bd sync: 2026-01-09 20:39:37

- CLI tool with check/enable/approve/reject/status commands
- Stop hook integration for Claude Code blocking
- Adversarial reviewer prompt template
- Dual-publish structure (SKILL.md + .claude-plugin)
- Session ID auto-detection from env vars or git state

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

.beads/issues.jsonl

@@ -119,7 +119,7 @@
 {"id":"skills-mjf","title":"Design: Portable adversarial reviewer","description":"Design a reviewer agent/skill that can run on any capable model.\n\nalice is Claude Opus with specific tools. We need:\n- Model-agnostic reviewer prompt/instructions\n- Tool requirements (read-only: Read, Grep, Glob, Bash)\n- Integration with orch for multi-model consensus\n- Decision format (APPROVED/ISSUES)\n- Issue filing (beads or tissue)\n\nKey principles from alice:\n- Work for the USER, not the agent\n- Assume errors exist, find them\n- Steel-man then attack\n- Seek second opinions\n\nOutput: Reviewer skill spec that works across agents","status":"closed","priority":3,"issue_type":"task","created_at":"2026-01-09T17:14:20.778647076-08:00","created_by":"dan","updated_at":"2026-01-09T19:59:37.80146821-08:00","closed_at":"2026-01-09T19:59:37.80146821-08:00","close_reason":"Covered in architecture design doc - adversarial reviewer section"}
 {"id":"skills-mx3","title":"spec-review: Define consensus thresholds and decision rules","description":"'Use judgment' for mixed results leads to inconsistent decisions.\n\nDefine:\n- What constitutes consensus (2/3? unanimous?)\n- How to handle NEUTRAL votes\n- Tie-break rules\n- When human override is acceptable and how to document it","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-15T00:23:24.121175736-08:00","updated_at":"2025-12-15T13:58:04.339283238-08:00","closed_at":"2025-12-15T13:58:04.339283238-08:00"}
 {"id":"skills-njb","title":"worklog: clarify or remove semantic compression references","description":"SKILL.md references 'semantic compression is a planned workflow' multiple times but it's not implemented. Speculative generality - adds cognitive load for non-existent feature. Either implement or move to design notes. Found by smells lens review.","status":"closed","priority":4,"issue_type":"task","created_at":"2025-12-25T02:03:25.387405002-05:00","updated_at":"2025-12-27T10:11:48.169923742-05:00","closed_at":"2025-12-27T10:11:48.169923742-05:00","close_reason":"Closed"}
-{"id":"skills-nto","title":"Prototype: End-to-end cross-agent workflow","description":"Build a working prototype of cross-agent quality gate.\n\n## Scenario\n1. Worker agent (any) does task\n2. Posts status to message layer\n3. Reviewer agent (any) checks work\n4. Posts approval/issues to memory layer\n5. Gate checks memory, allows/blocks completion\n\n## Test Matrix\n\n| Orchestrator | Worker | Reviewer | Enforcement |\n|--------------|--------|----------|-------------|\n| Claude | Claude | Gemini | Hook |\n| Claude | Gemini | Claude | Hook |\n| OpenCode | Claude | Gemini | Orchestrator |\n| Manual | OpenCode | Claude | Protocol |\n\n## Components to Build\n1. Message layer interface (post/read status)\n2. Memory layer interface (review state)\n3. Gate check CLI (for hooks and manual)\n4. Reviewer skill/prompt\n\n## Success Criteria\n- At least 2 agent combinations working\n- Gate actually blocks when review fails\n- State persists across agent boundaries","status":"open","priority":2,"issue_type":"task","created_at":"2026-01-09T19:32:57.579195169-08:00","created_by":"dan","updated_at":"2026-01-09T19:32:57.579195169-08:00"}
