Conversation
Add a shared Pane context menu (Rename, Pin/Unpin, Archive) to the expanded, compact, and pinned sidebar entries, backed by a rename dialog that trims input and rejects blanks. Renames go through one main-process helper that writes only the display name plus a new name_manually_set provenance bit, which the AI namer now respects; the stored name is the label everywhere and PR titles stay metadata, so the dormant auto-rename-to-PR-title setting is removed. Expanded rows lose the coloured accent bar and label animation and show agent state through SessionStatusBadge instead. Adds the DB columns both workstreams need and extracts runAgentDoctor into a service module. Claude-Session: https://claude.ai/code/session_01A6ystbZhfNVGeSUc7x8DNm
The interactive Add Repository dialog now discloses the configured default agent and sends launchDefaultAgent: true; projects:create then runs a main-owned, single-flight one-shot that validates platform and executable, creates the main session and a terminal panel, spawns the PTY, waits for CLI readiness, and only then writes a durable receipt. Any failure destroys the panel, leaves the project blank, writes no receipt, is memoised so nothing retries, and surfaces a non-modal notice with a manual Open agent action. The default agent setting copy now names both uses. Claude-Session: https://claude.ai/code/session_01A6ystbZhfNVGeSUc7x8DNm
|
React Doctor found 1 new issue in 1 file · 1 warning · score 74 / 100 (Needs work) · 1 fixed · vs Reviewed by React Doctor for commit |
Suppress the default Terminal that session-created auto-creates when the one-shot launch owns the session, so a launch yields exactly one agent tab and a failed launch leaves none. Verify provisional-panel removal with a retry and DB fallback and report any stale panel. Send the launch flag only when the dialog actually disclosed the agent, ignore dialog close while creating, and keep the failure notice if Open agent fails. Make the Pane context menu keyboard-operable (focus, arrows, Home/End, Escape restores the opener). Return raw rename errors only for the typed validation error, fix the stale single-flight entry on skip paths, drop the unused menu label field, and de-flake the Playwright text assertion. Claude-Session: https://claude.ai/code/session_01A6ystbZhfNVGeSUc7x8DNm
Pre-allocate the agent panel id so a persist-then-reject createPanel is still cleaned up, create the suppressed-terminal main session before the agent doctor runs so every attempted launch that fails leaves Explorer and Diff only, and bound the spawn plus readiness wait with one 45 s deadline so Add Repository can never hang. Context-menu opener falls back to the row element; Playwright covers the pinned expanded copy. Claude-Session: https://claude.ai/code/session_01A6ystbZhfNVGeSUc7x8DNm
When the 45 s launch deadline fires while the terminal is still queued for a spawn slot, the failure path now arms a late guard that destroys a terminal registering after cleanup, and initializeTerminal revalidates that its panel still exists after acquiring the slot and before spawning, so a deleted panel never runs the permissive agent command. The dialog also sends the agent it disclosed and main refuses to launch a different one. Cleanup probes are guarded so a DB read cannot escape the catch. Claude-Session: https://claude.ai/code/session_01A6ystbZhfNVGeSUc7x8DNm
An automatic launch now needs both the launch flag and the exact agent the dialog disclosed; a request without a disclosed agent, or with a different one, is skipped before any session or panel exists. Claude-Session: https://claude.ai/code/session_01A6ystbZhfNVGeSUc7x8DNm
Whitespace only (git diff -w is empty); applied from the refactor pass. Claude-Session: https://claude.ai/code/session_01A6ystbZhfNVGeSUc7x8DNm
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
QA drive evidence — PR #551Driver: the repo's own Playwright harness (Chromium + J1 — Rename from the sidebar (M1–M4, AC1/2/4/5)
📹 Journey video: J1-rename-propagation.webm (4.5 s; local J2 — Add repository, launch once (M5 renderer half, AC6)
📹 Journey video: J2-launch-once.webm (3.2 s; local J3 — Failed launch → notice → Open agent / Dismiss (M8 renderer half, M9, AC10)
📹 Journey video: J3-failed-launch-open-agent.webm (1.1 s; local J5 — No default agent (M10, AC9) and J6 — Clone from GitHub sends no flag (M11)
📹 Journey video: J6-clone-from-github.webm (0.7 s; local AC13 — five agent states, expanded and compact (M13)
AC12 — neutral selected / hover / focus across themes (M12)
Rubric 4 — axe scans (0 violations)
Other videos: J1-cancel-escape-pr-refresh.webm, J3-dismiss.webm, J5-no-default.webm. Playwright traces for all 18 tests are hosted as Console (rubric 5): every journey asserted zero console errors after filtering one pre-existing mock artifact — the mock leaves Passed automatedM1, M3, M4, M10, M13 in full; M2, M9, M12 for their renderer halves; AC1, AC2, AC4, AC5, AC6 (renderer), AC9 (mock approximation), AC10 (renderer), AC12, AC13. Remaining for the human
|
tbrownio
left a comment
tbrownio
left a comment
There was a problem hiding this comment.
