Bug
Sub-agents spawned via sessions_spawn are accepted but never announce completion back to the parent session. This affects ALL providers — not model-specific.
Structured Test (Mar 19, 2026)
Spawned 3 sub-agents simultaneously with identical tasks (simple generation + file read + computation) to isolate whether this is model-specific or systemic:
| Model |
Alias |
Session Key |
Accepted |
Completed |
Announced |
| Claude Opus 4 |
opus |
76d34b79 |
✅ |
❌ Unknown |
❌ No |
| Claude Sonnet 4 |
sonnet |
bf76a90b |
✅ |
❌ Unknown |
❌ No |
| GPT 5.4 |
gpt54 |
889a1734 |
✅ |
❌ Unknown |
❌ No |
All three returned {"status": "accepted"} with valid session keys. None announced completion after 2+ hours.
Additional Observations
subagents list shows nothing — no active, no recent (even within 30min of spawn)
sessions_history on child sessions returns forbidden — Session history visibility is restricted to the current session tree
sessions_list with kinds: ["subagent"] returns 0 results
- A 4th sub-agent (GPT 5.4, spawned ~15min earlier for a coding task) also failed identically — accepted, never completed
This Is a Regression
Sub-agents worked reliably as recently as Mar 17-18. Multiple GPT 5.4 sub-agents completed successfully (Bybit logger fix, bot code reviews). Something changed between Mar 18 and Mar 19.
Possibly Related
Impact
Sub-agents are a core workflow for us — code reviews, parallel research, bot audits, coding tasks. With 4/4 spawns failing silently, the feature is effectively broken.
Environment
- OpenClaw: 2026.3.13 (61d171a)
- Channel: Telegram
- Parent model: Claude Opus 4 (also tested from GPT 5.4 session)
- Sub-agent models tested: Opus, Sonnet, GPT 5.4
- Host: macOS, M1 Ultra
Expected
Sub-agents should either:
- Complete and announce results back to parent, OR
- Fail with a visible error in the parent session
Actual
Sub-agents accept the spawn, then vanish. No completion, no error, no trace. Parent session has no way to know what happened.
Bug
Sub-agents spawned via
sessions_spawnare accepted but never announce completion back to the parent session. This affects ALL providers — not model-specific.Structured Test (Mar 19, 2026)
Spawned 3 sub-agents simultaneously with identical tasks (simple generation + file read + computation) to isolate whether this is model-specific or systemic:
opus76d34b79sonnetbf76a90bgpt54889a1734All three returned
{"status": "accepted"}with valid session keys. None announced completion after 2+ hours.Additional Observations
subagents listshows nothing — no active, no recent (even within 30min of spawn)sessions_historyon child sessions returns forbidden —Session history visibility is restricted to the current session treesessions_listwithkinds: ["subagent"]returns 0 resultsThis Is a Regression
Sub-agents worked reliably as recently as Mar 17-18. Multiple GPT 5.4 sub-agents completed successfully (Bybit logger fix, bot code reviews). Something changed between Mar 18 and Mar 19.
Possibly Related
ac720778) showedstatus: failedwith runtime of 396ms — suggesting it crashed immediately rather than running and failing to announce.Impact
Sub-agents are a core workflow for us — code reviews, parallel research, bot audits, coding tasks. With 4/4 spawns failing silently, the feature is effectively broken.
Environment
Expected
Sub-agents should either:
Actual
Sub-agents accept the spawn, then vanish. No completion, no error, no trace. Parent session has no way to know what happened.