role-os 2.3.1 → 2.5.0

package/CHANGELOG.md

@@ -1,437 +1,472 @@
-# Changelog
-
-## 2.3.0
-
-### Added
-
-#### Dogfood Swarm Mission — Multi-Pass Health + Feature Convergence
-
-- **Dogfood swarm mission** — 9th mission in the library. Three-stage health pass (bug/security → proactive → humanization) then iterative feature pass with exclusive file ownership, build gates, and user checkpoints. Moves a repo from "works" to "production-ready." Proven on claude-collaborate (35→129 tests, 106 findings fixed, v1.1.0 shipped).
-- **7 new roles** — Swarm Coordinator, Swarm Backend Agent, Swarm Bridge Agent, Swarm Tests Agent, Swarm Infra Agent, Swarm Frontend Agent, Swarm Synthesizer (61 total roles)
-- **Swarm team pack** — 10th pack, 8 roles (7 swarm + Critic Reviewer), with mismatch guards and trial evidence
-- **Two new mission primitives**:
-  - `waveLoops` — iterative convergence with exit conditions, max iterations, build gates, and user approval flags
-  - `exclusiveOwnership` — strict domain file boundaries enforced by manifest
-- **Dynamic domain dispatch** — scales agent count based on repo structure via `swarm-manifest.json`
-- **`roleos swarm` CLI** — first-class entry point with subcommands: `swarm`, `swarm manifest`, `swarm manifest --generate`, `swarm status`, `swarm findings`, `swarm approve`, `swarm verify`
-- **Domain detection** (`src/swarm/domain-detect.mjs`) — auto-detects repo type (CLI, web, desktop, MCP, monorepo) and generates domain manifests with non-overlapping file ownership
-- **Build gate** (`src/swarm/build-gate.mjs`) — auto-detects build system (Node, Rust, Python, Go) and runs lint → typecheck → test verification after every wave
-- **Evidence persistence bridge** (`src/swarm/persist-bridge.mjs`) — optional connection back to dogfood-labs, converts wave results to dogfood submission + audit DB payloads
-- **7 artifact contracts** — `swarm-gate`, `wave-report` (×5 with domain-specific sections), `swarm-final-report`
-- **Pack handoff contract** for swarm flow
-
-### Tests
-- 97 new tests (swarm core, domain detection, build gate, persist bridge) — total: 1150
-
-## 2.2.1
-
-### Added
-- **`roleos audit` CLI** — first-class entry point for deep audit with subcommands: `audit`, `audit manifest`, `audit manifest --generate`, `audit status`, `audit verify`
-- **Shared state machine** (`src/state-machine.mjs`) — canonical step/run transitions shared by both runners
-- **Shared tool profiles** (`src/tool-profiles.mjs`) — extracted from dispatch.mjs to break trial→dispatch coupling
-
-### Fixed
-- **P3-1:** Cycle detection in composite execution (`detectCycles` + visited-set guard in `findUnreachable`)
-- **P3-2:** Dual-active guard in `startNext`/`startNextStep` prevents two steps active simultaneously
-- **P3-3:** Atomic persistence — `saveRun` writes to temp file then renames
-- **P4-1:** Dependency Auditor has own artifact contract (`dependency-audit`), pack handoff corrected
-- **P4-2:** `partitionBrief` returns topic-only for unknown roles instead of full brief
-- **P4-3:** Atom kind normalization layer bridges scout `.kind` and atom `.claim_kind`
-- **P4-4:** `/dev/stdin` → `readFileSync(0)` for Windows compatibility in all 5 hooks
-- **P4-5:** TOOL_PROFILES extracted to shared module, eliminating trial→dispatch coupling
-- Node 18 compatibility fix for `import.meta.dirname` in deep-audit-proof test
-
-### Tests
-- 18 new tests (audit-cmd, audit-p5, deep-audit-proof) — total: 954
-
-## 2.2.0
-
-### Added
-
-#### Deep Audit Mission — Runner-Native Componentized Repo Audit
-
-- **Deep audit mission** — 8th mission in the library. Decomposes a repo into bounded components, dispatches one auditor per component, inspects seams from the dependency graph, assesses test truth, then synthesizes into a ranked verdict and action plan.
-- **Dynamic dispatch** — missions with `dynamicDispatch` field now expand from a manifest at runtime. `createRun("deep-audit", task, { manifest })` creates N + M + K + 3 steps from the repo graph instead of a fixed static chain. A 6-component / 8-boundary repo produces 23 steps; a 10-component / 5-boundary repo produces 28.
-- **4 new audit roles** — Component Auditor, Seam Auditor, Test Truth Auditor, Audit Synthesizer. Each with full artifact contracts, tool profiles, and role definitions in starter-pack.
-- **Deep-audit pack** — 9th team pack with scaling chain order, dispatch defaults, and mismatch guards.
-- **Artifact validation at execution boundaries** — `validateArtifact()` now runs on every step completion in both `run.mjs` and `mission-run.mjs`. Validation results are attached to the step object. Warn, don't block.
-- **Proof run test suite** — `test/deep-audit-proof.test.mjs` proves the full runner-native lifecycle against the real audit-manifest.json: step creation, parcel identity, validation, escalation, partial failure, scaling formula, and report generation.
-
-### Fixed
-
-- **Critical: "approve" vs "accept" verdict mismatch** — `evidence.mjs:195` checked `!== "approve"` but the enum defines `"accept"`. Every accept verdict generated a spurious warning. Tests masked it via substring matching. Fixed to `"accept"` with hardened exact-assertion tests.
-- **Dead imports removed** — `TEAM_PACKS` and `ROLE_ARTIFACT_CONTRACTS` in mission-run.mjs, `TEAM_PACKS` in run.mjs, `scoreRole` and `MIN_SCORE_THRESHOLD` in trial.mjs were imported but never used.
-- **Warning message terminology** — all evidence warning messages now use "accept" instead of "approve" consistently.
-
-### Changed
-
-- Mission count: 7 → 8
-- Role count: 50 → 54 (4 deep audit roles)
-- Pack count: 8 → 9
-- Artifact contract count: 30 → 34 (4 new audit role contracts)
-- Test count: 905 → 936
-
-### Evidence
-
-- Self-audit dogfood: 128 findings (1 critical, 11 high, 39 medium) across 6 component parcels, 8 boundary seams, and 31 test files
-- Runner-native proof run: 23 dynamic steps from real manifest, full lifecycle, all green
-- Scaling formula verified: 2N + K + 3 holds for manifests of 3, 6, 10, and 15 components
-
-## 2.1.0
-
-### Added
-
-#### Brainstorm Mission (v0.4) — Structured Inquiry with Traceable Disagreement
-
-- **Brainstorm mission** — 7th mission in the library, 9-role chain with two-layer architecture
-- **Layer 1 (truth):** 4 analyst roles emit role-native schemas (ContextMap, UserValueMap, MechanicsMap, PositioningMap), not shared prose. Blindspot enforcement: forbidden phrases, forbidden claim kinds, filtered input partitions per role. Provenance-preserving atoms carry source_role, claim_kind, allowed_challengers. Cross-examination permission matrix (directed graph). Rebut phase: original analysts defend, narrow, or retract under pressure.
