[Cuarroc]

Ungesendete Chatentwürfe bleiben innerhalb der laufenden Oberfläche pro Projekt erhalten.
Beim Zurückkehren wird der echte Projektverlauf gelesen, ohne automatisch eine Nachricht zu senden.
Die lokalen Prüfungen sind grün; unabhängige Reviews und Abnahme stehen noch aus.

Report

Plan-ID: KR-04b
Ticket: 54a480e5-5c04-4387-b15b-ae60d7e62e6c
Operator: UMSETZEN. Execution owner: srv-kr-04b-chat-state. Parent: Nutzbar1 / Classic removal; Root owns acceptance, reviews and integration.

Candidate and contract

• Base: 60aa6ac17afb89cc5f88e220db1f9bf7dbb7d7fb; own clean HEAD, origin/main and remote main matched before edits. Predecessor feat(leitstand): mount orchestrator chat in Verlauf (KR-04a) #1569 was independently confirmed MERGED at 3ad56a8a2c231d5f4d56730a00b26101c6b29d95.
• Head: babf5f9586b5e12973c1bc998e76e7284b285fd3, branch codex/kr-04b-chat-state.
• Author: OpenAI gpt-6-astra, effort high, observed in this session's native turn_context at 2026-10-24 23:49:17 UTC.
• Original 15 October wave-plan/inventory sections 2/7, the current v2 subplan section 10.2, and feat(leitstand): mount orchestrator chat in Verlauf (KR-04a) #1569's deferred draft-lifetime obligation were compared with current source. The bounded contract remains applicable. PLAN updates belong to the coordinator under the dispatch exception.
• Pre-edit forecast: within 300 raw lines. Actual git diff --numstat 60aa6ac17afb89cc5f88e220db1f9bf7dbb7d7fb: +212 / -36 = 248 raw lines, including tests, five assigned files.
• Native get_goal({}) initially returned goal: null, remainingTokens: null, completionBudgetReport: null. create_goal returned an active bounded worker goal covering draft/history behavior, allowed files, 300-line limit, predecessor, red-first/gates/push/DRAFT/report acceptance and this owner. The checked handover is its endpoint; parent acceptance is separate. No quota inference from these null fields.

Changes and lifetime

• Keep one current draft per project in UI-session memory. Route unmounts and project switches preserve text, pending-send state and send errors. Empty idle entries without errors are removed. There is no draft revision history, silent eviction, hard count/text-size cap, or persistence to disk; memory is proportional to outstanding drafts and lasts until this UI document is reloaded/closed.
• An accepted send clears only its exact submitted project and text revision, including when the user edits away and back to identical text. Refused/failed sends retain the draft and error, including completion while the route is absent. A pending send remains guarded when returning to the route.
• Reuse the existing orchestrator lookup, backend message reader, 200-message limit and 10-second polling. When a first send completes after route return and the new view still has no orchestrator, repeat the existing lookup. No second history store/poller or navigation-triggered send/start.
• The shared chat hook adds an optional failure callback; classic callers retain the boolean send contract. Existing controls, keyboard behavior, no-project disabling and scroll handling remain covered by unchanged tests.
• Architecture drift: gate passed; no new allowlist entry, native seam, schema, routing, design, dependency or policy changes. Source tracing followed ShellMount → VerlaufRoute → VerlaufChat/useOrchestratorChat and the App caller; no project-graph completeness claim.

Evidence

All final evidence is for the head above in the assigned worktree. Commands ran in the foreground, with Cargo slot 2 and one build job; no hook bypass or history rewrite.

