refactor: migrate review-orchestrator to skill-based architecture

CHANGELOG.md

@@ -862,7 +862,7 @@ All command names have been simplified for clarity:
   - `ci-fixer.md` (sonnet): Fix CI failures and PR comments, called by ci-monitor
   - `simple-fixer.md` (haiku): Execute pre-defined code fixes mechanically
 - **Workflow Enforcement Gates** - Explicit STOP gates in all agents
-  - Agents cannot skip review-orchestrator, delivery-validator, docs-updater
+  - Agents cannot skip Phase 9 review loop, delivery-validator, docs-updater
   - Agents cannot create PRs - only /ship creates PRs
   - SubagentStop hooks enforce mandatory workflow sequence
 - **State Schema Files**

CLAUDE.md

@@ -49,7 +49,7 @@ lib/                    # Shared library (canonical source)
 └── index.js            # Main exports
 
 plugins/                # Claude Code plugins
-├── next-task/          # Master workflow (14 agents)
+├── next-task/          # Master workflow (12 agents)
 ├── ship/               # PR workflow
 ├── deslop/      # AI slop cleanup
 ├── audit-project/     # Multi-agent review
@@ -91,7 +91,7 @@ Platform-aware: `.claude/` (Claude), `.opencode/` (OpenCode), `.codex/` (Codex)
 Cannot skip in /next-task:
 - `exploration-agent` → before planning
 - `planning-agent` → before implementation
-- `review-orchestrator` → before shipping
+- **Phase 9 review loop** → MUST use orchestrate-review skill, spawns parallel reviewers, iterates until clean
 - `delivery-validator` → before /ship
 
 ## PR Auto-Review
@@ -137,6 +137,6 @@ Choose the appropriate model based on task complexity and quality multiplier eff
 **Examples**:
 - `/enhance:agent` uses opus - false positives damage agent quality across entire codebase
 - `simple-fixer` uses haiku - mechanically applies pre-defined fixes with no judgment
-- `review-orchestrator` uses opus - review quality affects entire workflow
+- Phase 9 review loop spawns sonnet reviewers - multiple focused agents reduce rubber-stamping
 - `worktree-manager` uses haiku - scripted git commands with no decision-making
 

README.md

@@ -75,7 +75,7 @@ This means you can run `/deslop apply` and trust that it won't break things.
 
 ### 2. Review Loops With Safeguards
 
-The review-orchestrator agent runs core review passes (code quality, security, performance, test coverage) plus conditional specialists until there are no open issues. Then it runs deslop-work on its own fixes to catch any AI artifacts it introduced.
+The Phase 9 review loop spawns parallel reviewer agents (code quality, security, performance, test coverage) plus conditional specialists, iterating until no open issues remain. It runs deslop-work after each iteration to catch any AI artifacts.
 
 ### 3. Workflow Enforcement
 
@@ -122,7 +122,7 @@ Frontier models write good code. That's solved. What's not solved:
 
 **1. One agent, one job, done extremely well**
 
-Same principle as good code: single responsibility. The exploration-agent explores. The implementation-agent implements. The review-orchestrator coordinates reviews. No agent tries to do everything. 29 specialized agents, each with narrow scope and clear success criteria.
+Same principle as good code: single responsibility. The exploration-agent explores. The implementation-agent implements. Phase 9 spawns multiple focused reviewers. No agent tries to do everything. 28 specialized agents, each with narrow scope and clear success criteria.
 
 **2. Pipeline with gates, not a monolith**
 
@@ -239,7 +239,6 @@ With proper workflow structure, context management, and quality gates, AI agents
 | implementation-agent | opus | Writes the actual code |
 | deslop-work | sonnet | Removes AI artifacts before review |
 | test-coverage-checker | sonnet | Validates tests exist and are meaningful |
-| review-orchestrator | opus | Coordinates parallel review agents |
 | delivery-validator | sonnet | Final checks before shipping |
 | docs-updater | sonnet | Updates documentation |
 | ci-monitor | haiku | Watches CI status |
@@ -552,7 +551,7 @@ implementation-agent writes code
     ↓
 deslop-work cleans AI artifacts
     ↓
-review-orchestrator iterates until approved
+Phase 9 review loop iterates until approved
     ↓
 delivery-validator checks requirements
     ↓

