Summary

In `2026.4.26`, the cron lane fails with `Error: No credentials found for profile "anthropic:claude-cli"` when the persisted `auth-profiles.json` token has expired, even though the macOS Keychain (`Claude Code-credentials`) holds fresh OAuth tokens. This is despite `EXTERNAL_CLI_SYNC_PROVIDERS` including `CLAUDE_CLI_PROFILE_ID` (the supposed fix from #71026, commit `8bbb143ab87e`).

The original #71026 was about a stale status display; this is an actual runtime failure — the cron throws and model-fallback dead-ends (`reason=auth next=none`).

Repro

1. Let the persisted `~/.openclaw/agents/main/agent/auth-profiles.json` `anthropic:claude-cli.expires` go past now (i.e., wait ~8h after the last `claude /login`).
2. Keep `claude` CLI fresh — keychain `Claude Code-credentials` still has a valid `expiresAt`.
3. Trigger a cron whose model resolves to provider `anthropic` (e.g., `anthropic/claude-haiku-4-5-20251001`) routed via `order.anthropic = ["anthropic:claude-cli"]`.

Observed

``` [diagnostic] lane task error: lane=cron-nested … error="Error: No credentials found for profile \"anthropic:claude-cli\"." [model-fallback/decision] decision=candidate_failed reason=auth next=none detail=No credentials found for profile "anthropic:claude-cli". ```

`gateway.log` shows the gateway itself reads the keychain successfully on its own cadence (`[agents/auth-profiles] read anthropic credentials from claude cli keychain`), but the cron lane sub-context either doesn't reach that path or reaches it and the bootstrap is rejected silently.

Root cause hypothesis

`refreshOAuthCredential` in `oauth.ts` is a dead-end for `claude-cli` creds. The provider has no entry in pi-ai's OAuth registry (`getOAuthProviders()` only knows `anthropic`, `github-copilot`, `antigravity`, `gemini-cli`, `openai-codex`):

```ts const oauthProvider = resolveOAuthProvider(credential.provider); // "claude-cli" → null if (!oauthProvider || typeof getOAuthApiKey !== "function") return null; ```

So when the persisted credential is expired, the only escape is the keychain bootstrap. If that bootstrap fails for any reason in the cron sub-context (cache state, identity check, sub-process state), refresh returns `null`, `resolveApiKeyForProfile` returns `null`, and `resolveApiKeyForProvider` throws `No credentials found` — no fallback, no diagnostic.

Manual fix that worked

```bash

1. Read keychain

security find-generic-password -s "Claude Code-credentials" -w

2. Replace the anthropic:claude-cli profile in auth-profiles.json with {access, refresh, expires} from claudeAiOauth

3. Re-run cron → succeeds

```

I've also set up a launchd LaunchAgent to do this every 10 minutes as a workaround.

Suggested fix directions

1. Don't make refresh a silent dead-end — when `refreshOAuthCredential` returns `null` for a provider that has no adapter, surface a clearer error and explicitly re-attempt the keychain bootstrap.
2. Always prefer keychain at runtime on darwin for `claude-cli` — skip the "is local usable?" gate and adopt unconditionally when `EXTERNAL_CLI_SYNC_PROVIDERS` resolves fresh creds.
3. Add diagnostic logging when the bootstrap path is skipped (currently silent — `log$1.debug` only).

Environment

• OpenClaw `2026.4.26` (`be8c246`)
• macOS 25.3.0 (Darwin)
• Node `v22.16.0`
• Claude CLI keychain auth (`subscriptionType: max`)