crucible

Scaffold a new Crucible subagent at agents/<name>.md with role-appropriate frontmatter and tool grants. Roles: planner, reviewer, oracle, validator, analyst, generic.

Run the Final Oracle evidence audit independently against the current evidence/ tree. Convenes 3 Oracle auditors, computes quorum, writes decision.md. Idempotent — safe to re-run. PRD §1.16.2 CMD-3.

/crucible:forge in a refusal-driven retry loop. Runs forge; on REFUSED, parses REFUSAL.md, invokes /crucible:remediate, re-gates. Stops on COMPLETE or after --max-attempts (default 3). Iron Rule preserved at every iteration.

Scaffold a new Crucible slash command at commands/<name>.md with proper frontmatter (name, description, allowed-tools), pipeline section, and refusal-modes section.

Verify Crucible installation health. Compares the plugin manifest (plugin.json) against the Claude Code plugin record and reports drift. Verifies SDK reachability. Read-only diagnostic. PRD §1.16.2 CMD-5.

Print the DAG of a Crucible command or skill — which skills/agents/hooks fire, in what order, with what evidence outputs. Read-only education tool. Useful for "what does /crucible:forge actually do?" queries.

Idempotent auto-repair for common Crucible drift — regenerate stale evidence/INDEX.md files, sync plugin.json ↔ marketplace.json versions, re-create missing evidence subdirs, re-link orphaned trial directories. Read-write but safe (creates only; never deletes user content).

End-to-end Crucible pipeline. One command: codebase-analysis → docs-research → planning → oracle-plan-review → execute → validation → evidence-indexing → 3-reviewer consensus → 3-oracle quorum → completion-gate. Refuses on any cited blocker. The conductor that PRD §1.16.2 implied but never named.

Render the current evidence-tree state as a Mermaid graph — which MSCs are sealed (PASS), which are pending, which failed. Optional --run-id scopes to one forge run. Read-only.

Scaffold a new Claude Code hook for Crucible. Creates bin/<name>.sh with canonical stdin/stderr/exit-code protocol and patches hooks/hooks.json to register it. Supports SessionStart, PreToolUse, PostToolUse, Stop events.

Run Crucible's Comprehensive Planning + Execution mode. Builds a plan via the planner subagent, submits it for Oracle plan-review, executes only after approval, and seals every step into evidence/. PRD §1.16.2 CMD-1.

Read REFUSAL.md, produce a delta plan targeting ONLY the failing MSCs/blockers, execute it, and prepare the next forge iteration. Used by /crucible:autopilot but also runnable standalone.

Resume a halted /crucible:forge or /crucible:autopilot run by inspecting the evidence tree and continuing from the first missing phase artifact. Evidence-tree-as-state — no separate state file required.

Scaffold a new Crucible rule fragment under templates/rules/<name>.md, then recompose docs/CRUCIBLE-CLAUDE-MD.md so the new rule is included on next /crucible:setup. Reminder to re-install with --force.

Scaffold a new Crucible skill. Creates skills/<name>/SKILL.md with proper frontmatter, evidence-path conventions, refusal modes section, and an optional scripts/ subdir. Plugin reload required after to surface the new skill.

Bootstrap a project for Crucible. Creates evidence/ tree (16 standard subdirs + INDEX.md), .crucible/active sentinel, and runs /crucible:setup --local. The "first time using Crucible in this project" command.

Report the current completion-gate state. Pretty-prints the MSC table, reviewer consensus, Oracle quorum, and overall verdict from evidence/completion-gate/report.json. Read-only. PRD §1.16.2 CMD-4.

Run /crucible:forge inside a named trial subdirectory under evidence/robust-trials/trial-NN/. Fulfills PRD §1.13.5 FR-TRIAL-1..5 (≥4 trials, mix of planning + validation + SDK-driven). Use this when you need a labeled, isolated forge run.

Run Crucible's Validation-Only mode. Accepts a target artifact, claim, or codebase region; runs non-mutating validation; produces per-item PASS/FAIL with cited evidence. Refuses to issue a verdict when evidence is insufficient. PRD §1.16.2 CMD-2.

Use this subagent to build repo-wide context before any modification in Crucible's planning mode. Activate whenever a planning task starts, whenever the planner subagent needs a module map, or whenever a refactor task requires understanding existing code. Read-only — never modifies source. Always runs before the planner builds the executable plan. Outputs a structured evidence/codebase-analysis/ directory.

Use this subagent to fetch and cite current upstream documentation for every external dependency in scope. Activate before writing code against any SDK/framework/API/CLI, whenever training data may be outdated, or whenever a fact must be sourced rather than recalled. Produces evidence/documentation-research/ with raw markdown sources, ISO-8601 fetch timestamps, and a SUMMARY.md citing 3-5 verified facts per source pointing to local sources/ filenames. Refuses memory-only references.

Use this subagent as the FIRST of at least three Final Oracle auditors in Crucible's quorum-gated final evidence audit (VG-14). Oracle 1's emphasis is COMPLETENESS-AND-CITATION — do reviewer-consensus + every MSC have approved verdicts and citations? Activate when the final evidence audit phase begins. Read-only access to evidence/. Issues APPROVE or BLOCK with cited blockers. Never shares context with Oracle 2 or 3. Quorum requires ≥2 APPROVE.

Use this subagent as the SECOND of at least three Final Oracle auditors in Crucible's quorum-gated final evidence audit (VG-14). Oracle 2's emphasis is STRUCTURAL INTEGRITY — does every directory have README.md + INDEX.md, are gate-receipt files (vg0-* through vg15-*) all present, and does the report.json schema parse? Activate when the final evidence audit phase begins. Read-only access to evidence/. Issues APPROVE or BLOCK with cited blockers. Never shares context with Oracle 1 or 3.

Use this subagent as the THIRD of at least three Final Oracle auditors in Crucible's quorum-gated final evidence audit (VG-14). Oracle 3's emphasis is ADVERSARIAL SKEPTICISM — try to find what a hostile reviewer would point at to BLOCK completion. Activate when the final evidence audit phase begins. Read-only access to evidence/. Issues APPROVE or BLOCK with cited blockers. Never shares context with Oracle 1 or 2. Designed to catch what completeness/integrity audits miss.

Use this subagent for Crucible's comprehensive planning mode. Activate whenever a user invokes /crucible:planning, when a complex feature/refactor/migration is requested, or whenever execution must be Oracle-pre-approved. The planner consumes codebase-analysis and documentation-research outputs and produces an executable plan with per-step skill/subagent/hook attribution. Refuses to execute any step until the Oracle plan-review approves the plan.

Use this subagent as the FIRST of three independent verification reviewers in Crucible's three-reviewer consensus (VG-13). Reviewer A's emphasis is COMPLETENESS — for every Mandatory Success Criterion (MSC-1..MSC-21), does evidence exist at the cited path, and is it non-empty? Activate when reviewer-consensus is required. Read-only access to evidence/. Refuses to PASS any MSC without an evidence citation. Never shares context with reviewers B or C.

Use this subagent as the SECOND of three independent verification reviewers in Crucible's three-reviewer consensus (VG-13). Reviewer B's emphasis is INTEGRITY — does each evidence file's content actually match its claim? E.g., does session.jsonl contain real PreToolUse/PostToolUse messages, or is it a fabrication? Activate when reviewer-consensus is required. Read-only access to evidence/. Refuses to PASS without content verification. Never shares context with reviewers A or C.

Use this subagent as the THIRD of three independent verification reviewers in Crucible's three-reviewer consensus (VG-13). Reviewer C's emphasis is IRON-RULE COMPLIANCE — does any artifact, anywhere in evidence/, contain mocks, fakes, fixtures, or test files? Activate when reviewer-consensus is required. Read-only access to evidence/. Refuses to PASS if any Iron-Rule violation is detected. Never shares context with reviewers A or B.

Use this subagent for Crucible's validation-only mode. Activate whenever the user invokes /crucible:validation, asks to verify a deliverable, asks for a review without changes, or when planning hands off after execution. Iron Rule applies — NO mocks, NO fakes, NO fixtures, NO test files. The validator exercises the REAL system and captures real-system artifacts. Refuses to issue a verdict on insufficient evidence.

Build repo-wide context before any modification. Use this skill whenever starting a planning task in a real codebase, refactoring across multiple files, surveying module boundaries, identifying hot paths, or understanding existing code before changing it. Produces a structured evidence/codebase-analysis/ artifact (file inventory, module map, dependency manifests, hot-path identification). Read-only — never modifies source. Always runs before the planning skill in comprehensive mode.

