Repo History Training

Kitsoki treats repository history as training material for stories, not as ambient memory. A history lesson becomes useful only when it is represented as a typed case, armed by a deterministic oracle, and promoted into a story artifact: prompt text, Starlark glue, graph structure, flow fixture, agent eval, docs, or a selection policy.

The v1 contract is history_task.v1. It is intentionally lane-neutral:

• bugfix cases use the external bakeoff RED/GREEN oracle discipline.
• agent_eval cases use story-local eval datasets and comparators.
• flow_fixture, static_check, and artifact-comparator cases can be added without changing the corpus model.
• human_review_required is only a queue state; it is not autonomous training evidence until promoted into a deterministic check.

Validate committed or drafted cases with:

go run ./cmd/kitsoki history task-cases validate tools/history-training/examples

Adapt an existing bugfix-bakeoff manifest into lane-neutral task cases with:

go run ./cmd/kitsoki history task-cases adapt-bugfix \
  tools/bugfix-bakeoff/external/projects/query-string/manifest.yaml \
  --bug qs1,qs2 \
  --validate

make history-smoke runs that adapter for the selected bakeoff rows and writes the JSON review artifact under .artifacts/history-training/readiness/ before it performs preflight, RED/GREEN arming, handoff preparation, and repo-bakeoff flow validation. The command does not call a live LLM.

Manifest Contract

Each case records:

• source provenance (source.corpus_ref, repo, baseline/success refs);
• the target story and entrypoint;
• the trainable surface that would receive a promoted improvement;
• the operator-facing input or ticket;
• the oracle or comparator that arms the case;
• the cost policy for live cells;
• an artifact root under .artifacts/.

The validator rejects unarmed autonomous cases. For example, red_green requires baseline and success refs plus an oracle command, while static checks require a command or comparator. This keeps mining volume from becoming training noise.

Promotion Rule

A case is not "trained" when it is mined. It is trained only when the accepted outcome lands in a durable Kitsoki artifact and that artifact passes its deterministic validation gate.