adapters/codex/install.sh

@@ -124,8 +124,47 @@ for mapping in "${SKILL_MAPPINGS[@]}"; do
   fi
 done
 
+# Install native skills (already have SKILL.md format)
+echo
+echo "📚 Installing native skills..."
+
+# Native skill mappings: skill_name:plugin:skill_name_in_source
+NATIVE_SKILL_MAPPINGS=(
+  "orchestrate-review:next-task:orchestrate-review"
+)
+
+for mapping in "${NATIVE_SKILL_MAPPINGS[@]}"; do
+  IFS=':' read -r SKILL_NAME PLUGIN SOURCE_SKILL <<< "$mapping"
+  SOURCE_SKILL_DIR="$REPO_ROOT/plugins/$PLUGIN/skills/$SOURCE_SKILL"
+  TARGET_SKILL_DIR="$CODEX_SKILLS_DIR/$SKILL_NAME"
+
+  if [ -d "$SOURCE_SKILL_DIR" ]; then
+    # Create skill directory
+    mkdir -p "$TARGET_SKILL_DIR"
+
+    # Copy SKILL.md
+    if [ -f "$SOURCE_SKILL_DIR/SKILL.md" ]; then
+      cp "$SOURCE_SKILL_DIR/SKILL.md" "$TARGET_SKILL_DIR/SKILL.md"
+      echo "  ✓ Installed native skill: \$${SKILL_NAME}"
+    else
+      echo "  ⚠️  Skipped \$${SKILL_NAME} (SKILL.md not found)"
+      continue
+    fi
+
+    # Copy optional subdirectories (references/, scripts/, assets/)
+    for subdir in references scripts assets; do
+      if [ -d "$SOURCE_SKILL_DIR/$subdir" ]; then
+        cp -r "$SOURCE_SKILL_DIR/$subdir" "$TARGET_SKILL_DIR/"
+        echo "    ✓ Copied $subdir/ directory"
+      fi
+    done
+  else
+    echo "  ⚠️  Skipped \$${SKILL_NAME} (source not found: $SOURCE_SKILL_DIR)"
+  fi
+done
+
 # Remove old/deprecated skills and prompts
-OLD_SKILLS=("deslop" "review" "reality-check-set" "pr-merge")
+OLD_SKILLS=("deslop" "review" "reality-check-set" "pr-merge" "review-orchestrator")
 for old_skill in "${OLD_SKILLS[@]}"; do
   if [ -d "$CODEX_SKILLS_DIR/$old_skill" ]; then
     rm -rf "$CODEX_SKILLS_DIR/$old_skill"

adapters/opencode-plugin/index.ts

@@ -34,7 +34,6 @@ const AGENT_THINKING_CONFIG: Record<string, { budget: number; description: strin
 
   // Reasoning tier - extended thinking (16k)
   "planning-agent": { budget: 16000, description: "Implementation planning" },
-  "review-orchestrator": { budget: 16000, description: "Multi-agent review coordination" },
   "delivery-validator": { budget: 16000, description: "Delivery validation" },
   "implementation-agent": { budget: 16000, description: "Code implementation" },
 

