Claude Code RETRY_WATCHDOG gives up after 3 timeouts: Sume jobs
Unattended Claude Code sessions using CLAUDE_CODE_RETRY_WATCHDOG now stop after three timeouts. Persist Sume job ids so a restart resumes, not resubmits.
The Claude Code 2.1.288 release notes in the changelog, read 2026-10-03, change how unattended sessions behave. Sessions using CLAUDE_CODE_RETRY_WATCHDOG previously retried for hours after a failed long stream; now they stream again and give up after three timeouts. The same release says non-interactive sessions and subagents continue from the partial response after a mid-response API timeout, and that MCP calls could run twice on a result over 16 MB or an unparseable one.
What this means for a Sume run
An unattended session that submits a Sume job and then waits can now end earlier than before. That is good for cost and bad if you assumed the session would wait it out. The job does not stop when the session does: the Sume jobs guide says a client timeout does not cancel the job and it keeps billing.
| Change | Effect on a Sume run |
|---|---|
| Gives up after three timeouts | Session may end while a job runs |
| Continues from a partial response | Do not assume the create call was repeated or skipped; check the job |
| MCP calls could run twice on a very large or unparseable result | Same reason to use an idempotency_key |
| Re-authenticate prompt for more OAuth scope | Unattended runs need the scope up front |
Make the session restartable
Write the job id to a file the moment the create tool returns it, along with the idempotency key. On a restart, read the file first. If it holds a job id, call jobs_wait or jobs_result instead of creating again. If a crash happened before the job id was written, repeating the create with the same idempotency_key returns the same job rather than a second one.
The Claude Code MCP docs list related limits: CLAUDE_CODE_MCP_TOOL_IDLE_TIMEOUT is 5 minutes for HTTP, and MAX_MCP_OUTPUT_TOKENS is 25,000. Keep result envelopes small with include_results only when you need them.
Checklist
- Choose the idempotency key before the session starts, for example from the batch id.
- Cap spend with
max_spend_usd. - Request
mcp:writeup front if the run must create, since an unattended session cannot answer a re-authenticate prompt. - Reuse the same ids in
jobs_waitafterwait_slice_expired. - Check balance and job state at the start of every resumed run.
Sources
Related posts
More in Agents
- Claude Sonnet 4.5 retires Nov 30: what to move to on Sume
Anthropic deprecated claude-sonnet-4-5-20250929 on Sep 30, retiring it Nov 30, 2026 in favor of Sonnet 5.5. Sume's catalog has no Sonnet 4.5 row.
- Claude thinking blocks are model-bound: one model per thread
Sonnet 5.5 and Fable 5.1 thinking blocks only work for the model (and account) that made them. Why an agent thread should keep one model, and how to log it.
- Codex 0.160 Guardian review reads prior instructions: Sume args
Codex CLI 0.160 adds opt-in Guardian review that can fetch earlier user instructions. Write Sume calls so a reviewer can approve them from the arguments alone.
- Codex config.toml enabled_tools and required for Sume's read tools
Codex's MCP config can allowlist tools with enabled_tools and fail startup with required. Give a task only Sume's read and wait tools and a longer timeout.
Written by Sume