Surviving Should-Fix / Nice-to-Have findings from the automated review + refactor passes (none block merge; details inline).
type: wrap-up-report
|
type: postmortem
|
| Step / dispatch | Start | End | Dur | Tokens | Est. cost | Note |
|---|---|---|---|---|---|---|
| Step 0 preflight + load + brief.html | 18:24 | 18:28 | 04:04 | overseer | — | gh/codex/pnpm probes, install no-op, intake artifact |
| code-researcher (Codex) | 18:28 | 18:30 | 02:09 | 95,070 | unknown (blended) | ran while Overseer read anchors |
| Overseer anchor reading + plan.md | 18:28 | 18:40 | 12:06 | overseer | — | serial with research (overlapped) |
| plan-reviewer pass 1 (Codex) | 18:40 | 18:43 | 02:57 | 133,573 | unknown | parallel with Claude lane |
| plan-reviewer pass 1 (Claude) | 18:40 | 18:48 | 07:22 | 144,459 | unknown (no class split) | 7:22 — the slower lane |
| plan revision 1 (Overseer) | 18:48 | 18:52 | 04:37 | overseer | — | 14 MF folded |
| plan-reviewer pass 2 (Codex) | 18:52 | 18:55 | 02:26 | 100,767 | unknown | parallel |
| plan-reviewer pass 2 (Claude) | 18:53 | 18:59 | 06:27 | 104,699 | unknown | 6:27 |
| plan revision 2 (Overseer) | 18:59 | 19:02 | 02:37 | overseer | — | |
| plan-reviewer pass 3 (Codex) | 19:02 | 19:05 | 03:20 | 120,406 | unknown | parallel |
| plan-reviewer pass 3 (Claude) | 19:02 | 19:09 | 06:54 | 113,234 | unknown | 6:54 |
| plan revision 3 + implementer launch | 19:09 | 19:10 | 01:21 | overseer | — | cap reached |
| implementer (Codex, all 22 tasks) | 19:10 | 19:27 | 17:05 | 400,698 | unknown | 17:05; diagram authored in parallel |
| Excalidraw diagram + hosting (Overseer) | 19:11 | 19:20 | 08:41 | overseer | — | overlapped implementer |
| backend-verifier (Codex) | 19:28 | 19:30 | 02:26 | 66,526 | unknown | 2:26 |
| build gate + commits + push + PR open | 19:30 | 19:34 | 03:06 | overseer | — | frontend build in background |
| code-reviewer pass 1 (Codex) | 19:34 | 19:37 | 03:44 | 172,211 | unknown | 4 MF incl. events.ts double terminal |
| code-reviewer pass 1 (Claude) | 19:34 | 19:41 | 07:03 | 140,096 | unknown | 0 MF / 3 SF |
| implementer fix round 1 (resume) | 19:41 | 19:48 | 07:01 | 412,786 | unknown | 7:01 |
| code-reviewer pass 2 (Codex) | 19:49 | 19:52 | 02:38 | 134,187 | unknown | 2 new MF |
| code-reviewer pass 2 (Claude) | 19:49 | 19:57 | 08:04 | 146,135 | unknown | 0 MF / 2 SF |
| implementer fix round 2 (resume) | 19:58 | 20:05 | 07:03 | 103,336 | unknown | 7:03 |
| code-reviewer pass 3 (Codex) | 20:06 | 20:09 | 02:59 | 134,843 | unknown | 1 MF (orphan PTY) |
| code-reviewer pass 3 (Claude) | 20:06 | 20:09 | 03:31 | 87,277 | unknown | 0 MF / 1 SF (same orphan PTY) |
| implementer fix round 3 (resume) | 20:10 | 20:13 | 03:18 | 108,591 | unknown | 3:18 |
| scoped code-reviewer 1 (Codex) | 20:14 | 20:16 | 01:49 | 91,257 | unknown | 1 MF (disclosure omitted) |
| implementer fix round 4 (resume) | 20:16 | 20:19 | 02:36 | 65,267 | unknown | 2:36 |
| scoped code-reviewer 2 (Codex) | 20:19 | 20:21 | 01:04 | 68,538 | unknown | clean |
| refactor-simple (Codex) | 20:19 | 20:21 | 01:55 | 117,464 | unknown | parallel with scoped review |
| refactor-deep (Codex) | 20:19 | 20:24 | 04:20 | 200,525 | unknown | 2 'Critical' = owner decisions |
| implementer refactor apply (whitespace) | 20:24 | 20:26 | 01:24 | 20,726 | unknown | git diff -w empty |
