Skip to content

feat: add return-triggered warm task briefs - #20422

Merged
dodgeblaster merged 1 commit into
mainfrom
feat/20156-return-warm-task-brief
Oct 16, 2026
Merged

dodgeblaster merged 1 commit into
mainfrom
feat/20156-return-warm-task-brief

Conversation

@iamwhatever

@iamwhatever iamwhatever commented Oct 14, 2026 •

Copy link
Copy Markdown
Contributor

Problem / Motivation

Goal: Replace the immediate static warm resume with a grounded task brief shown only after a real return.

Opening an active crewmate thread currently shows fixed task status at once. It does not start with the request that caused the work, and it can race a newer answer.

Why it matters

People returning after a long break can recover the request, durable task truth, and next actions without rereading the thread. Unrelated opens and short absences spend nothing.

Not a goal

  • Changing the new-thread or six-hour cold welcome.
  • Treating visible idle time, window blur, socket loss, reconnect, or elapsed time alone as a return.
  • Adding a route, listener, model provider, data destination, or retry loop.
  • Merging this PR or either related refactor automatically.

What changed (motivation → approach → change)

After at least 30 minutes away, a person returning to the same active crewmate thread gets a short catch-up from the crewmate in its own voice: what they asked for, where it stands, and one to three suggestions. This replaces fixed copy so they recover the task as it stands now, while ordinary opens and short absences stay quiet and free.

The trigger design is RFC #20157. Its plain-message amendment is RFC PR #20778 (merged). #20979 records that posted messages stay as plain history and must merge before this implementation. The client marks only an explicit document.hidden leave and reports the mounted crewmate thread on the later visible edge. The gateway keeps the away episode in memory on a suspend-counting wall clock, refuses a backward clock step, requires the same thread, then durably claims a task fingerprint before showing status or spending. Any later owner focus frame, including another connection, consumes the process-local away episode. That owner-only claim is an empty display-only row in the existing transcript, so reset and deletion remove it with the task.

The gateway first posts a short fixed crewmate opener: “Welcome back, let me catch you up on where we are.” It is picked by the script of the user's request (English, Chinese, Japanese, Korean, Russian, Hindi, Bengali; Latin script gets English) and costs no model call. It then makes exactly one bounded, tool-free call through the existing governed background model path. The result is a second ordinary crewmate message in first-person prose: what was asked, where it stands, and one to three concrete suggestions. The prompt forbids headings, bold, lists and task-title dumps, and the gateway strips bold markers. There is no card, header, chip or Dismiss button; both rows draw in the normal assistant bubble with no special renderer. Generated messages are excluded from the task fingerprint, session summaries, consolidation prompts and model replay. Posted messages are ordinary history and stay; nothing retracts or rewrites them. User, assistant, surface, task, reset, close, and replacement changes cancel the attempt or drop its stale result unposted. Once model acquisition starts, cancellation lets the governed call finish safely, then discards stale output. Failure is a second short fixed crewmate line (“Sorry, I couldn't pull that together just now. Ask me where things stand.”) in the same language as the opener. The RFC gets a 2026-10-16 UX amendment line recording this shape.

The static MateResumeCard and its warm-only copy were removed. The board read remains only to preserve the cold-welcome gate. User data stays inside the existing model and account path.

Implementation spans src/kiro_crew/autonudge_judge.py, src/kiro_crew/context.py, src/kiro_crew/decisions, src/kiro_crew/history_consolidation.py, src/kiro_crew/history_projection.py, src/kiro_crew/warm_task_brief_contract.py, src/kiro_crew/dashboard, src/kiro_crew/session_summary.py, src/kiro_crew/security_posture.py, website/src/app-sdk, website/src/hooks, website/src/lib, and website/src/store. The security posture registry counts the final brief scrub as a transcript-and-browser egress boundary. Text and render coverage live in website/src/i18n and website/src/test. The owning contracts are updated under docs/system-specs/modules, with the user path indexed in docs/feature-map; the earlier warm-card RFC is marked partially superseded while its cold-welcome decision stays current.