Command / stage	Exit / observation
bash scripts/sync.sh start	0; assigned clean branch at the pinned base.
npm run dev:start-check -- --max-cargo 2 --min-free-gb 3.5	0 before admission to final build, prepush and push. Additional process/ancestry guards observed load1 7.176 / 6.170 / 7.738 respectively, available RAM above 3,500 MiB, no active heavy roots at those admissions. Slot 1 was observed on the other worker during an earlier check; slot 2 remained assigned here. See procedural deviation below.
npx --no-install vitest run src/shell/mount/verlauf.mount.test.tsx — initial RED	1; six existing tests passed, new draft regression failed because returning to the route erased the draft. Test committed first in 117bdc99.
Same command — first-send RED	1; with draft retention applied, the added first-send case failed because accepted history was never loaded after return. Test committed before the route fix in b15d956f.
npx --no-install vitest run src/shell/mount/verlauf.mount.test.tsx src/screens/leitstand/verlauf/VerlaufChat.test.tsx src/lib/orchestratorChat.test.ts src/lib/orchestratorChat.switch.test.ts	0; 32 tests.
npm run typecheck	0 standalone; also 0 in the final precommit, prepush and normal push hook.
npm run build	0 at the final source/head; frontend production build completed.
bash scripts/ci/red-first.sh	0; both exact regression names missing at merge-base and green at head. The unchanged gate used its authorized existing-lockfile npm-ci in an isolated checkout.
bash scripts/ci/secret-scan.sh	0 before each of the three normal commits; hooks also passed.
bash scripts/ci/gates.sh lane prepush	0; 1,716 frontend tests, 2,327 Rust tests, 33 Denkraum browser tests passed; HQ tests 995 passed and Denkraum Node tests 390 passed.
git push -u origin codex/kr-04b-chat-state	0; normal full prepush hook also passed at the same head, including 1,716 frontend / 2,327 Rust / 33 browser tests.
git ls-remote origin codex/kr-04b-chat-state	0; exact remote SHA babf5f9586b5e12973c1bc998e76e7284b285fd3, observed 2026-10-25 00:08 UTC.
git diff --check / git status --short --branch	0 / 0; clean, tracking the pushed branch.

The new functional proof uses the public ShellMount route and real composer, active-project module, IPC adapters and chat hook. Only Tauri invoke/listen are mocked. This is browser MOCK functional proof in Vitest/jsdom, not native execution or a Chromium chat drill. Literal assertions cover route return, project isolation, fresh backend history, no empty claim at first read, Shift+Enter, no restoration send/start, pending-send protection, identical-text/new-revision survival, late refusal/error preservation, exact accepted clearing and first-conversation discovery. Existing tests were not weakened; no skip, timeout or retry was added.

Procedural deviation

An orchestration mistake started the initial npm run build after the extra load guard had returned 1 (load1 = 10.96875, threshold <8), although the RAM/Cargo startcheck passed. That build completed with exit 0. Subsequent heavy commands were made conditional on both successful checks; a foreground 30-second wait preceded the next admission. The final-head build, full prepush and normal push each had a fresh passing guard. No threshold/configuration was changed. Root must disposition this process deviation; the later green runs do not erase it.

NICHT ABGEDECKT

• Linux does not compile #[cfg(windows)] tests (run_in_pty, npm_shim, ConPTY, DPAPI) or check Windows clippy branches. The Windows lane remains a merge-queue obligation owned by Root.
• Existing skips remain: seven Rust, three HQ Node and one Denkraum Node test. No new skips.
• Prepush is not the full Linux workflow or application browser-smoke lane. Frontend production build was run separately. No chat screenshot/visual, full Glass acceptance, user UAT, native/provider, release or effective-product claim.
• External services (updater endpoint, provider routing, mirror) were not verified. No real messages or provider workers were started for this package.
• No browser restart/disk-sync guarantee, draft-size cap or persisted-state policy change. Send/reload error cases are mocked functional evidence.
• Plan-check ran offline: no GitHub/plan-age comparison or running-job check from that gate. Coordinator owns the agreed PLAN update and final row reconciliation.
• Project-memory mirror/connector and project graph were unavailable here; only known-source and live Git/GitHub checks are claimed. Provider usage was not independently available; Root must refresh native usage/rules at publication.

Review and handover

Review: ausstehend (Stufe B)

Root arranges two different non-OpenAI review families under current rule 5/E1, dispositions the procedural deviation, and owns acceptance, ready/review-ok and Mergify. This is one checked DRAFT handover: implemented, locally tested, committed and remote-SHA verified; independent reviews, parent acceptance, merge and effective/native operation remain unproven. No follow-on dispatch by this worker.

Review disposition (Root, 2026-10-25 ~00:30Z, head babf5f9586b5e12973c1bc998e76e7284b285fd3)

Tier B (author OpenAI gpt-6-astra). Two reviewers from two different non-author families, both on this exact head: Anthropic (Claude subagent, read-only) — APPROVE WITH CONDITIONS; Moonshot (Kimi CLI, read-only, diff-only) — APPROVE. No high or medium finding.

