Empirical max_turns sizing for
delegate_task (default
10 — fine for read-only Q&A but tight for writes).
The interacting cap is [backends.claude] timeout_local:
~10 s per turn, so max_turns=80 wants
timeout_local ≥ 800.
5–1015–2550–80100+
A run that hits the cap surfaces as
code=exit_nonzero: claude -p exit 1: (no stderr) with
duration much less than timeout_local. Diagnose by
checking the session JSONL — see incident-playbook entry #1.
No delegated jobs yet.
Call delegate_task(..., mode="async")
from an MCP-connected agent to enqueue work — it lands here
while the worker processes it.
| Project | Task | Inbox | Status | Duration | Age | |
|---|---|---|---|---|---|---|
| writes | ||||||
|
Loading full payload…
job_id:
cid:
model:
deliverable:
outputerror |
||||||