| fresh-eyes PR body: Overseer + 2 agents | 20:25 | 20:29 | 04:22 | 86,776 (agents) | unknown | pass 3 no change |
| frontend-verifier QA drive (Claude) | 20:27 | 20:35 | 08:07 | 126,951 | unknown | 18 tests, 45 shots, 7 videos |
| QA hosting + body + evidence comment + inline comments | 20:35 | 20:38 | 03:00 | overseer | — | 70 assets to qa-assets |
| wrap-up + fresh-eyes ×2 + label + notify | 20:38 | 20:41 | 02:37 | overseer | — |
Aggregates (share of 2:17): Step 0 + research + plan writing 16 m (12 %) · plan-review loop 30 m (22 %) · implement 17 m (12 %) · verify + PR 6 m (4 %) · post-PR review + fix loop 46 m (34 %) · scoped/refactor 9 m (7 %) · fresh-eyes + QA drive + evidence 12 m (9 %) · wrap-up 3 m (2 %). Summed Overseer turnaround gaps between a report landing and the next dispatch: ≈ 14 m total, never more than ~4 m each (plan revisions were the largest). Human-idle 0.
Ranked stalls: none in-run. The longest productive waits were the implementer's first dispatch (17:05) and each Claude review lane (6–8 m per pass, consistently the slower lane by 3–5 m).
Per-phase pacing from commits (PDT): 12:32 feature commits ×2 → 12:49 fix 1 → 13:06 fix 2 → 13:14 fix 3 → 13:19 fix 4 → 13:26 whitespace refactor. Fix rounds every 8–17 m, each gated by a full dual-lane review pass.
Blocker inventory: AskUserQuestion gates 0 · rate-limit hits 0 · red gates 0 · environment notes 2 (no Playwright MCP registered for Claude sub-agents → QA drive used the repo's Playwright harness; the Claude reviewer's node ABI mismatched better-sqlite3, so it could not run three Vitest suites and read them instead — Codex lanes and the verifier ran them green).
Tokens & cost. Overseer (claude-fable-5, 1 h cache TTL): input 23,619 + output 222,302 + cache read 67,938,206 + cache write 598,829 → ≈ $91 (0.24 + 11.12 + 67.94 + 11.98). Claude sub-agents 949,627 (notification totals; per-class split not harvested → cost unknown). Codex 2,546,771 blended tokens used → cost unknown (no class split survives the blend). The overseer's cache reads are 94 % of raw volume: a 186-message turn loop over a ~370 k-token context.
The single change that would have cut the most time: the post-PR loop spent 46 m on three full dual-lane passes plus two scoped passes because the first implementer pass shipped a defect the plan could not have known about only by reading (events.ts auto-creates a Terminal on every session-created). A cheap, targeted "listener/side-effect sweep" in the code-researcher dispatch (grep every sessionManager.on(...) / panelManager event consumer touched by the plan) would have surfaced it at plan time, collapsing at least one review pass and one fix round (~15 m, ~700 k Codex tokens).
What we asked for
Owner-approved brief: display-only Pane rename from the sidebar; once-only default-agent launch on Add Repository with blank state on skip/failure; neutral sidebar rows with the status badge instead of the accent bar. Zone 1, escalated to effective 0.
Outcome vs intended
On-target as far as the run can tell — every AC has automated or QA-drive evidence except the real-agent manual items (AC6–AC8) and three owner decisions recorded in the PR. The outcome half of this postmortem runs after PR review.
Why the gap happened
Omitted — operations-only.
