cli-agent-diagnose

To diagnose AI agent CLI failures, you provide a structured trace JSON with command, stdout, stderr, and exit code, then run the classifier to identify the matching failure mode. This process returns an actionable workaround for the specific command error.

The best way to classify AI agent error modes is using a taxonomy-based diagnostic tool that parses command output signals and optionally applies LLM classification for ambiguous cases. This yields a specific failure mode and a skill patch to prevent recurrence.

You prevent recurring CLI failures by running a diagnostic classifier on failed traces to emit a memory update and a reusable skill patch. This workflow patches the agent's behavior to avoid repeating the same command execution mistake.

Yes, you can proactively intercept risky Bash calls using an optional Claude Code PreToolUse hook. This hook warns the agent before executing potentially risky commands, preventing failures before they happen.

No, the CLI failure classifier uses deterministic matching for clear command output signals and only optionally uses an Anthropic LLM classifier for ambiguous cases. This dual approach ensures accurate diagnosis without always requiring LLM overhead.

Deterministic trace classification is limited when command output signals are ambiguous or do not match known failure taxonomy entries. In these cases, the optional LLM-based classifier is required to infer the correct failure mode and workaround.