Backwards compatibility

Breaking: opening an active crewmate thread no longer shows the immediate static warm card. RFC #20157 records the accepted replacement. Writer sweep on origin/main @ b413800: the only warm greeting producer was useMateGreeting; the only component consumer was MembersPage (now MembersPageThread); the shared crewmateGreeting host remains for the unchanged cold welcome.

Undo: reverting the feature leaves completed rows as ordinary crewmate messages; empty claim rows draw nothing. Existing reset and delete paths remove both.

Reader: website/src/pages/members/MembersPageThread.tsx:useMateGreeting -- crew page -- still reads the board only to decide whether the unchanged cold recap may run.
Reader: website/src/components/ChatPane.tsx:crewmateGreeting -- crew page -- still draws MateWelcomeCard for the unchanged cold path.
Reader: website/src/components/chat/crewmateBubbles.ts:isCrewmateChatRow -- crew page -- draws generated opener, result, and failure rows as ordinary crewmate messages; the fingerprint claim is a notice row, which crewmate chat never draws.
Reader: src/kiro_crew/dashboard/ws.py:note_foreground -- slot retention -- consumes the focus frame's mounted member slot on the Crew page.

Tests

  • Backend tests cover the opener language pick, prose-only result parsing, and visible, unfocused, hidden, short, wrong-thread, duplicate-tab, socket-cleanup, restart, terminal, privacy, claim, failure, cancellation, ordering, and stale-result cases.
  • Frontend tests cover mounted member presence, hidden and reconnect frames, non-chat and placeholder frames, ordinary crewmate opener/result rows, claim hiding, and no immediate warm card.
  • Existing cold welcome, slot retention, eager spawn, reconnect, ChatPane, type, lint, i18n, and Electron visibility checks pass.
  • Determinism proof: focused backend and frontend groups pass repeatedly and under shuffled test order; no test depends on wall-clock waiting.

Manual verification

The Electron source check confirms the dashboard keeps Chromium's default background throttling. Minimize and hide-to-tray smoke could not run on this headless Linux host; Windows and macOS desktop behavior remains release verification. Per RFC #20157, a platform that reports no document.hidden edge fails silent and adds no fallback trigger.

Screenshots / video

Committed under temp-screenshots/20422-warm-brief-plain/, captured from the real Members page capture entry with a stubbed API. Each frame asserts no "Where we left off" text, no Dismiss button and no bold section headings. The side panel's "dashboard could not be loaded" line is the capture stub, not this change.

Ordinary open of an active thread: no card, no Dismiss:

Ordinary open of an active thread: no card, no Dismiss

While the one model call runs: the opener stands alone as an ordinary message. By design there is no spinner, typing bubble or status row; the opener itself says the crewmate is catching the user up, and the brief call is not a chat turn:

Opener alone while the model call runs

Return: opener then prose catch-up, light:

Return: opener then prose catch-up, light

Return: opener then prose catch-up, dark:

Return: opener then prose catch-up, dark

Chinese request: Chinese opener and catch-up:

Chinese request: Chinese opener and catch-up

Model failure: opener then the fixed failure line:

Model failure: opener then the fixed failure line

Related Issues

Part of #20156

Implements the accepted RFC in #20157.

Related open PRs

Checklist

  • At most two commits, with one Conventional Commit
  • Existing focused tests pass and new tests cover the feature
  • Self-review completed; code follows project style guidelines
  • Documentation updated
  • No secrets, credentials, or internal references in the diff

Contribution License Agreement

Contributions are accepted under the project's Apache-2.0 license.