+{"id":"skills-nto","title":"Prototype: End-to-end cross-agent workflow","description":"Build a working prototype of cross-agent quality gate.\n\n## Scenario\n1. Worker agent (any) does task\n2. Posts status to message layer\n3. Reviewer agent (any) checks work\n4. Posts approval/issues to memory layer\n5. Gate checks memory, allows/blocks completion\n\n## Test Matrix\n\n| Orchestrator | Worker | Reviewer | Enforcement |\n|--------------|--------|----------|-------------|\n| Claude | Claude | Gemini | Hook |\n| Claude | Gemini | Claude | Hook |\n| OpenCode | Claude | Gemini | Orchestrator |\n| Manual | OpenCode | Claude | Protocol |\n\n## Components to Build\n1. Message layer interface (post/read status)\n2. Memory layer interface (review state)\n3. Gate check CLI (for hooks and manual)\n4. Reviewer skill/prompt\n\n## Success Criteria\n- At least 2 agent combinations working\n- Gate actually blocks when review fails\n- State persists across agent boundaries","status":"closed","priority":2,"issue_type":"task","created_at":"2026-01-09T19:32:57.579195169-08:00","created_by":"dan","updated_at":"2026-01-09T20:39:24.013666826-08:00","closed_at":"2026-01-09T20:39:24.013666826-08:00","close_reason":"Prototype complete: review-gate CLI with hooks.json, adversarial reviewer prompt, and dual-publish structure"}
 {"id":"skills-oes","title":"Define skill manifest format","description":"Skills need to declare their interface so beads can validate.\n\nManifest should include:\n- Required inputs (args, env vars)\n- Optional inputs with defaults\n- Expected outputs (files, artifacts)\n- Preconditions (tools, repos, permissions)\n\nLocation: SKILL.md frontmatter or separate manifest.yaml\n\nEnables: Proto validation before spawning, better error messages.","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-23T19:49:30.673372413-05:00","updated_at":"2025-12-23T20:55:04.427620449-05:00","closed_at":"2025-12-23T20:55:04.427620449-05:00","close_reason":"ADRs revised with orch consensus feedback"}
 {"id":"skills-p2o","title":"Refactor update-agent-context.sh: array+loop for agents","description":"File: .specify/scripts/bash/update-agent-context.sh (772 lines)\n\nIssues:\n- 12 nearly-identical if-blocks in update_all_existing_agents() (lines 632-701)\n- Should be refactored into loop with array of agent configurations\n- Current pattern repeats: if [[ -f \"$CLAUDE_FILE\" ]]; then update_agent_file...\n\nFix:\n- Create AGENTS array with (file, name, format) tuples\n- Replace 12 if-blocks with single for loop\n- Estimated reduction: 60 lines\n\nSeverity: HIGH","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-24T02:50:57.385820971-05:00","updated_at":"2025-12-25T01:44:58.370191619-05:00","closed_at":"2025-12-25T01:44:58.370191619-05:00","close_reason":"update-agent-context.sh is .specify upstream code, not maintained here"}
 {"id":"skills-p3v","title":"Cross-language FFI wormholes via LSP","description":"Bridge FFI boundaries where standard LSPs go blind:\n- Rust extern C → clangd lookup\n- Go CGO → match C symbols\n- Python FFI → trace bindings\n\nGenerate synthetic go-to-definition maps. When hovering over C call in Rust, intercept hover request, query C LSP, inject C definition into Rust tooltip.\n\nEnables seamless polyglot navigation.","status":"closed","priority":4,"issue_type":"feature","created_at":"2025-12-24T02:29:57.597602745-05:00","updated_at":"2025-12-29T14:37:35.354771695-05:00","closed_at":"2025-12-29T14:37:35.354771695-05:00","close_reason":"Parked: waiting on gastown (Steve Yegge's orchestration layer for beads). Revisit when gastown lands."}