-- **Layer 2 (render):** 5 distinct voices — Boundary Memo (taxonomist), Field Notes (ethnographer), System Sketch (whiteboard), Claim Brief (strategist), Cross-Exam Transcript (litigator). Lexical bans prevent voice convergence. Debate transcript generator. Both layers always available.
-- **Trace links:** Every rendered sentence maps to a truth-layer atom. Synthesis cites atoms, never prose.
-- **Golden run proof:** Full artifact chain for MCP server marketplace topic — truth artifacts, dispute graph (4 challenges, 3 narrowed, 1 unresolved), rendered artifacts, trace map (16+ links). Published as `examples/golden-run.md`.
-- **Result formatter:** `formatBrainstormResult()` produces saveable markdown with verdict, directions, dispute, tensions, rendered artifacts (opt-in), and evidence trail. Layer parameter controls truth-only vs both.
-- **Artifact contracts:** 9 brainstorm role contracts (replacing 3 v0.1 scout contracts) with completion rules, required evidence, and consumer mapping.
-- **Pack update:** Brainstorm pack updated from v0.1 scouts to v0.3/v0.4 analysts with correct chain order and required artifacts.
-
-### Changed
-
-- Mission count: 6 → 7
-- Role count: 31 → 50 (brainstorm analysts, contrarian, plus existing)
-- Artifact contract count: 20 → 30
-- Test count: 617 → 905
-
-## 2.0.1
-
-### Added
-
-- 4 version consistency tests (semver, >= 1.0.0, CHANGELOG, help output)
-
-## 2.0.0
-
-### Added
-
-#### Operator Friction Pass (Phase U)
-- `roleos run "<task>"` — one command from task description to active execution
-- Persistent disk-backed runs in `.claude/runs/` — survives session interruptions
-- Entry level auto-selection: mission, pack, or free routing with force overrides (`--mission=`, `--pack=`)
-- Step-local operator guidance at every step: role, artifact, required sections, completion rule, stop conditions
-- `roleos resume [id]` — continue interrupted runs from disk
-- `roleos next` — start the next step or show what's active
-- `roleos explain [id]` — full run state with guidance, escalations, interventions
-- `roleos complete <artifact> [note]` — complete the active step with artifact reference
-- `roleos fail <partial|failed> <reason>` — fail with honest downstream blocking
-- `roleos run list` — list all runs with status icons
-- `roleos run show <id>` — full run detail
-
-#### Intervention Shortcuts
-- `roleos retry <step>` — retry a failed/partial step, unblock downstream
-- `roleos reroute <step> <role> <reason>` — swap a step to a different role
-- `roleos escalate <from> <to> <trigger> <action>` — escalate between roles with step re-opening
-- `roleos block <step> <reason>` — manually block a step
-- `roleos reopen <step> <reason>` — reopen a completed step for re-execution
-
-#### Friction Measurement
-- `roleos report [id]` — generate completion report with honest-partial
-- `roleos friction [id]` — measure operator touches: interventions, escalations, manual steps
-- Friction score: low/medium/high based on touch count vs step count
-
-### Evidence
-- 613 tests, zero failures (86 new)
-- 6 friction trials validated: clean run, reroute, retry, pack-level, free-routing, disk resume
-- All entry levels produce low/medium friction scores
-- Disk round-trip verified: create → pause → load → resume → complete
-
-## 1.9.0
-
-### Added
-
-#### Unified Entry Path (Phase T)
-- `roleos start <task>` — auto-decides mission vs pack vs free routing
-- Three-level fallback ladder with confidence scores and alternatives
-- Composite task detection warns when a task should be decomposed
-- `--json` flag for machine-readable entry decisions
-- 46 new tests: entry engine, comparison trials, CLI integration
-
-#### Handbook Updates
-- New Missions handbook page with full mission documentation
-- Updated Getting Started to lead with `roleos start`
-- Updated Reference with all CLI commands (start, mission, packs, artifacts, status, doctor)
-- Updated handbook index with entry levels and 9 operating layers
-
-#### README Overhaul
-- "How it works" section leads with `roleos start` examples
-- Quick Start updated with mission and start commands
-- Added 6 Missions table
-- Updated project structure with all 18 source modules
-- Updated status history through v1.9.0
-
-### Evidence
-- 527 tests, zero failures (46 new)
-- Entry path trials validated against 20+ real task descriptions
-- Fallback ladder tested: mission, pack, free-routing, composite, empty input
-
-## 1.8.0
-
-### Added
-
-#### Mission Library (Phase S — Mission Hardening)
-- 6 named, repeatable mission types: feature-ship, bugfix, treatment, docs-release, security-hardening, research-launch
-- Each mission declares: pack, role chain, artifact flow, escalation branches, honest-partial definition, stop conditions, dispatch defaults, trial evidence
-- Mission runner: create → step through → complete/fail → generate completion report
-- Completion proof reporter with honest-partial and formatted text output
-- `roleos mission list` — list all missions
-- `roleos mission show <key>` — full mission detail
-- `roleos mission suggest <text>` — signal-based mission suggestion
-- `roleos mission validate [key]` — validate mission wiring against packs/roles
-
-#### Mission Runner Engine
-- `createRun()` — instantiate a mission with tracked steps
-- `startNextStep()` / `completeStep()` / `failStep()` — step lifecycle
-- `recordEscalation()` — re-opens completed steps on escalation loops
-- `getRunPosition()` / `getArtifactChain()` — run introspection
-- `generateCompletionReport()` / `formatCompletionReport()` — honest outcome reporting
-
-### Evidence
-- 465 tests, zero failures (67 new)
-- All 6 missions validate against live pack/role catalog
-- Full lifecycle tests: end-to-end runs, escalation loops, partial completions, failure reporting
-
-## 1.7.0
-
-### Added
-
-#### Completion Proof (Phase R)
-- `roleos artifacts` CLI command: list, show, validate, chain subcommands
-- 13 new CLI integration tests for artifact inspection
-- Real task completion missions through the full stack
-
-#### Completion Proof Evidence
-- R1-1 Feature mission: `roleos artifacts` command shipped through feature pack
-  - Pack: feature (high confidence, correct)
-  - Chain: 5 roles, 0 escalations, 1 minor correction
-  - Artifact contracts: all 4 used and valid
-- R1-2 Bugfix mission: README.zh.md npm anomaly
-  - Diagnosed correctly: npm auto-includes README* regardless of files field
-  - Escalated honestly: fix requires structural decision (translation file organization)
-  - Not force-closed: deferred to treatment pass
-
-### Evidence
-- 398 tests, zero failures
-- 3 missions run through the full stack
-- Completion metrics recorded per mission
-
-## 1.6.0
-
-### Added
-
-#### Artifact Spine (Phase Q)
-- 20 per-role artifact contracts: each defines artifact type, required sections, evidence references, downstream consumers, and completion rules
