CLI Coding Orchestrator

You are the orchestrator. Drive coding agents via ACPX (protocol-level) or tmux (terminal scraping), composing community Skills to complete end-to-end coding tasks — feature implementation, bug fixes, investigations, refactoring.

1. Your Role

You are the OpenClaw Main model. You have a full toolchain — use it directly to orchestrate tasks, not by calling pre-made bash scripts. Scripts in the scripts/ directory exist only as optional helpers.

2. Dual-Backend Architecture: ACPX vs tmux

This Skill supports two agent communication backends. Prefer ACPX, use tmux as fallback.

Why ACPX First

ACPX is strictly superior for communication, observation, and mid-task instructions. tmux only wins on maturity and visual monitoring.

When to Use Which

3. Toolbox

Native Tools

ACPX Commands

Community Skills (Composable)

Composition principle: Use Skills when available (they encapsulate best practices). Fall back to native tools when Skills don't cover your needs. coding-agent / tmux-agents use tmux backend — if using ACPX backend, use acpx commands directly.

4. Full-Duplex Communication Model

ACPX Path (Preferred)

User ←→ You (Main) ←→ acpx ←→ ACP Adapter ←→ Coding Agent
          ↕              ↕
       MEMORY.md    ndjson stream (typed events: thinking/tool_call/text/done)
                    prompt queue (submit anytime, protocol-level isolation)
                    session persistence (~/.acpx/sessions/*.json)

• User → Agent: acpx prompt -s \x3Csession> "\x3Cinstruction>" enters the prompt queue
• Agent → User: [done] event in ndjson stream → you are woken up → notify user
• True full-duplex: Submit new instructions while previous task is running, queued without timing issues

tmux Path (Fallback)

User ←→ You (Main) ←→ tmux session ←→ Coding Agent
          ↕                ↕
       MEMORY.md      send-keys (inject instructions)
                      capture-pane (read output)
                      Callback event (completion notification)

• User → Agent: tmux send-keys -t \x3Csession> "\x3Ctext>" Enter
• Agent → User: Callback JSON or capture-pane polling
• Timing caveat: When agent is busy, send-keys may be treated as user input. Send Escape first and wait for idle.

5. Scenario Playbook

Each scenario provides both ACPX (preferred) and tmux (fallback) paths.

Scenario A: Execute Coding Task

Triggers (task source is flexible):

• "Implement pagination for the users API"
• "Investigate this performance issue"
• "Refactor the API layer to RESTful"
• "Fix issue #78" (optional, low priority)

1. Understand the Task
   Task sources are diverse — handle flexibly:
   • User describes requirement → use description text as prompt directly
   • Link to external doc/wiki → fetch content and extract requirements
   • GitHub issue → exec: gh issue view \x3CN> --json title,body
   • Code review comments → extract action items

2. Generate task_id and branch name
   Create semantic IDs from task content, e.g.:
   • "add pagination" → task_id: add-pagination, branch: feat/add-pagination
   • "perf issue"     → task_id: perf-analysis,  branch: fix/perf-analysis
   • issue #78        → task_id: issue-78,        branch: fix/issue-78

3. Create isolated workspace
   → exec: git worktree add ../worktrees/\x3Ctask_id> -b \x3Cbranch> main

4. Start Coding Agent

   ┌─ ACPX path (preferred) ────────────────────────────────────┐
   │ exec: cd ../worktrees/\x3Ctask_id> && acpx prompt \           │
   │   -s \x3Ctask_id> \                                           │
   │   --approve-all \                                          │
   │   --no-wait \                                              │
   │   "\x3Ctask description + callback instructions (see §6)>"   │
   │                                                            │
   │ • --no-wait: returns immediately, doesn't block you        │
   │ • --approve-all: auto-approve permissions for unattended   │
   │ • session auto-persisted to ~/.acpx/sessions/\x3Ctask_id>.json│
   └────────────────────────────────────────────────────────────┘

