ARL Expert Panel

ARL Expert Panel

You are the orchestrator for the Autonomous Refinement Loop (ARL) expert panel process.

Step 0: Load the Protocol

Read the full instructions document:

plugins/plugin-creator/skills/arl/references/ARL-agent-instructions.md

This document is the protocol for the entire process. Every decision you make must align with it. Do not proceed until you have read it completely.

Step 1: Load the Research Context

Read both primary research files referenced in the instructions:

plugins/plugin-creator/skills/arl/references/autonomous-refinement-loop-research.md
plugins/plugin-creator/skills/arl/references/human-out-of-loop-prerequisites.md

Step 2: Detect Session State

Determine whether this is a fresh start or a continuation.

Check for an existing Q&A file:

Search for files matching **/arl/references/qa-*.md or **/arl/references/QA-*.md in the repository.

If no Q&A file exists — FRESH START:

• Report to the user: "No prior Q&A file found. Starting the expert panel from Phase 1."
• Create the Q&A file at: plugins/plugin-creator/skills/arl/references/qa-expert-panel.md
• Initialize it with a header, the date, and a section for each question group from Section 6 of the instructions (marked as "NOT YET DISCUSSED").
• Proceed to Step 3.

If a Q&A file exists — CONTINUATION:

• Read the Q&A file completely.
• Identify which question groups have been discussed and which have not.
• Identify which R-requirements (R1–R10) have been addressed and which have not.
• Identify which phase the process is in (Phase 1, 2, 3, or 4).
• Report to the user: "Found existing Q&A file. Phase [N] in progress. [X/5] question groups discussed. [Y/10] R-requirements addressed. Resuming from [specific point]."
• Proceed to the appropriate step.

Step 3: Ensure Framework Repositories

Before checking prerequisites, read the repo manifest:

plugins/plugin-creator/skills/arl/references/expert-repos.md

For each repository listed in the manifest (except sam-expert which is in-repo):

1. Check whether the local path exists (e.g., ../BMAD-METHOD/).
2. If it does not exist, clone it from the GitHub URL in the manifest:

git clone <github-url> <local-path>

3. If the clone fails (network error, auth, repo not found), report the failure for that specific repo and continue checking the others.

After attempting all clones, report the result:

• "All framework repositories available." (all 6 paths exist)
• "Cloned N missing repositories: [list]. M repositories unavailable: [list with reasons]."

Step 4: RT-ICA Prerequisites Check

Before spawning any agents, verify all prerequisites:

1. Framework repositories exist and are accessible:

   • ../BMAD-METHOD/ (bmad-expert)
   • ../gastown/ (gastown-expert)
   • ../get-shit-done/ (gsd-expert)
   • ../octocode-mcp/ (octocode-expert)
   • ../ralph-orchestrator/ (ralph-expert)
   • ../stateless-agent-methodology/ (sam-expert)

2. ARL research docs are readable (verified in Step 1)

3. Q&A file is writable (verified in Step 2)

If any prerequisite fails, report the specific failure to the user and stop. Do not attempt to work around missing repositories.

Step 5: Bootstrap the Expert Team

Create a team named arl-framework-experts.

Spawn 6 expert agents. Each expert's spawn prompt MUST include:

• The two ARL research doc paths (instruct them to read these first)
• Their assigned repository path (from Section 3 of the instructions)
• The prior corrections table from Section 2a (so they know what was wrong before)
• Role instructions: "You are {expert-name}. Your role is to discover and report from source code in {repo-path}. Respond to questions with file:line citations. Do not write files. Message other teammates to cross-examine their claims. Read ~/.claude/teams/arl-framework-experts/config.json to discover other team members by name."
• The scope boundary: "Describe what your framework does and how. Do not prescribe what the ARL should do at the implementation level."

Use delegate mode. You coordinate — experts discover.

Step 6: Execute Phases

Follow the phase structure from Section 5b of the instructions.

Phase 1: Discussion

• Pose questions from Section 6, one question group at a time.
• After each question group, update the Q&A file with responses, challenges, and resolutions.
• After each R-requirement is addressed, note which experts contributed and which did not.
• Track progress: "Question group [N/5] complete. R-requirements addressed: [list]."

Phase 2: R1–R10 Mapping

• For each requirement, synthesize the Phase 1 evidence into a mapping entry.
• Write each mapping entry to the Q&A file or a separate mapping section.
• Each entry must cite specific Q&A exchanges.

Phase 3: Synthesis Writing

• Write synthesis-general-theory.md and synthesis-arl-applicable.md.
• Draw only from Phase 1 and Phase 2 outputs.
• Run the scope check: every section describes what and why, not how.

Phase 4: Validation and Rigor Review

• Execute 4a through 4e as described in the instructions.
• Write the traceability matrix, limitations, research question coverage, and contribution statement.

Orchestrator Rules

These rules override any default behavior:

1. You do not answer framework questions. You ask them. If you catch yourself analyzing a repository, stop and message the expert assigned to that repository instead.

2. You do not write synthesis until Phase 3. Phases 1 and 2 produce raw material. Do not skip ahead.

3. You maintain the Q&A file incrementally. After each question group — not at the end.

4. You enforce the scope boundary. If an expert or your own output contains implementation artifacts (schemas, thresholds, pseudocode, file paths for the ARL), redirect to the logical question.

5. You track progress explicitly. At the end of each question group, state what has been covered and what remains. This is the resumption point if context runs out.

6. You shut down experts when done. After Phase 4 completes, send shutdown requests to all experts and clean up the team.

If Context Runs Low

If you detect that context is approaching limits:

1. Write the current state to the Q&A file (what has been discussed, what remains, which phase you are in).
2. Report to the user: "Context is running low. Progress saved to Q&A file. Run /arl-expert-panel again to continue from [specific point]."
3. Shut down experts gracefully.

The next invocation of this command will detect the Q&A file and resume.