-- `validateArtifact(role, content)` — structural validation against role contracts (missing sections, evidence references, content depth)
-- 7 pack-level handoff contracts: define the expected artifact flow between steps for each pack (e.g., strategy-brief → implementation-spec → change-plan → test-package → verdict)
-- `validatePackChain(pack, artifacts)` — validates an entire pack's artifact chain for completeness
-- `getArtifactContract(role)` / `getHandoffContract(pack)` — lookup APIs
-- `formatArtifactValidation()` / `formatPackChain()` — display formatters
-
-#### Artifact contract coverage
-- Product Strategist → strategy-brief (problem-framing, scope, non-goals, tradeoffs)
-- Spec Writer → implementation-spec (acceptance-criteria, edge-cases, interface-spec)
-- Backend/Frontend Engineer → change-plan (files-to-change, implementation-approach, risk-notes)
-- Test Engineer → test-package (test-plan, test-cases, false-confidence-assessment)
-- Security Reviewer → security-findings (findings, severity-assessment, recommendations)
-- Critic Reviewer → verdict (verdict, evidence, required-corrections)
-- And 14 more roles with full contracts
-
-### Evidence
-- 385 tests, zero failures
-- 27 new artifact tests
-
-## 1.5.0
-
-### Added
-
-#### Hook Spine / Runtime Enforcement (Phase R)
-- 5 lifecycle hooks: SessionStart, UserPromptSubmit, PreToolUse, SubagentStart, Stop
-- `scaffoldHooks()` generates all 5 hook scripts in .claude/hooks/
-- `roleos init claude` now scaffolds hooks + settings.local.json with hook config
-- `roleos doctor` now checks for hook scripts (check 7) and settings hooks (check 8)
-
-#### SessionStart hook
-- Establishes session contract on every new session
-- Records session ID, timestamp, initializes state tracking
-- Adds context reminding Claude to use /roleos-route for non-trivial tasks
-
-#### UserPromptSubmit hook
-- Classifies prompts as substantial (>50 chars + action verbs)
-- After 2+ substantial prompts without a route card, adds context reminder
-- Does not block — advisory enforcement
-
-#### PreToolUse hook
-- Records all tool usage in session state
-- Flags write tools (Bash, Write, Edit) used without route card after substantial work
-- Advisory, not blocking — preserves operator control
-
-#### SubagentStart hook
-- Injects active role contract into delegated agents
-- Ensures subagents inherit the Role OS session context
-
-#### Stop hook
-- Warns when substantial sessions end without route card or outcome artifact
-- Advisory — does not block session exit
-- Trivial sessions (< 2 substantial prompts) are exempt
-
-### Evidence
-- 358 tests, zero failures
-- 23 new hook tests covering all 5 lifecycle hooks
-
-## 1.4.0
-
-### Added
-
-#### Session Spine (Phase Q)
-- `roleos init claude` — scaffolds Claude Code integration: CLAUDE.md instructions, /roleos-route + /roleos-review + /roleos-status slash commands
-- `roleos doctor` — verifies repo is correctly wired for Role OS sessions (6 checks: .claude/ dir, CLAUDE.md section, /roleos-route command, context files, role contracts, packets)
-- Route card generation — session header artifact proving Role OS was engaged (task type, pack, confidence, composite status, success artifact)
-- CLAUDE.md template instructs Claude to route through Role OS before non-trivial work
-- /roleos-route command produces structured route cards
-- /roleos-review command guides structured verdict production
-- /roleos-status command shows active work and context health
-- Appends to existing CLAUDE.md without overwriting (detects Role OS section)
-- --force flag overwrites existing command files
-
-### Evidence
-- 335 tests, zero failures
-
-## 1.3.0
-
-### Added
-
-#### Outcome Calibration (Phase M)
-- Run outcome ledger — append-only JSONL recording pack selection, confidence, overrides, escalations, corrections, completion status
-- `computeCalibration()` — pack usage rates, high-confidence accuracy, operator override rates, per-pack performance
-- `computePackBoosts()` — weight tuning from clean completed runs (+0.5/run, capped at 2.0)
-- `computeConfidenceAdjustment()` — raises threshold when high-confidence is often overridden, lowers when medium is often accepted
-- Auto-generated calibration suggestions when metrics drift
-- Safety constraint: calibration never overrides mismatch guards, conflict rules, escalation honesty, or evidence requirements
-
-#### Mixed-Task Decomposition (Phase N)
-- `detectComposite()` — 7 subtask categories (build, bugfix, security, docs, research, launch, treatment) with signal-based detection
-- Structural connector detection ("and then", "after that", "plus", "also")
-- Confidence levels: high (3+ categories or 2+ with connectors), medium, low
-- `decompose()` — generates linked child packets sorted by phase order
-- `createRunPlan()` — dependency-aware parent plan with child tracking
-- Honest fallback: medium/low confidence shows uncertainty warning with `--no-split` override
-
-#### Composite Execution (Phase O)
-- `initExecution()` / `advance()` — dependency-driven child execution with artifact passing
-- 7 artifact contracts defining what each category produces and expects
-- Artifact ledger tracking all cross-packet handoffs
-- `blockChild()` / `recoverChild()` / `failChild()` — branch recovery with transitive cascade
-- `invalidateDownstream()` — resets stale children when upstream changes, removes stale artifacts
-- `synthesize()` — truthful parent-level completion report
-- Independent branches continue unaffected when a sibling fails
-
-#### Adaptive Replanning (Phase P)
-- 6 structured change event types: scope-change, artifact-changed, new-requirement, review-finding, dependency-discovered, priority-change
-- `analyzeImpact()` — identifies valid/stale children, stale artifacts, whether new children or reorder needed
-- `replan()` — selective replanning: invalidates only affected branches, inserts new children, updates dependencies
-- Plan diff: shows what changed, what stayed valid, what reopened, what was inserted
-- Execution resumes from next valid child after replan — no restart required
-
-### Evidence
-- 317 tests, zero failures
-- Calibration, decomposition, composite execution, and replanning each have dedicated test suites
-
-## 1.2.0
-
-### Added
-- Pack auto-selection in `roleos route` — suggests best pack when confidence is high
-- `roleos route --pack=<name>` — use a specific pack for routing
-- Pack mismatch detection — warns when a pack doesn't fit the task, suggests the correct alternative
-- Pack fallback — mismatched or unknown packs fall back to free routing automatically
-- `checkPackMismatch()` API with 7 guard sets covering all pack×task-type combinations
-- `getPackRoles()` API with conditional Orchestrator support
-
-### Changed
-- Docs pack: Support Triage Lead now opens (was Feedback Synthesizer). Feedback Synthesizer is second. Release Engineer + Deployment Verifier moved to optional (overhead for docs-only tasks).
