dev-orchestrator（开发协调器）

Dev Orchestrator Skill

Purpose

This is the default orchestrator for all non-trivial development and investigation tasks in amplihack. It replaces the ultrathink-orchestrator skill.

When a user asks you to build, implement, fix, investigate, or create anything non-trivial, this skill ensures:

1. Task is classified — Q&A / Operations / Investigation / Development
2. Goal is formulated — clear success criteria identified
3. Workstreams detected — parallel tasks split automatically
4. Recipe runner used — code-enforced workflow execution
5. Outcome verified — reflection confirms goal achievement

How It Works

User request
     │
     ▼
[Classify] ──→ Q&A ──────────────────→ analyzer agent (technical/code questions)
     │
     ├──────→ Ops ────────────────────→ builder agent
     │
     └──→ Development / Investigation
             │
         [Recursion guard] (AMPLIHACK_SESSION_DEPTH vs AMPLIHACK_MAX_DEPTH=3)
             │         │
           ALLOWED   BLOCKED → [announce-depth-limited banner]
                                    ↓
                           [execute-single-fallback-blocked]
                                    ↓
                           [Execute round 1 (single-session)]
             │
         [Decompose]
             │         │
             1 ws     N ws ──→ [multitask parallel] + tree context in env
             │
         [Execute round 1]
             │
         [Reflect] ──→ ACHIEVED ──→ [Summarize]
             │
           PARTIAL/NOT_ACHIEVED
             │
         [Execute round 2]
             │
         [Reflect] ──→ ACHIEVED ──→ [Summarize]
             │
           PARTIAL/NOT_ACHIEVED
             │
         [Execute round 3 (final)]
             │
         [Final reflect + Summarize]

Session tree enforcement (prevents infinite recursion):

• Each subprocess inherits AMPLIHACK_TREE_ID, AMPLIHACK_SESSION_DEPTH, AMPLIHACK_MAX_DEPTH
• Depth >= 3: recursion guard blocks sub-workstream spawning, adapts to single-session
• Max 10 concurrent sessions per tree; extras queue in FIFO order
• State tracked in /tmp/amplihack-session-trees/{tree_id}.json

Activation

This skill auto-activates for development and investigation keywords. It is also the default behavior per CLAUDE.md — invoke it for any non-trivial task.

The UserPromptSubmit hook reinforces this by injecting a classification prompt that instructs Claude to invoke Skill(skill="dev-orchestrator") for dev/investigation tasks.

Execution Instructions

YOUR NEXT ACTION after reading this skill MUST include a Bash tool call that executes run_recipe_by_name("smart-orchestrator"). If your runtime requires a report_intent call (e.g. Copilot CLI), emit it in parallel with the Bash launch — both tool calls in the same response. The constraint is that no *other* substantive action (reading files, calling TodoWrite, invoking Agent, or typing a response) may precede or replace the Bash launch.

If you find yourself doing anything else instead of launching the recipe runner, you are bypassing the workflow. Stop. Go back. Run the recipe runner.

Development and Investigation tasks have exactly ONE execution path: the recipe runner. There is no "Option B." There is no "I'll follow the steps manually." There is no "Let me just quickly..." The recipe runner IS the workflow. Skipping it means skipping the 23-step process, the recursion guards, the goal verification, and the quality gates that exist for a reason.

A PostToolUse hook monitors every tool call after this skill activates. If it detects 3 tool calls without recipe runner evidence, it will fire a hard WARNING. Do not wait for the warning — run the recipe runner immediately.

When this skill is activated:

REQUIRED: Execute via Recipe Runner — IMMEDIATELY

Your next tool call(s) must include the recipe runner launch (alongside report_intent if your runtime requires it). CRITICAL: Claude Code's run_in_background kills processes after ~10 minutes. Recipe workstreams can take hours. You MUST use a tmux session for execution:

