Headless and scheduled runs
Agent jobs that work in your terminal and die under cron, CI, or a background worker. Credential scope, TTY assumptions, and the launchd path on macOS.
Agent Rescue
Your agent setup is broken. Let's fix it on a call.
Claude Code, Codex, OpenClaw, MCP. When the install fails, the tool calls silently do nothing, the scheduled run dies but the manual one works, or the bill arrives and nobody can explain it. You share your screen, I find the cause, we fix it while you watch.
Direct answer
Agent Rescue
Rescue a broken agent setup by isolating the cause before changing anything: confirm what the process can actually see (credentials, stdin, config path), then check auth end to end, then verify the failing step in isolation. Most of these failures are silent by design — the tool reports success or nothing at all — so the fix is diagnosis first, not another restart. Book the fixed-scope 60-minute rescue session if it is blocking you today.
Fast triage
What we check first
Silent tool and config failures
Tool calls that never fire, instruction files the agent is not reading, and nested config quietly overriding the rules you wrote.
Restore access
Recover dashboard, chat, gateway, tunnel, and VPS access without turning the session into a broad rebuild.
Token and auth repair
Trace missing or mismatched tokens across environment variables, browser storage, config, and service restarts. Check OpenAI, Codex, Claude, or Ollama auth enough to prove model access works end to end.
Cost and limit surprises
A bill that does not match your plan, caching that is not caching, context that refills every turn, or a retry loop paying full price on the failure path.
Multi-agent collisions
Parallel agents overwriting each other, a lead that reports before teammates finish, and staging that sweeps up work the agent never wrote.
Workspace safety
Avoid destructive fixes. Verify project files are present and separate your application issue from the agent issue.
Recovery notes
Leave with what broke, what changed, restart commands, known risks, and what to check if the same symptom returns.
Before the call
Bring the symptoms, not the secrets
Have these ready and redact tokens before pasting logs anywhere.
- The exact command you ran, what you expected, and what happened instead.
- The last time it worked and what changed right before it failed.
- Which tool and version: Claude Code, Codex, OpenClaw, or an MCP server.
- How it runs: interactive terminal, cron or launchd, CI, Docker, or a VPS.
- Redacted logs for the failing step, plus any provider or gateway errors.
- A list of anything you do not want changed during the session.
Choose the right path
Related OpenClaw help
Debug OpenClaw
Use this when the install half-works and you need a structured live diagnosis.
Agent setup and training
Use this for new installs, model routing, channel setup, and training.
Works in your terminal, fails in cron
Start here if your job only breaks when it is scheduled.
Gateway token error
Start here if your exact symptom is a missing or mismatched gateway token.
FAQ
Common questions
What is an agent rescue session?
It is a fixed-scope 60-minute remote session for broken or unstable agent setups: Claude Code, Codex, OpenClaw, and MCP servers. Scheduled runs that fail headless, silent tool-call failures, config that is being ignored, gateway and token problems, auth, cost surprises, and multi-agent conflicts.
Why is it not OpenClaw-only any more?
Because the failures are not OpenClaw-specific. The same causes — credentials unreachable outside your login session, a CLI blocking on a terminal that is not attached, config loaded from the nearest file rather than the one you edited — show up identically across Claude Code, Codex, and MCP servers. Scoping the session to one tool meant turning away the same bug wearing a different name.
Do you need access to my machine or my repo?
No. The session is a screen-share and you drive. Nothing is installed for you and no repository access is required. If you would rather not share a screen at all, send logs and config and take the async written diagnosis instead.
Is the fix guaranteed?
The cause is guaranteed or you do not pay. The fix is best effort, because some causes sit in a vendor bug or in infrastructure neither of us controls. If we cannot fix it in the session, you leave knowing exactly what is broken and what the options are, and you are not charged if the cause was never identified.
What if we run out of time?
If it is still open at the end and it is the same problem, we keep going or finish it async at no extra charge. A second session is for a second problem, not for the first one taking longer than an hour.
How is rescue different from setup?
Rescue starts from a broken existing install and focuses on recovery. Setup starts from a new or planned install and focuses on getting the environment configured correctly.
Ready when this is worth fixing live
Book Agent Rescue Session
Book through CloudYeti. We will use the time to stabilize, isolate, or document the next exact recovery step.
Book Agent Rescue Session