All Articles
Every Tool Call Succeeds, Yet Nothing Moves Forward: Detecting Stagnation in Unattended Agents
No errors, yet the agent keeps replaying the same move while your budget quietly drains. Here is how to detect a success-but-no-progress loop using a progress oracle and action fingerprints, with a working Python implementation that halts safely.
A Silent Drop to a Weaker Model Is Scarier Than an Error: Designing a Capability Floor for Claude API Fallback
When a model becomes unavailable in an unattended pipeline, automatically dropping to a weaker model is dangerous. Drawing on years of running automated indie pipelines, this is how to use per-task capability contracts and a degradation budget to decide where to stop.
When Your Cowork × GitHub MCP Triage Quietly Drifts Each Run — Field Notes on Idempotency and Label Boundaries
Cowork + GitHub MCP issue triage looks perfect on the first run, then quietly breaks when left unattended: duplicate comments, reclassification churn, rate exhaustion. Field notes on idempotent prompts, label-boundary self-audits, and request budgets that keep weekly triage stable.
Stop Leaving a Static API Key in Your Unattended Jobs — Move to Short-Lived WIF Credentials
Claude Code is moving from static API keys toward short-lived, scoped credentials via WIF (Workload Identity Federation). Here is how to translate that idea to a small unattended pipeline so a leaked key has a much smaller blast radius — with working code and the failure modes to watch.
When an OAuth Token Expires, Your Unattended Run Has Nowhere to Go — A Token-Lifecycle Design That Keeps Remote MCP Alive
Remote MCP connectors are authorized via OAuth, but access tokens are short-lived. Interactive sessions can re-authorize in a browser; an unattended scheduled run has nobody to click the dialog. Here is a token-lifecycle design that owns expiry and refreshes ahead of time.
Will It Stay Light When You Run It Unattended? Observing and Capping Claude Code's Long-Session Memory
How to keep long, unattended Claude Code sessions from slowly getting heavier — with a tiny ps-based RSS sampler, a rolling-baseline watchdog, and session segmentation, shown with working scripts and a before/after comparison.
Designing the Give-Up Condition in Self-Repair Loops: Four Error Classes, Four Retry Budgets
LLM self-repair loops break on the fantasy that 'if you keep fixing, it eventually passes.' Classify errors into four classes, give each its own retry budget. Working TypeScript and real cost numbers included.
Logged as success, but it produced nothing — stopping silent failures in Cowork scheduled tasks with end-of-run assertions
A Cowork scheduled task exits 0, yet not a single artifact was produced. Trusting the exit code alone hides this silent failure. Here is how to turn your definition of done into end-of-run assertions that fail loudly with an evidence log.
When Claude API Streaming Stops Without an Error: Detecting Silent Stalls and Resuming Mid-Stream
How to catch the 'silent stall' where Claude API streaming stops with no exception at all, using a content-level watchdog that times the gap between tokens, plus a resume path that carries received text forward as an assistant prefill, and a four-layer timeout budget for long-running automation.
When You've Gone Five Steps in the Wrong Direction: Rewinding Conversation and Code with Claude Code's /rewind
Claude Code's /rewind is a safety net that returns both conversation and code to an earlier checkpoint. Here is how it differs from git, what it restores, and the pitfalls that bite in automated runs.
Stop the Bill Before It Balloons: Designing API Key Blast Radius for Unattended Pipelines
Designing for leaks instead of pretending they won't happen: workspace-scoped keys, zero-downtime rotation, and a usage watchdog that flags spikes with a rolling baseline and median absolute deviation — wired into a scheduled run.
When the Same Model Name Starts Behaving Differently: A Startup Canary for Unattended Pipelines
An in-place Opus upgrade can change your output, and an unattended publishing pipeline will never notice. Here is a lightweight startup canary that fingerprints behavior, catches drift, and halts the batch — with measured cost and latency.