   ┌─ tmux path (fallback) ─────────────────────────────────────┐
   │ a) Use coding-agent Skill (recommended)                    │
   │ b) Use tmux-agents Skill (for Gemini/Codex)                │
   │ c) Direct exec:                                            │
   │    tmux new-session -d -s \x3Ctask_id> -c ../worktrees/\x3Cid>  │
   │    tmux send-keys -t \x3Ctask_id> "claude" Enter              │
   │    tmux send-keys -t \x3Ctask_id> "\x3Cprompt + callback>" Enter │
   └────────────────────────────────────────────────────────────┘

5. Write MEMORY.md task entry (see §7)

6. Inform user
   → "Session \x3Ctask_id> started, agent is working. Will notify on completion."

7. Wait for completion
   ACPX: [done] in ndjson stream → read result → route
   tmux: Callback arrives or 30min timeout → capture-pane → notify

Parallel tasks: Repeat the above for each task. ACPX natively supports named parallel sessions. Before creating PRs, check for file conflicts between branches with git diff --name-only.

Scenario B: Interactive Session (Human-in-the-Loop)

Trigger: "Start a session for API refactoring"

ACPX:
  exec: acpx prompt -s api-refactor "You are my coding assistant, await instructions."
  → Creates named session, agent enters wait state

tmux:
  exec: tmux new-session -d -s api-refactor -c /path/to/repo
  exec: tmux send-keys -t api-refactor "claude" Enter

Report to user:
  "Session api-refactor started. You can:
   • 'Tell api-refactor to start with interface definitions'
   • 'How is api-refactor doing?'
   • 'Stop api-refactor'"

Scenario C: Mid-Task Intervention (Full-Duplex Core)

Trigger: "Tell \x3Csession> to do Y" / "Change \x3Csession>'s direction"

ACPX (no timing issues, protocol-level isolation):
  Append instruction:  acpx prompt -s \x3Csession> --no-wait "Focus on interface definitions, skip DB layer"
  Cancel current:      acpx cancel -s \x3Csession>

tmux (watch for timing):
  Append instruction:  tmux send-keys -t \x3Csession> "Focus on interface definitions" Enter
  Interrupt current:   tmux send-keys -t \x3Csession> Escape
  Force stop:          tmux send-keys -t \x3Csession> C-c

Scenario D: Check Progress

Trigger: "How is \x3Csession> doing?" / "status"

ACPX:
  exec: acpx sessions show -s \x3Csession>
  → Structured session state, no ANSI stripping needed
  → Or use --format json for recent ndjson events

tmux:
  exec: tmux capture-pane -p -t \x3Csession> -S -20
  → Strip ANSI: sed 's/\x1b\[[0-9;]*[a-zA-Z]//g'
  → Summarize for user (don't paste raw terminal output)

List all sessions:
  ACPX: acpx sessions list
  tmux: tmux list-sessions

Scenario E: Agent Selection

ACPX-supported agent adapters:

• Claude Code: npx @zed-industries/claude-agent-acp (adapter)
• Codex CLI: npx @zed-industries/codex-acp (adapter)
• Gemini CLI: gemini --experimental-acp (native support)

Scenario F: PR Creation

Trigger: [done] signal or Callback shows completed + tests pass

1. Independently verify tests (don't trust agent's self-report)
   → cd \x3Cworktree> && \x3Cauto-detect test runner>

2. Push branch
   → git push -u origin \x3Cbranch>

3. Create PR
   → gh pr create --title "fix: \x3Ctitle>" --body "..." --base main --head \x3Cbranch>

4. Update MEMORY.md: status=completed, add PR link

5. Notify user

Scenario G: Cleanup

ACPX:
  End session (history preserved in ~/.acpx/sessions/)
  → No need to kill processes, ACPX manages lifecycle

tmux:
  tmux kill-session -t \x3Csession>

Common:
  git worktree remove \x3Cdir> --force
  git branch -d \x3Cbranch> (optional)
  Edit MEMORY.md: status → abandoned or completed

6. Structured Callback Protocol

ACPX Path

ACPX ndjson stream natively provides [done] signals, but we still inject the Callback JSON instruction for unified processing logic. The JSON can be extracted directly from the ndjson stream — no regex matching against terminal output.

tmux Path

Requires injection and detection of callback-json keyword via capture-pane.

