Problem

Sub-agent runs can sit in running state indefinitely with zero output. From the parent's perspective this is indistinguishable from a healthy slow start. The actual cause is openclaw:prompt-error with error: "aborted" being mapped to { status: "timeout" } — losing the distinction between startup abort and mid-run abort.

We observed 70 failing transcripts matching this pattern across agents (62 run_aborted, 8 startup_aborted, 6 empty_output_startup).

Root cause

The embedded run handler emits only three lifecycle phases: start, end, error. There is no first-progress signal, so the parent cannot distinguish a working child from a silently aborting one. A 45s abort looks identical to a 90-min hang until the wall-clock timeout fires.

Proposed addition (additive, non-breaking)

Two new lifecycle event phases:

1. first-progress — emitted on first qualifying tool call or non-empty assistant text
2. startup-failed — typed failure with reason: startup_aborted, startup_timeout, empty_output_startup, auth_failed

Both are additive. Existing listeners use explicit string comparisons and already silently drop unknown phase strings — confirmed from source inspection.

Update safety

• Additive only — no existing consumer changes required
• No breakage on update — unknown phases are silently dropped
• Plugin-SDK boundary not crossed
• External workspace watchdog (polling-based) works as interim without core changes

Evidence

70 failing transcripts captured from live corpus. Representative fixture: chunker/sessions/8cb62941. Full design with 20 test cases and 4-phase rollout strategy available.