Orchestrator

Dependencies

This skill requires Python 3.8+ and standard library only. No external packages needed.

To install this skill's dependencies:

pip-compile ./requirements.in
pip install -r ./requirements.txt

See ./requirements.txt for the dependency lockfile (currently empty — standard library only).

Orchestrator: Loop Router & Lifecycle Manager

The Orchestrator assesses the incoming trigger, selects the right loop pattern, and manages the shared closure sequence (seal, persist, retrospective, self-improvement).

The Core Loop

Ecosystem Context

• Patterns: learning-loop | red-team-review | dual-loop | agent-swarm | triple-loop-learning
• Inner Loop Reference: cli-agent-executor.md — Persona configs for specialized CLI execution.

Routing Decision Tree

Use this to select the correct loop pattern:

1. Does the trigger mention unguided friction evaluation, tests, and self-optimization?
   └─ YES → Pattern 5: triple-loop-learning
   └─ NO → continue

2. Is this work I can do entirely myself (research, document, iterate)?
   └─ YES → Pattern 1: learning-loop
   └─ NO → continue

3. Does it need adversarial review before proceeding?
   └─ YES → Pattern 2: red-team-review
   └─ NO → continue

4. Can the work be split into parallel independent tasks?
   └─ YES → Pattern 4: agent-swarm
   └─ NO → Pattern 3: dual-loop (sequential inner/outer delegation)

Process Flow

1. Plan (Strategy): You define the work (Spec → Plan → Tasks). When planning scripts/pipelines, default to a "Modular Building Blocks" architecture (CLI wrappers + independent core modules).
2. Delegate (Handoff): You pack the context into a Task Packet and assist the user in handing off to the Inner Loop.
3. Execute (Tactics): The Inner Loop agent (which has no git access) writes code and runs tests.
4. Verify (Review): You verify the output against acceptance criteria.
5. Correct (Feedback): If verification fails, you generate a Correction Packet and loop back to step 3.
6. Retrospective (Learning): You assess the loop's success and document learnings.
7. Primary Agent Handoff (Closure): You signal the repository environment to seal the session, update databases, and commit to Git.

Roles

You (Outer Loop / Director)

• Responsibilities: Planning, Git Management, Verification, Correction, Retrospective.
• Context: Full repo access, strategic constraints (ADRs), long-term memory.
• Tools: agent-orchestrator, git, and optionally any upstream planning tool.

Inner Loop (Executor / Worker)

• Responsibilities: Coding, Testing, Debugging.
• Context: Scoped to the Task Packet ONLY. No distractions.
• Constraints: NO GIT COMMANDS. Do not touch .git.
• Tools: Editor, Terminal, Test Runner.

Commands

You orchestrate workflows by natively executing the agent_orchestrator.py script provided by this skill (located in scripts/).

1. Planning Status

Use the scan command to inspect the state of the spec and readiness for delegation.

python ./scripts/agent_orchestrator.py scan --spec-dir <PATH>

Tip: Always ensure you have a clear plan or spec before delegating tasks.

2. Delegation (Handoff)

When a task is ready for implementation, generate a Task Packet using the packet command.

python ./scripts/agent_orchestrator.py packet --wp <WP-ID> --spec-dir <PATH>

This generates a markdown file in the handoffs/ directory. You must then instruct the user/system to launch the Inner Loop with this file.

3. Verification & Correction

Check the Inner Loop's work against the packet using the verify command.

python ./scripts/agent_orchestrator.py verify --packet handoffs/task_packet_NNN.md --worktree <PATH>

If the work fails criteria, use the Severity-Stratified Output schema to generate a structured correction packet:

• 🔴 CRITICAL: The code fails to compile, tests fail, or the requested feature is entirely missing. (Action: Hard reject, return to Inner Loop with exact error logs).
• 🟡 MODERATE: The feature works, but violates project architecture, ADRs, or performance standards. (Action: Flag for revision, return to Inner Loop with the specific ADR reference).
• 🟢 MINOR: The feature works and follows architecture, but has minor naming or stylistic issues. (Action: Do not return to Inner Loop. The Orchestrator fixes it directly and proceeds).

Generate the correction packet to send back to the Inner Loop:

python ./scripts/agent_orchestrator.py correct --packet handoffs/task_packet_NNN.md --feedback "Specific failure reason"

4. Parallel Execution (Agent Swarm)

For bulk operations or partitioned tasks, use the swarm_run.py script from the agent-swarm skill.

python ./swarm_run.py --job <JOB_FILE> [--resume]

This is the designated route for all Pattern 4 triggers.

