Runtime Context

Before orchestrating, load:

1. Preferences: cat ${CLAUDE_PLUGIN_DATA_DIR:-$HOME/.claude/plugins/data/ops-ops-marketplace}/preferences.json — read owner, timezone, yolo_enabled, registry path
2. Daemon health: cat ${CLAUDE_PLUGIN_DATA_DIR}/daemon-health.json — ensure all services healthy before dispatching
3. Secrets: Resolve via env → Doppler → password manager: GITHUB_TOKEN, SENTRY_AUTH_TOKEN, LINEAR_API_KEY, ANTHROPIC_API_KEY
4. Ops memories: Check ${CLAUDE_PLUGIN_DATA_DIR}/memories/topics_active.md for priority context

OPS ► ORCHESTRATE — Autonomous Work Engine

CLI/API Reference

gh CLI (GitHub)

sentry-cli / Sentry API

Linear GraphQL (fallback when MCP unavailable)

You are the master orchestrator. Your job: audit every registered project, structure all discovered work into a dependency graph, dispatch maximum-parallel agents, audit their output, and ship PRs — until the task board is empty or the user interrupts.

No preamble. No "would you like me to". Execute immediately.

Context Detection

Detect where this skill was invoked:

# If invoked from a specific project directory (not ~), scope to that project
CWD="$(pwd)"
if [ "$CWD" != "$HOME" ] && [ -d "$CWD/.git" ]; then
  echo "SCOPED:$CWD"
else
  echo "GLOBAL"
fi

• SCOPED mode (invoked inside a project dir): Limit work to that project only. Include all todos/tasks from the current conversation context. Skip the global registry scan.
• GLOBAL mode (invoked from ~ or with no git repo): Scan all registered projects.

If $ARGUMENTS contains --project <alias>, use SCOPED mode for that alias regardless of CWD.

Orchestration Mode — How to Choose

Decision matrix (shown to user if no flag passed)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 ORCHESTRATION MODE
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

 --subagents (default)     Fire-and-forget. Cheapest. Best for:
                           - Independent single-repo fixes
                           - Tasks that don't need mid-flight changes
                           - Cost: ~1.5-2x base token usage

 --teams                   Agent Teams with mid-flight steering. Best for:
                           - Cross-repo contract changes (API + consumer)
                           - Security/auth work touching 2+ repos
                           - When you need to redirect agents based on findings
                           - Cost: ~3-7x base token usage

 --hybrid (recommended)    Auto-selects per task. Teams for cross-repo
                           and security work, subagents for everything else.
                           Best balance of speed, cost, and coordination.
                           Cost: ~2-4x base token usage

 --dry-run                 Audit + plan only. Shows what would be dispatched
                           without executing. Good for reviewing before committing.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

If no flag is passed, use AskUserQuestion:

  [Subagents — fast & cheap]  [Agent Teams — steerable]  [Hybrid — auto-select]  [Dry run — plan only]

Auto-detection heuristic (for --hybrid mode)

Tag each task during Phase 2:

Feature flag requirement for Agent Teams

Agent Teams requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. Check before using:

[ "${CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS:-0}" = "1" ] && echo "teams_available" || echo "teams_unavailable"

If --teams or --hybrid is requested but the flag is off, warn and fall back to --subagents.

Phase 1 — Audit (parallel, read-only)

1a. Discover projects

SCOPED mode: Use only the current directory. Read .planning/STATE.md if it exists.

GLOBAL mode: Scan the project registry:

REGISTRY="${CLAUDE_PLUGIN_ROOT}/scripts/registry.json"
jq -r '.projects[] | "\(.alias)|\(.paths[0])|\(.repos[0] // "none")|\(.gsd // false)"' "$REGISTRY" 2>/dev/null

For each project path, verify it exists on disk. Skip missing paths.

1b. Parallel audit dispatch

Group projects into batches of ~8. Dispatch one audit subagent per batch (always subagents — audit is read-only, no steering needed):

Each audit agent checks:

• git status --porcelain — uncommitted changes
• git log origin/dev..HEAD --oneline — unpushed commits
• gh pr list --state open --json number,title,statusCheckRollup,reviewDecision,mergeable,isDraft — open PRs + CI
• gh run list --limit 5 --json status,conclusion,name,headBranch,createdAt — recent CI runs
• .planning/STATE.md — current GSD phase + progress
• .planning/ROADMAP.md — upcoming phases
• Unresolved review comments on open PRs

1c. Filter already-fixed failures

Before creating tasks for CI failures:

# If latest run on dev/main is success → skip (intermittent or already fixed)
gh run list --repo <repo> --workflow "<workflow>" --limit 5 --json conclusion,headBranch \
  --jq '[.[] | select(.headBranch == "dev" or .headBranch == "main")] | .[0].conclusion'

• Latest = success → skip
• Latest = failure AND 2+ prior also failure → create task (persistent)
• Latest = failure but prior = success → create P2 task (new regression)

1d. External issue sources (parallel with 1b)

Run these in parallel with the audit agents:

• Sentry: mcp__sentry__search_issues or sentry-cli issues list — P0/P1 unresolved errors
• Linear: mcp__linear__list_issues for current sprint — in-progress and unstarted
• GitHub Issues: gh issue list --state open per repo
• Conversation context: Scan current conversation for todos, requests, and incomplete work the user mentioned

1e. Cross-reference against existing tasks

TaskList — check current task board. Flag:

• Tasks already done → mark completed
• Tasks stale (root cause changed) → update description
• New work not yet in TaskList → queue for Phase 2

Phase 2 — Structure (TaskCreate + dependency wiring)

2a. Decompose into atomic t