agentfdr

Debugging infinite loops

Catch agents stuck in an "Edit → test fails → same edit" cycle that runs for 40 minutes.

Detecting goal drift

Spot when an agent asked for a bugfix instead refactored the router.

Analyzing token waste

Identify when huge tool results crowd the context window, causing cache misses and 200k token re-reads per turn.

Investigating bad landings

Understand why an agent says "Done!" with failing tests or never stops at all.

CI pipeline monitoring

Put a tripwire on autonomous pipelines with `agentfdr assert` to block loops or excessive token usage.

Cost tracking

Aggregate transcripts across projects to monitor per-model usage and estimated costs against subscription plans.

Post-mortem reporting

Generate Markdown autopsy reports ready to paste into issue trackers or Slack threads.

Timeline viewer

See the whole session on one screen—every turn's tool calls, context-window composition, output tokens, with prompt and compaction markers.

Anomaly detection

Automatically flags tool loops, error streaks, context bloat, token spikes, cache thrash, file churn, and refusals—one click from the evidence.

Turn dissection

A resizable side panel showing usage breakdown, each tool call's duration, result size, and snippet. Step through turns with ←/→ keys.

Live watch mode

`agentfdr watch` follows a session that's still running, showing context climb in real time.

CI gate

`agentfdr assert --no-loops --max-tokens 2M` exits with code 1 on violation, acting as a tripwire for autonomous pipelines.

Markdown autopsy

`agentfdr blame` renders the whole analysis as Markdown, ready to paste into an issue or Slack thread.

Usage tracking

Aggregates every project's transcripts into the same shape as subscription metering—current 5-hour window, daily history, rolling week—with per-model breakdown and estimated cost.

Budget calibration

Auto-detects your plan tier, set budgets, calibrate once against Claude Code's /usage screen, and get warning bars before hitting limits.