agent-docs/OPENCODE-REFERENCE.md

@@ -531,7 +531,7 @@ OpenCode supports thinking/reasoning across **multiple providers** with differen
 **Via Plugin Hook (runtime):**
 ```typescript
 "chat.params": async (input, output) => {
-  if (input.agent === "review-orchestrator") {
+  if (input.agent === "planning-agent") {
     output.options.thinking = { type: "enabled", budgetTokens: 16000 }
   }
 }
@@ -662,7 +662,7 @@ export const WorkflowPlugin: Plugin = async (ctx) => {
 
     "chat.params": async (input, output) => {
       // Use higher reasoning for complex agents
-      if (["review-orchestrator", "planning-agent"].includes(input.agent)) {
+      if (["planning-agent", "implementation-agent"].includes(input.agent)) {
         output.options.thinking = { type: "enabled", budgetTokens: 16000 }
       }
     }
@@ -890,10 +890,6 @@ Add to user's `opencode.jsonc`:
     },
 
     // Complex agents - extended thinking
-    "review-orchestrator": {
-      "model": "anthropic/claude-sonnet-4",
-      "options": { "thinking": { "type": "enabled", "budgetTokens": 16000 } }
-    },
     "planning-agent": {
       "model": "anthropic/claude-sonnet-4",
       "options": { "thinking": { "type": "enabled", "budgetTokens": 16000 } }

agent-docs/workflow.md

@@ -25,7 +25,7 @@ The main orchestrator **MUST spawn these agents in order**:
 | 7 | `implementation-agent` | opus | Read, Write, Edit, Bash | Execute plan |
 | 8 | `deslop-work` | sonnet | Read, Grep, Edit, Bash(git:*) | Clean AI slop (uses pipeline.js) |
 | 8 | `test-coverage-checker` | sonnet | Bash(npm:*), Read, Grep | Validate test coverage |
-| 9 | `review-orchestrator` | opus | Task(review) | Multi-pass review loop |
+| 9 | Phase 9 review loop | sonnet reviewers | Task(general-purpose) | Multi-pass review with parallel agents |
 | 10 | `delivery-validator` | sonnet | Bash(npm:*), Read | Validate completion |
 | 11 | `docs-updater` | sonnet | Read, Edit, Task(simple-fixer) | Update documentation |
 | 12 | `/ship` command | - | - | PR creation and merge |
@@ -34,13 +34,13 @@ The main orchestrator **MUST spawn these agents in order**:
 
 - **`exploration-agent`** - Required for understanding codebase before planning
 - **`planning-agent`** - Required for creating implementation plan
-- **`review-orchestrator`** - Required for code review before shipping
+- **Phase 9 review loop** - Required for code review before shipping (uses orchestrate-review skill)
 - **`delivery-validator`** - Required before calling /ship
 
 ### Review Decision Gate
 
-If review-orchestrator reports `blocked: true` (iteration limit or stall), /next-task must decide:
-- Re-run review-orchestrator with `--resume`, or
+If Phase 9 review loop reports `blocked: true` (iteration limit or stall), /next-task must decide:
+- Re-run Phase 9 review loop, or
 - Override and continue if issues are non-blocking (clear the queue file).
 
 ---

docs/CROSS_PLATFORM.md

@@ -81,7 +81,7 @@ claude --plugin-dir /path/to/awesome-slash/plugins/next-task
 | exploration-agent | opus | Deep codebase analysis |
 | planning-agent | opus | Design implementation plans |
 | implementation-agent | opus | Execute plans with quality code |
-| review-orchestrator | opus | Multi-agent review iteration |
+| Phase 9 review loop | sonnet reviewers | Multi-pass review with parallel agents |
 | deslop-work | sonnet | Clean AI slop from changes |
 | test-coverage-checker | sonnet | Validate test coverage |
 | delivery-validator | sonnet | Autonomous delivery validation |
@@ -150,7 +150,7 @@ The native plugin (`~/.opencode/plugins/awesome-slash/`) provides deep integrati
 | Execution | 0 | worktree-manager, simple-fixer, ci-monitor |
 | Discovery | 8k | task-discoverer, docs-updater |
 | Analysis | 12k | exploration-agent, deslop-work, ci-fixer |
-| Reasoning | 16k | planning-agent, implementation-agent, review-orchestrator |
+| Reasoning | 16k | planning-agent, implementation-agent |
 | Synthesis | 20k | plan-synthesizer, enhancement-orchestrator |
 
 **Provider-Specific Thinking:**

docs/README.md

@@ -42,7 +42,7 @@ AI models can write code. The bottleneck is everything else—picking tasks, man
 
 | Document | Description |
 |----------|-------------|
-| [reference/AGENTS.md](./reference/AGENTS.md) | All 32 agents: purpose, model, tools, restrictions. <!-- AGENT_COUNT_TOTAL: 32 --> |
+| [reference/AGENTS.md](./reference/AGENTS.md) | All 31 agents: purpose, model, tools, restrictions. <!-- AGENT_COUNT_TOTAL: 31 --> |