.beads/last-touched

@@ -1 +1 @@
-skills-hf1
+skills-nto

.claude-plugin/marketplace.json

@@ -86,6 +86,11 @@
       "name": "worklog",
       "source": "./skills/worklog",
       "description": "Create structured worklogs documenting work sessions"
+    },
+    {
+      "name": "review-gate",
+      "source": "./skills/review-gate",
+      "description": "Quality gate for cross-agent review enforcement"
     }
   ]
 }

.gitignore

@@ -26,3 +26,6 @@ Thumbs.db
 # Environment
 .env
 .env.local
+
+# Review gate state (transient)
+.review-state/

skills/review-gate/.claude-plugin/plugin.json

@@ -0,0 +1,13 @@
+{
+  "name": "review-gate",
+  "version": "0.1.0",
+  "description": "Quality gate for cross-agent review enforcement",
+  "repository": "https://git.clarun.xyz/dan/skills",
+  "author": "dan",
+  "license": "MIT",
+  "claude_code": {
+    "skill_paths": [
+      "../skills/review-gate.md"
+    ]
+  }
+}

skills/review-gate/SKILL.md

@@ -0,0 +1,121 @@
+# review-gate
+
+Quality gate for cross-agent review enforcement. Blocks agent completion until work is reviewed and approved.
+
+## Purpose
+
+When agents work autonomously (unattended, CI/CD, batch jobs), they may:
+- Claim "done" when work is incomplete
+- Miss requirements or make incorrect assumptions
+- Take shortcuts that don't serve user intent
+
+review-gate enforces an external quality check before the agent can exit.
+
+## Usage
+
+### Enable review requirement
+
+At the start of work that needs review:
+
+```bash
+review-gate enable [session_id]
+```
+
+This creates a "pending" review state. The agent will be blocked from exiting until approved.
+
+### Approve review
+
+After reviewing work (manually or via reviewer agent):
+
+```bash
+review-gate approve [session_id]
+```
+
+### Reject review
+
+If issues are found:
+
+```bash
+review-gate reject [session_id] "reason" ["issue1" "issue2" ...]
+```
+
+### Check status
+
+```bash
+review-gate status [session_id]
+review-gate list  # all sessions
+```
+
+## Session ID
+
+Session ID is auto-detected in this order:
+1. Provided argument
+2. `REVIEW_SESSION_ID` env var
+3. `CLAUDE_SESSION_ID` env var
+4. Generated from git branch + commit
+
+## Hook Integration
+
+For Claude Code, add to your project's `.claude/hooks.json`:
+
+```json
+{
+  "hooks": {
+    "Stop": [{
+      "hooks": [{
+        "type": "command",
+        "command": "review-gate check",
+        "timeout": 30
+      }]
+    }]
+  }
+}
+```
+
+The Stop hook runs when the agent tries to exit. If review is pending or rejected, exit is blocked.
+
+## State Storage
+
+Review state is stored in `.review-state/` as JSON files. This directory should be gitignored (transient state, not code).
+
+## Commands
+
+| Command | Description | Exit Code |
+|---------|-------------|-----------|
+| `check [id]` | Check if exit allowed | 0=allowed, 1=blocked |
+| `enable [id]` | Enable review requirement | 0 |
+| `approve [id]` | Mark as approved | 0 |
+| `reject [id] [reason] [issues...]` | Mark as rejected | 0 |
+| `status [id]` | Show review state | 0 |
+| `list` | List all sessions | 0 |
+| `clean [age]` | Remove old states (default: 7d) | 0 |
+
+## Environment Variables
+
+| Variable | Description |
+|----------|-------------|
+| `REVIEW_SESSION_ID` | Override session ID |
+| `REVIEW_STATE_DIR` | Override state directory (default: `./.review-state`) |
+| `CLAUDE_SESSION_ID` | Fallback session ID (set by Claude Code) |
+
+## Cross-Agent Pattern
+
+For agents without Stop hooks, use orchestrator pattern:
+
+```bash
+# Orchestrator script
+review-gate enable "$SESSION_ID"
+<spawn worker agent>
+# Worker completes...
+<spawn reviewer agent>
+# Reviewer approves or rejects...
+review-gate check "$SESSION_ID" || {
+    echo "Review not passed, looping back to worker"
+    # ... retry logic
+}
+```
+
+## Dependencies
+
+- `jq` - JSON parsing
+- `bash` 4.0+

skills/review-gate/hooks/hooks.json

@@ -0,0 +1,16 @@
+{
+  "hooks": {
+    "Stop": [
+      {
+        "matcher": "",
+        "hooks": [
+          {
+            "type": "command",
+            "command": "review-gate check",
+            "timeout": 30
+          }
+        ]
+      }
+    ]
+  }
+}