-- Pack calibration applied from comparison evidence: conditional Orchestrator, Security Reviewer in Treatment, Product Strategist opens Research, mismatch guards on all 7 packs.
-
-### Evidence
-- Pack comparison: calibrated packs now win or tie 6/7 (was 2/7 pre-calibration)
-- Misfit honesty: 0 full bluffs, 0 undetected partial bluffs (was 1 + 3)
-- 230 tests, zero failures
-
-## 1.1.0
-
-### Added
-
-#### Routing
-- Full 31-role catalog — all roles scored by keyword, trigger phrase, packet type bias, and deliverable affinity
-- Dynamic chain builder — phase-ordered assembly replacing static templates
-- Routing confidence assessment (high/medium/low)
-- `excludeWhen` enforcement — roles suppressed when exclusion patterns match packet content
-- `detectType` false-positive prevention — "integration testing" no longer triggers integration type
-- `--verbose` flag for `roleos route` — hides scoring noise by default
-
-#### Conflict Detection
-- 4-pass conflict engine: hard conflicts, sequence, redundancy, coverage gaps
-- Per-role constraint registry: lateOnly, requiresBeforePacks
-- Overlap pair detection
-- Repair suggestions on every finding
-
-#### Escalation Auto-Routing
-- Blocked/rejected/conflict/split work auto-routes to named resolver
-- Every escalation includes: target role, recovery type, required artifact, handoff context
-
-#### Structured Evidence
-- 12 evidence kinds, 4 statuses, closed 4-verdict enum (accept/accept-with-notes/reject/blocked)
-- Role-aware evidence requirements for 15 roles
-- Sufficiency checks with contradiction detection
-
-#### Runtime Dispatch
-- Execution manifests for multi-claude with per-role tool profiles and budgets
-- 8 execution states with auto-advance
-- Escalation packet generation for blocked/rejected steps
-
-#### Proven Team Packs
-- 7 battle-tested packs: feature, bugfix, security, docs, launch, research, treatment
-- `roleos packs list` — show all packs with role counts
-- `roleos packs suggest <packet>` — suggest best pack for a packet
-- `roleos packs show <name>` — show pack details (roles, artifacts, stop conditions)
-- Pack suggestion engine with confidence levels
-
-#### Trials
-- Full roster proven: 30/30 gold-task trials + 5/5 negative (wrong-task honesty) trials
-- 7 pack execution trials — all packs ran full chains with honest Critic verdicts
-- Trial framework: buildClusterTrials, evaluateTrialOutput, formatTrialReport
-
-### Changed
-- 32 → 31 roles: Information Architect merged into Docs Architect
-- Verdict vocabulary unified: evidence.mjs now uses accept/reject/blocked (matching review.mjs)
-- "worker" terminology replaced with "role" in dispatch.mjs
-
-### Fixed
-- `excludeWhen` was declared on 14 roles but never enforced — now active in scoreRole
-- `detectType` false-positived on "integration testing" — now uses word-boundary regex
-- "Not triggered: N roles" noise hidden by default (shown with --verbose)
-- Handbook: Team Packs page added, reference sidebar reordered
-
-## 1.0.2
-
-### Fixed
-- Fix double-nested `.claude/.claude/` directory created by `roleos init` — `starter-pack/.claude/workflows/full-treatment.md` moved to `starter-pack/workflows/`
-- Read VERSION from `package.json` at runtime instead of hardcoded constant — prevents version drift between CLI and package metadata
-
-### Added
-- `roleos init --force` — update canonical scaffolded files while always protecting user-filled `context/` files
-- 4 regression tests: no double-nesting, correct workflow placement, version sync, --force context protection
-
-## 1.0.0
-
-### Added
-- `roleos init` — scaffold Role OS starter pack into `.claude/`
-- `roleos packet new <type>` — create feature, integration, or identity packets
-- `roleos route <packet-file>` — recommend smallest valid role chain with dependency verification
-- `roleos review <packet-file> <verdict>` — record accept/reject/blocked verdicts
-- Full starter pack: 8 role contracts, 3 schemas, 4 policies, 3 workflows
-- Guided context templates with inline prompts
-- 3 canonical example packets (feature, integration, identity)
-- Adoption handbook
+# Changelog
+
+## 2.5.0
+
+### Added
+
+#### Local-panel seat — a second, family-different verifier for `verify-citations`, runnable locally for free
+
+- **`roleos verify-citations --local-panel`** — adds a local grounded-entailment PANEL (the `offload` CLI on **Qwen3-4B + Qwen3-14B + Mistral-Nemo-12B** via llama-swap) as a SECOND verifier seat, decorrelated from the Claude generator by construction (no Anthropic model in the panel) and from prism's single groundedness model (3 seats, ≥2 families, conservative majority). It re-checks each citation prism marked `supported` against prism's own retrieved evidence (`source_title` + `supporting_span`).
+- **Monotone-tightening** — the panel can only downgrade a passing gate to **escalate** (`local_panel_disagreement`), never loosen one; the deterministic existence floor (`fabricated` → blocking) always dominates, and a non-passing gate is left untouched. A panel that is requested but unreachable **escalates** (`local_panel_unreachable`) — the same closed-gate invariant prism uses.
+- **Why it earns its seat (EXTERNAL_VERIFIER, now local + zero-cost):** the panel's measured property is **zero false-confirms** — a 3-seat conservative majority never stamps a false claim "supported." On a 16-case real-arXiv citation set, `mistral-nemo-12b` solo false-confirmed a refuted claim (inverting arXiv:2404.13076's finding); the panel held it at `insufficient`. Receipt + dataset: `tensor-engine-knowledge/verifier/citation-panel-receipt.json` (study-swarm wave-6, recipe #156).
+- **Receipt** gains a `local_panel` block (PIN_PER_STEP): the exact seat models used, per-citation panel verdicts, and any disagreement that downgraded the gate — folded into the receipt's hash chain via the verdict + a panel digest.
+- New module `src/citation-panel.mjs` (`runOffloadPanel`, `applyLocalPanel`, `buildEvidence`); injectable `offloadExec` for tests. **Off by default** — opt in with `--local-panel` (needs llama-swap up + `offload.py` on the rig; `OFFLOAD_PYTHON` / `OFFLOAD_SCRIPT` / `--llamaswap-base` configurable).