LOG_FILE=$(mktemp /tmp/recipe-runner-output.XXXXXX.log)
chmod 600 "$LOG_FILE"
tmux new-session -d -s recipe-runner "env -u CLAUDECODE PYTHONPATH=src python3 -c \"
import os
os.environ.pop('CLAUDECODE', None)

from amplihack.recipes import run_recipe_by_name

result = run_recipe_by_name(
    'smart-orchestrator',
    user_context={
        'task_description': '''TASK_DESCRIPTION_HERE''',
        'repo_path': '.',
    },
    progress=True,
)
print(f'Recipe result: {result}')
\" 2>&1 | tee \"$LOG_FILE\""
echo \"Recipe runner log: $LOG_FILE\"

Key points:

• env -u CLAUDECODE — unset so nested Claude Code sessions can launch
• PYTHONPATH=src python3 — uses the interpreter on PATH while forcing imports from the checked-out repo source tree (do NOT hardcode .venv/bin/python)
• run_recipe_by_name — delegates to the Rust binary; the adapter parameter is no longer needed
• progress=True — streams recipe-runner stderr live so tmux logs show nested step activity
• chmod 600 "$LOG_FILE" — keeps the tmux log private to the current user
• tmux new-session -d — detached session, no timeout, survives disconnects
• Monitor with: tail -f "$LOG_FILE" or tmux attach -t recipe-runner

Restarting a stale tmux session: Some runtimes (e.g. Copilot CLI) block tmux kill-session because it does not target a numeric PID. Use one of these shell-policy-safe alternatives instead:

# Option A (preferred): use a unique session name per run to avoid collisions
tmux new-session -d -s "recipe-$(date +%s)" "..."

# Option B: locate the tmux server PID and terminate with numeric kill
tmux list-sessions -F '#{pid}' 2>/dev/null | xargs -I{} kill {}

# Option C: let tmux itself handle it — send exit to all panes
tmux send-keys -t recipe-runner "exit" Enter 2>/dev/null; sleep 1

If using Option A, update the tail -f / tmux attach commands to use the same session name.

DO NOT use run_in_background for recipe execution — it will be killed after ~10 minutes (Issue #2909).

The recipe runner is the required execution path for Development and Investigation tasks. Always try smart-orchestrator first.

Common rationalizations that are NOT acceptable:

• "Let me first understand the codebase" — the recipe does that in Step 0
• "I'll follow the workflow steps manually" — NO, the recipe enforces them
• "The recipe runner might not work" — try it first, report errors if it fails
• "This is a simple task" — simple or complex, the recipe runner handles both

Q&A and Operations only may bypass the recipe runner:

• Q&A: Respond directly (analyzer agent)
• Operations: Builder agent (direct execution, no workflow steps)

Error Recovery: Adaptive Strategy (NOT Degradation)

When smart-orchestrator fails, failures must be visible and surfaced — never swallowed or silently degraded. The recipe handles error recovery automatically via its built-in adaptive strategy steps, but if you observe a failure outside the recipe, follow this protocol:

1. Surface the error with full context:

Report the exact error, the step that failed, and the log output. Never say "something went wrong" — always include the specific failure details.

2. File a bug with reproduction details:

For infrastructure failures (import errors, missing env vars, binary not found, decomposition producing invalid output), file a GitHub issue:

gh issue create \
  --title "smart-orchestrator infrastructure failure: <one-line summary>" \
  --body "<full error context, reproduction command, env details>" \
  --label "bug"

3. Evaluate alternative strategies:

If smart-orchestrator fails at the infrastructure level (not because the task is wrong), you MAY invoke the specific workflow recipe directly. This is an adaptive strategy — it must be announced explicitly, not done silently:

Example:

# ANNOUNCE the strategy change first — never do this silently
print("[ADAPTIVE] smart-orchestrator failed at parse-decomposition: <error>")
print("[ADAPTIVE] Switching to direct investigation-workflow invocation")
run_recipe_by_name("investigation-workflow", user_context={...}, progress=True)

This is NOT a license to bypass smart-orchestrator. Always try it first. Direct invocation is only permitted when smart-orchestrator fails at the infrastructure level. "The task seems simple" is NOT an infrastructure failure.

