Skip to content

README.md

Latest commit

 

History

History
115 lines (83 loc) · 4.27 KB

README.md

File metadata and controls

115 lines (83 loc) · 4.27 KB
Raw

Cron Timeout Override Shim

Issue: #4808 — Lane B of #4755.

Problem

Non-trivial isolated agentTurn cron payloads time out per-provider during the OpenClaw sequential fallback chain. Observed: 5 providers × ~2.5min ≈ 13min exceeds each provider's per-call timeout for complex multi-step workloads (release-please pre-flight). The cron fails with FallbackSummaryError: All models failed (5).

Why this script exists

The root fix is upstream (openclaw/openclaw#95408 — per-agent model.requestTimeoutSeconds, Lane C, Hermes). Until that merges + ships + this host upgrades, we need a workaround on the Aegis side.

The workaround: bump models.providers.<provider>.timeoutSeconds for the 3 unique providers used by ag-hermes (the agent that runs the release-please cron). OpenClaw 2026.5.7 reads this knob at model-f6pqrkVH.js:348 (applyConfiguredProviderOverrides).

This script applies the override idempotently.

Why it's safe (global per-provider, not per-agent)

The OpenClaw 2026.5.7 schema only honors timeoutSeconds at the models.providers.<provider> level, not per-agent. Setting it raises the ceiling for every agent that uses those providers. This is acceptable:

  • Simple-payload crons (watchdog, qa-scan, sentinel) complete in ~30s, well under any reasonable timeoutSeconds value. The bump is invisible.
  • Outer cron-level bound (payload.timeoutSeconds) is unchanged. Each cron still has its own outer timeout (e.g., 120s for watchdog, 900s for release-please). Bumping the inner per-provider timeout doesn't extend those.
  • Cost ceiling is the same — the LLM call still pays per token, just gets more wall-clock before giving up.

The shim is documented as a workaround. Once Lane C merges + ships + this host upgrades, the override can be reverted by deleting the timeoutSeconds field from each provider in ~/.openclaw/openclaw.json.

Usage

# DRY-RUN (default) — show what would change
bash scripts/devops/add-cron-timeout-overrides.sh

# Apply the default 600s (10min) override
APPLY=1 bash scripts/devops/add-cron-timeout-overrides.sh

# Apply a custom timeout
TIMEOUT_SECONDS=900 APPLY=1 bash scripts/devops/add-cron-timeout-overrides.sh

# Apply to a subset of providers
TARGET_PROVIDERS="minimax-portal zai" APPLY=1 bash scripts/devops/add-cron-timeout-overrides.sh

# Non-default install path
OPENCLAW_CONFIG=/path/to/openclaw.json APPLY=1 bash scripts/devops/add-cron-timeout-overrides.sh

Default timeout: 600s (10min) — 4× the observed ~2.5min per-provider ceiling, giving headroom for ~2× LLM round-trip variance.

Re-enabling the release-please cron

After applying the override, the ad1ab50a-dba8-40e2-a3de-ca2d2d09dba5 cron (release-please dispatch) can be re-enabled. The current state has it disabled with sessionTarget: "session:agent:ag-hermes:..." (named session, from Hephaestus's prior failed workaround on the named-session lock-in bug).

The cron config update is manual at the ~/.openclaw/cron/jobs.json level. Two changes required:

  1. Set enabled: true
  2. Change sessionTarget back to "isolated"
  3. Update the prompt to a current release-please dispatch (the current one references issue #4708 and a 2026-06-16 memory file)

The cron daemon picks up the change on its next read cycle (< 60s).

Restart the OpenClaw gateway

The new timeoutSeconds takes effect on the next gateway reload. To pick up immediately:

openclaw gateway restart

Then trigger one manual isolated agentTurn run on ad1ab50a to verify the new ceiling holds.

Tests

scripts/devops/__tests__/add-cron-timeout-overrides.test.ts covers:

  1. DRY-RUN does not modify the config
  2. APPLY=1 sets timeoutSeconds on each target provider
  3. TIMEOUT_SECONDS env var overrides the default
  4. Idempotency (re-running is a no-op)
  5. Skip semantics (providers already at-or-above target)
  6. Scope (TARGET_PROVIDERS env var)
  7. Error paths (missing config, invalid timeout, malformed config)
  8. Partial success (missing target provider doesn't abort other updates)

Run with:

npx vitest run scripts/devops/__tests__/add-cron-timeout-overrides.test.ts