mirror of
https://github.com/openclaw/openclaw.git
synced 2026-05-24 15:11:04 +00:00
9e9feb52f4
The schema.help text for `models.providers.*.timeoutSeconds` documents the key as the user-facing knob for "slow local or self-hosted model servers". In practice the option is also the only configurable lever for the LLM idle/first-token watchdog. However `resolveLlmIdleTimeoutMs` was still running the explicit provider timeout through `clampImplicitTimeoutMs`, clamping it back down to the implicit ~120s `DEFAULT_LLM_IDLE_TIMEOUT_MS` ceiling for any non-cron, non-local provider. Consequence (matches #77744 and #78361): - User sets `models.providers.llamacpp.timeoutSeconds: 14400` (or 600 for a slow Gemini/Opus turn with a large tool payload). - Hot reload accepts the value, runtime resolves `modelRequestTimeoutMs = 14_400_000`. - Idle watchdog still trips at ~120s with "LLM idle timeout (120s): no response from model", aborting an otherwise-healthy upstream that is mid-prefill or buffering thinking tokens. Fix: when the caller passes an explicit `modelRequestTimeoutMs` (sourced from `models.providers.<id>.timeoutSeconds` / `model.requestTimeoutMs`), treat it as a deliberate ceiling for cloud providers too. The run-timeout / agent-timeout bounds still apply via `timeoutBounds`, so a shorter explicit run timeout always wins. The implicit default watchdog still kicks in when the user has not set a provider timeout, preserving the network-silence-as-hang guard for default configs. Updated the two corresponding test cases that asserted the old clamp-on-cloud behavior; all 71 tests in `llm-idle-timeout.test.ts` and the wider 430-test `src/agents/pi-embedded-runner/run/` lane pass. Schema help text refreshed to call out that the same knob raises the idle watchdog ceiling. Refs: #77744, #78361