| name | refactoring-campaign |
| description | Drive a behavior-preserving refactoring campaign end-to-end through striatum — instantiate the three-stage workflow chain (goal selection, falsified plan gate, sliced execution), prepare and start each run, hand artifacts between stages, and stop at refusal gates. Use when the user asks to run a refactoring campaign, invokes /refactoring-campaign, or wants a goal-adjudicated refactoring with provenance against a registered target repository. |
Refactoring Campaign
Orchestrate the three-stage campaign preserved at
examples/refactoring-campaign/ in the striatum repo: stage 0 selects one
named goal (implementation-panel graph), stage 1 gates a falsified plan
(falsification-gate graph), stage 2 executes bounded slices with a
preservation review (code-change graph). Each stage is a separate striatum
run; each terminal artifact is the next stage's input contract.
This skill is the campaign driver. For run mechanics defer to the
generated skills: striatum-scaffold (prepare/branch), striatum-supervise
(lane driving), striatum-recover (wedges), striatum-claim-loop (packet
work). The daemon is a hard prerequisite for every verb.
Inputs
- target repo — a registered repository (default: current repo).
- campaign slug — short kebab-case name for this campaign.
- goal (optional) — if the user names a concrete refactoring goal,
skip stage 0 and author
GOAL_DECISION.md directly as an
operator-owned decision (owner: human — honest: the human named it).
- lanes (optional) — default multi-family: claude authors/synthesizes,
codex + agy on panels and reviews. See REFERENCE.md.
Quick start
scripts/instantiate.sh --slug <slug> --target-repo <path>
striatum --repo <path> workflow validate --allow-same-model-pairing \
striatum/workflows/refactoring-campaign-<slug>/stage-N-*/workflow.json
striatum --repo <path> run prepare --workflow <that path>
striatum --repo <path> run start --run-id <id>
scripts/wait-run.sh <run-id> 60 <path>
The campaign loop
- Stage 0 — goal selection. Run it; read
striatum/refactoring/<slug>/00-goal/GOAL_DECISION.md. If the
arbitrator refused all candidates → campaign stop, report why.
Skipped entirely when the user supplied the goal.
- Stage 1 — plan gate. Run it; read
01-plan/GATE_SUMMARY.md and
the adjudicator ledger. Cleared → continue. needs_revision exhausted
or gate refused → campaign stop. At most ONE fresh stage-1 rerun,
and only for substance failures, never for a refused gate.
- Stage 2 — execution. Before
run prepare, set
execute_slices.write_scope.allowed_paths from the committed plan's
blast radius and add path-shaped frozen surfaces to forbidden_paths
(workflows are frozen at prepare — edit first). Run it; read
02-execution/PRESERVATION_REVIEW.md and FINAL_REPORT.md.
- Integrate. On an
accept verdict, land the run worktree via the
serialized run.integrate (never hand-merge; merge_conflict is an
operator decision). Report the final report path and commit map.
Stop matrix
| Signal | Action |
|---|
| Stage 0 arbitrator refuses all goals | Stop; surface the refusal rationale |
| Stage 1 gate refused (undischargeable constraint) | Stop; this is the gate working |
Stage 1 substance needs_revision exhausted | One fresh stage-1 run max, then stop |
| Stage 2 ledger records an honest stop condition | Report it; the operator decides; do not stretch slices |
| Run wedges (stale lease, dead lane) | striatum-recover; never advance state by hand |
An honest stop with a truthful ledger is a successful campaign outcome,
not a failure. Drive autonomously between gates — do not insert human
confirmation prompts the gates don't require — and report outcomes
faithfully from daemon reads, never from assumptions.
Details, lane bindings, and driving discipline: REFERENCE.md.