4. Dynamic Routing (Model Agnostic)

As the Orchestrator, you can route tasks to ANY capable CLI agent based on complexity:

flowchart LR
    Router{Task Router} -->|Complex| High["High-Reasoning CLI (Opus/Ultra)"]
    Router -->|Routine| Fast["Fast CLI (Haiku/Flash)"]
    Router -->|Audit| Spec["Specialist CLI (Security/QA)"]

5. Red Team / Peer Review

Use the bundle command to compile files for a human or 3rd-party agent review.

python ./scripts/agent_orchestrator.py bundle --files <file1> <file2> --output <OUTPUT_BUNDLE.md>

This creates a single markdown bundle ideal for "paste-to-chat" reviews.

6. Retrospective (Post-Loop Learning)

Generate a retrospective template to close the cognitive loop with structured learning, before signaling the environment to seal.

python ./scripts/agent_orchestrator.py retro

This creates a template in the retros/ directory.

Meta-Improvement: After reviewing what went right/wrong, the orchestrator should also assess whether any loop infrastructure needs updating:

• Skills — Were SKILL.md instructions unclear or incomplete? Fix them.
• Scripts — Did any automation break or cause friction? Patch it.
• Templates — Were task templates missing fields or overly rigid? Refine them.
• Diagrams — Does the architecture diagram still reflect reality? Update it.
• Personas — Did a persona produce poor results? Tune it.

This makes each loop iteration smoother than the last.

7. Handoff to Primary Agent (Ecosystem Sealing)

Once the loop is complete and learning has been extracted, the Orchestrator MUST pass control to the environment's global sovereign (e.g., the primary agent plugin).

The Orchestrator explicitly DOES NOT:

• Trigger context-bundler:context-bundling or update recursive language model (RLMs) summaries in caches.
• Trigger persist_soul or sync to HuggingFace.
• Execute Vector DB ingestion scripts.
• Execute Git commands (git commit, git push).

These are environment-specific actions owned entirely by the Primary Agent.

Chained Command Handoff

When the Orchestrator loop is complete, it must use Chained Command Invocation to offer the user the explicit next steps to seal the ecosystem. Output this block:

## Orchestration Complete. Offer Next Steps:

The Inner Loop has successfully executed and verified the task. Please trigger the closure sequence:
- **"/sanctuary-seal"**: To capture the learning snapshot and update RLMs.
- **"/sanctuary-persist"**: To backup the soul to HuggingFace.
- **"Run Retrospective"**: If you wish to review the friction log.

8. Sub-Agent Limitations

• Be aware that claude-cli-agent has a hard stop on passing massive context bundles (~5MB+) either natively via stdin or --file. If your payload exceeds context windows, you must write a semantic chunking script instead of blindly dumping a context-bundler package into a prompt!
• Automated sub-agent invocations will silently fail or throw an interactive block if you do not use --dangerously-skip-permissions or if the user is not authenticated natively using claude login.

Lifecycle State Tracking

The orchestrator must verify these gates at each phase:

No phase may be skipped. If a gate fails, the orchestrator must resolve it before proceeding.

Loop Controls (Ralph-Inspired)

Automation

Best Practices

1. One WP at a Time: Do not delegate multiple WPs simultaneously unless you are running a swarm.

2. Explicit Context: The Inner Loop only knows what is in the packet. If it needs a file, list it.

3. No Git in Inner Loop: This is a hard constraint to prevent state corruption.

4. Correction is Learning: Do not just "fix it yourself" if the Inner Loop fails. Generate a correction packet. This trains the system logic.

5. Never Abandon Closure: The orchestrator must shepherd Review → Accept → Retro → Merge. Stopping after delegation is a protocol violation.

6. Merge from Main Repo: Always merge from the project root, never from inside a worktree.

Research Basis

This skill implements the "Dual-Loop Agent Architecture" inspired by:

1. Self-Evolving Recommendation System (arXiv:2602.10226):
   • Defines the specialized roles of Planner (Outer) vs Executor (Inner).
2. FormalJudge (arXiv:2602.11136):
   • Provides the theoretical framework for "Scalable Oversight" via structured verification rather than just human inspection.