joenandez/caspar/caspar-clean

clean

End-to-end cleanup orchestrator. The primary agent owns scope, sequencing, test risk assessment, and final synthesis. Phase work runs in subagents using the existing focused skills: caspar-prune, caspar-test, and caspar-sweep.

Inputs

• $ARGUMENTS — optional scope hint: commit/SHA, unstaged/staged, context/session, task dir, or files.
• target_out_dir — optional OUT_DIR override.

Working Set

• Late-bind at runtime: branch = git rev-parse --abbrev-ref HEAD; OUT_DIR = target_out_dir || docs/tasks/{branch}.
• Resolve the same full working set used by prune/test/sweep from committed changes, staged, unstaged, and untracked files. If a provided ref/scope is invalid or ambiguous, stop and ask.
• Write/update {OUT_DIR}/working_set.json with scope, files, and the primary's P0-P3 risk tiers before dispatching test work.

Outputs + DONE

• caspar-prune cleanup completed by a subagent; manual-review items preserved.
• Primary-authored P0-P3 risk assessment and test plan recorded in working_set.json.
• caspar-test work completed by @caspar:tester test-lead subagents using the primary's risk plan.
• caspar-sweep completed by a subagent, including final hygiene, verification, and conventional commits.
• Final report includes: prune summary path, manual-review items, risk-tier counts, tests added/verified, lint/test status, sweep commit list, and any blockers.
• DONE when: each phase skill actually ran in a subagent; primary did the test risk assessment itself; no phase rules were duplicated or hand-written in place of loading the focused skill; manual-review items are surfaced; sweep either committed cleanly or stopped on an explicit blocker.

Method / guardrails

1. Resolve scope once. Establish files, task dir, and OUT_DIR. Keep dynamic details out of the prompt body; read them live.
2. Prune phase. Dispatch a prune-lead subagent instructed to load and execute Skill(caspar-prune) for the resolved scope. It returns cleanup edits, summary path, validation status, and manual-review items.
3. Primary risk assessment. After prune returns, the primary classifies every changed file P0-P3:

• P0 Critical: auth/payment/security/crypto/session/token, PII, permissions, user-data mutation, external API handlers, DB migrations, @critical.
• P1 Core: feature components, API handlers, state/business logic, fetch/cache, user-visible error paths.
• P2 Supporting: exported utilities, validators, transformers, adapters, hooks with real logic.
• P3 Skip: docs, styles, config, types, constants/enums, re-export barrels, pass-through wrappers, generated files.

Write a compact plan: - [P{tier}] {file}: {behavior or SKIP reason}.

1. Test phase. Dispatch @caspar:tester test-lead subagents in parallel. Each subagent loads Skill(caspar-test), consumes the primary risk plan for its batch, and does the test/verification work. P0 gets dedicated focus; P1/P2 may be grouped; P3 is skipped with reason.
2. Sweep phase. Dispatch a sweep-lead subagent instructed to load and execute Skill(caspar-sweep) on the resulting diff. Sweep owns final hygiene, verification, and commits.
3. Synthesize. Read phase returns and git state; report only the final state and blockers.

Guardrails:

• Do not inline the bodies of prune/test/sweep; call the skills.
• Do not let the primary perform prune edits, test authoring, or sweep commits except to recover from a surfaced blocker with user approval.
• CONFIRMED_SAFE cleanup may be applied; UNCERTAIN/UNSAFE cleanup stays untouched and appears in final manual review.
• --no-verify, lint/type suppressions, and forced green are forbidden unless the user explicitly permits them.

Handoff

If sweep committed successfully, report commit hashes/messages and suggest /caspar:rebase or push/PR. If blocked, report the exact phase and blocker, plus the manual-review list.

Escalate-If

• Scope is ambiguous, a ref is invalid, or no meaningful working set exists.
• A phase skill conflicts with this orchestration contract; surface the conflict instead of improvising.
• P0 coverage cannot be made behavioral and mutation-resistant.
• Sweep finds secrets/PII or cannot commit without bypassing verification.