@iamwhatever
iamwhatever requested review from a team and KiroCrewCaptain as code owners October 14, 2026 18:41
@github-actions github-actions Bot added the readiness: checking Automated validation is still running label Oct 14, 2026
@github-actions

github-actions Bot commented Oct 14, 2026 •

Copy link
Copy Markdown
Contributor

Design Review (Opus 5.5) — 🟡 CONCERNS

Design-level review of 3d28ced7c3a238a10b4e5702f1655b742f440a26 — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

Design-Verdict: CONCERNS

The crewmate never sees the catch-up it posts, so a reply to one of its suggestions ("do the second one") reaches a model with no record of them.

Watch

  • The brief is cut off from the conversation it lives in. The brief is generated outside the chat turn by a separate background model call (run_bg_oneliner). It is then removed from replay and recall (messages = _without_warm_task_brief_messages(messages) in context.py), and it is never sent to the live ACP session. The prompt still asks for "one to three concrete things they could do next", and the description says posted messages "stay as plain history". So the user sees an ordinary crewmate bubble offering next steps and answers it, but the crewmate never saw that bubble. It will ask what the user means or guess. Leaving it out of the fingerprint and summaries makes sense. Leaving it out of the next turn's context undoes the purpose of the feature, which is to make resuming easy.
    Clears when: the next real user turn after a brief includes the posted result text (for example, as a marked injected preamble or by keeping RESULT_KIND in replay), with a test that a reply to a suggestion gets that text in its prompt; or the prompt stops asking for suggestions.

[DESIGN-REVIEWED] 3d28ced

Description read

[DESCRIPTION-READ] 7caa5f8a068d28e19661eecf610c657867d2c55b1b2981cc0588d8cf55df6bf9

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 .github/scripts/pr-description-capture.sh over the same inputs; if it differs, the description or that evidence changed after this verdict was formed and any finding drawn from either is unproven.

@github-actions

github-actions Bot commented Oct 14, 2026 •

Copy link
Copy Markdown
Contributor

UX Review (Opus 5.5) — 🟡 CONCERNS

UX-level review of 3d28ced7c3a238a10b4e5702f1655b742f440a26 — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

UX-Verdict: CONCERNS

The opener promises a catch-up but nothing shows it is coming, and a reply typed while waiting silently cancels it.

Watch

  • Pending state looks stuck. The fixed opener posts first, then nothing shows until the single model call returns. The PR says "no spinner, typing bubble or status row" on purpose. On shot-06 the blind reader said: "Nothing shows it is working: no dots, no spinner -> no idea whether to wait or type… I can't tell if it is stuck." This happens on every return, and the call can take seconds, so the user's first moment back is spent unsure whether to wait. Smallest fix: show the existing assistant typing indicator under the opener while the call runs.
  • Typing while waiting cancels the brief. Per the PR, "User, assistant… changes cancel the attempt or drop its stale result unposted." A user who types during the wait gets no catch-up and no failure line. The opener stays in the chat history saying "let me catch you up on where we are." with nothing after it. That is a broken promise left on screen permanently. It hits the impatient users the pending gap above creates. Fix: when the attempt is cancelled, post the fixed failure line, or skip the opener if no result comes.

Suggestions

  • Failure line "Sorry, I couldn't pull that together just now. Ask me where things stand.": the blind reader pointed out that the dashboard error right next to it has a "Try again" button and this one doesn't. Keep the line, but have "Ask me where things stand" match a phrase the crewmate reliably answers with a status, so the suggested reply works.

[UX-REVIEWED] 3d28ced

Description read

[DESCRIPTION-READ] 921a7e6524f68dd3f492f59c9628d725a6fa1bbe2899758d78ac9eb9a7cf33e0

That sha256 names everything this verdict read: the PR description, plus the 9 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 .github/scripts/pr-description-capture.sh over the same inputs; if it differs, the description or that evidence changed after this verdict was formed and any finding drawn from either is unproven.

@github-actions

github-actions Bot commented Oct 14, 2026 •

Copy link
Copy Markdown
Contributor

First Principles Review (Opus 5.5) — 🟡 CONCERNS

Premise-level review of 3d28ced7c3a238a10b4e5702f1655b742f440a26 — why this exists and whether the shipped surface is the smallest honest version. Updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

First-Principles-Verdict: CONCERNS

The diff overrides the accepted RFC's verbatim-quote and English-fixed-line rules, with a 7-language opener table and an amendment that exists only in this PR.