skills/review-gate/scripts/review-gate

@@ -0,0 +1,238 @@
+#!/usr/bin/env bash
+# review-gate - Quality gate CLI for cross-agent review enforcement
+#
+# Usage:
+#   review-gate check [session_id]    - Check if review is approved (for hooks)
+#   review-gate enable [session_id]   - Enable review requirement
+#   review-gate approve [session_id]  - Mark review as approved
+#   review-gate reject [session_id] [reason]  - Mark review as rejected
+#   review-gate status [session_id]   - Show current review status
+#
+# Exit codes:
+#   0 - Review approved or not required (allow exit)
+#   1 - Review pending or rejected (block exit)
+
+set -euo pipefail
+
+# Session ID: use provided, or env var, or generate from git
+get_session_id() {
+    local provided="${1:-}"
+    if [[ -n "$provided" ]]; then
+        echo "$provided"
+    elif [[ -n "${REVIEW_SESSION_ID:-}" ]]; then
+        echo "$REVIEW_SESSION_ID"
+    elif [[ -n "${CLAUDE_SESSION_ID:-}" ]]; then
+        echo "$CLAUDE_SESSION_ID"
+    else
+        # Generate from git state
+        local branch=$(git branch --show-current 2>/dev/null || echo "main")
+        local short_sha=$(git rev-parse --short HEAD 2>/dev/null || echo "unknown")
+        echo "review-${branch}-${short_sha}"
+    fi
+}
+
+# State file location (simple file-based for prototype)
+STATE_DIR="${REVIEW_STATE_DIR:-${PWD}/.review-state}"
+mkdir -p "$STATE_DIR"
+
+get_state_file() {
+    local session_id="$1"
+    echo "${STATE_DIR}/${session_id}.json"
+}
+
+# Commands
+cmd_check() {
+    local session_id=$(get_session_id "${1:-}")
+    local state_file=$(get_state_file "$session_id")
+
+    if [[ ! -f "$state_file" ]]; then
+        # No review registered - allow exit
+        echo "No review required for session: $session_id"
+        exit 0
+    fi
+
+    local status=$(jq -r '.status' "$state_file" 2>/dev/null || echo "unknown")
+
+    case "$status" in
+        approved)
+            echo "✓ Review approved for session: $session_id"
+            exit 0
+            ;;
+        pending)
+            echo "⏳ BLOCKED: Review pending for session: $session_id"
+            echo ""
+            echo "To proceed, spawn a reviewer agent:"
+            echo "  /task Review the work in this session. Check git diff and file changes."
+            echo "  Then run: review-gate approve $session_id"
+            echo ""
+            echo "Or to skip review: review-gate approve $session_id"
+            exit 1
+            ;;
+        rejected)
+            local reason=$(jq -r '.reason // "No reason provided"' "$state_file")
+            local issues=$(jq -r '.issues[]? // empty' "$state_file" | head -5)
+            echo "❌ BLOCKED: Review rejected for session: $session_id"
+            echo "Reason: $reason"
+            if [[ -n "$issues" ]]; then
+                echo "Issues:"
+                echo "$issues" | sed 's/^/  - /'
+            fi
+            echo ""
+            echo "Fix the issues and request re-review."
+            exit 1
+            ;;
+        *)
+            echo "⚠ Unknown review status: $status"
+            exit 1
+            ;;
+    esac
+}
+
+cmd_enable() {
+    local session_id=$(get_session_id "${1:-}")
+    local state_file=$(get_state_file "$session_id")
+
+    cat > "$state_file" <<EOF
+{
+    "session_id": "$session_id",
+    "status": "pending",
+    "enabled_at": "$(date -Iseconds)",
+    "enabled_by": "${USER:-unknown}",
+    "worker_agent": "${CLAUDE_MODEL:-unknown}",
+    "issues": []
+}
+EOF
+
+    echo "✓ Review enabled for session: $session_id"
+    echo "  State file: $state_file"
