fix(client-runtime): imported threads that never ran no longer sit in Working as Waiting#33042

Problem

Imported Claude Code and Codex sessions become threads with a bound provider thread and no runs. Once such a thread is un-settled (or its snooze expires), the sidebar shows it in the Working shelf with the grey Waiting label indefinitely, although nothing runs in it. Reproduced on main ade473018d in the web app: import a project through /welcome → both imported threads land in Settled → Un-settle thread on one → a Working 1 shelf appears with that thread as Waiting. The same shared code drives the mobile thread list.

Change

The server reports status: "idle" for any thread without runs. On the client, shellRuntime (packages/client-runtime/src/state/models.ts) and deriveThreadStructureRuntime (packages/client-runtime/src/state/threadExecution.ts) returned no runtime for a thread that never ran only when no provider thread was bound, so an imported thread got a runtime with status idle. On the client, idle means "stopped with background work still open", which the Working shelf and the Waiting label pick up.

Both guards now return no runtime for any thread without a run, unless background work holds it. The web Workspace selector used that idle runtime to stay locked for a bound thread that has no messages yet (an imported ACP session); it now checks the bound provider thread itself, so that lock is unchanged. A never-ran thread with a pull request watch keeps parking at idle as before. Provider-native Codex, OpenCode and Grok subagent children, which also bind a provider thread without runs of their own, now match Claude and Cursor children, which already had no runtime; their status comes from the parent's subagent records, and they are never listed in the sidebar shelves.

Scope and approval

Fixes pingdotgg/t3code#33024

Verification

Tests. packages/client-runtime/src/state/entities.test.ts › keeps an imported thread that never ran out of Working fails on main (expected true to be false from isThreadWorking) and passes with the fix. threadExecution.test.ts › times a provider-native subagent from its runless root turn while it works now binds a provider thread on its runless projection; it fails (expected { status: 'idle', … } to be null) with only threadExecution.ts reverted. Both were re-checked failing with the fix reverted. The client-runtime suites entities, threadExecution, threadInbox, orchestrationV2Projection, subagentRuntime and threadSubagents pass. vp run check:changed on the final head passes: format, lint, knip, typecheck (7 packages) and tests. apps/web/src/components/ChatView.logic.test.ts › isThreadEnvironmentLocked locks an imported session bound before its first run; with the old lock condition that row returns false.

Real app, before (4c361199f4) and after (28ee80fa3a). Each side ran on its own fresh database. Synthetic Claude Code transcripts were imported through the server, so both imported threads start in Settled. Then one was un-settled. Web: Chromium 157 on Linux, DPR 2, font scale 1.0, clock fixed. Native: pooled iPhone 17 Pro simulator, iOS 27.2, 3×, dark. Diff0 marks every changed pixel; DiffT marks channel deltas above 16.

Recordings (before | after):

• Un-settle on web, 1440 dark: before | after. Settled at 0s, Un-settle at about 1s, new placement at about 2s.
• Un-settle on iOS, via long-press then Un-settle: before | after.
• Watch removed, web 1440 dark: before | after. Unwatch at about 2s; the after side shows the age at about 6s.

Round 2 added current-head web captures for the snooze-wake folder row, the opened thread, the held-only queue, the silent-compaction boundary, the stale-error boundary and the workspace choices: round 2 web index (seed, driver, recipe), before d1aeed2492, after a64a6511ee. Native round 2 (Android Pixel Tablet split-view sidebar with selection, collapsed shelf, shelf disabled and pin/unpin; Pixel 9 at font scale 1.3; iPhone 17 Pro content search, watch removal and Request placement): native index (seed, rpc driver, scripts). After the Workspace lock fix, on the final head: lock recapture index, before d313b2aff3, after 20ef152693: the un-settled import leaves Working (sidebar-only diff), the bound empty thread's Workspace selector stays locked and opens no menu (before | after), and the empty native child is pixel-identical. These supersede the round 2 A13 and A16 Workspace frames. Final rows on the same SHAs: final index (bundle): the silent-compaction boundary with no changed pixels outside the sidebar (before | after), a held queued run released into a real Replay run (before | after), and Un-settle on Android Pixel 9 at font scale 1.3 (before | after). Earlier evidence: the fixture matrix and the evidence indexes for web and harness, watch, opened thread, subagent child and snooze, and native. Seed and driver scripts: web, watch and opened thread (recipe), and native (recipe).

What else changed:

• Sidebar and mobile list: Working-shelf membership and counts, row position, and Waiting replaced by the age. Snooze expiry shows Woke and returns to the age after Dismiss.
• Row emphasis: the mobile row gets brighter text and a heavier title. On web, the shelf-to-list move changes the row's opacity and the truncation of compact titles.
• Unchanged at zero threshold: the opened imported thread (history and composer), an unwatched thread in Settled or snoozed, and a never-ran thread held by a pull request watch.
• The Workspace selector stays locked for a thread with a bound provider thread and no messages or runs, such as an imported ACP session before its first turn: the lock now checks the bound provider thread directly instead of relying on the idle runtime this PR removes (isThreadEnvironmentLocked in apps/web/src/components/ChatView.logic.ts).
• A bound runless thread carrying a stale provider-session error (labelled harness, not reachable through import) no longer shows the error banner and tooltip; the transcript moves down 8 CSS px because the timeline header spacing without a top banner is 24 instead of 16 (the banner itself is absolutely positioned); with the banner dismissed, the before transcript matches the after side exactly.
• A provider-native subagent child with no runs of its own (labelled harness): status, timer (running, completed, waiting, failed), parent relationship and the locked Workspace selector are unchanged.
• Native settled starting state: identical at threshold 16. At zero threshold, 173 pixels on the iOS glass toolbar edge differ by channel delta 1–2, the same on both takes. The endpoint diff is confined to the un-settled row and the rows it shifts. Only the status bar was normalised, identically on both sides.

Not verified:

• Native snooze acknowledgement and manual wake; only the native snooze expiry is captured (web covers the full wake sequence).
• Electron desktop, which wraps the web UI.
• Live sync with a real external pull request watch; the watch here pointed at a synthetic target.
• A live Codex subagent child.
• Native (iOS/Android) detail panes for the labelled harness states (runless child, held-only queue): the native app has no harness injection, so these are shown on web only; the shared client-runtime code and its tests cover both.
• Android Pixel Tablet sidebar pairs: zero-threshold diffs also light scattered glyph-edge antialiasing outside the sidebar; the threshold-16 diffs are confined to the changed rows.
• A thread with only held queued runs and the silent-compaction boundary: captured in a labelled web harness only (round 2 index), not through a live provider.

Review. Three rounds of two independent reviews. None found a defect in the two guards. Round 3 found that an imported ACP session (bound provider thread, no messages, no runs) lost its Workspace lock with the idle runtime; isThreadEnvironmentLocked now keeps that lock, with a test that fails on the old condition. The other findings were evidence gaps and body wording, addressed above. No sensitive area. vp run check:changed passes on 20ef152693.

AI-assisted; verified with the tests and tools listed above.