Not justified as shipped

  • Item 3 — inherited: base RFC §3.5 says "The gateway… supplies the exact source quote itself… The model never chooses a different origin". The diff's own amendment "supersedes the three-part shape and gateway-supplied quote", so the model now paraphrases what the user asked. The amendment is proposed here, not recorded on base.
    Subtraction: drop the in-diff 2026-10-16 amendment and keep the gateway-supplied verbatim request in the brief (an existing base-RFC rule).
    Clears when: a maintainer-recorded decision exists on base (a merged amendment, then push or rebase), or /ai-review override first-principles 3d28ced7c3a238a10b4e5702f1655b742f440a26: <decision>.
  • Item 4 — inherited: base RFC says the fixed lines "are English, like other gateway-authored lines". _language_of plus two 7-entry tables in warm_task_brief.py add a second, server-side localization mechanism outside the i18n catalog. It covers 7 of the dashboard's 13 locales, so de/es/fr/it/pt still get English.
    Subtraction: delete _language_of and the non-en entries of _OPENER_TEXT/_FAILURE_TEXT.
    Clears when: the tables are deleted, or a recorded decision covers script-picked fixed lines.

What this change ships

Inventory (8 items) — 6 justified

Intent: let a person returning after 30+ minutes recover the task without rereading the thread. ADDITION.

  1. Opening an active crewmate thread no longer shows the static "Where we left off" card — justified
  2. A hidden-then-return after 30 min on the same thread posts an opener and one model-written catch-up — justified
  3. Catch-up is model prose with no verbatim request quote or headings — inherited (contradicts base RFC §3.5)
  4. Opener and failure lines are picked by the request's script, in 7 languages — inherited
  5. Mounted Crewmates thread now drives slot_focused (and its retention) — justified
  6. Hidden frame carries a new hidden: true flag — justified
  7. Generated rows are excluded from summaries, replay, consolidation, compaction, nudge and transfer — justified
  8. Empty claim row persisted per task fingerprint, hidden by the renderer — justified

[FIRST-PRINCIPLES-REVIEWED] 3d28ced

Description read

[DESCRIPTION-READ] 7bfb07ddb944bd5646aeab8c8840fd361c370bba82d52b1d1cc18e8427b9ae99

That sha256 names the PR description this verdict read. Recompute it with .github/scripts/pr-description-capture.sh; if it differs, the description changed after this verdict was formed and any finding drawn from the description is unproven.

@github-actions

github-actions Bot commented Oct 14, 2026 •

Copy link
Copy Markdown
Contributor

Opus 5.5 Review — ✅ no blocking findings

Reviewed 3d28ced7c3a238a10b4e5702f1655b742f440a26 — this comment is updated in place on each push.

Review details

No findings.

[OPUS-REVIEWED] 3d28ced

Verdict parsed from the review's SHA-scoped output markers for commit 3d28ced7c3a238a10b4e5702f1655b742f440a26.

False positive or not applicable? A repository writer can comment:
/ai-review override fable 3d28ced7c3a238a10b4e5702f1655b742f440a26: <one-sentence reason>

@github-actions

github-actions Bot commented Oct 14, 2026 •

Copy link
Copy Markdown
Contributor

GPT 6.1 Review — ✅ no blocking findings

GPT 6.1 completed its review of 3d28ced7c3a238a10b4e5702f1655b742f440a26 and found no blocking issues.