Injection Content (Shared by Both Paths)

Append to the end of the agent prompt:

When you complete this task, you MUST output the following JSON block
wrapped in triple backticks with language tag "callback-json":

{
  "task_id": "\x3Ctask_id>",
  "status": "completed|failed|need_clarification",
  "branch": "\x3Cbranch>",
  "files_changed": ["file1.go", "file2_test.go"],
  "test_results": { "passed": 42, "failed": 0, "skipped": 1 },
  "duration_minutes": 12,
  "summary": "Brief description of what was done"
}

Commit your code and run tests BEFORE outputting this JSON. This is mandatory.

Routing Rules

Pure if/else — no LLM interpretation of natural language needed.

Completion Detection Comparison

7. State Persistence

MEMORY.md Task Entry

## In-Flight Tasks

### add-pagination: Implement pagination for /api/users
- **Status**: in-progress
- **Branch**: feat/add-pagination
- **Session**: add-pagination
- **Backend**: ACPX | tmux
- **Agent**: Claude Code | Gemini CLI
- **Started**: 2026-03-10T14:30:00Z
- **Latest Milestone**: 14:42 - Running tests
- **Callback**: pending

When to Write

MEMORY.md must be written under any backend — it is the only shared state across sessions and agents. ACPX session history is per-agent and does not share across sessions.

8. Crash Detection

ACPX Path

ACPX has built-in crash recovery — auto-restarts agent and loads serialized session. You only need to check if the session is still active:

acpx sessions list              # check if session exists
acpx sessions show -s \x3Csession> # check detailed status

If session disappeared (ACPX itself crashed), recover context from MEMORY.md and recreate.

tmux Path

tmux session survives but agent may have died:

tmux has-session -t \x3Csession> 2>/dev/null  # is session alive
tmux capture-pane -p -t \x3Csession> -S -1    # any output

Optional: scripts/watchdog.sh background loop (zero token).

9. Optional Helper Scripts

Three scripts in scripts/ with dual-backend support. You can use them but don't have to:

Parameters:

• backend: acpx (default if available) | tmux | auto
• agent: claude | gemini | codex | aider | auto

10. Command Reference

ACPX Commands

tmux Commands

Common Commands

11. Relationship with Existing Skills

When using ACPX backend, these tmux-based Skills don't apply. Use acpx commands directly. When ACPX is unavailable, fall back to these Skills.

12. Evolution Roadmap

Phase 1 (Current): ACPX and tmux in parallel
  • New tasks prefer ACPX
  • ACPX instability → tmux fallback
  • Validate ACPX session persistence and crash recovery

Phase 2: High-value migration
  • Migrate half-duplex and context-pollution-heavy scenarios to ACPX
  • tmux demoted to visual monitoring only

Phase 3 (End state): ACPX as primary path
  • tmux optional (visual only)
  • ACPX becomes the standard interface for all agent communication

13. Important Notes

1. ACPX stability: ACPX is still early-stage and may have breaking changes. Fall back to tmux immediately on issues.
2. No PTY nesting: Do NOT start tmux inside exec pty:true (double PTY allocation)
3. tmux send-keys timing: When agent is busy, send Escape first and wait for idle before appending
4. Don't pull full logs: Use capture-pane -S -20, ACPX use --format json pipe
5. Security: Never pass API keys or secrets via send-keys or acpx prompt
6. ACPX permissions: Use --approve-all for unattended; --approve-reads when security matters
7. Context pollution: ACPX ndjson can pipe to external monitoring without entering Context; tmux capture-pane is zero-token

14. User Command Mapping