Evaluate the completion gate — refuse on any missing criterion. Use this skill ONLY when invoked as the final step of a Crucible run, when VG-15 executes, or when a user attempts to claim completion. Reads the entire evidence/ tree, evaluates every Mandatory Success Criterion (MSC-1..MSC-21) against cited evidence, AND requires three-reviewer consensus PASS plus Oracle quorum APPROVED. Emits machine-readable evidence/completion-gate/report.json. Has NO override flag. NO force-complete. Refusal is a feature.

Deactivate Crucible enforcement in the current project. Use this when Crucible's hooks are blocking a session in a project that is not actively running a Crucible workflow, or when you want to step out of enforcement temporarily. This is the explicit opt-out — Crucible's hooks become silent no-ops. Removes the .crucible/active sentinel. Reversible via /crucible:enable. Does not delete any evidence/ artifacts.

Fetch and cite current upstream documentation for every external dependency in scope. Use this skill before writing any code that calls an external SDK, framework, API, or CLI. Use whenever training data might be outdated. Use whenever a fact must be sourced rather than recalled. Produces evidence/documentation-research/ with raw markdown sources, ISO-8601 fetch timestamps, and a SUMMARY.md citing 3-5 verified facts per source pointing to local sources/ filenames. Refuses memory-only references — every fact must cite a sources/ file.

Activate Crucible enforcement in the current project. Use this when you intentionally want Crucible's hooks (PreToolUse, PostToolUse, Stop) to enforce evidence-gated completion. Without activation, Crucible is silent in this project. This is the explicit opt-in step before starting any /crucible:planning or /crucible:validation workflow. Creates a sentinel file at .crucible/active in the project root. Reversible via /crucible:disable. Safe to invoke multiple times — idempotent.

Maintain README.md (purpose) and INDEX.md (artifact enumeration) in every evidence directory. Use this skill whenever new artifacts land in evidence/, whenever a gate completes and produces receipts, whenever a directory grows beyond 10 files, or whenever a reviewer/Oracle is about to inspect the evidence tree. Produces and refreshes the human-readable indexes that make the evidence package navigable. Refuses to leave any evidence directory un-indexed at gate-completion time.

Convene Oracle plan-review verdicts (pre-execution) AND final Oracle evidence-audit verdicts (post-execution, quorum-gated). Use this skill whenever the planning skill submits a plan for approval, whenever VG-14 of a Crucible run executes, or whenever a quorum decision is required. Spawns 3+ independent Oracle subagents in isolated contexts; aggregates verdicts; refuses completion if quorum (≥2 APPROVE, 0 unresolved critical blockers) is not met. Produces evidence/oracle-plan-reviews/ AND evidence/final-oracle-evidence-audit/.

Build the executable plan for a Crucible comprehensive planning task. Use this skill whenever the user invokes /crucible:planning, asks to plan a feature/refactor/migration, or requires Oracle-pre-approved execution. Builds on codebase-analysis and documentation-research outputs. Attributes every step to a skill, subagent, or hook. Submits the plan to Oracle plan-review BEFORE execution per FR-PLAN-5/6. Refuses to start execution without approval. Produces evidence/oracle-plan-reviews/ and evidence/robust-trials/trial-NN/ structured artifacts.

Locate and inspect the JSONL session log for a robust trial; emit per-trial line citations proving the required behaviors fired. Use this skill whenever a trial completes, whenever an Oracle requests behavior verification, whenever VG-11 of a Crucible run executes, or whenever a reviewer needs to confirm "did the hook actually fire?" Reads from the canonical Claude Code session log path (~/.claude/projects/ENCODED-CWD/SESSION-ID.jsonl). Produces evidence/session-logs/TRIAL-ID/INDEX.md with cited line numbers per behavior. Never edits session logs — fix the plugin and re-run instead.

Run validation-only mode — non-mutating verification of an artifact, claim, or codebase region. Use this skill whenever the user invokes /crucible:validation, asks to verify a deliverable, asks for a review without changes, or when the planning skill hands off after execution. Iron Rule applies — NO mocks, NO test-doubles, NO fixtures, NO test files. Validation exercises the REAL system and captures real-system artifacts (CLI stdout, screenshots, API responses, build outputs). Produces evidence/validation-artifacts/ with per-item PASS/FAIL verdicts and refusal-on-insufficient-evidence.

4 hooks across 4 events