This comment is updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] 3d28ced

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 3d28ced7c3a238a10b4e5702f1655b742f440a26: <one-sentence reason>

@github-actions github-actions Bot added readiness: action required A blocking check or review needs attention merge conflict Branch has merge conflicts with its base — author must resolve before merge and removed readiness: checking Automated validation is still running labels Oct 14, 2026
@iamwhatever
iamwhatever force-pushed the feat/20156-return-warm-task-brief branch from e2935bd to 0250844 Compare October 14, 2026 23:50
@github-actions github-actions Bot added readiness: checking Automated validation is still running readiness: action required A blocking check or review needs attention and removed readiness: action required A blocking check or review needs attention merge conflict Branch has merge conflicts with its base — author must resolve before merge readiness: checking Automated validation is still running labels Oct 14, 2026
@iamwhatever
iamwhatever force-pushed the feat/20156-return-warm-task-brief branch from 0250844 to 65fa29a Compare October 15, 2026 01:10
@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • Active task original request span=58da0f5bfce5

fixed: head 65fa29a selects the last genuine request before the earliest open work item, with a prior finished-task regression test.

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • Durable claim storage span=ee2d856df384

fixed: the PR body now records the empty owner-only display row as the protected derived claim and names its transcript reset/delete lifecycle.

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • Feature map span=ea1ed556230c

fixed: head 65fa29a replaces the stale immediate-resume entry with the return-triggered task brief contract.

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • Dismissible failure span=aa1e0a931a43

fixed: head 65fa29a adds a dismiss action while preserving the RFC-required failure sentence.

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • Real pane placement evidence span=c468aa1953b0

fixed: the PR now shows preparing and unavailable above the composer, and the durable brief in transcript order.

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • Current task heading span=cba8b18bd78e

rebutted: RFC #20157 requires exactly Original request, Current task, and Next steps. The bullets are bounded current-state facts.

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • Preparing copy span=869a4f14da2a

rebutted: RFC #20157 fixes this deterministic sentence exactly. It promises waiting here only; leaving clears the status as designed.

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • Where it stands wording span=975a784eaf2e

rebutted: the accepted RFC fixes the visible section name as Current task; changing it would diverge from the recorded decision.

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • Longer preparing sentence span=f9eaea007288

rebutted: the accepted RFC requires the exact shorter deterministic status and states that it costs no model call.

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • Suggested next steps wording span=379fd5a6d105

rebutted: RFC #20157 names the required third part Next steps and already defines them as one to three suggestions.

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • Original request truncation span=4c690302a83f

fixed: head 65fa29a keeps genuine user text whole until the single 4,000-character bound adds its omission mark; a long-request test pins it.

@github-actions github-actions Bot added the readiness: checking Automated validation is still running label Oct 15, 2026
@iamwhatever

Copy link
Copy Markdown
Contributor Author

Sensitive change review

  • Head commit: 66de427
  • What rule changed: Registered return-triggered task briefs as a scrubbed transcript egress boundary.
  • Before -> after: Warm task briefs did not exist; now original requests and generated summaries are scrubbed before persistent browser delivery.
  • Worst case if wrong: An unsanitized credential or exfiltration URL could appear in persisted chat and browser rendering.
  • Evidence checked: No sandbox behavior changed; inspected the final warm-brief scrub calls and security posture sink registration.
  • Tools still work: The governed one-call model path remains tool-free and existing chat tools are unchanged.
  • Backward compatible: Cold welcome and ordinary chat behavior remain unchanged; only the replaced warm static card differs.
  • Existing tests: 498 focused backend and 485 focused frontend tests pass, with mypy and typecheck green.
  • How to undo: Revert commit 66de427 to restore the prior warm greeting behavior.

@iamwhatever
iamwhatever force-pushed the feat/20156-return-warm-task-brief branch from 66de427 to 6f4fc09 Compare October 15, 2026 10:20
@iamwhatever

Copy link
Copy Markdown
Contributor Author

Sensitive change review

  • Head commit: 6f4fc09
  • What rule changed: Registered return-triggered task briefs as a scrubbed transcript egress boundary.
  • Before -> after: Warm task briefs did not exist; now original requests and generated summaries are scrubbed before persistent browser delivery.
  • Worst case if wrong: An unsanitized credential or exfiltration URL could appear in persisted chat and browser rendering.
  • Evidence checked: No sandbox behavior changed; inspected the final warm-brief scrub calls and security posture sink registration.
  • Tools still work: The governed one-call model path remains tool-free and existing chat tools are unchanged.
  • Backward compatible: Cold welcome and ordinary chat behavior remain unchanged; only the replaced warm static card differs.
  • Existing tests: 498 model-path, 294 transfer/projection, and 485 frontend tests pass, with mypy and typecheck green.
  • How to undo: Revert commit 6f4fc09 to restore the prior warm greeting behavior.

