bd audit record|label — an append-only log of agent interactions at .beads/interactions.jsonl, “intended to be versioned in git”, with two stated purposes:
- auditing (“why did the agent do that?”)
- dataset generation (SFT/RL fine-tuning)
An entry is one event with a kind (llm_call, tool_call, label), and the fields to match: --model, --prompt, --response, --error for a model call; --tool-name, --exit-code for a tool call; --issue-id to tie it to the work. bd audit label <entry-id> --label good --reason "…" appends a new entry referencing the parent — labels never mutate history.
The second purpose is unique in this wiki. No other tool here captures its own execution traces as training data. Every framework’s answer to “learn from what happened” is prose for a human or a later agent to read — ce-compound’s solution docs, gstack-learn’s corpus, bmad-retrospective’s action items. Beads records structured, labelled prompt/response pairs keyed to the work item they were produced for, which is a fine-tuning dataset accumulating as a side effect of ordinary work. That is pattern-knowledge-compounding pointed at the model rather than at the agent’s context, and it is the sort of thing only the state layer can do, because only the state layer is still there afterwards.
Its pattern-evidence-before-claims edge is the audit half: a claim an agent made is checkable against the recorded call that produced it. Beads records; it does not verify — contrast bernstein’s janitor, which “does not trust agent claims—verifies them”. Commits carry a matching Agent-Signature: trailer with model and reasoning metadata (or unknown-model when unavailable).
Maps to no canonical SDLC stage.
See Also
- beads-comment — the events journal’s
actorcolumn, the same attribution question inside the database. - bernstein — verification rather than recording.
- pattern-knowledge-compounding — compounding aimed at context; this aims at weights.