research-auto

Run repeated Experiment cycles inside one active Package. Use research-run for a single known Experiment.

Preconditions

• one active Project and Package;
• one current active Direction;
• at least one current ExperimentSpec with a structured numeric gate;
• execution_mode == AUTONOMOUS;
• declared compute, time, and external-action boundaries.

If any precondition is absent, return the smallest owning handoff. Do not create parallel approval records or infer intent from the Control Panel.

Context

Start each cycle from:

python3 -m lib.research_state.cli \
  --workspace <workspace> \
  context <package-id> \
  --experiment <experiment-id>

Use the packet’s current versions, Rules, decisions, evidence summary, and next_refs. Fetch a referenced Run or note only when the current action needs it. Do not carry the full prior transcript into the next cycle.

Loop

read compact packet
  -> choose one READY ExperimentSpec
  -> verify implementation and resource readiness
  -> launch through lib.experiments.launch
  -> monitor structured status
  -> finalize evidence
  -> evaluate the exact declared metric
  -> route: continue | revise plan | halt

The launcher freezes current authority. The campaign never supplies its own Run context or weakens the verifier.

Plan changes

Use one RESEARCH_PLAN_UPDATE transaction when evidence justifies a Direction or ExperimentSpec revision. Include only changed nodes. The transaction must preserve earlier versions, Runs, and Evidence.

Pause for the user when a proposed change alters the Project charter, increases budget or external authority, requires a destructive action, or changes the terminal Package decision.

Deterministic routing

The helper exposes:

python3 skills/research-auto/scripts/conductor.py status ...
python3 skills/research-auto/scripts/conductor.py gate-eval ...
python3 skills/research-auto/scripts/conductor.py append-cycle ...
python3 skills/research-auto/scripts/conductor.py pack ...

Use the exact ExperimentSpec metric and operator. Never substitute another measurement because it is the only one present.

Routes:

• CONTINUE: a current legal Experiment remains;
• REVISE_PLAN: evidence supports a typed Plan update;
• HALT_GATE: the declared success gate is met;
• HALT_BUDGET: budget is exhausted;
• HALT_NO_CANDIDATE: no legal Experiment remains;
• HALT_BLOCKED: required authority, evidence, or compute is unavailable.

Completion

Stop when one route is durably recorded and no launched Run remains unreconciled. The user still owns the terminal Package outcome.