What to change so it doesn't recur
claude/agents/code-researcher.md(dcouple/orchestra) — add to Method step 2: "For every event the plan will emit or listen to (emit('…'),.on('…')), enumerate all existing listeners and their side effects withpath:line; a listener that creates, deletes, or launches something is a load-bearing finding." Evidence:main/src/events.ts:64-75was missed by the researcher and by three plan-review passes; it cost one review pass + one fix round.claude/skills/do/SKILL.md— Step 1, plan-review dispatch: when the item touches process/PTY lifecycle, instruct reviewers to trace the full lifecycle including deadline/cancellation paths ("what happens if the awaited call completes after the caller gave up?"). The orphan-PTY-after-deadline defect (pass 3 + scoped pass) is the generic "late completion after cancellation" class; naming it in the dispatch would have caught it in pass 2 when the deadline was introduced.references/zones.md— data point, no edit proposed yet: at effective zone 0 the Claude review lane found 0 Must Fix in all three post-PR passes (its findings were Should Fix that Codex escalated or duplicated), while costing 373 k tokens; the Codex lane found every Must Fix. Two more runs like this would justify making the second lane plan-review-only at zone 0 for single-phase items.references/notify.md/claude/skills/do/SKILL.mdStep 0 — preflight should probe for a registered Playwright MCP server (claude mcp list) when the item is browser-required and say explicitly which driver the QA drive will use; this run discovered the gap at preflight but only by reading the agent charter.
Dial record & right-sizing
zone: 1 # item; Overseer-escalated to effective 0 (permissive agent launch = trust boundary)
lanes: dual
passes: {plan: 3/3, post_pr: 3/3, post_pr_scoped_extra: 2}
findings: {plan: {pass1: {codex: 9, claude: 5}, later: {codex: 5, claude: 7}},
post_pr: {pass1: {codex: 4, claude: 0}, later: {codex: 3, claude: 0}}} # Must Fix counts; Claude lane raised 3+2+1 Should Fix
verifiers: {frontend: ran, qa_pass: ran}
qa_findings: 0
cleanup: {disposition: none_created, markers: [agent-e2e-20260829], registered: 0}
wall_clock: 2:13 (transcript first event 2026-08-29T18:24Z → wrap-up 20:37Z; single session, no resume)
deviations: "escalated 1→0: permissive agent auto-launch is a trust boundary; dual review lanes"
pr_size: {files_changed: 53, additions: 2412, deletions: 711}
tokens:
codex: {total: 2546771, by_role: {implementer: 1111404, plan_reviewer: 354746, code_reviewer: 601036, code_researcher: 95070, backend_verifier: 66526, refactor_simple: 117464, refactor_deep: 200525}}
claude_subagents: 949627 # plan-reviewer 362392, code-reviewer 373508, fresh-eyes 86776, frontend-verifier 126951 (harness usage summaries)
overseer: 68782956 # transcript JSONL, dedup by message.id: input 23619 + cache_create 598829 + cache_read 67938206 + output 222302
total: 72279354
spend_ratio: 23144.2 # tokens.total ÷ 3123 changed lines; cache reads dominate the overseer figure
agents:
- {role: code-researcher, model: gpt-5.6-sol, effort: low, dispatches: 1, wall_clock: 2:09, tokens: 95070}
- {role: plan-reviewer, model: gpt-5.6-sol, effort: medium, dispatches: 3, wall_clock: 8:43, tokens: 354746}
- {role: plan-reviewer, model: claude-opus, effort: thinking, dispatches: 3, wall_clock: 20:42, tokens: 362392}
- {role: implementer, model: gpt-5.6-sol, effort: medium, dispatches: 6, wall_clock: 38:27, tokens: 1111404}
- {role: backend-verifier, model: gpt-5.6-sol, effort: low, dispatches: 1, wall_clock: 2:26, tokens: 66526}
- {role: code-reviewer, model: gpt-5.6-sol, effort: medium, dispatches: 5, wall_clock: 12:14, tokens: 601036}
- {role: code-reviewer, model: claude-opus, effort: thinking, dispatches: 3, wall_clock: 18:38, tokens: 373508}
- {role: refactor-simple, model: gpt-5.6-sol, effort: medium, dispatches: 1, wall_clock: 1:55, tokens: 117464}
- {role: refactor-deep, model: gpt-5.6-sol, effort: medium, dispatches: 1, wall_clock: 4:20, tokens: 200525}
- {role: fresh-eyes, model: claude-opus, effort: thinking, dispatches: 2, wall_clock: 2:37, tokens: 86776}
- {role: frontend-verifier, model: claude-sonnet, effort: thinking, dispatches: 1, wall_clock: 8:07, tokens: 126951}Judgment: review effort slightly overdone — right-sized on the Codex lane, overdone on the Claude diff-review lane. Plan review (3 passes, both lanes) paid for itself: 14 Must Fix folded before code. Post-PR: Codex passes 1–3 found 4/2/1 Must Fix (all real; pass 3's orphan PTY was security-relevant), so the cap was earned; the Claude diff lane caught 0 Must Fix across three passes for 373 k tokens — the single dial that would have changed it is review_lanes: single for the post-PR loop at this zone (keep dual for plan review).
