plat 067

← Pulse platform issue index

PLAT-067 — scheduled parent continues after its coding-agent terminal has disappeared

Coordination	Value
Assigned agent	unassigned
Ticket state	implemented — dead-transport dispatch stopped; tmux-disappearance root cause still open
Last synchronized	2026-08-10

• Priority: P0 — a successful child step can make the rest of a scheduled producing run impossible, while consuming its full timeout budget.
• Owner: scheduled coding-agent session recovery, background-completion delivery, and scheduler turn dispatch.
• Source workflow: RTS Latency.
• Source run/session: cron occurrence at 05:00 IST on 2026-08-10; schedule-cron--42eca39a_1786318255358919000.
• Related tickets: PLAT-004 (correct terminal run status), PLAT-020 (same-session continuation), PLAT-035 (retained turn settlement), PLAT-052 (known scheduler turns retain one native CLI), and PLAT-054 (do not expire live child work).

What happened

1. The scheduled parent dispatched step-ingest-notion-feedback as a background child at 05:48 IST.
2. The child itself completed successfully: it scanned 29 tickets, wrote its output, and its final validation completed.
3. At 05:53:11, delivery of that completion to the parent Claude Code session failed with context deadline exceeded; it was queued instead.
4. At 05:53:28, the parent session's tmux server was already absent (no server running).
5. At 05:58:38, the backend tried to prelaunch a restored Claude Opus session, waited five minutes for a prompt, and timed out. It nevertheless called StreamWithEvents for the next production collector turn.
6. That request produced no usable response and was cancelled at 06:26:12. Pulse then repeated the same failed restored-session pattern with Claude Sonnet.

The scheduler ultimately recorded the run as an error, which is correct under PLAT-004. However, the production voice collector and all later producing work did not run. The fault is not the Notion child or an RTS workflow artifact: the child completed before the parent transport disappeared.

Root-cause boundary

The exact reason the original Claude tmux server disappeared is not yet isolated. The proven defect is the recovery decision after that loss: a failed/missing parent transport is treated as usable enough to receive another scheduler turn. The next call then waits until cancellation rather than recovering or failing decisively.

Required behaviour

Before dispatching any next scheduled turn or queued child completion:

1. verify that the parent transport/pane exists and can accept input;
2. if it does not, attempt one same-conversation replacement/resume;
3. if that replacement does not reach a verified ready prompt, do not call StreamWithEvents for the next turn;
4. preserve the completed child's queued notification and report a specific recoverable scheduler error, with the failed transport identity and last terminal evidence;
5. retry/recover only the parent continuation, never re-run an already successful child merely to recreate its completion.

What shipped (2026-08-10)

The defect was a single swallowed error, not a missing recovery system. server.go:5850:

if handle, err := mcpagent.StartAgentTransportSession(agentCtx, underlyingAgent); err != nil {
    logfWithContext(queryLogCtx, "[CHAT_HISTORY] Failed to prelaunch restored ... : %v", err)
    // ← nothing else; execution continued to StreamWithEvents 57 lines later
} else if handle != nil && strings.TrimSpace(handle.TmuxSession) != "" {

Required behaviours 1 and 2 were already implemented. StartAgentTransportSession is the verify-and-single-replacement step: it relaunches the session with --resume and waits for a verified ready prompt. The five-minute timeout in the evidence below is that mechanism working correctly and concluding the transport is unusable. The platform therefore already knew the answer — it just stepped over it.

The fix honours that answer, using the same sendError(..., true) + return abort the very same function already uses for a StreamWithEvents failure 57 lines later:

sendError(fmt.Sprintf("parent_transport_unavailable: could not restore the coding-agent terminal for this session: %v", err), true)
return

This satisfies requirement 3 (never stream into a dead transport), 4 (a specific named parent_transport_unavailable error carrying the underlying transport failure), and 5 (the queued child completion is untouched, so recovery retries only the parent continuation and never re-runs a successful child).

Blast radius is confined to the case where all three already hold: restoredNativeCodingResume is set, the runtime requires a launchable terminal transport, and the relaunch has already failed — i.e. exactly the state where there is provably nothing to send to. Verified structurally: the new return sits in handleQuery with no intervening closure, so it aborts the turn identically to the existing reference pattern.

Deliberately not changed: the err == nil but nil-handle / empty-TmuxSession case also falls through today. It plausibly should not, but there is no evidence it occurs, and widening an abort on speculation risks converting this fix into an outage. Left for evidence.

Still open: why the parent tmux server disappeared in the first place. Unchanged by this fix — this stops the wasted hour and the killed schedule that followed it, not the disappearance itself.

Acceptance

A real scheduled run with a background child must either:

• deliver the child's completion and continue the same parent conversation; or
• stop promptly with an explicit parent_transport_unavailable / recovery-failed outcome.

It must not spend a further turn timeout on a dead session, and it must retain the child's successful result for the retry path.

Evidence

agent_go/logs/server_debug.log for the source session contains:

• 05:53:11 — Live steer delivery failed ... context deadline exceeded — falling back to queue;
• 05:53:28 — tmux capture failed: no server running;
• 05:58:38 — restored Claude Opus prelaunch timed out after five minutes, followed by a new StreamWithEvents call;
• 06:26:12 — that turn completed only after cancellation;
• 06:31:16 and 06:58:50 — equivalent failed Pulse resume/cancellation sequence.

The durable conversation record is workspace-docs/Workflow/rtslatency/builder/conversation/2026-08-10/session-schedule-cron--42eca39a_1786318255358919000-conversation.json.