+}
+
+cmd_approve() {
+    local session_id=$(get_session_id "${1:-}")
+    local state_file=$(get_state_file "$session_id")
+
+    if [[ ! -f "$state_file" ]]; then
+        echo "No review state for session: $session_id"
+        echo "Creating approved state..."
+    fi
+
+    cat > "$state_file" <<EOF
+{
+    "session_id": "$session_id",
+    "status": "approved",
+    "approved_at": "$(date -Iseconds)",
+    "approved_by": "${USER:-unknown}",
+    "reviewer_agent": "${CLAUDE_MODEL:-unknown}"
+}
+EOF
+
+    echo "✓ Review approved for session: $session_id"
+}
+
+cmd_reject() {
+    local session_id=$(get_session_id "${1:-}")
+    local reason="${2:-Review failed}"
+    local state_file=$(get_state_file "$session_id")
+
+    # Collect any additional args as issues
+    shift 2 2>/dev/null || true
+    local issues=("$@")
+
+    local issues_json="[]"
+    if [[ ${#issues[@]} -gt 0 ]]; then
+        issues_json=$(printf '%s\n' "${issues[@]}" | jq -R . | jq -s .)
+    fi
+
+    cat > "$state_file" <<EOF
+{
+    "session_id": "$session_id",
+    "status": "rejected",
+    "rejected_at": "$(date -Iseconds)",
+    "rejected_by": "${USER:-unknown}",
+    "reviewer_agent": "${CLAUDE_MODEL:-unknown}",
+    "reason": "$reason",
+    "issues": $issues_json
+}
+EOF
+
+    echo "❌ Review rejected for session: $session_id"
+    echo "Reason: $reason"
+}
+
+cmd_status() {
+    local session_id=$(get_session_id "${1:-}")
+    local state_file=$(get_state_file "$session_id")
+
+    if [[ ! -f "$state_file" ]]; then
+        echo "No review state for session: $session_id"
+        exit 0
+    fi
+
+    echo "Review state for session: $session_id"
+    echo "---"
+    jq . "$state_file"
+}
+
+cmd_list() {
+    echo "Active review sessions:"
+    echo "---"
+    shopt -s nullglob
+    for f in "$STATE_DIR"/*.json; do
+        local session=$(basename "$f" .json)
+        local status=$(jq -r '.status' "$f" 2>/dev/null || echo "unknown")
+        printf "  %-40s %s\n" "$session" "$status"
+    done
+    shopt -u nullglob
+}
+
+cmd_clean() {
+    local max_age="${1:-7d}"
+    echo "Cleaning review states older than $max_age..."
+    find "$STATE_DIR" -name "*.json" -mtime +${max_age%d} -delete 2>/dev/null || true
+    echo "Done."
+}
+
+# Main
+case "${1:-help}" in
+    check)
+        cmd_check "${2:-}"
+        ;;
+    enable)
+        cmd_enable "${2:-}"
+        ;;
+    approve)
+        cmd_approve "${2:-}"
+        ;;
+    reject)
+        cmd_reject "${2:-}" "${3:-}" "${@:4}"
+        ;;
+    status)
+        cmd_status "${2:-}"
+        ;;
+    list)
+        cmd_list
+        ;;
+    clean)
+        cmd_clean "${2:-7d}"
+        ;;
+    help|--help|-h)
+        echo "review-gate - Quality gate CLI for cross-agent review"
+        echo ""
+        echo "Commands:"
+        echo "  check [session]     Check if review approved (exit 0) or blocked (exit 1)"
+        echo "  enable [session]    Enable review requirement for session"
+        echo "  approve [session]   Mark review as approved"
+        echo "  reject [session] [reason] [issues...]  Mark as rejected"
+        echo "  status [session]    Show review status"
+        echo "  list                List all review sessions"
+        echo "  clean [age]         Clean old review states (default: 7d)"
+        echo ""
+        echo "Session ID is auto-detected from REVIEW_SESSION_ID, CLAUDE_SESSION_ID,"
+        echo "or generated from git branch/commit."
+        ;;
+    *)
+        echo "Unknown command: $1"
+        echo "Run 'review-gate help' for usage."
+        exit 1
+        ;;
+esac