Acceptance
Still awaiting review (operations-only postmortem).
System changes
- PR comment: (filled after posting)
- Work item: local-only intake artifact (no tracker) — no second anchor.
- Proposals 1–4 above: pending human verdict; nothing applied.
Make the AC11 import-guard test separator-agnostic so it passes on windows-2022. Drop the React Doctor prop-change effects by keying the context-menu body, rename dialog body, and dialog error state instead. Extract the duplicated sidebar menu/rename/focus lifecycle into usePaneContextMenu, split launchDefaultAgentOnce into attempt, readiness, and cleanup helpers, and forget a project's launch memo when the project is deleted. Claude-Session: https://claude.ai/code/session_01A6ystbZhfNVGeSUc7x8DNm
…eted Each launch attempt carries an AbortController; projects:delete aborts it before teardown and forgets the memo regardless of the delete result. The attempt re-checks cancellation and the project row before every stage, runs the normal provisional-panel cleanup, and never writes a receipt after cancellation. Rejected attempts are memoised as a failed result, and the rename dialog selects its prefilled text once. Claude-Session: https://claude.ai/code/session_01A6ystbZhfNVGeSUc7x8DNm
Final readiness evidence —
|
Sidebar rows and headers now use the navigation text family instead of body text-primary, and the dark/light navigation-primary tokens sit a step below body text so the rail reads as navigation. The pane hover card drops the PR body, status prose, and 10px type for a compact card: name, branch, last activity, PR number + state pill + diff size, title, and a View on GitHub link. Claude-Session: https://claude.ai/code/session_01LfUbWdfJ7xSx8a3TPq4XWC
Follow-up — sidebar contrast + hover card (
|
| dark | light |
|---|---|
 |
 |
Checks: typecheck 0 errors, pnpm lint 0, frontend Vitest 30 files / 298 passed, focused Playwright (add-repository, sidebar-compact, sidebar-pane-actions, window-title-bar) 26/26 after fixing the pill label to "Open" (the spec pins the capitalised text).
The archive and pin buttons were transparent but still occupied their slots, so every idle row reserved 48px to the right of its git stats. They now sit left of the stats and collapse to zero width until the row is hovered or focused, which leaves the stats on the row's right edge and gives the pane name that width at rest. A pinned pane keeps its pin visible. Claude-Session: https://claude.ai/code/session_01LfUbWdfJ7xSx8a3TPq4XWC
|
Additional customer evidence from a September 4 feedback call: the user explicitly prefers their manually assigned pane label, but reported that it does not always persist and can be replaced by the latest PR or commit label. For this PR, the desired precedence is: an explicit user-assigned name remains authoritative until the user changes or clears it. Automatic PR/commit-derived labels should only act as a fallback when no custom name exists. A regression test covering refresh/restart and PR/commit updates would capture the reported failure mode. |
tbrownio
changed the title
|
Closeout QA verdict: partial. Automated gates and the browser journeys pass; native agent/restart checks and the product decisions below remain open. Tested browser commit: In response to parsakhaz’s naming feedback: a real SQLite close/reopen plus fresh SessionManager keeps
Current QA files are local in
Harness limits: Electron APIs, repository creation, PR metadata and agent launch are mocked in browser QA. The history panel’s “Unable to load commit history” and the manual-tab loading placeholder are mock limitations. SQLite tests use temporary databases that are closed and deleted. No real repository registration, agent process, external message or service data was created.
Remaining decisions / checks:
#573 is still open and only overlaps the rename affordance plus the shared-store defect. Its double-click editor remains separate; #551 is not superseded. All applicable CI checks are green. The dedicated Vite process was stopped and the worktree is clean. No merge, close, release or deployment was performed. |
2 participants
Manually named Panes should keep their labels as work progresses. This addresses the reported PR/commit label replacement and the workspace-entry brief: give users a display-only rename action, open a default agent when adding a repository, and make sidebar selection easier to distinguish from agent status.