@github-actions github-actions Bot added readiness: action required A blocking check or review needs attention and removed readiness: checking Automated validation is still running labels Oct 15, 2026
@github-actions github-actions Bot added the merge conflict Branch has merge conflicts with its base — author must resolve before merge label Oct 15, 2026
@iamwhatever
iamwhatever force-pushed the feat/20156-return-warm-task-brief branch from 6f4fc09 to d3a5789 Compare October 15, 2026 22:42
@github-actions github-actions Bot added readiness: checking Automated validation is still running and removed readiness: action required A blocking check or review needs attention labels Oct 15, 2026
@iamwhatever

Copy link
Copy Markdown
Contributor Author

Disposition for head d3a5789 (rebased on main 0f98bd9):

GPT BLOCK — unbounded waits in cancellation tests: fixed. The two flush-cancellation tests were removed with the retraction code they covered. The surviving test, test_cancel_during_model_suppresses_result_without_cancelling_acquisition, bounds every wait with asyncio.wait_for(..., timeout=5): the model-stub release, the entered signal, the task join, and the cleanup gather.

Opus advisory — English fixed copy: accepted, not changed. The model now writes current status and suggested actions in the language the user writes in (prompt line added). The two fixed lines ("Preparing…", "I couldn't prepare…") and the three headings stay English, like other gateway-authored lines. Localizing them by meta.kind would need the special client renderer the RFC rules out. The original request is quoted as the user wrote it. This trade-off is now recorded in the RFC (#20979).

Scope cut (First Principles item 9). Posted brief messages are ordinary history and stay. Removed: _retract_persisted_result, _surface_persisted_result, _discard_unsurfaced, invalidate_warm_task_brief and its chat_message_update path, the slot-detail _unpersisted_warm_brief withholding, the flush_slot_now -> bool change, and the tests for them. Kept: fingerprint claim dedupe, generated-row exclusions from model inputs, trigger rules, the egress registry entry. Net diff went from +2386/-769 to +1833/-753.

@github-actions github-actions Bot removed the merge conflict Branch has merge conflicts with its base — author must resolve before merge label Oct 15, 2026
@iamwhatever

Copy link
Copy Markdown
Contributor Author

Sensitive change review

  • Head commit: d3a5789
  • What rule changed: The posture inventory registers dashboard/warm_task_brief.py as a redaction egress sink. Runtime redaction rules did not change.
  • Before -> after (what was allowed/blocked, now): The omission gate blocked the unclassified call site; it now counts the existing transcript-and-browser scrub. No runtime input becomes newly allowed or blocked.
  • Worst case if wrong (breaks commands / leaks secret / locks user out): The posture panel could overstate coverage. Runtime still applies both redactors before the brief is posted.
  • Evidence checked (which sandbox result proves it): test/test_security_posture.py passed on this head with the focused warm-brief, session-summary, history-projection and session-transfer suites (522 passed).
  • No regression:
    • Tools still work (which tools/commands were run, result): Warm-brief, slot-reader, persistence-order and unflushed-tail pytest passed 73/73; focused Vitest passed 401/401; tsc, i18n lint, black, isort and flake8 passed.
    • Backward compatible (old config / old data / old callers still work, how checked): The registry entry is metadata only. Existing transcript, API and config shapes remain readable; slot-detail and persistence code are back to main.
    • Existing tests (which suites passed on this head): Trigger, plain messages, cancellation, member page, WebSocket, cold welcome, security posture, session transfer, formatting and docs passed.
  • How to undo: Revert d3a5789cf0bc; posted messages remain readable ordinary crewmate messages and empty claims draw nothing.