+
+### Tests
+- 16 new tests (evidence building; the panel runner — agree / refuted / insufficient / no-evidence / unreachable / garbage; monotone `applyLocalPanel`; end-to-end `runCitationGate --local-panel` incl. the PIN'd receipt + blocking-skips-panel). **1196 total, all green.**
+
+## 2.4.0
+
+### Added
+
+#### Citation-Verification Gate — defers citation truthfulness to prism (an external, family-different verifier)
+
+- **`roleos verify-citations <dispatch.md|.json>` CLI** — extracts a research dispatch's citations, shells the external `prism verify` CLI (a family-different, reasoning-stripped citation verifier), and gates on the verdict. Exit `0` = accept, `20` = blocking (a cited paper did not resolve in arXiv/Crossref — likely fabricated), `10` = advisory (revise / escalate), `2` = no resolvable citations found.
+- **Citation gate module** (`src/verify-citations.mjs`, peer to `build-gate.mjs`) — deterministic, copy-only extraction (`extractCitations` — never invents an identifier); a three-tier gate keyed to the failure source (`gateCitations`): existence `fabricated` → **BLOCKING** hard halt, soft groundedness `contradicted` → advisory revise, low-confidence → advisory escalate. An unreachable verifier **escalates, never default-accepts** ("an unreachable gate is a closed gate"). Emits a receipt chained to prism's HMAC receipt (per-citation `source_sha256` pins → drift-detectable on re-run).
+- **Critic Reviewer** gains a citation-verification clause — for a research dispatch it runs the gate, treats blocking as reject and advisory as accept-with-notes / escalate, and never grades the citations itself.
+- **Design doc** (`design/citation-verification-runner.md`) — research-grounded by a 4-question study-swarm (`wf_20651368-297`), with a Standards-compliance section scoring 15/15 on the applicable standards (NAMED_COMPENSATORS a documented read-only skip): EXTERNAL_VERIFIER, ANDON_AUTHORITY, PIN_PER_STEP, DECOMPOSE_BY_SECRETS, and UNCERTAINTY_GATED_HUMANS (the contrastive escalate-to-human path).
+- Pairs with prism **v0.3.2**'s `prism verify --gate` (verdict-coded exit status).
+
+### Tests
+- 14 new tests (extraction; three-tier gate; runner with injected prism — accept / block / escalate / unreachable). Module is testable with no real prism shell-out.
+
+## 2.3.1
+
+### Changed
+- Version bump for dogfood swarm mission release
+
+## 2.3.0
+
+### Added
+
+#### Dogfood Swarm Mission — Multi-Pass Health + Feature Convergence
+
+- **Dogfood swarm mission** — 9th mission in the library. Three-stage health pass (bug/security → proactive → humanization) then iterative feature pass with exclusive file ownership, build gates, and user checkpoints. Moves a repo from "works" to "production-ready." Proven on claude-collaborate (35→129 tests, 106 findings fixed, v1.1.0 shipped).
+- **7 new roles** — Swarm Coordinator, Swarm Backend Agent, Swarm Bridge Agent, Swarm Tests Agent, Swarm Infra Agent, Swarm Frontend Agent, Swarm Synthesizer (61 total roles)
+- **Swarm team pack** — 10th pack, 8 roles (7 swarm + Critic Reviewer), with mismatch guards and trial evidence
+- **Two new mission primitives**:
+  - `waveLoops` — iterative convergence with exit conditions, max iterations, build gates, and user approval flags
+  - `exclusiveOwnership` — strict domain file boundaries enforced by manifest
+- **Dynamic domain dispatch** — scales agent count based on repo structure via `swarm-manifest.json`
+- **`roleos swarm` CLI** — first-class entry point with subcommands: `swarm`, `swarm manifest`, `swarm manifest --generate`, `swarm status`, `swarm findings`, `swarm approve`, `swarm verify`
+- **Domain detection** (`src/swarm/domain-detect.mjs`) — auto-detects repo type (CLI, web, desktop, MCP, monorepo) and generates domain manifests with non-overlapping file ownership
+- **Build gate** (`src/swarm/build-gate.mjs`) — auto-detects build system (Node, Rust, Python, Go) and runs lint → typecheck → test verification after every wave
+- **Evidence persistence bridge** (`src/swarm/persist-bridge.mjs`) — optional connection back to dogfood-labs, converts wave results to dogfood submission + audit DB payloads
+- **7 artifact contracts** — `swarm-gate`, `wave-report` (×5 with domain-specific sections), `swarm-final-report`
+- **Pack handoff contract** for swarm flow
+
+### Tests
+- 97 new tests (swarm core, domain detection, build gate, persist bridge) — total: 1150
+
+## 2.2.1
+
+### Added
+- **`roleos audit` CLI** — first-class entry point for deep audit with subcommands: `audit`, `audit manifest`, `audit manifest --generate`, `audit status`, `audit verify`
+- **Shared state machine** (`src/state-machine.mjs`) — canonical step/run transitions shared by both runners
+- **Shared tool profiles** (`src/tool-profiles.mjs`) — extracted from dispatch.mjs to break trial→dispatch coupling
+
+### Fixed
+- **P3-1:** Cycle detection in composite execution (`detectCycles` + visited-set guard in `findUnreachable`)
+- **P3-2:** Dual-active guard in `startNext`/`startNextStep` prevents two steps active simultaneously
+- **P3-3:** Atomic persistence — `saveRun` writes to temp file then renames
+- **P4-1:** Dependency Auditor has own artifact contract (`dependency-audit`), pack handoff corrected
+- **P4-2:** `partitionBrief` returns topic-only for unknown roles instead of full brief
+- **P4-3:** Atom kind normalization layer bridges scout `.kind` and atom `.claim_kind`
+- **P4-4:** `/dev/stdin` → `readFileSync(0)` for Windows compatibility in all 5 hooks
+- **P4-5:** TOOL_PROFILES extracted to shared module, eliminating trial→dispatch coupling
+- Node 18 compatibility fix for `import.meta.dirname` in deep-audit-proof test
+
+### Tests
+- 18 new tests (audit-cmd, audit-p5, deep-audit-proof) — total: 954
+
+## 2.2.0
+
+### Added
+
+#### Deep Audit Mission — Runner-Native Componentized Repo Audit
+
+- **Deep audit mission** — 8th mission in the library. Decomposes a repo into bounded components, dispatches one auditor per component, inspects seams from the dependency graph, assesses test truth, then synthesizes into a ranked verdict and action plan.
+- **Dynamic dispatch** — missions with `dynamicDispatch` field now expand from a manifest at runtime. `createRun("deep-audit", task, { manifest })` creates N + M + K + 3 steps from the repo graph instead of a fixed static chain. A 6-component / 8-boundary repo produces 23 steps; a 10-component / 5-boundary repo produces 28.
