1.6 KiB
1.6 KiB
| title | description |
|---|---|
| bd audit | Record and label agent interactions (append-only JSONL) |
{/* AUTO-GENERATED: do not edit manually */}
Generated from bd help --doc audit.
Record explicit agent/tool interaction audit entries in .beads/interactions.jsonl.
This optional JSONL sidecar is disabled by default. Enable it with:
bd config set audit.enabled true
Issue history is always recorded in the database and is visible with bd history <id> --events. The JSONL sidecar is for explicit interaction capture:
- auditing ("why did the agent do that?")
- dataset generation (SFT/RL fine-tuning)
Entries are append-only. Labeling creates a new "label" entry that references a parent entry.
bd audit [command]
bd audit label
Append a label entry referencing an existing interaction
bd audit label <entry-id> [flags]
Flags:
--label string Label value (e.g. "good" or "bad")
--reason string Reason for label
bd audit record
Append an audit interaction entry
bd audit record [flags]
Flags:
--error string Error string (llm_call/tool_call)
--exit-code int Exit code (tool_call) (default -1)
--issue-id string Related issue id (bd-...)
--kind string Entry kind (e.g. llm_call, tool_call, label)
--model string Model name (llm_call)
--prompt string Prompt text (llm_call)
--response string Response text (llm_call)
--stdin Read a JSON object from stdin (must match audit.Entry schema)
--tool-name string Tool name (tool_call)