Bug type

Behavior bug (background task uses wrong model/provider and creates repeated auth/model errors)

Beta release blocker

Possibly — this can silently create repeated background LLM failures after enabling the commitments feature.

Summary

After enabling the commitments feature on an OpenClaw install whose normal/default model is configured for Codex OAuth (openai-codex/gpt-5.5), the commitments extractor spawned repeated background lanes but attempted to use direct OpenAI provider auth (openai) instead of the configured Codex provider (openai-codex).

Because this install has Codex OAuth but no direct OPENAI_API_KEY, every extractor run failed with:

FailoverError: No API key found for provider "openai". You are authenticated with OpenAI Codex OAuth. Use openai-codex/gpt-5.5, or set OPENAI_API_KEY for direct OpenAI API access.

The feature produced repeated background auth/model failures rather than inheriting the agent/default model/provider or circuit-breaking after the first terminal auth/config error.

Steps to reproduce

1. Run OpenClaw with default/main model configured to a Codex OAuth model, e.g. openai-codex/gpt-5.5.
2. Ensure Codex OAuth is authenticated, but no direct OPENAI_API_KEY is configured.
3. Enable commitments / inferred commitment extraction.
4. Send normal chat messages that trigger commitment extraction.
5. Inspect gateway logs.

Expected behavior

Commitment extraction should either:

1. Use the configured agent/default model and provider (openai-codex/gpt-5.5 in this case), or
2. Have an explicit commitments model setting that supports provider-qualified model IDs, or
3. Fail once with a clear configuration error and circuit-break/disable extraction until configuration is fixed.

It should not silently fall back to direct openai when the system is configured for openai-codex OAuth.

Actual behavior

Gateway logs showed repeated commitments lanes failing on direct OpenAI auth:

[diagnostic] lane task error: lane=session:agent:main:commitments:commitments-20254e69-c966-424b-a43a-4042c05f507d durationMs=1317 error="FailoverError: No API key found for provider "openai". You are authenticated with OpenAI Codex OAuth. Use openai-codex/gpt-5.5, or set OPENAI_API_KEY for direct OpenAI API access."
[commitments] commitment extraction failed

[diagnostic] lane task error: lane=session:agent:main:commitments:commitments-07ac6deb-2ec0-4cf6-b3e3-cff0afaf23fb durationMs=1260 error="FailoverError: No API key found for provider "openai". You are authenticated with OpenAI Codex OAuth. Use openai-codex/gpt-5.5, or set OPENAI_API_KEY for direct OpenAI API access."
[commitments] commitment extraction failed

[diagnostic] lane task error: lane=session:agent:main:commitments:commitments-6c3a9bf4-347c-4860-bb1a-f67f0fe428cf durationMs=1236 error="FailoverError: No API key found for provider "openai". You are authenticated with OpenAI Codex OAuth. Use openai-codex/gpt-5.5, or set OPENAI_API_KEY for direct OpenAI API access."
[commitments] commitment extraction failed

This continued across multiple background lanes until commitments were disabled/restarted.

OpenClaw version

2026.4.29

Operating system

macOS arm64

Install method

Global npm / local gateway install

Model

Configured default/main model: openai-codex/gpt-5.5

The commitments extractor appears to use provider openai instead.

Provider / routing chain

Expected:

agent/main -> openai-codex -> gpt-5.5 -> Codex OAuth profile

Observed commitments background lane:

agent/main -> commitments extractor -> openai -> missing OPENAI_API_KEY -> repeated FailoverError

Additional context

This looks related to the broader background-run fragility cluster, but appears to be a distinct commitments-specific provider/model selection bug:

• Heartbeat model leaks into user session status/runtime model #74217 — heartbeat/internal runtime model leakage into user session status
• Feature: Reason-aware cron guardrails (quota/auth/rate-limit aware backoff + circuit breaker) #14376 — reason-aware cron/background guardrails for auth/quota/rate-limit failures
• github-copilot claude-opus-4.7: output_config.effort="low" rejected, poisons auth profile #70322 — provider/model request errors can poison auth profile availability
• [Bug]: abortable(activeSession.prompt()) creates zombie Agent loop when signal is pre-aborted #74859 — zombie background LLM loop after aborted run
• Gateway 7-second event-loop stalls on 2026.4.27 — createOpenClawTools synchronously rebuilds plugin registry + provider-auth env-var resolution per tool-creation pass #74860 — gateway degradation from repeated provider/auth/tool resolution
• [Bug]: Configuring a paid Google Gemini key implicitly turned existing heartbeat traffic into paid background usage #64129 — heartbeat/background work unexpectedly inheriting paid model usage

Impact

• Repeated background LLM/auth failures
• Log noise and potential gateway degradation
• Confusing auth diagnostics because the main model/auth path is healthy
• Users with only Codex OAuth may see commitments fail immediately even though normal chat works

Suggested fix direction

• Ensure commitments extraction resolves model/provider through the same configured model routing path as the owning agent/session, including provider-qualified IDs like openai-codex/gpt-5.5.
• Add a commitments-specific model config only if needed, but make provider family explicit.
• Add a circuit breaker for terminal background auth/config errors (missing_api_key, auth, format) so the extractor does not repeatedly retry an impossible provider path.
• Add regression coverage for Codex OAuth-only installs with no direct OPENAI_API_KEY.