4. Detect hollow success:

A recipe can complete structurally (all steps exit 0) but produce empty or meaningless results — agents reporting "no codebase found" or reflection marking ACHIEVED when no work was done. After execution, check that:

• Round results contain actual findings or code changes (not "I could not access...")
• PR URLs or concrete outputs are present for Development tasks
• At least one success criterion was verifiably evaluated

If results are hollow, report this to the user with the specific empty outputs. Do not declare success when agents produced no meaningful work.

Required Environment Variables

The recipe runner requires these environment variables to function:

If AMPLIHACK_HOME is not set and auto-detection fails, parse-decomposition and activate-workflow will fail with "orch_helper.py not found". Set it to the directory containing amplifier-bundle/.

After Execution: Reflect and verify

After execution completes, verify the goal was achieved. If not:

• For missing information: ask the user
• For fixable gaps: re-invoke with the remaining work description
• For infrastructure failures: file a bug and try adaptive strategy

Enforcement: PostToolUse Workflow Guard

A PostToolUse hook (workflow_enforcement_hook.py) actively monitors every tool call after this skill is invoked. It tracks:

• Whether /dev or dev-orchestrator was called (sets a flag)
• Whether the recipe runner was actually executed (clears the flag)
• How many tool calls have passed without workflow evidence

If 3+ tool calls pass without evidence of recipe runner execution, the hook emits a hard WARNING. This is not a suggestion — it means you are violating the mandatory workflow. State is stored in /tmp/amplihack-workflow-state/.

Task Type Classification

• Hybrid is not a distinct task_type — the orchestrator classifies as Development and decomposes into multiple workstreams (one investigation, one development).

Workstream Decomposition Examples

Override Options

Single workstream override: Pass force_single_workstream: "true" in the recipe user_context to prevent automatic parallel decomposition regardless of task structure. This is a programmatic option (not directly settable from /dev):

run_recipe_by_name(
    "smart-orchestrator",
    user_context={
        "task_description": task,
        "repo_path": ".",
        "force_single_workstream": "true",  # disables parallel decomposition
    }
)

To force single-workstream execution without modifying recipe context: Set AMPLIHACK_MAX_DEPTH=0 before running /dev. This causes the recursion guard to block parallel spawning and fall back to single-session mode for all tasks:

export AMPLIHACK_MAX_DEPTH=0  # set in your shell first
/dev build a webui and an api  # then type in Claude Code

Note: The env var must be set in your shell before starting Claude Code — it cannot be prefixed inline on the /dev command. This affects all depth checks, not just parallel workstream spawning.

Canonical Sources

• Recipe: amplifier-bundle/recipes/smart-orchestrator.yaml
• Parallel execution: .claude/skills/multitask/orchestrator.py
• Development workflow: amplifier-bundle/recipes/default-workflow.yaml
• Investigation workflow: amplifier-bundle/recipes/investigation-workflow.yaml
• CLAUDE.md: Defines this as the default orchestrator

Relationship to Other Skills

Entry Points

• Primary: /dev <task description>
• Auto-activation: Via CLAUDE.md default behavior + hook injection
• Legacy: /ultrathink <task> (deprecated alias → redirects to /dev)

Status Signal Reference

The orchestrator uses two status signal formats:

Execution status (from builder agents)

Appears at the end of round execution steps:

• STATUS: COMPLETE — the round's work is fully done
• STATUS: CONTINUE — more work remains after this round
• STATUS: PARTIAL — the final round (round 3) reached partial completion
• STATUS: DEPTH_LIMITED — (legacy, no longer emitted; use BLOCKED path instead)

Goal status (from reviewer agents)

Appears at the end of reflection steps:

• GOAL_STATUS: ACHIEVED — all success criteria met, task is done
• GOAL_STATUS: PARTIAL -- [description] — some criteria met, more work needed
• GOAL_STATUS: NOT_ACHIEVED -- [reason] — goal not met, another round needed

The goal-seeking loop uses GOAL_STATUS signals to decide whether to run round 2 or 3.

BLOCKED path (recursion guard): When multi-workstream spawning is blocked by the depth limit, the orchestrator adapts to single-session execution:

1. announce-depth-limited — prints a warning banner with remediation info
2. execute-single-fallback-blocked — executes the full task as a single builder agent session (announced, not silent — the banner makes the strategy change visible)