Behavior
A Pane is a workspace; its display name is separate from its Git branch and worktree path. Right-click a Pane in the expanded sidebar, compact sidebar, or the pinned section in either mode → Rename → enter
Human label→ Save. The visible name becomesHuman label; branch and directory remain unchanged. Blank input is rejected, Escape cancels, and duplicate labels are allowed.The rename writes both the name and a persistent
name_manually_setflag. Reloading sessions or reopening SQLite preserves both. A late AI-generated name cannot replace an explicit name, even when the user deliberately chooses the original fallback text. PR titles and commit statistics remain metadata. The active main-repository Pane is stored both in the sidebar list and as the current selection; both now update together, avoiding a stale sidebar label.Adding a repository through the desktop dialog discloses the configured default agent and command. Creating the repository authorizes that launch with the existing permissive preset and no initial prompt. The repository is saved before main validates the disclosed agent, platform, and executable. Main then creates one agent tab, waits until its process is alive and the command has been sent, and writes a success receipt. An agent showing its own login/trust screen counts as live; authentication is not checked. Concurrent requests share one attempt. Navigation, tab closure, or app restart does not initiate another automatic launch: only the create-repository request triggers it. Failed attempts are remembered for the current process; after restart, reopening an existing repository does not call that creation path.
With no configured agent, no agent is launched. The ordinary project-opening path may still create a shell terminal; this differs from the brief’s zero-terminal blank state and needs owner acceptance.
If executable validation or launch fails, the saved repository remains available. Cleanup removes the provisional agent tab, leaves Explorer/Diff, and displays a non-blocking notice with Open agent and Dismiss. The manual button uses normal panel creation; automatic retry is disabled. Executable validation can take about five seconds, followed by a 45-second launch deadline. Deleting the project cancels an in-flight launch, and late completion cannot leave an orphan terminal.
Sidebar rows use neutral hover/selected surfaces and a compact accessible status badge. This also includes softer navigation text, right-aligned Git statistics with hover/focus row actions, and a simpler hover card with branch, activity, PR state/title and a GitHub link. Settings copy explains the default-agent use; the unused PR-title auto-rename setting is removed.
Integration and review
Current main was merged into this branch, preserving its existing
createDefaultTerminalOnCreatecontract and external-worktree labels. An independent integration reviewer found no must-fix issues. GitHub marks all 10 existing review threads resolved; that does not settle the product decisions listed below.#573 remains open and is not superseded: its double-click inline editor is a separate affordance. Its relevant dual-copy store fix is also included here; its inline UI is not imported. The broader launch and sidebar changes remain specific to #551.
Visual overview
The existing diagram shows the original and proposed launch flow. It is historical design evidence, not a screenshot of the current head.
Historical rename UI, from the earlier mock-browser QA:
Current QA media is local only; the images above document the earlier implementation.
Validation
Browser-tested commit:
71cc03240237cc6972fca118b9838344ad6794a6. Product/unit/build checks ran on2ae78b37; the later commit only corrects the hover sequence in the browser regression.QA verdict: partial. All 28 focused browser regressions pass on
71cc0324. Two visual journeys pass on the identical product code at2ae78b37, covering rename/validation/hover card/narrow compact mode and agent disclosure/pending/failure/manual recovery. Ten screenshots and two videos are retained locally in/Users/tbrownio/.pane/closeout-2026-09-16/sidebar-evidence/; no new media was uploaded.Marker:
agent-e2e-551-20260916; dedicated port 4551; Electron APIs and agent outcomes are mocked. Real-agent/native restart and Windows/WSL checks remain manual. Product decisions are listed below. Detailed evidence and remaining checks. Hosted CI is green: Quality Checks + Smoke, macOS/Windows main suites, React Doctor and Socket. The wrapper matrix was path-filtered out. Full main suites: 1,037 passed / 2 skipped on macOS and Linux; 1,036 passed / 3 skipped on Windows. CI also passed 335 frontend tests, 47 preload routing cases and 29 functional smoke tests.Remaining decisions and manual QA
projects:createcallers, not restricted by transport to the desktop dialog. Clone/onboarding/RunPane registration do not send that flag. Confirm this authorization boundary.Two additive SQLite columns are migrated at startup (
name_manually_set,default_agent_launched_at). No new dependencies or release steps.