@iamwhatever

Copy link
Copy Markdown
Contributor Author

Sensitive change review

  • Head commit: 286249c
  • What rule changed: The posture inventory registers dashboard/warm_task_brief.py as a redaction egress sink. Runtime redaction rules did not change.
  • Before -> after (what was allowed/blocked, now): The omission gate blocked the unclassified call site; it now counts the existing transcript-and-browser scrub. No runtime input becomes newly allowed or blocked.
  • Worst case if wrong (breaks commands / leaks secret / locks user out): The posture panel could overstate coverage. Runtime still applies both redactors before the brief is posted.
  • Evidence checked (which sandbox result proves it): test/test_security_posture.py passed with the focused warm-brief, session-summary, history-projection and session-transfer suites (522 passed).
  • No regression:
    • Tools still work (which tools/commands were run, result): Warm-brief and wait-discipline ratchet pytest passed 67/67 on this head; focused Vitest passed 401/401; tsc, i18n lint, black, isort and flake8 passed.
    • Backward compatible (old config / old data / old callers still work, how checked): The registry entry is metadata only. Existing transcript, API and config shapes remain readable; slot-detail and persistence code are back to main.
    • Existing tests (which suites passed on this head): Trigger, plain messages, cancellation, wait ceilings, member page, WebSocket, cold welcome, security posture, session transfer, formatting and docs passed.
  • How to undo: Revert 286249c2e585; posted messages remain readable ordinary crewmate messages and empty claims draw nothing.

@iamwhatever

iamwhatever commented Oct 16, 2026 •

Copy link
Copy Markdown
Contributor Author

Sensitive change review

  • Head commit: 250eb44
  • What rule changed: The posture inventory registers dashboard/warm_task_brief.py as a redaction egress sink. Runtime redaction rules did not change.
  • Before -> after (what was allowed/blocked, now): The omission gate blocked the unclassified call site; it now counts the existing transcript-and-browser scrub. No runtime input becomes newly allowed or blocked. The model's prose message is still passed through both redactors before it is posted.
  • Worst case if wrong (breaks commands / leaks secret / locks user out): The posture panel could overstate coverage. Runtime still applies both redactors before the brief is posted.
  • Evidence checked (which sandbox result proves it): test/test_security_posture.py passed on this head with the warm-brief, session-summary, history-projection and session-transfer suites (561 passed).
  • No regression:
    • Tools still work (which tools/commands were run, result): Warm-brief and consolidation pytest passed 496/496; focused Vitest passed 411/411; tsc, flake8, isort, mypy and docs lint passed.
    • Backward compatible (old config / old data / old callers still work, how checked): The registry entry is metadata only. Existing transcript, API and config shapes remain readable.
    • Existing tests (which suites passed on this head): Trigger, plain messages, opener language, prose parsing, cancellation, member page, WebSocket, cold welcome, consolidation facade, security posture, session transfer and docs passed.
  • How to undo: Revert 250eb44c044f; posted messages remain readable ordinary crewmate messages and empty claims draw nothing.

