[Juliusolsson05]

Problem and behavior

Fixes #746. Reduce repeated screen IPC by comparing recognized complete spinner/status lines, while retaining raw cached snapshots and emitting real content and picker changes. Duplicate screen strings are aliased only on the renderer wire and restored in preload.

Review corrections

Global timer/token/whitespace substitutions are removed. Duration, token-count, /rc and whitespace edits in normal content remain meaningful; multiline composer regions are preserved even when pasted text resembles a status line. Unknown chrome formats take the full path.

Successful session recovery seeds only the requesting renderer from the latest raw screen cache. A reloaded or newly attached local renderer no longer waits for a non-spinner change or idle backend repaint. Failed/conflicting recovery and destroyed requesters receive no seed. The global gate is not reset. Headless composer/condition parsing remains independent; Claude package #52 retains attribute-aware parsing on every emitted frame.

Verification and limits

26 focused gate, real-manager, recovery IPC and wire tests pass. Current main is integrated; full CI runs on the final head. No live app restart or FPS claim. Spinner/timer text in structured mirrors remains frozen between meaningful frames; the raw PTY terminal and process working status are unaffected. This is signature-based chrome recognition, not a general terminal rendering-state parser.

Final verification: both full host CI gates pass on 2e561bf (run 33938037152). Ready for owner merge confirmation; no live app restart.

[Juliusolsson05]

Review applied in the latest commit:

• Picker in the gate key: GatedScreenFrame.picker is compared (JSON of the small parsed state), so a selection move or close emits even on a spinner tick — test added.
• Supersession reset: screenFrameGate.forget now also runs at the lastGateEvaluation clearing site after this.sessions.set(...).
• Manager-level test (sessionManager.screenGate.test.ts): a spinner tick never reaches manager.on('screen') but updates getScreenSnapshot.
• Glyph set gains ✺; timer-regex scope documented; plan and PR body corrected (no forwarder/preload suites claimed; remote-client frozen-spinner limitation stated).

[Juliusolsson05]

Adversarial review (re-run after the pause; worktree node_modules present).

VERDICT: APPROVE-WITH-NITS

The gate's failure direction is safe (any diff outside the enumerated volatile shapes emits; a missing rule costs one extra frame, never a lost one), the rework items are complete (picker key screenFrameGate.ts:104-106 with tests; supersession reset sessionManager.ts:2852 after sessions.set plus the cleanup path at :862), and every wire/replay/redaction consumer traced handles the aliased form.

1. SHOULD-FIX — Renderer reload/late-subscribe during a thinking phase gets no screen frame until the agent's next real change. Pre-gate, the next 6 Hz spinner tick repainted a reloaded pane within ~166 ms; post-gate, runtime.screen/latestScreenRef (useIpcSubscriptions.ts:751,810) stay empty for minutes (DebugPanel/ReaderView fallback, TileLeaf Enter-baseline). Remote late-joiners are covered (RemoteServer.ts:200 pushes the cached raw snapshot); the renderer is not. Fix: push the cached raw snapshot through aliasScreenSnapshotForWire on window attach, mirroring the remote path.
2. NIT — trailing drop runs never journaled: a drop run only produces a session.screen.gate.dropped sample when a later emit closes it (sessionManager.ts:2688-2696); long-thinking sessions that end while thinking lose their final run — biasing the metric away from exactly the sessions it measures.
3. NIT — snapshotStats() documented "for the perf journal / tests" (screenFrameGate.ts:127-129) but nothing in production reads it; wire it to session-exit/quit journaling (would also fix #2) or drop the claim.
4. NIT — the picker key works only by runtime accident: AgentScreenSnapshot doesn't declare picker (session.ts:202-207) and GatedScreenFrame.picker is unknown (screenFrameGate.ts:74). Typing the real SlashPickerState would make drift a compile error.
5. NIT — ELAPSED_TIMER/TOKEN_COUNTER rewrite those shapes anywhere in the frame (screenFrameGate.ts:49-51); an in-place-updating figure inside agent output would freeze. Theoretical for current TUIs; one line under limitations.

Defect hunts that came up clean: spinner-adjacent-to-new-output always emits (non-volatile byte breaks key equality; tested); markdown-only diffs dropped matches pre-existing renderer semantics (useIpcSubscriptions.ts:793-801 already excluded markdown from invalidation); string aliasing can't break the renderer's === fast path (value comparison of primitives); replay unaffected (reconstructSlices.ts:173,276 no-ops session:screen); redaction caps keys only when present; remote server takes the full raw payload.

Commands: screenFrameGate + screenGate + wire + coalescer suites 15/15 PASS; sessionFeed 4/4 PASS; rendering replay 20/20 PASS; tsc -b clean.