+- **4 new audit roles** — Component Auditor, Seam Auditor, Test Truth Auditor, Audit Synthesizer. Each with full artifact contracts, tool profiles, and role definitions in starter-pack.
+- **Deep-audit pack** — 9th team pack with scaling chain order, dispatch defaults, and mismatch guards.
+- **Artifact validation at execution boundaries** — `validateArtifact()` now runs on every step completion in both `run.mjs` and `mission-run.mjs`. Validation results are attached to the step object. Warn, don't block.
+- **Proof run test suite** — `test/deep-audit-proof.test.mjs` proves the full runner-native lifecycle against the real audit-manifest.json: step creation, parcel identity, validation, escalation, partial failure, scaling formula, and report generation.
+
+### Fixed
+
+- **Critical: "approve" vs "accept" verdict mismatch** — `evidence.mjs:195` checked `!== "approve"` but the enum defines `"accept"`. Every accept verdict generated a spurious warning. Tests masked it via substring matching. Fixed to `"accept"` with hardened exact-assertion tests.
+- **Dead imports removed** — `TEAM_PACKS` and `ROLE_ARTIFACT_CONTRACTS` in mission-run.mjs, `TEAM_PACKS` in run.mjs, `scoreRole` and `MIN_SCORE_THRESHOLD` in trial.mjs were imported but never used.
+- **Warning message terminology** — all evidence warning messages now use "accept" instead of "approve" consistently.
+
+### Changed
+
+- Mission count: 7 → 8
+- Role count: 50 → 54 (4 deep audit roles)
+- Pack count: 8 → 9
+- Artifact contract count: 30 → 34 (4 new audit role contracts)
+- Test count: 905 → 936
+
+### Evidence
+
+- Self-audit dogfood: 128 findings (1 critical, 11 high, 39 medium) across 6 component parcels, 8 boundary seams, and 31 test files
+- Runner-native proof run: 23 dynamic steps from real manifest, full lifecycle, all green
+- Scaling formula verified: 2N + K + 3 holds for manifests of 3, 6, 10, and 15 components
+
+## 2.1.0
+
+### Added
+
+#### Brainstorm Mission (v0.4) — Structured Inquiry with Traceable Disagreement
+
+- **Brainstorm mission** — 7th mission in the library, 9-role chain with two-layer architecture
+- **Layer 1 (truth):** 4 analyst roles emit role-native schemas (ContextMap, UserValueMap, MechanicsMap, PositioningMap), not shared prose. Blindspot enforcement: forbidden phrases, forbidden claim kinds, filtered input partitions per role. Provenance-preserving atoms carry source_role, claim_kind, allowed_challengers. Cross-examination permission matrix (directed graph). Rebut phase: original analysts defend, narrow, or retract under pressure.
+- **Layer 2 (render):** 5 distinct voices — Boundary Memo (taxonomist), Field Notes (ethnographer), System Sketch (whiteboard), Claim Brief (strategist), Cross-Exam Transcript (litigator). Lexical bans prevent voice convergence. Debate transcript generator. Both layers always available.
+- **Trace links:** Every rendered sentence maps to a truth-layer atom. Synthesis cites atoms, never prose.
+- **Golden run proof:** Full artifact chain for MCP server marketplace topic — truth artifacts, dispute graph (4 challenges, 3 narrowed, 1 unresolved), rendered artifacts, trace map (16+ links). Published as `examples/golden-run.md`.
+- **Result formatter:** `formatBrainstormResult()` produces saveable markdown with verdict, directions, dispute, tensions, rendered artifacts (opt-in), and evidence trail. Layer parameter controls truth-only vs both.
+- **Artifact contracts:** 9 brainstorm role contracts (replacing 3 v0.1 scout contracts) with completion rules, required evidence, and consumer mapping.
+- **Pack update:** Brainstorm pack updated from v0.1 scouts to v0.3/v0.4 analysts with correct chain order and required artifacts.
+
+### Changed
+
+- Mission count: 6 → 7
+- Role count: 31 → 50 (brainstorm analysts, contrarian, plus existing)
+- Artifact contract count: 20 → 30
+- Test count: 617 → 905
+
+## 2.0.1
+
+### Added
+
+- 4 version consistency tests (semver, >= 1.0.0, CHANGELOG, help output)
+
+## 2.0.0
+
+### Added
+
+#### Operator Friction Pass (Phase U)
+- `roleos run "<task>"` — one command from task description to active execution
+- Persistent disk-backed runs in `.claude/runs/` — survives session interruptions
+- Entry level auto-selection: mission, pack, or free routing with force overrides (`--mission=`, `--pack=`)
+- Step-local operator guidance at every step: role, artifact, required sections, completion rule, stop conditions
+- `roleos resume [id]` — continue interrupted runs from disk
+- `roleos next` — start the next step or show what's active
+- `roleos explain [id]` — full run state with guidance, escalations, interventions
+- `roleos complete <artifact> [note]` — complete the active step with artifact reference
+- `roleos fail <partial|failed> <reason>` — fail with honest downstream blocking
+- `roleos run list` — list all runs with status icons
+- `roleos run show <id>` — full run detail
+
+#### Intervention Shortcuts
+- `roleos retry <step>` — retry a failed/partial step, unblock downstream
+- `roleos reroute <step> <role> <reason>` — swap a step to a different role
+- `roleos escalate <from> <to> <trigger> <action>` — escalate between roles with step re-opening
+- `roleos block <step> <reason>` — manually block a step
+- `roleos reopen <step> <reason>` — reopen a completed step for re-execution
+
+#### Friction Measurement
+- `roleos report [id]` — generate completion report with honest-partial
+- `roleos friction [id]` — measure operator touches: interventions, escalations, manual steps
+- Friction score: low/medium/high based on touch count vs step count
+
+### Evidence
+- 613 tests, zero failures (86 new)
+- 6 friction trials validated: clean run, reroute, retry, pack-level, free-routing, disk resume
+- All entry levels produce low/medium friction scores
+- Disk round-trip verified: create → pause → load → resume → complete
+
+## 1.9.0
+
+### Added
+
+#### Unified Entry Path (Phase T)
+- `roleos start <task>` — auto-decides mission vs pack vs free routing
+- Three-level fallback ladder with confidence scores and alternatives
+- Composite task detection warns when a task should be decomposed
+- `--json` flag for machine-readable entry decisions
+- 46 new tests: entry engine, comparison trials, CLI integration
+
+#### Handbook Updates
+- New Missions handbook page with full mission documentation
+- Updated Getting Started to lead with `roleos start`
+- Updated Reference with all CLI commands (start, mission, packs, artifacts, status, doctor)
+- Updated handbook index with entry levels and 9 operating layers
+
+#### README Overhaul
+- "How it works" section leads with `roleos start` examples
+- Quick Start updated with mission and start commands
+- Added 6 Missions table
+- Updated project structure with all 18 source modules
+- Updated status history through v1.9.0
+
+### Evidence
+- 527 tests, zero failures (46 new)