skills/review-gate/skills/review-gate.md

@@ -0,0 +1,109 @@
+# review-gate
+
+Quality gate for cross-agent review enforcement. Blocks agent completion until work is reviewed and approved.
+
+## Purpose
+
+When agents work autonomously (unattended, CI/CD, batch jobs), they may:
+- Claim "done" when work is incomplete
+- Miss requirements or make incorrect assumptions
+- Take shortcuts that don't serve user intent
+
+review-gate enforces an external quality check before the agent can exit.
+
+## Usage
+
+### Enable review requirement
+
+At the start of work that needs review:
+
+```bash
+review-gate enable [session_id]
+```
+
+This creates a "pending" review state. The agent will be blocked from exiting until approved.
+
+### Approve review
+
+After reviewing work (manually or via reviewer agent):
+
+```bash
+review-gate approve [session_id]
+```
+
+### Reject review
+
+If issues are found:
+
+```bash
+review-gate reject [session_id] "reason" ["issue1" "issue2" ...]
+```
+
+### Check status
+
+```bash
+review-gate status [session_id]
+review-gate list  # all sessions
+```
+
+## Session ID
+
+Session ID is auto-detected in this order:
+1. Provided argument
+2. `REVIEW_SESSION_ID` env var
+3. `CLAUDE_SESSION_ID` env var
+4. Generated from git branch + commit
+
+## Hook Integration
+
+For Claude Code, add to your project's `.claude/hooks.json`:
+
+```json
+{
+  "hooks": {
+    "Stop": [{
+      "hooks": [{
+        "type": "command",
+        "command": "review-gate check",
+        "timeout": 30
+      }]
+    }]
+  }
+}
+```
+
+The Stop hook runs when the agent tries to exit. If review is pending or rejected, exit is blocked.
+
+## Commands
+
+| Command | Description | Exit Code |
+|---------|-------------|-----------|
+| `check [id]` | Check if exit allowed | 0=allowed, 1=blocked |
+| `enable [id]` | Enable review requirement | 0 |
+| `approve [id]` | Mark as approved | 0 |
+| `reject [id] [reason] [issues...]` | Mark as rejected | 0 |
+| `status [id]` | Show review state | 0 |
+| `list` | List all sessions | 0 |
+| `clean [age]` | Remove old states (default: 7d) | 0 |
+
+## Cross-Agent Pattern
+
+For agents without Stop hooks, use orchestrator pattern:
+
+```bash
+# Orchestrator script
+review-gate enable "$SESSION_ID"
+<spawn worker agent>
+# Worker completes...
+<spawn reviewer agent>
+# Reviewer approves or rejects...
+review-gate check "$SESSION_ID" || {
+    echo "Review not passed, looping back to worker"
+    # ... retry logic
+}
+```
+
+## Dependencies
+
+- `jq` - JSON parsing
+- `bash` 4.0+

skills/review-gate/templates/reviewer-prompt.md

@@ -0,0 +1,78 @@
+# Adversarial Reviewer
+
+You are reviewing work done by another agent. Your job is to find problems.
+
+## Your Mandate
+
+You work for the USER, not the worker agent. The worker may have:
+- Misunderstood the request
+- Taken shortcuts
+- Made incorrect assumptions
+- Produced incomplete work
+
+Your job is to catch these issues before the work is accepted.
+
+## Ground Truth
+
+The USER originally requested:
+
+```
+{user_prompt}
+```
+
+Read these words carefully. The worker's interpretation may differ from what the user actually asked for.
+
+## Your Methodology
+
+1. **Read the user's EXACT words** - not the agent's summary or interpretation
+2. **Examine ALL changes** - `git diff`, file contents, test results
+3. **ASSUME errors exist** - actively look for problems
+4. **Steel-man first** - give benefit of doubt, understand what worker tried to do
+5. **Then attack systematically** - check requirements, edge cases, correctness
+
+## Review Checklist
+
+- [ ] Does the work address what the user ACTUALLY asked for?
+- [ ] Are there any missing requirements?
+- [ ] Does the code compile/run without errors?
+- [ ] Are there obvious bugs or logic errors?
+- [ ] Are there security concerns?
+- [ ] Is anything incomplete or TODO-marked?
+
+## Your Tools (READ-ONLY)
+
+You have access to:
+- `Read`, `Grep`, `Glob` - examine files
+- `Bash` - run `git diff`, `git log`, test commands (read-only operations)
+- `orch` - get second opinions from other models (for non-trivial concerns)
+
+You should NOT:
+- Modify any files
+- Make commits
+- "Fix" issues yourself
+
+## Your Decision
+
+After thorough review, run one of:
+
+**If APPROVED** (work meets user's actual request):
+```bash
+review-gate approve {session_id}
+```
+
+**If REJECTED** (problems found):
+```bash
+review-gate reject {session_id} "Brief summary of issues" "issue 1" "issue 2" ...
+```
+
+Then report your findings to the user.
+
+## Second Opinions
+
+For non-trivial concerns, use orch to get consensus:
+
+```bash
+orch consensus "Is this implementation correct? [describe concern]" gemini deepseek claude
+```
+
+If you're uncertain, err on the side of rejection. Better to have the worker fix potential issues than to let bugs through.