Collaboration spawn: thread-store reports ENOSPC with 170 GiB free

Codex collaboration thread-store reports ENOSPC with local disk available

Summary

Creating a collaboration sub-agent fails immediately with:

collab spawn failed: thread-store internal error: No space left on device (os error 28)

The failure happens before the child agent starts or reads the workspace.

Reproduction

1. Use Codex desktop collaboration mode.
2. Request creation of a minimal read-only child agent.
3. Observe the error above.

Two attempts in the same parent thread failed identically. The second request was a
no-op probe: Read-only probe: report whether you start successfully. Do not inspect
or modify workspace.

Local environment evidence

| Check | Observed result |
|---|---|
| Date/time zone | 2026-07-29, America/Chicago |
| macOS root volume free space | 170 GiB available |
| macOS data volume free space | 170 GiB available |
| Inodes | approximately 1.78 billion free |
| Installed CLI | codex-cli 0.145.0 |
| ~/.codex/logs_2.sqlite | 43 MiB |
| ~/.codex/logs_2.sqlite-wal | 6 MiB |
| ~/.codex/goals_1.sqlite + WAL | 52 KiB total |
| External history persistence | disabled (history.persistence = "none") |
| OpenTelemetry exporters | disabled (otel.exporter = "none", metrics_exporter = "none") |

No matching thread-store, ENOSPC, or No space left on device record was found in
the recent macOS Codex app-log directory. No session transcript or application log is
included in this package because either can contain sensitive prompts, paths, or tokens.

What this rules out

• Host-disk exhaustion.
• Host inode exhaustion.
• Docker image/cache consumption as the direct cause (9.1 GB had already been pruned).
• A workspace-specific failure: the child fails before workspace action.
• A visible oversized local Codex SQLite file.

Hypothesis / requested investigation

thread-store appears to be a coordinator-side or app-internal storage layer distinct
from the visible local Codex SQLite files. Please inspect the collaboration spawn path
for a per-thread quota, failed temporary-store allocation, or an ENOSPC error being
misclassified from another storage backend.

Privacy

This report deliberately excludes transcripts, authentication material, repository
contents, and raw logs. A session ID can be supplied through in-app feedback if needed.

Follow-up diagnostic (Codex Doctor)

codex doctor (0.145.0) reports all visible Codex SQLite state databases healthy, but
reports two active rollout/session JSONL files with no usable header record. Both files
are zero bytes and were created on 2026-07-28. The doctor command has no repair option.

This makes malformed zero-byte rollout files a stronger local reproduction lead than
disk capacity. The next safe remediation is to preserve their metadata in this report,
move the two zero-byte files to Trash, restart Codex, and retry collaboration spawn.

Follow-up: codex doctor identified two zero-byte active rollout JSONL files. I moved only those empty files to local Trash, after which codex doctor --summary reports threads: rollout files and state DB thread inventory agree. A fresh collaboration-spawn probe still fails with the same thread-store internal error: No space left on device (os error 28). Therefore corrupted rollout metadata was repaired but is not sufficient to resolve the collaboration thread-store failure.