+- Entry path trials validated against 20+ real task descriptions
+- Fallback ladder tested: mission, pack, free-routing, composite, empty input
+
+## 1.8.0
+
+### Added
+
+#### Mission Library (Phase S — Mission Hardening)
+- 6 named, repeatable mission types: feature-ship, bugfix, treatment, docs-release, security-hardening, research-launch
+- Each mission declares: pack, role chain, artifact flow, escalation branches, honest-partial definition, stop conditions, dispatch defaults, trial evidence
+- Mission runner: create → step through → complete/fail → generate completion report
+- Completion proof reporter with honest-partial and formatted text output
+- `roleos mission list` — list all missions
+- `roleos mission show <key>` — full mission detail
+- `roleos mission suggest <text>` — signal-based mission suggestion
+- `roleos mission validate [key]` — validate mission wiring against packs/roles
+
+#### Mission Runner Engine
+- `createRun()` — instantiate a mission with tracked steps
+- `startNextStep()` / `completeStep()` / `failStep()` — step lifecycle
+- `recordEscalation()` — re-opens completed steps on escalation loops
+- `getRunPosition()` / `getArtifactChain()` — run introspection
+- `generateCompletionReport()` / `formatCompletionReport()` — honest outcome reporting
+
+### Evidence
+- 465 tests, zero failures (67 new)
+- All 6 missions validate against live pack/role catalog
+- Full lifecycle tests: end-to-end runs, escalation loops, partial completions, failure reporting
+
+## 1.7.0
+
+### Added
+
+#### Completion Proof (Phase R)
+- `roleos artifacts` CLI command: list, show, validate, chain subcommands
+- 13 new CLI integration tests for artifact inspection
+- Real task completion missions through the full stack
+
+#### Completion Proof Evidence
+- R1-1 Feature mission: `roleos artifacts` command shipped through feature pack
+  - Pack: feature (high confidence, correct)
+  - Chain: 5 roles, 0 escalations, 1 minor correction
+  - Artifact contracts: all 4 used and valid
+- R1-2 Bugfix mission: README.zh.md npm anomaly
+  - Diagnosed correctly: npm auto-includes README* regardless of files field
+  - Escalated honestly: fix requires structural decision (translation file organization)
+  - Not force-closed: deferred to treatment pass
+
+### Evidence
+- 398 tests, zero failures
+- 3 missions run through the full stack
+- Completion metrics recorded per mission
+
+## 1.6.0
+
+### Added
+
+#### Artifact Spine (Phase Q)
+- 20 per-role artifact contracts: each defines artifact type, required sections, evidence references, downstream consumers, and completion rules
+- `validateArtifact(role, content)` — structural validation against role contracts (missing sections, evidence references, content depth)
+- 7 pack-level handoff contracts: define the expected artifact flow between steps for each pack (e.g., strategy-brief → implementation-spec → change-plan → test-package → verdict)
+- `validatePackChain(pack, artifacts)` — validates an entire pack's artifact chain for completeness
+- `getArtifactContract(role)` / `getHandoffContract(pack)` — lookup APIs
+- `formatArtifactValidation()` / `formatPackChain()` — display formatters
+
+#### Artifact contract coverage
+- Product Strategist → strategy-brief (problem-framing, scope, non-goals, tradeoffs)
+- Spec Writer → implementation-spec (acceptance-criteria, edge-cases, interface-spec)
+- Backend/Frontend Engineer → change-plan (files-to-change, implementation-approach, risk-notes)
+- Test Engineer → test-package (test-plan, test-cases, false-confidence-assessment)
+- Security Reviewer → security-findings (findings, severity-assessment, recommendations)
+- Critic Reviewer → verdict (verdict, evidence, required-corrections)
+- And 14 more roles with full contracts
+
+### Evidence
+- 385 tests, zero failures
+- 27 new artifact tests
+
+## 1.5.0
+
+### Added
+
+#### Hook Spine / Runtime Enforcement (Phase R)
+- 5 lifecycle hooks: SessionStart, UserPromptSubmit, PreToolUse, SubagentStart, Stop
+- `scaffoldHooks()` generates all 5 hook scripts in .claude/hooks/
+- `roleos init claude` now scaffolds hooks + settings.local.json with hook config
+- `roleos doctor` now checks for hook scripts (check 7) and settings hooks (check 8)
+
+#### SessionStart hook
+- Establishes session contract on every new session
+- Records session ID, timestamp, initializes state tracking
+- Adds context reminding Claude to use /roleos-route for non-trivial tasks
+
+#### UserPromptSubmit hook
+- Classifies prompts as substantial (>50 chars + action verbs)
+- After 2+ substantial prompts without a route card, adds context reminder
+- Does not block — advisory enforcement
+
+#### PreToolUse hook
+- Records all tool usage in session state
+- Flags write tools (Bash, Write, Edit) used without route card after substantial work
+- Advisory, not blocking — preserves operator control
+
+#### SubagentStart hook
+- Injects active role contract into delegated agents
+- Ensures subagents inherit the Role OS session context
+
+#### Stop hook
+- Warns when substantial sessions end without route card or outcome artifact
+- Advisory — does not block session exit
+- Trivial sessions (< 2 substantial prompts) are exempt
+
+### Evidence
+- 358 tests, zero failures
+- 23 new hook tests covering all 5 lifecycle hooks
+
+## 1.4.0
+
+### Added
+
+#### Session Spine (Phase Q)
+- `roleos init claude` — scaffolds Claude Code integration: CLAUDE.md instructions, /roleos-route + /roleos-review + /roleos-status slash commands
+- `roleos doctor` — verifies repo is correctly wired for Role OS sessions (6 checks: .claude/ dir, CLAUDE.md section, /roleos-route command, context files, role contracts, packets)
+- Route card generation — session header artifact proving Role OS was engaged (task type, pack, confidence, composite status, success artifact)
+- CLAUDE.md template instructs Claude to route through Role OS before non-trivial work
+- /roleos-route command produces structured route cards
+- /roleos-review command guides structured verdict production
+- /roleos-status command shows active work and context health
+- Appends to existing CLAUDE.md without overwriting (detects Role OS section)
+- --force flag overwrites existing command files
+
+### Evidence
+- 335 tests, zero failures
+
+## 1.3.0
+
+### Added
+
+#### Outcome Calibration (Phase M)
+- Run outcome ledger — append-only JSONL recording pack selection, confidence, overrides, escalations, corrections, completion status
+- `computeCalibration()` — pack usage rates, high-confidence accuracy, operator override rates, per-pack performance
+- `computePackBoosts()` — weight tuning from clean completed runs (+0.5/run, capped at 2.0)
+- `computeConfidenceAdjustment()` — raises threshold when high-confidence is often overridden, lowers when medium is often accepted