ID	Source	Sev	Finding	Disposition
R1894-C1 / R1894-K2	Claude, Kimi	low	VerlaufRoute.tsx:23-26 re-lookup also after a failed first send: loading swap, composer focus lost, extra list_workers	Follow-up (no push in this PR; run lookup only after an accepted send)
R1894-C2	Claude	low	possible loading flash after an in-place successful first send (hypothesis, React lanes)	Follow-up together with C1
R1894-C3	Claude	low	retained sending:true never clears if the IPC call never settles (ipc.ts:323-325 has no timeout)	Accepted as limit, added to NICHT ABGEDECKT below; follow-up
R1894-C4	Claude	low	ProjectChat re-renders on each keystroke	Follow-up (performance only)
R1894-C5 / R1894-K3	Claude, Kimi	low	non-retained mode keeps drafts across p1->p2->p1	Rejected: test-only mode, per-project semantics intended
R1894-K1	Kimi	low	failure reported in both draft.error and chat.error; stale draft error can mask a newer chat error until dismissed	Follow-up
R1894-K4	Kimi	low	module-level draft store without eviction	Rejected: bounded by project count, entries deleted when empty/idle (chatDraft.ts:13)
R1894-K5	Kimi	low	~80-line monolithic test	Rejected: style only, coverage confirmed by both reviewers
PROC-1894-1	Chief	process	first npm run build started after load guard exit 1 (load1 10.97 > 8)	Accepted as one-off deviation; acceptance relies only on the final build/prepush/push runs after passing guards (log: load1 7.18 / 6.17 / 7.74); spec rule guard && phase sent to Stabschef

NICHT ABGEDECKT (Root addendum): if send_to_orchestrator never settles, that project's composer stays on "Sendet …" until the UI reloads (R1894-C3). Evidence is browser MOCK (Tauri invoke/listen mocked); native behaviour not shown.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-25T00:21:35.262106Z	babf5f9	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[Cuarroc]

Root: review-ok set for head babf5f9586b5e12973c1bc998e76e7284b285fd3 (tier B: two non-author families on this head — Anthropic APPROVE WITH CONDITIONS (all low, dispositioned in PR body), Moonshot APPROVE; all required checks green; Codex observer no findings). Any later push voids it.

[mergify]

Merge Queue Status

• ✅ Entered queue — 2026-10-25 00:29 UTC · Rule: default · triggered by merge protections
• ✅ Checks passed · on draft merge queue: checking #1894 on main (60aa6ac) #1895
• ✅ Merged — 2026-10-25 00:38 UTC · at ae9abb5f0b1388d317f7dfb81dd86e387bebf780 · merge

This pull request spent 8 minutes 39 seconds in the queue, including 8 minutes running CI.

Required conditions to merge

• check-success = gates (linux)
• check-success = gates (windows)
• check-success = red-first
• github-review-approved [🛡 GitHub repository ruleset rule main-pr-required-checks]
  • fix: preserve project chat drafts across routes (KR-04b) #1894
• label != do-not-merge
  • fix: preserve project chat drafts across routes (KR-04b) #1894
• any of [🛡 GitHub repository ruleset rule main-pr-required-checks]:
  • check-success = @github-actions/gates (linux)
  • check-neutral = @github-actions/gates (linux)
  • check-skipped = @github-actions/gates (linux)
• any of [🛡 GitHub repository ruleset rule main-pr-required-checks]:
  • check-success = @github-actions/gates (windows)
  • check-neutral = @github-actions/gates (windows)
  • check-skipped = @github-actions/gates (windows)
• any of [🛡 GitHub repository ruleset rule main-pr-required-checks]:
  • check-success = @github-actions/red-first
  • check-neutral = @github-actions/red-first
  • check-skipped = @github-actions/red-first
• any of [🛡 GitHub repository ruleset rule main-pr-required-checks]:
  • check-neutral = @mergify/Mergify Merge Protections
  • check-skipped = @mergify/Mergify Merge Protections
  • check-success = @mergify/Mergify Merge Protections

[Cuarroc]

Report addendum — review evidence limits (Root, 2026-10-25 ~00:42Z, post-merge)

Tier B counted two non-author families for head babf5f9 (author: OpenAI Codex gpt-6-astra/high):

• Anthropic: Claude subagent inside Root's own harness (claude-opus-5-5), separate agent with its own reads; Root only adjudicated.
• Moonshot: Kimi CLI v2.1.1 (Kimi Code subscription).

Basis: origin/main AGENTS.md rule 5 + E1 (Elias 2026-10-20: while OpenRouter credit <= 0 a second subscription family may replace OpenRouter). The DOC-mission rule "Root alone uses Claude" (FINAL-20261020T2156Z) is scoped to that docs mission; KR-04b is Glass product work, covered by the "reviews outside this mission: E1" clause.

Evidence gaps, disclosed (not measured after the fact):

1. The exact Kimi model ID was not logged; the family is inferred from the vendor CLI only.
2. The E1 condition "OpenRouter credit <= 0" rests on a memory note dated 2026-10-17, not on a fresh observation; usage_overview shows no OpenRouter balance.

Future review starts record the observed reviewer model identity and the E1 condition evidence before counting a review.