@iamwhatever

Copy link
Copy Markdown
Contributor Author

Sensitive change review

  • Head commit: 35f1dde
  • What rule changed: The posture inventory registers dashboard/warm_task_brief.py as a redaction egress sink. Runtime redaction rules did not change.
  • Before -> after (what was allowed/blocked, now): The omission gate blocked the unclassified call site; it now counts the existing transcript-and-browser scrub. No runtime input becomes newly allowed or blocked. The model's prose message is still passed through both redactors before it is posted.
  • Worst case if wrong (breaks commands / leaks secret / locks user out): The posture panel could overstate coverage. Runtime still applies both redactors before the brief is posted.
  • Evidence checked (which sandbox result proves it): test/test_security_posture.py passed on this head with the warm-brief, consolidation, session-summary, history-projection and session-transfer suites (1024 passed).
  • No regression:
    • Tools still work (which tools/commands were run, result): Those 1024 pytest passed; focused Vitest passed 411/411; tsc, flake8, isort, mypy and docs lint passed.
    • Backward compatible (old config / old data / old callers still work, how checked): The registry entry is metadata only. Existing transcript, API and config shapes remain readable; a non-string meta.kind on old rows is now ignored instead of raising.
    • Existing tests (which suites passed on this head): Trigger, plain messages, opener language, prose parsing, non-string kind, cancellation, member page, WebSocket, cold welcome, consolidation facade, security posture, session transfer and docs passed.
  • How to undo: Revert 35f1dde1fc46; posted messages remain readable ordinary crewmate messages and empty claims draw nothing.

Replace immediate static warm resume cards with a return-triggered summary after an explicit 30-minute away episode. Preserve the cold welcome and gate one background model call behind owner presence, a durable fingerprint claim, and a stale-result check.

The preparing, result and failure lines are ordinary crewmate messages and stay as ordinary history once posted. A stale result is dropped unposted. The model writes the status and suggested actions in the user's language.
@iamwhatever

iamwhatever commented Oct 16, 2026 •

Copy link
Copy Markdown
Contributor Author

Sensitive change review

  • Head commit: 3d28ced
  • What rule changed: The posture inventory registers dashboard/warm_task_brief.py as a redaction egress sink. Runtime redaction rules did not change.
  • Before -> after (what was allowed/blocked, now): The omission gate blocked the unclassified call site; it now counts the existing transcript-and-browser scrub. No runtime input becomes newly allowed or blocked. The model's prose message is still passed through both redactors before it is posted.
  • Worst case if wrong (breaks commands / leaks secret / locks user out): The posture panel could overstate coverage. Runtime still applies both redactors before the brief is posted.
  • Evidence checked (which sandbox result proves it): test/test_security_posture.py passed with the warm-brief, consolidation, session-summary, history-projection and session-transfer suites (1024 passed). This head adds one screenshot over 35f1dde and rebases onto main.
  • No regression:
    • Tools still work (which tools/commands were run, result): Those 1024 pytest passed; focused Vitest passed 411/411; tsc, flake8, isort, mypy and docs lint passed.
    • Backward compatible (old config / old data / old callers still work, how checked): The registry entry is metadata only. Existing transcript, API and config shapes remain readable; a non-string meta.kind on old rows is ignored instead of raising.
    • Existing tests (which suites passed on this head): Trigger, plain messages, opener language, prose parsing, non-string kind, cancellation, member page, WebSocket, cold welcome, consolidation facade, security posture, session transfer and docs passed.
  • How to undo: Revert 3d28ced7c3a2; posted messages remain readable ordinary crewmate messages and empty claims draw nothing.

@iamwhatever

Copy link
Copy Markdown
Contributor Author

Dispositions for head 3d28ced:

  • UX evidence gap (opener alone while the model call runs): temp-screenshots/20422-warm-brief-plain/pending-light.png is now committed and embedded. There is no spinner or typing bubble by design. The user asked for a plain crewmate message with no special renderer, and the brief call is not a chat turn.
  • UX watch, failure after the opener: the failure line now ends with "Ask me where things stand." That gives the user a next step. Holding the opener until the result would bring back a silent wait, which the accepted RFC rejects. Kept as is.
  • UX watch, time cue: kept as is. The owner asked for the brief to read as an ordinary message the crewmate sends on its own. The opener "Welcome back, ..." is the cue.
  • UX suggestion, jargon in the brief: noted, no change. The model writes from the user's own transcript, so it uses the words the user already uses.
  • First Principles, RFC shape: the 2026-10-16 amendment line in rfc-delayed-task-brief.md records the owner's UX decision (first-person opener, prose catch-up, no headings). It ships with the code it governs.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants