plat 029

← Pulse platform issue index

PLAT-029 — stale live metadata reconnects forever to a missing tmux

Coordination	Value
Assigned agent	Codex
Ticket state	implemented
Last synchronized	2026-08-04

• Priority: P0
• Owner: terminal live-attach lifecycle
• Evidence: the Social Media main terminal was recorded as state=completed, process_state=live, snapshot_kind=live, but tmux has-session proved its named session no longer existed. The frontend kept opening the live-stream websocket and appeared stuck instead of rendering the retained transcript as a scrollable archive.
• Root cause: the stream resolver trusted an in-memory terminal snapshot without verifying its tmux target. Missing-target verification existed only for restart recovery, after the in-memory snapshot was absent. The periodic watchdog was not a sufficient request-time correctness boundary.
• Implementation: before upgrading a stored terminal to a websocket, the backend now runs a bounded tmux has-session probe. A missing target is immediately reconciled through MarkProcessClosed, which retains content, clears the tmux identity, and publishes process_state=closed plus snapshot_kind=archived. An indeterminate probe returns 503 without opening a hanging websocket. Cross-site origin rejection now happens before lookup or reconciliation, so a forbidden request cannot inspect or mutate terminal lifecycle state.
• Primary files: agent_go/cmd/server/terminal_live_attach.go and agent_go/cmd/server/terminal_live_attach_test.go.
• Verification: the backend regression starts from a stored live snapshot, makes the tmux probe return can't find session, expects HTTP 410, and proves the retained transcript survives as a closed archive with no tmux target. The focused resolve/stream suite passes.
• Acceptance: a dead tmux can never be upgraded to a websocket. Selecting its terminal settles to retained static content rather than reconnecting indefinitely. Runtime confirmation requires a rebuilt server; development did not restart the user's running backend.