 | [reference/SLOP-PATTERNS.md](./reference/SLOP-PATTERNS.md) | All detection patterns by language, severity, auto-fix. |
 | [reference/MCP-TOOLS.md](./reference/MCP-TOOLS.md) | MCP server tools: parameters, returns, platform config. |
 

docs/reference/AGENTS.md

@@ -2,15 +2,15 @@
 
 Complete reference for all agents in awesome-slash.
 
-**TL;DR:** 32 agents across 5 plugins. opus for reasoning, sonnet for patterns, haiku for execution. Each agent does one thing well. <!-- AGENT_COUNT_TOTAL: 32 -->
+**TL;DR:** 31 agents across 5 plugins. opus for reasoning, sonnet for patterns, haiku for execution. Each agent does one thing well. <!-- AGENT_COUNT_TOTAL: 31 -->
 
 ---
 
 ## Quick Navigation
 
 | Plugin | Agents | Jump to |
 |--------|--------|---------|
-| next-task | 13 | [task-discoverer](#task-discoverer), [worktree-manager](#worktree-manager), [exploration-agent](#exploration-agent), [planning-agent](#planning-agent), [implementation-agent](#implementation-agent), [deslop-work](#deslop-work), [test-coverage-checker](#test-coverage-checker), [review-orchestrator](#review-orchestrator), [delivery-validator](#delivery-validator), [docs-updater](#docs-updater), [simple-fixer](#simple-fixer), [ci-monitor](#ci-monitor), [ci-fixer](#ci-fixer) |
+| next-task | 12 | [task-discoverer](#task-discoverer), [worktree-manager](#worktree-manager), [exploration-agent](#exploration-agent), [planning-agent](#planning-agent), [implementation-agent](#implementation-agent), [deslop-work](#deslop-work), [test-coverage-checker](#test-coverage-checker), [delivery-validator](#delivery-validator), [docs-updater](#docs-updater), [simple-fixer](#simple-fixer), [ci-monitor](#ci-monitor), [ci-fixer](#ci-fixer) |
 | audit-project | 10 | [code-quality-reviewer](#code-quality-reviewer), [security-expert](#security-expert), [performance-engineer](#performance-engineer), [test-quality-guardian](#test-quality-guardian), [architecture-reviewer](#architecture-reviewer), [database-specialist](#database-specialist), [api-designer](#api-designer), [frontend-specialist](#frontend-specialist), [backend-specialist](#backend-specialist), [devops-reviewer](#devops-reviewer) |
 | enhance | 7 | [enhancement-orchestrator](#enhancement-orchestrator), [plugin-enhancer](#plugin-enhancer), [agent-enhancer](#agent-enhancer), [claudemd-enhancer](#claudemd-enhancer), [docs-enhancer](#docs-enhancer), [prompt-enhancer](#prompt-enhancer), [enhancement-reporter](#enhancement-reporter) |
 | drift-detect | 1 | [plan-synthesizer](#plan-synthesizer) |
@@ -26,7 +26,7 @@ Complete reference for all agents in awesome-slash.
 
 ## Overview
 
-awesome-slash uses 32 specialized agents across 5 plugins. Each agent is optimized for a specific task and assigned a model based on complexity:
+awesome-slash uses 31 specialized agents across 5 plugins. Each agent is optimized for a specific task and assigned a model based on complexity:
 
 | Model | Use Case | Cost |
 |-------|----------|------|
@@ -35,7 +35,7 @@ awesome-slash uses 32 specialized agents across 5 plugins. Each agent is optimiz
 | haiku | Mechanical execution, no judgment | Low |
 
 **Agent types:**
-- **File-based agents** (22) - Defined in `plugins/*/agents/*.md` with frontmatter <!-- AGENT_COUNT_FILE_BASED: 22 -->
+- **File-based agents** (21) - Defined in `plugins/*/agents/*.md` with frontmatter <!-- AGENT_COUNT_FILE_BASED: 21 -->
 - **Role-based agents** (10) - Defined inline via Task tool with specialized prompts <!-- AGENT_COUNT_ROLE_BASED: 10 -->
 
 ---
@@ -200,39 +200,6 @@ awesome-slash uses 32 specialized agents across 5 plugins. Each agent is optimiz
 
 ---
 
-### review-orchestrator
-
-**Model:** opus
-**Purpose:** Coordinate multi-agent review until clean.
-
-**What it does:**
-1. Launches core review passes (parallel when nested subagents are supported; otherwise runs in-agent passes in series on Claude Code):
-   - Code quality (includes error handling)
-   - Security
-   - Performance
-   - Test coverage
-2. Adds conditional specialists (DB, architecture, API, frontend, backend, devops)
-3. Aggregates findings by severity
-4. Writes a review queue file in the platform state dir
-5. Fixes all non-false-positive issues
-6. Runs a deslop pass after each iteration (deslop-work when nested subagents are supported; inline slop scan on Claude Code)
-7. Loops until no open issues remain
-8. Stops early on iteration limit or stall and returns control for decision
-
-**Tools available:**
-- Task (only when nested subagents are supported)
-- Bash (git)
-- Read, Write, Edit
