v0.4.4 — Parliament Protocol, Implementation Loop, Daemon Control Center

Parliament Protocol (v1-v6)

• 3-role diverge-converge deliberation (advocate/devil/judge)
• Meeting log accumulation → convergence → CPS (Context-Problem-Solution)
• Amendment propose/vote/resolve with majority voting
• Confluence verification (4-point integrity check)
• Normal Form convergence tracking with regression detection
• Blueprint naming linter (blueprint_lint — 21st MCP tool)

5 Enforcement Gates

Gate	Blocks When	Bypass
Amendment	Pending amendments unresolved	--force
Verdict	Latest audit != approved	--force
Confluence	Verification failed	--force
Design	Design artifacts missing	--force
Regression	Normal-form stage regressed	alert only

Implementation Loop

• quorum orchestrate plan — interactive Socratic planner (replaces interview)
• quorum orchestrate run — full WB execution: agent → audit → approve → next → retro → merge
• Auto-plan from CPS if no WBs exist

Daemon Control Center

• ParliamentPanel: live sessions, convergence, amendments, normal form
• AgentChatPanel: multi-pane interactive agent relay (select, pin, type, send)
• ProcessMux.split(): backend-abstracted pane splitting
• Version auto-synced from package.json

CLI

• quorum parliament — deliberation with --rounds, --mux, --resume, --history
• quorum agent attach — interactive terminal relay to mux sessions
• quorum setup — locale selection (en/ko)
• Deprecated: interview (→ orchestrate plan), ask (no quorum context)

Infrastructure

• MuxAuditor: Auditor interface backed by ProcessMux (tmux/psmux)
• Planner Phase 0: CPS auto-intake before intent capture
• Design Phase mandatory for CPS-origin tracks
• i18n: all hardcoded messages → locale keys
• npm run version-sync — auto-sync package.json version to plugins
• 7-component fitness score, 12-factor trigger scoring

1004 tests, 0 fail.

v0.4.0 — Autonomous Quality Loop, Blast Radius, Structured Orchestration

What's New

95 files changed, +10,168 -2,388. Three-pass /simplify review (reuse, quality, efficiency) — all clean.

AST Hybrid Scanning

• Regex first pass (speed, <1ms/file) → TypeScript Compiler API second pass (precision, <50ms/file)
• ast-analyzer.ts: 5 single-file analyzers + 2 cross-file analyzers
• Program mode: unused export detection, import cycle detection via dependency graph DFS
• ast-bridge.mjs: fail-safe MJS↔TS bridge for hybrid scanning
• perf_scan is the first hybrid tool (regex detects while(true), AST verifies break/return)

Fitness Score Engine

• 5-component quality metric (typeSafety, testCoverage, patternScan, buildHealth, complexity)
• FitnessLoop gates LLM audit with 3 decisions:
  • auto-reject: score drop >0.15 → skip LLM (cost savings)
  • self-correct: mild drop (0.05–0.15) → warn agent
  • proceed: stable/improved → continue to audit
• Daemon FitnessPanel shows real-time fitness trends

Blast Radius Analysis

• BFS on reverse import graph (inEdges) for transitive dependent discovery
• buildRawGraph() extracted from dependency_graph for reuse (5s TTL cache)
• 10th trigger factor: ratio > 10% → score += up to 0.15 (auto-escalation)
• Pre-verify evidence includes blast radius section
• New MCP tool: blast_radius (19 tools total)

Structured Orchestration

• ClaimService: per-file ownership (INSERT...ON CONFLICT), TTL-based expiry
• ParallelPlanner: graph coloring for conflict-free execution groups
• OrchestratorMode: auto-selects serial / parallel / fan-out / pipeline / hybrid
• auto-learn.ts: repeat rejection pattern detection (3+) → CLAUDE.md rule suggestions
• New CLI command: quorum orchestrate

Event Reactor

• respond.mjs rewritten: SQLite verdict → side-effects only (-1043/+211 lines)
• MessageBus: finding-level SQLite communication replaces file-based IPC
• 39 event types including finding.detect/ack/resolve + fitness signals

Marketplace

• plugin.json: mcpServers + skills declarations added
• marketplace.json in separate repo (berrzebb/claude-plugins) — avoids #35842

Simplify Pass

• Dedup: inline OLS → computeTrend() reuse, pre-verify → computeBlastRadius() reuse
• Cache: tool-core module lazy-load cache, hasPlanDoc session cache
• Batch: submitFindings N+1 transactions → single batch
• Cleanup: dead recentDetect query, orphaned JSDoc, unused params

Tests

• 743 tests (20 new: blast-radius, claim, orchestrator, fitness, hybrid-scan, trigger-fitness)
• All passing, 0 failures

Full Changelog

v0.3.0...v0.4.0

v0.2.5 — Worktree Isolation, Integration Invariant, Language-Aware Quality Rules

What's New

5-round Codex audit approved. 54 files changed, +1144 -303.

Worktree Audit Isolation

• Per-worktree lock, log, session, debounce files
• deriveAuditCwd() resolves worktree root from watch_file path
• Codex -C flag uses auditCwd (not REPO_ROOT)
• All pre-verification runs in worktree context
• [INFRA_FAILURE] verdict with execution metadata (separated from [APPROVED])
• Worktree depth=1 invariant with nesting guard

Windows Compatibility

• cli-runner.mjs: PATHEXT extensions first, extensionless POSIX script fallback
• All providers (Codex/Claude/OpenAI/Gemini): .cmd wrapper resolution
• Prompt via stdin (not argv) — eliminates shell interpretation risk
• needsShell only for .cmd/.bat wrappers
• Fixes MinGW popup + Codex STATUS_HEAP_CORRUPTION crash

Integration Invariant

• PRD template: 5 categories + 4 typed refinements (state_transition, state_machine, identifier_normalization, schema_contract)
• RTM: Runtime column (consumed/instantiated/unwired) + integration-gap status
• Audit prompt: phase-aware judgment via integration_owner
• Work Breakdown: integration_owner field for convergence points

Language-Aware Quality Rules

• config.json presets: TypeScript, Python, Rust, Go, Node fallback
• All templates/prompts/skills reference quality_rules.presets (zero hardcoded tool names)
• setup.ts auto-detects project type
• verify TEMPLATE command detects hardcoded tool references
• verify RUNTIME command checks save/load boundary contracts

Protocol Improvements

• No-abandon policy: evidence submission mandatory regardless of Tier
• minimum_tier experiment setting (forces audit for A/B experiments)
• Completion gate: evidence existence check in task-completed.mjs
• infra_failure propagated through all auditors, consensus, daemon, bus events

Experimental Results (tetris A/B/C)

• tetris1 (no invariant): component 0/6, runtime 0/6
• tetris2 (WB invariant): component 5/6, runtime 2/6
• tetris3 (+ integration invariant): component 6/6, runtime 2/6 (Codex review)
• Conclusion: invariant + independent audit are complementary, not substitutes

Full Changelog

v0.2.3...v0.2.5

v0.2.3

Full Changelog: https://github.com/berrzebb/quorum/commits/v0.2.3

Full Changelog: https://github.com/berrzebb/quorum/commits/v0.2.3

Full Changelog: https://github.com/berrzebb/quorum/commits/v0.2.3

Full Changelog: https://github.com/berrzebb/quorum/commits/v0.2.3

Full Changelog: https://github.com/berrzebb/quorum/commits/v0.2.3

Full Changelog: https://github.com/berrzebb/quorum/commits/v0.2.3

Full Changelog: https://github.com/berrzebb/quorum/commits/v0.2.3