welcome card, the Board's PR facts and the preparing bubble, but attaches no image, so the blind read was not performed. Evidence gaps
[UX-REVIEWED] fc1db03 Description read[DESCRIPTION-READ] 779c0e25f3a761b2a8656ad7afac64b3f363b374fe6db6aabccf8a56f928e768 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 |
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 Review detailsNo findings. All four candidates fail under falsification: the [OPUS-REVIEWED] fc1db03 Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
24467b8 to
8ddc557
Compare
8ddc557 to
9f2f4e3
Compare
9f2f4e3 to
8afc006
Compare
8afc006 to
3b7c398
Compare
589aea0 to
e8d5f6f
Compare
Sensitive change review
|
e8d5f6f to
8807ecd
Compare
Sensitive change review
|
8807ecd to
e89ddc6
Compare
Sensitive change review
|
Sensitive change review
|
The warm welcome's model wrote prose about facts and a phrase check tried to catch it; a truth audit of 71 welcomes found 11 of 58 model sentences false. Gateway code now renders every state, count, tag, time and "since you left" row from its own reads, with read times. The conductor's why, options with outcomes and pick come from its widened goal-board asks record, shown as its words with their time. The model writes at most one fact-free framing line and picture. The Board reads pull-request facts that code recorded: the welcome's live read and the accept_eval tool emit object/observed (producers warm_brief, accept_eval) into a new objects slot fold, which the Ultra Mode Board and goal-board draw with read times. The preparing caption moves into a reply bubble in both chat surfaces. RFC rfc-delayed-task-brief.md section 3.11 records the owner's rulings.
Sensitive change review
|
Sensitive change review
|
|
Closing per owner decision 2026-11-09: the warm welcome restarts from main with a simpler four-part shape (goal, what needs you, what is being done now, what comes after). A new PR will replace this one. |
Problem / Motivation
Goal: The warm welcome and the Board state only facts that gateway code read itself, each with its read time, so an owner can act on a welcome without asking what is true.
A truth audit of the 71 warm welcomes posted 2026-11-02 to 2026-11-07 found the gateway's own pull-request rows right 34 of 34 times, the model's sentences false 11 of 58 times, and sentences copied from worker reports stale 5 of 9 times. The existing value check would have caught 1 of the 11. A UX review of the same welcomes found 2 of 71 an owner could act on within five seconds. The Board showed no pull-request state at all.
Owner-directed shape. The owner directed this as ONE pull request holding both the RFC amendment (
docs/request-for-change/rfc-delayed-task-brief.mdsection 3.11) and all code, on 2026-11-08, relayed by the conductor. Design and First Principles: the RFC change in this PR is the owner's own ruling, not an agent's proposal.Why it matters
The welcome is the first thing an owner reads on return. When one sentence in five is false, the owner has to re-check everything, and the welcome costs time instead of saving it. Owners asked a follow-up question in 32 of 55 replies.
Not a goal
What changed (motivation → approach → change)
Root cause: the model wrote prose about facts and a phrase list tried to catch it. The fix moves every fact to gateway code.
warm_task_brief_welcome.py,warm_task_brief.py).meta.briefversion 4: count line ("Waiting on you: N" / "N 件事等你") plus the first must-do; at most 3 must-do cards fromawaits_person(the Board's own test); tags from the ledger and this welcome's live PR read, bad first; a PR pill only for a PR this welcome read; at most 3 "since you left" rows, bad first, "+N more on the Board"; a Board link. The conductor's why, options with outcome and pick come from its goal-boardasksrecord and show under "Suggested HH:MM" / "建议(HH:MM)" and "Recommended" / "推荐". A merged or closed PR drops every ask about it. No model text reaches the welcome: the owner allowed one framing line "or none", and a word list cannot prove a sentence holds no fact (CI review: "The tests succeeded." passed every gate), soframingandbodystay empty as this PR's default. Owner ruling 2026-11-09: model pictures and the framing line stay out of this PR; RFC 3.11.1 records it. The per-return one-call model call is removed; only the RFC 3.8.1 wake still runs a model turn. Conductor options draw as plain text rows, and every gateway string is redacted on its whole scan views before any cut. The welcome no longer shows a running count; the Board shows each worker's liveness. Every gateway string passes the secret scan views and is redacted on every path. Facts are read again at post time; the retry-on-change loop is gone. Cards are static once posted and each says "checked HH:MM" (owner ruling 2026-11-09); live cards that hide once their item stops waiting are deferred to a follow-up. A model failure costs only the framing line.asksto{what, why, options: [{label, outcome}], pick, pick_why, written_at};written_atcan never be later than the field's write time. The goal-conductor skill says how to write it.object/observedgains one code-only producer,warm_brief(the welcome's live read, once per read, no extra GitHub call). Which PR belongs to an item has one key,ledger_ref(acceptance.pr, then item.pr); the welcome,pr_factsand goal-board (via a stampedpr_url) all use it. PR state, checks, review and mergeable come only from thewarm_briefproducer's newest read. Deferred: refreshing the Board on conductor wakes (recordingaccept_evalverdicts) is cut from this PR, because the evaluator maps an expiredghlogin tofailand the Board would show it as failed checks. RFC 3.11.1 records it; follow-up #33688. A newobjectsslot fold keeps the newest read per subject and producer (64 subjects, 2 KiB facts, row cost measured at 9,144 bytes and charged 9,200)./api/crew-boardaddspr_factsandask_records; goal-board gets aprsfield. Both draw the warm read's facts with its time.PreparingReplyBubble(lazy-loaded, so the App chunk does not grow) puts the caption inside a reply bubble inChatTranscriptViewandChatPane; it folds out in place as the welcome lands (website UX lens 13).board:409reference fixed; crew-log core, emitter and session-types specs updated.RFC: rfc-delayed-task-brief.md is
partialon main; section 3.11 is added by this PR at the owner's direction.Backwards compatibility
Compatible: a stored brief of version 3 or older keeps drawing in its old shape through
WarmBriefCard; a version 1 goal-boardasksentry (options: [string]) still reads as labels; the new producers widen a closed vocabulary and add no required field.Reader: src/kiro_crew/dashboard/warm_task_brief.py:_render -- chat -- replaces the model-written headline/status/lines fields for new welcomes; old rows keep their meta and draw via WarmBriefCard (test_warm_task_brief.py)
Reader: website/src/components/chat/WarmBriefCard.tsx -- chat -- reads version < 4 as before and hands version 4 to WarmWelcomeCard (WarmWelcomeCard.test.tsx)
Reader: src/kiro_crew/dashboard/handlers/work_ledger_board.py:_conductor_asks -- crew page -- still returns option labels per item from the widened record (test_crew_board_asks.py)
Reader: src/kiro_crew/dashboard_templates/builtin/goal-board/template.html:optionsFor -- crew page -- reads both string and {label, outcome} options (goalBoardPage.test.ts, test_dashboard_template_goal-board.py)
Reader: test/test_warm_task_brief_shape.py, test_warm_task_brief_links.py, test_warm_task_brief_asks.py, test_warm_task_brief_unverified.py -- release -- deleted with the model-written shape they tested; their redaction and marker checks moved to test_warm_task_brief_welcome.py
Tests
test/test_warm_task_brief_welcome.py(new): count line and 3-card cap; the conductor record shown with its time and hidden without one; a pick outside the labels dropped; a merged or closed PR dropping its asks; only read PRs named; failed read "not checked since"; changes bad first with "+N";objectsfold bounds, producer check and out-of-order reads; thewarm_briefproducer once per read; token, PEM and marker redaction; a secret in a gateway string redacted on every path; a remote worker's binding kept in the snapshot.test/test_crew_log_fold_eager_memory.py: theobjectsrow measured at its clamps in 4-byte text.test/test_warm_task_brief.py,_redesign.py,_wake.py,_crew_mode.py,_facts.py: adapted to the version 4 shape; race-redo tests removed with the race loop.WarmWelcomeCard.test.tsx,prFactTags.test.ts,goalBoardPage.test.ts(the warm read only, a read from another day dated),ChatPane.preparing.test.tsxandChatPage.views.test.tsx(caption inside a reply bubble).Proof: not provable -- other: feature change, not a fix; the tests pin new behaviour rather than a reverted defect
Manual verification
Captured with
website/scripts/capture-warm-welcome-gateway.mjsagainst the capture harness at 1280 and 320 px: the preparing bubble, then the welcome landing in its place; every frame asserted. Not run against a live gateway.Screenshots / video
Captured on head e8d5f6f (
website/scripts/capture-warm-welcome-gateway.mjs,website/scripts/capture-crew-pr-facts.mjs; every frame asserted).Welcome: preparing bubble folding into the welcome
Ultra Mode Board "Needs your input" ask (left) and Crew Board page row (right): facts named by the PR they were read for (acme/widgets#7, not the worker's #9), tags with read time, neutral labels. Captured before the checker line was removed: the dim "checker HH:MM" line in these frames is no longer drawn.
320 / 360 px
Related Issues
Refs #25306. Follow-up: #33688 (Board refresh on conductor wakes, deferred). Fixes #32844 (the preparing hint now sits in the reply bubble). Supersedes #30767 (its survivors are folded in; that PR is not pushed to).
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; see the Licensing section of CONTRIBUTING.md.