+- Auto-generated calibration suggestions when metrics drift
+- Safety constraint: calibration never overrides mismatch guards, conflict rules, escalation honesty, or evidence requirements
+
+#### Mixed-Task Decomposition (Phase N)
+- `detectComposite()` — 7 subtask categories (build, bugfix, security, docs, research, launch, treatment) with signal-based detection
+- Structural connector detection ("and then", "after that", "plus", "also")
+- Confidence levels: high (3+ categories or 2+ with connectors), medium, low
+- `decompose()` — generates linked child packets sorted by phase order
+- `createRunPlan()` — dependency-aware parent plan with child tracking
+- Honest fallback: medium/low confidence shows uncertainty warning with `--no-split` override
+
+#### Composite Execution (Phase O)
+- `initExecution()` / `advance()` — dependency-driven child execution with artifact passing
+- 7 artifact contracts defining what each category produces and expects
+- Artifact ledger tracking all cross-packet handoffs
+- `blockChild()` / `recoverChild()` / `failChild()` — branch recovery with transitive cascade
+- `invalidateDownstream()` — resets stale children when upstream changes, removes stale artifacts
+- `synthesize()` — truthful parent-level completion report
+- Independent branches continue unaffected when a sibling fails
+
+#### Adaptive Replanning (Phase P)
+- 6 structured change event types: scope-change, artifact-changed, new-requirement, review-finding, dependency-discovered, priority-change
+- `analyzeImpact()` — identifies valid/stale children, stale artifacts, whether new children or reorder needed
+- `replan()` — selective replanning: invalidates only affected branches, inserts new children, updates dependencies
+- Plan diff: shows what changed, what stayed valid, what reopened, what was inserted
+- Execution resumes from next valid child after replan — no restart required
+
+### Evidence
+- 317 tests, zero failures
+- Calibration, decomposition, composite execution, and replanning each have dedicated test suites
+
+## 1.2.0
+
+### Added
+- Pack auto-selection in `roleos route` — suggests best pack when confidence is high
+- `roleos route --pack=<name>` — use a specific pack for routing
+- Pack mismatch detection — warns when a pack doesn't fit the task, suggests the correct alternative
+- Pack fallback — mismatched or unknown packs fall back to free routing automatically
+- `checkPackMismatch()` API with 7 guard sets covering all pack×task-type combinations
+- `getPackRoles()` API with conditional Orchestrator support
+
+### Changed
+- Docs pack: Support Triage Lead now opens (was Feedback Synthesizer). Feedback Synthesizer is second. Release Engineer + Deployment Verifier moved to optional (overhead for docs-only tasks).
+- Pack calibration applied from comparison evidence: conditional Orchestrator, Security Reviewer in Treatment, Product Strategist opens Research, mismatch guards on all 7 packs.
+
+### Evidence
+- Pack comparison: calibrated packs now win or tie 6/7 (was 2/7 pre-calibration)
+- Misfit honesty: 0 full bluffs, 0 undetected partial bluffs (was 1 + 3)
+- 230 tests, zero failures
+
+## 1.1.0
+
+### Added
+
+#### Routing
+- Full 31-role catalog — all roles scored by keyword, trigger phrase, packet type bias, and deliverable affinity
+- Dynamic chain builder — phase-ordered assembly replacing static templates
+- Routing confidence assessment (high/medium/low)
+- `excludeWhen` enforcement — roles suppressed when exclusion patterns match packet content
+- `detectType` false-positive prevention — "integration testing" no longer triggers integration type
+- `--verbose` flag for `roleos route` — hides scoring noise by default
+
+#### Conflict Detection
+- 4-pass conflict engine: hard conflicts, sequence, redundancy, coverage gaps
+- Per-role constraint registry: lateOnly, requiresBeforePacks
+- Overlap pair detection
+- Repair suggestions on every finding
+
+#### Escalation Auto-Routing
+- Blocked/rejected/conflict/split work auto-routes to named resolver
+- Every escalation includes: target role, recovery type, required artifact, handoff context
+
+#### Structured Evidence
+- 12 evidence kinds, 4 statuses, closed 4-verdict enum (accept/accept-with-notes/reject/blocked)
+- Role-aware evidence requirements for 15 roles
+- Sufficiency checks with contradiction detection
+
+#### Runtime Dispatch
+- Execution manifests for multi-claude with per-role tool profiles and budgets
+- 8 execution states with auto-advance
+- Escalation packet generation for blocked/rejected steps
+
+#### Proven Team Packs
+- 7 battle-tested packs: feature, bugfix, security, docs, launch, research, treatment
+- `roleos packs list` — show all packs with role counts
+- `roleos packs suggest <packet>` — suggest best pack for a packet
+- `roleos packs show <name>` — show pack details (roles, artifacts, stop conditions)
+- Pack suggestion engine with confidence levels
+
+#### Trials
+- Full roster proven: 30/30 gold-task trials + 5/5 negative (wrong-task honesty) trials
+- 7 pack execution trials — all packs ran full chains with honest Critic verdicts
+- Trial framework: buildClusterTrials, evaluateTrialOutput, formatTrialReport
+
+### Changed
+- 32 → 31 roles: Information Architect merged into Docs Architect
+- Verdict vocabulary unified: evidence.mjs now uses accept/reject/blocked (matching review.mjs)
+- "worker" terminology replaced with "role" in dispatch.mjs
+
+### Fixed
+- `excludeWhen` was declared on 14 roles but never enforced — now active in scoreRole
+- `detectType` false-positived on "integration testing" — now uses word-boundary regex
+- "Not triggered: N roles" noise hidden by default (shown with --verbose)
+- Handbook: Team Packs page added, reference sidebar reordered
+
+## 1.0.2
+
+### Fixed
+- Fix double-nested `.claude/.claude/` directory created by `roleos init` — `starter-pack/.claude/workflows/full-treatment.md` moved to `starter-pack/workflows/`
+- Read VERSION from `package.json` at runtime instead of hardcoded constant — prevents version drift between CLI and package metadata
+
+### Added
+- `roleos init --force` — update canonical scaffolded files while always protecting user-filled `context/` files
+- 4 regression tests: no double-nesting, correct workflow placement, version sync, --force context protection
+
+## 1.0.0
+
+### Added
+- `roleos init` — scaffold Role OS starter pack into `.claude/`
+- `roleos packet new <type>` — create feature, integration, or identity packets
+- `roleos route <packet-file>` — recommend smallest valid role chain with dependency verification
+- `roleos review <packet-file> <verdict>` — record accept/reject/blocked verdicts
+- Full starter pack: 8 role contracts, 3 schemas, 4 policies, 3 workflows
+- Guided context templates with inline prompts
+- 3 canonical example packets (feature, integration, identity)
+- Adoption handbook