plat 351

← Pulse platform issue index

PLAT-351 — Retained-turn observer settles on durable runner outcome when the pane never reports idle

Field	Value
Status	implemented and pushed; live acceptance pending (localhost restart)
Priority	P1 retained-turn correctness
Owner	server agent runtime
Reported	2026-09-22
Related	PLAT-179 (shared finalResponse tool-call guard), PLAT-346 (retained lifecycle timers)

Problem

observeRetainedMainTurnStream settled a retained main-agent turn only on llmproviders.CodingAgentPaneReady pixels, with a stability gate. That dispatch has no Muse case and returns false forever for muse-cli, and the observer loop had no deadline and never consulted the provider's durable turn sidecar (it only did so in handleRetainedMainTurnStreamClosed, after the tmux control stream died).

A quiet-but-finished Muse retained turn therefore stayed in retainedMainTurns indefinitely: the session stayed busy, the global activity monitor showed a false "running" (1hr+ in the reported News Monitor case), and no log line distinguished it from a genuinely active turn. The provider's own completion contract already existed and was correct — museRetainedTurnReady (5s log quiet + TUI at prompt + pane stable, with pending-question auto-answer serviced first) feeds musecli.ReadRetainedTurnMessages → retainedturn.FinalResponse — but nothing on the live path asked it.

Fix

• After retainedMainTurnDurableQuietWindow (15s) of tmux-output quiet with a never-idle pane, the observer consults retainedturn.FinalResponse (at most every retainedMainTurnDurableRecheckWindow, 5s) and settles completed on runner outcome instead of pixels. The completion event carries the sidecar's final text via the existing emit path.
• Providers without a retained-turn reader return "" and keep exact pane-only behavior. Silent long tool runs cannot settle early: while a tool is in flight the newest assistant message carries the ToolCall (PLAT-179 guard) or the TUI is not at its prompt, so the sidecar stays empty.
• The live-loop settle is additionally allowlisted to readers that assert completion (Muse, Codex, Claude, Cursor). Pi is excluded: its reader returns all in-turn messages with no completion gate, so a question awaiting a slow user would settle early. Follow-up: gate Pi's reader on completion (or a provider-contract capability flag) and re-admit it.
• One diagnostic WARNING per observer at 30m when neither pane-idle nor a durable final response has fired. It never settles the turn: marathon turns are untouched by design (no hard observer deadline).
• Consolidated the two inline final-response seam blocks (stream-closed reconcile, completion emit) onto one retainedTurnFinalResponse helper. Windows are vars, not consts, so tests can shrink them.

Commit: 7e887ec3d.

Verification

• New TestRetainedMainTurnSettlesOnDurableFinalResponseWhenPaneNeverIdles: Muse pane stays busy forever; an empty sidecar keeps the turn open (consult observed), then a durable final settles it to completed with unified_completion, live tmux kept. Passes.
• Full Retained|LiveAttach|LiveInput|MCPAgentSession|TerminalStream|TmuxStream set passes; go vet and gofmt clean.
• Full cmd/server suite shows only the 3 pre-existing failures, byte-identical on the pristine tree (a 4th flaked once under load and passed on rerun with the change).
• Live acceptance pending: restart localhost, confirm the next quiet Muse retained turn settles within ~15s of its final response and the global monitor clears, and that a still-unsettled turn emits the 30m warning with provider/elapsed/quiet.