Swarm workers: non-Claude-route spawns die at boot; headless workers die mid-run reported as 'unknown done'

Bug report: swarm workers die at boot / mid-run (v0.52.0)

Session: coordinator session_bear_1784454767013_cc7393423d202025, 2026-07-19, linux x86_64.

Symptom 1: non-Claude-route spawns die at boot

Every swarm spawn on a non-Claude model route died immediately with journal stuck at 4K (only the initial user message, msgs=1):

• horse: gemini-3.5-flash-medium (agy openai-compatible) — status "failed"
• koala: gemini-3-flash-agent (Antigravity) — stalled, never failed officially
• turkey: gpt-5.5 (openai-oauth) — status "failed"
• parrot: gemini-3.1-pro-low (Antigravity) — stalled at boot
  Meanwhile claude-sonnet-4-6 (agy + Antigravity) and claude-opus-4-6-thinking spawns booted fine every time.

Symptom 2: headless Claude workers died mid-run without status

First wave spawned with spawn_mode=headless: frog/lion/wolf (all sonnet/opus) worked for ~30 min (journals 60-170K, real edits in the working tree) then processes vanished. await_members reported them "done" with status "unknown" instead of failed. Second wave without spawn_mode=headless (inline default) ran fine to completion.

Symptom 3: await/summary UX issues

• "All members done" fired while statuses were all "unknown" and work was demonstrably unfinished (uncommitted mid-edit state).
• swarm summary/stop on spawned-child session IDs returns "not in the same swarm as requester" / "Unknown swarm session" even for sessions this coordinator spawned.
• bg wait latest=true finds no task for a background await_members watch.

Start by reproducing the non-Claude boot failures and headless exits on Linux, then trace the spawn_mode=headless, await_members, swarm summary/stop, and bg wait paths. Done means non-Claude workers boot reliably, headless workers report failure or completion accurately, and await/summary recognize spawned sessions without declaring unfinished work done.