robcsaszar/crucible

crucible は何をしますか？

Parallel domain recon → adversarial kill → plan → verify → SPEC → implement → PR.

File Ownership Rule

Subagents NEVER write files. They return results as text. You (the orchestrator) write every artifact. Files are the cross-phase handoff — subagent memory is not.

Setup

Parse the invocation for:

• Domains (required) — user-facing areas to audit, e.g. player, curator
• Scope hint (optional) — narrow focus within a domain, e.g. lobby SSE, role selection

Determine output directory: .crucible/run-<N>/ where N is the next unused integer.

Prior-run check: scan .crucible/ for existing run directories. Read each findings.json. Feed their confirmed IDs as a skip list into Phase 1 agents — don't re-discover known fixed issues; weight effort toward gaps.

Files written this run (all by orchestrator):

File	Phase	Role
codebase-map.md	1	Context injected verbatim into all downstream agents
recon-<domain>.json	1	Per-domain raw findings
findings.json	1→2→4	Merged findings; verdicts appended each phase
PLAN.md	3	Prioritized fix plan
SPEC.md	5	Before/after code + test cases per finding

Phase 1 — Recon

MANDATORY READ references/recon-phase.md before launching agents.

Launch one research agent per declared domain, all in parallel. Feed each: codebase path, domain name, scope hint (if any), prior known IDs to skip.

Agents return JSON text. If an agent returns prose instead of a JSON array, re-prompt once with the exact schema shape from assets/findings-schema.json. If it fails again, write an empty recon-<domain>.json ([]) and note the failure in codebase-map.md. Continue — one failed domain is recoverable; do not abort the run.

Collect valid responses, then:

1. Write recon-<domain>.json per agent response
2. Synthesize codebase-map.md — key file paths, domain boundaries, shared modules, conventions
3. Merge all domain findings into findings.json

Completion criterion: every domain has a recon-<domain>.json, codebase-map.md written, findings.json written. Do NOT load references/adversarial-phase.md until Phase 2.

Phase 2 — Adversarial Review

MANDATORY READ references/adversarial-phase.md before launching agents.

Launch one general agent per finding (batch findings from the same domain). Each agent gets the finding JSON and codebase-map.md — NOT the recon agent's reasoning. Independence boundary is the quality gate.

Agents return CONFIRMED / REJECTED / DOWNGRADED with code evidence. Append each verdict to the corresponding entry in findings.json. Run node scripts/validate-findings.cjs findings.json — fix any schema errors before Phase 3.

When the validator itself can't run — node not on PATH, the script or its schema missing, a usage error, or any failure before validation begins (Failed to load schema, Failed to parse JSON) — this is not a pass. Note ⚠ findings unvalidated — validator unavailable: <reason> in the findings output and at the top of the SPEC handoff, then continue. A schema rejection is the opposite case and always blocks: fix the findings, re-run, do not proceed on a rejection. Both currently exit 1, so read the error text, not the exit code.

Completion criterion: every finding has an adversarial_verdict field. Schema validation passes. Do NOT reload references/recon-phase.md from this point forward.

Phase 3 — Plan Synthesis

Zero-findings check: before synthesizing, count confirmed + downgraded entries in findings.json. If zero, write .crucible/run-<N>/REPORT.md with the single line "No actionable findings after adversarial review." Stop — do not open a PR.

Synthesize confirmed + downgraded findings from findings.json into PLAN.md:

• P1 (critical/high): broken core flows, data loss, permanent disconnection
• P2 (medium): user-visible errors, missing recovery paths
• P3 (low): polish, doc drift, copy issues

Group findings that share a root cause — one fix may eliminate several. When grouping: the highest severity among the grouped findings sets the group's priority. If grouped findings have conflicting fix approaches, split them — shared root cause does not require a shared fix. Declare fix order where dependencies exist. Rejected findings excluded entirely.

Completion criterion: PLAN.md written; every confirmed finding appears exactly once.

Phase 4 — Independent Verification

Launch one research agent per P1 + P2 finding, all in parallel. Each agent is told: "You did NOT write this finding. Verify every factual claim against current source code."

Each agent checks: file exists at path, line number matches described code, scope name is correct, pattern described actually appears. Returns VERIFIED / CORRECTED [field: wrong → right] / REJECTED [reason].

Apply all corrections to PLAN.md and findings.json. Run schema validation again. PLAN.md and findings.json must agree — no disagreement between deliverables.

Completion criterion: all P1+P2 findings verified. All corrections applied. Schema passes.

Phase 5 — SPEC

Write SPEC.md with one section per confirmed finding (P1 → P2 → P3 order):

• Exact file path + line range (verified)
• Before code (real source, not pseudocode)
• After code (concrete fix)
• Test cases: one sentence per behavior the test must verify

Run node scripts/validate-findings.cjs findings.json final check.

Completion criterion: every PLAN.md entry has a SPEC section. Validation passes.

Phase 6 — Handoff

1. Update flow docs (docs/flows/) for any finding that changes a user-facing behaviour
2. Invoke the implement skill with SPEC.md
3. Open a PR — this is the HITL gate; no intermediate approval is requested

PR body: link to PLAN.md, finding count by severity, link to SPEC.md, list of flow docs updated.

NEVER

• NEVER ask a subagent to write a file Instead: Have the agent return the content as text; write the file yourself. Why: Cross-phase state lives in files. If an agent writes directly, the orchestrator loses the canonical copy and cannot reconcile deliverables after Phase 4 corrections.

• NEVER give an adversarial agent the recon agent's chain of reasoning Instead: Pass only the structured finding (JSON) and codebase-map.md. Why: Seeing the recon agent's logic anchors the adversarial agent to the same assumptions — the independence boundary is what kills false positives.

• NEVER proceed past Phase 2 with schema validation failures Instead: Fix the failing fields in findings.json before synthesizing the plan. Why: A PLAN built on malformed findings produces a SPEC with wrong file paths and line numbers that the implement skill cannot use.

• NEVER write SPEC.md while any P1/P2 finding has verification: "pending" Instead: Run node scripts/validate-findings.cjs findings.json --pre-spec and resolve all failures before Phase 5. Why: The --pre-spec guard exists in the validator script but is invisible unless invoked — an agent that skips this call will SPEC unverified file paths and silently pass wrong line numbers to implement.