Dev Orchestrator

Dev Orchestrator Skill

Workflow Graph

flowchart TD
    A[User Request] --> B{Classify Task}
    B -->|Q&A| C[analyzer agent]
    B -->|Operations| D[builder agent]
    B -->|Development| E{Recursion Guard}
    B -->|Investigation| E

    E -->|BLOCKED| F[derive-recursion-guard emits BLOCKED:*]
    F --> G["execute-single-fallback-blocked-{development,investigation}"]
    G --> R1_BLOCKED[Round 1 result]

    E -->|ALLOWED| H{Decompose Workstreams}
    H -->|"1 workstream<br/>(default-workflow or investigation-workflow)"| I[Execute recipe]
    H -->|N workstreams| J[multitask parallel]
    I --> R1[Round 1 result]
    J --> R1

    R1_BLOCKED --> REFLECT
    R1 --> REFLECT{Reflect on Round 1}
    REFLECT -->|ACHIEVED| SUM[Summarize]
    REFLECT -->|PARTIAL / NOT_ACHIEVED| K[Execute Round 2]
    K --> REFLECT2{Reflect on Round 2}
    REFLECT2 -->|ACHIEVED| SUM
    REFLECT2 -->|PARTIAL / NOT_ACHIEVED| L[Execute Round 3 - final]
    L --> REFLECT_FINAL[Final Reflect]
    REFLECT_FINAL --> VAL{Outside-In Testing Validation}
    REFLECT --> VAL

    VAL --> SUM
    SUM --> DONE[Complete Session + Cleanup]

    %% Adaptive error recovery path
    R1 -.->|empty result| DET[detect-execution-gap]
    DET --> FILE_BUG[file-routing-bug]
    FILE_BUG --> ADAPT{Adaptive Recipe}
    ADAPT -->|Investigation| INV_RECIPE[investigation-workflow]
    ADAPT -->|Development| DEV_RECIPE[default-workflow]
    INV_RECIPE --> REFLECT
    DEV_RECIPE --> REFLECT

    C --> DONE
    D --> DONE

Purpose

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

Development Routing Invariant

Development classification always routes to default-workflow; model-provided recipe fields do not override that invariant.

Hybrid decompositions route each workstream by its own normalized classification.

This invariant is enforced by the Rust orchestration helper before workstream execution. Prompt instructions guide the model, but model-emitted recipe values are not trusted for Development workstreams.

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 → [derive-recursion-guard prints BLOCKED:* to stderr]
                                    ↓
                  [execute-single-fallback-blocked-development]
                  [execute-single-fallback-blocked-investigation]
                                    ↓
                           [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 the recipe runner. 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).

Default: Direct Execution (Rust recipe-runner-rs via CLI)

Use the Rust recipe-runner-rs through the amplihack recipe run CLI:

cd /path/to/repo && env -u CLAUDECODE \
  AMPLIHACK_HOME=/path/to/amplihack \
  amplihack recipe run amplifier-bundle/recipes/smart-orchestrator.yaml \
    -c task_description="TASK_DESCRIPTION_HERE" \
    -c repo_path="." \
    --verbose

Key points:

• amplihack recipe run — invokes Rust recipe-runner-rs; no Python runner or wrapper needed
• -c key=value — passes context variables (equivalent to user_context dict)
• --verbose — streams recipe-runner stderr live so you see nested step activity
• The recipe runner manages its own child processes (agent sessions, bash steps) as direct subprocesses

This is the preferred execution mode for most scenarios. It is simpler, has no external dependencies beyond the Rust binary, works on all platforms, and makes output capture straightforward.

For tmux durable execution, environment variables, error recovery strategies, fallback recipes, and enforcement details, see reference.md.

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):

amplihack recipe run amplifier-bundle/recipes/smart-orchestrator.yaml \
  -c task_description="TASK_DESCRIPTION_HERE" \
  -c repo_path="." \
  -c force_single_workstream="true"

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 invoke /dev in your agent session

Note: The env var must be set in your shell before starting the agent session — 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: native amplihack orch / amplihack multitask commands
• 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. derive-recursion-guard — classifies the guard result and prints the reason to stderr (BLOCKED:registration_failed, BLOCKED:unknown). There is no separate announcement step; the classification is the announcement.
2. execute-single-fallback-blocked-development or execute-single-fallback-blocked-investigation — executes the full task as a single builder agent session. The step chosen depends on the workstream's classification, which is why there are two.

Both live in amplifier-bundle/recipes/smart-execute-routing.yaml.

Note the scope of this path: it handles a run that is already executing and is refused when it tries to descend further. It does not apply when amplihack recipe run is itself invoked at the ceiling — there the guard refuses before any recipe starts, and the correct response is to complete the work inline rather than to launch a recipe (the refusal message says so).

A leftover AMPLIHACK_SESSION_DEPTH in a surviving shell does not wedge you: an inherited depth that nothing corroborates — no tree id, no live orchestrator ancestor — is discarded and the run starts a fresh tree at depth 0 (issue #1326).