-
-**Restrictions:**
-- MUST NOT create PR
-- MUST NOT push
-- MUST NOT invoke delivery-validator
-
-**Why opus:** Review coordination requires judgment. Which findings are real? Which can be auto-fixed? How to prioritize? Opus handles this well.
-
----
-
 ### delivery-validator
 
 **Model:** sonnet
@@ -666,7 +633,6 @@ Agents have restricted tool access for safety:
 | Agent | Restricted From | Why |
 |-------|-----------------|-----|
 | implementation-agent | PR creation, git push | Workflow enforces order |
-| review-orchestrator | PR creation, git push | Must complete review first |
 | delivery-validator | PR creation, git push | Must pass validation first |
 | worktree-manager | Most tools | Only needs git |
 | simple-fixer | Most tools | Only needs edit |

docs/workflows/NEXT-TASK.md

@@ -182,29 +182,29 @@ Both agents run in parallel:
 
 ### Phase 9: Review Loop
 
-**Agent:** review-orchestrator (opus)
+**Execution:** Inline in main orchestrator (uses orchestrate-review skill)
 **Human interaction: No**
 
-The agent:
-1. Launches core review passes (parallel when nested subagents are supported; otherwise runs in-agent passes in series on Claude Code):
-   - Code quality (includes error handling)
-   - Security
-   - Performance
-   - Test coverage
-2. Adds conditional specialists (DB, architecture, API, frontend, backend, devops)
-3. Aggregates findings by severity (critical/high/medium/low)
-4. Fixes all non-false-positive issues
-5. Writes a review queue file in the platform state dir
-6. **Runs a deslop pass after EACH iteration** (uses deslop-work where nested subagents are supported; inline slop scan on Claude Code)
-7. Repeats until no open issues remain
-8. Stops early if iteration limit or stall detected; control returns to /next-task for decision
+The orchestrator:
+1. Reads orchestrate-review skill for guidance
+2. Detects content signals (database, API, frontend, backend, devops, architecture)
+3. Spawns parallel Task agents (general-purpose, sonnet) - one per review pass:
+   - Core (always): code quality, security, performance, test coverage
+   - Conditional: database, architecture, api, frontend, backend, devops
+4. Aggregates findings by severity (critical/high/medium/low)
+5. Fixes all non-false-positive issues
+6. Commits fixes
+7. Runs deslop-work after each iteration
+8. Re-reviews changed files
+9. Repeats until no open issues remain
+10. Stops early if iteration limit or stall detected
 
 The loop continues until clean, but stops early if iteration limits or stall detection trigger.
 
 **Restrictions enforced:**
 - MUST NOT create PR
 - MUST NOT push to remote
-- MUST NOT invoke delivery-validator (handled by workflow)
+- MUST NOT skip review loop
 
 ---
 
@@ -318,7 +318,7 @@ A SubagentStop hook enforces the workflow sequence. When any agent completes, th
 
 **Enforced rules:**
 - Cannot skip deslop-work or test-coverage-checker
-- Cannot skip review-orchestrator
+- Cannot skip Phase 9 review loop
 - Cannot skip delivery-validator
 - Cannot skip docs-updater
 - Cannot create PR before `/ship` is invoked
@@ -330,7 +330,7 @@ A SubagentStop hook enforces the workflow sequence. When any agent completes, th
 
 | Model | Agents | Why |
 |-------|--------|-----|
-| **opus** | exploration-agent, planning-agent, implementation-agent, review-orchestrator | Complex reasoning, quality-critical phases |
+| **opus** | exploration-agent, planning-agent, implementation-agent | Complex reasoning, quality-critical phases |
 | **sonnet** | task-discoverer, deslop-work, test-coverage-checker, delivery-validator, docs-updater, ci-fixer | Moderate reasoning, structured tasks |
 | **haiku** | worktree-manager, simple-fixer, ci-monitor | Mechanical execution, no judgment needed |
 

docs/workflows/SHIP.md

@@ -184,7 +184,7 @@ mutation($threadId: ID!) {
 
 ### Phase 7: Internal Review (Standalone Only)
 
-**Skipped when called from `/next-task`** (review already completed by review-orchestrator).
+**Skipped when called from `/next-task`** (review already completed by Phase 9 review loop).
 
 When standalone, launches core review passes in parallel:
 - Code quality (includes error handling)
@@ -335,7 +335,7 @@ Outputs detailed logging for each phase.
 When called from `/next-task` (via `--state-file` argument):
 
 **Skipped phases:**
-- Phase 7 (internal review) - Already done by review-orchestrator
+- Phase 7 (internal review) - Already done by Phase 9 review loop
 - Deslop cleanup - Already done by deslop-work
 
 **Still runs:**