Skip to content

CLAUDE.md

Latest commit

 

History

History
89 lines (67 loc) · 3.83 KB

CLAUDE.md

File metadata and controls

89 lines (67 loc) · 3.83 KB
Raw

consensus-loop (v2.5.1)

Claude Code hook plugin — cross-model audit gate with structural enforcement. edit → audit → agree → retro → commit.

Quick Commands

node --test tests/              # all tests
node --test tests/cl1-verify.test.mjs  # single file
FEEDBACK_HOOK_DRY_RUN=1 node index.mjs # dry run
rm -rf ~/.claude/plugins/cache/berrzebb-plugins  # clear cache

Module Map

index.mjs          ← PostToolUse: watch_file → audit trigger
  ├→ context.mjs   ← Config, paths, parsers, i18n (single source)
  ├→ audit.mjs     ← Background audit (Codex/GPT, pre-verification)
  ├→ respond.mjs   ← Tag sync (promote/demote)
  └→ retrospective.mjs ← Retro marker setup

session-gate.mjs    ← PreToolUse: blocks Bash/Agent until retro completes
session-start.mjs   ← SessionStart: handoff sync + resume detection
session-stop.mjs    ← Stop: handoff sync + auto-commit
subagent-start.mjs  ← SubagentStart: inject audit state into agents
subagent-stop.mjs   ← SubagentStop: worker completion + deferred retro
pre-compact.mjs     ← PreCompact: save audit state
post-compact.mjs    ← PostCompact: restore audit state
worktree-create.mjs ← WorktreeCreate: auto-configure + merge permissions
worktree-remove.mjs ← WorktreeRemove: preserve evidence + cleanup
teammate-idle.mjs   ← TeammateIdle: CQ gate (agent teams)
task-completed.mjs  ← TaskCompleted: done-criteria gate (agent teams)

Skills (9) + Agents (3)

Skill Shortcut Purpose
consensus-loop:orchestrator /cl-orch Distribute, track, correct, merge
consensus-loop:planner /cl-plan PRD + 10 design doc types
consensus-loop:verify /cl-verify 8 done-criteria checks
consensus-loop:merge /cl-merge Squash-merge + emergency rollback
consensus-loop:retrospect /cl-retro Memory extraction (Quick/Full)
consensus-loop:audit /cl-audit Manual audit trigger
consensus-loop:status /cl-status Current state check
consensus-loop:guide /cl-guide Evidence writing guide
consensus-loop:tools /cl-tools CLI for 9 MCP tools
Agent Model Role
implementer Sonnet Headless worker (worktree, completion gate)
scout Opus Read-only RTM generator (3-way)
ui-reviewer Sonnet Browser-based UI verification

All skills support headless mode — no interactive prompts in subagent context.

MCP Tools (9)

All available via CLI: node scripts/tool-runner.mjs <tool> --param value

code_map · dependency_graph · audit_scan · coverage_map · rtm_parse · rtm_merge · audit_history · fvm_generate · fvm_validate

Config

config.json — tags, paths, quality rules, hook toggles. See examples/config.example.json.

Key: plugin.locale, consensus.watch_file, consensus.trigger_tag/agree_tag/pending_tag, quality_rules[]

Key Patterns

  • Facade prompts: templates/*.md → references/{locale}/ policy files
  • Pre-verification: audit.mjs runs CQ/T/CC locally, injects {{PRE_VERIFIED}} into auditor prompt
  • Worktree chain: --watch-file propagates through index → audit → respond
  • Tier routing: T1 (micro, no worktree) → T2 (standard, worktree+audit) → T3 (complex, scout+full audit)
  • Policy as Data: edit references/*.md to change audit criteria, no code changes
  • Fail-open: all hooks pass through on error (no system lockout)

Resume (auto-recovery)

session-start.mjs auto-detects: stale audit.lock, pending verdicts, incomplete retro, active agents, compaction snapshot.

Tests

node --test tests/                          # all (context, hooks, MCP, FVM, enforcement)
node --test tests/hooks-lifecycle.test.mjs  # 17 hook tests
node --test tests/fvm-integration.test.mjs  # full FVM pipeline (5 roles)