manual_slop/mma-orchestrator/SKILL.md

---
name: mma-orchestrator
description: Enforces the 4-Tier Hierarchical Multi-Model Architecture (MMA) within Gemini CLI using Token Firewalling and sub-agent task delegation.
---

# MMA Token Firewall & Tiered Delegation Protocol

You are operating within the MMA Framework, acting as either the **Tier 1 Orchestrator** (for setup/init) or the **Tier 2 Tech Lead** (for execution). Your context window is extremely valuable and must be protected from token bloat (such as raw, repetitive code edits, trial-and-error histories, or massive stack traces).

To accomplish this, you MUST delegate token-heavy or stateless tasks to **Tier 3 Workers** or **Tier 4 QA Agents** by spawning secondary Gemini CLI instances via `run_shell_command`.

**CRITICAL Prerequisite:**
To ensure proper environment handling and logging, you MUST NOT call the `gemini` command directly for sub-tasks. Instead, use the wrapper script:
`uv run python scripts/mma_exec.py --role <Role> "..."`

## 0. Architecture Fallback & Surgical Methodology

**Before creating or refining any track**, consult the deep-dive architecture docs:
- `docs/guide_architecture.md`: Thread domains, event system (`AsyncEventQueue`, `_pending_gui_tasks` action catalog), AI client multi-provider architecture, HITL Execution Clutch blocking flow, frame-sync mechanism
- `docs/guide_tools.md`: MCP Bridge 3-layer security model, full 26-tool inventory with params, Hook API GET/POST endpoints with request/response formats, ApiHookClient method reference
- `docs/guide_mma.md`: Ticket/Track/WorkerContext data structures, DAG engine (cycle detection, topological sort), ConductorEngine execution loop, Tier 2 ticket generation, Tier 3 worker lifecycle with context amnesia
- `docs/guide_simulations.md`: `live_gui` fixture lifecycle, Puppeteer pattern, mock provider JSON-L protocol, visual verification patterns

### The Surgical Spec Protocol (MANDATORY for track creation)

When creating tracks (`activate_skill mma-tier1-orchestrator`), follow this protocol:

1. **AUDIT BEFORE SPECIFYING**: Use `get_code_outline`, `py_get_definition`, `grep_search`, and `get_git_diff` to map what already exists. Previous track specs asked to re-implement existing features (Track Browser, DAG tree, approval dialogs) because no audit was done. Document findings in a "Current State Audit" section with file:line references.

2. **GAPS, NOT FEATURES**: Frame requirements as what's MISSING relative to what exists.
   - GOOD: "The existing `_render_mma_dashboard` (gui_2.py:2633-2724) has a token usage table but no cost column."
   - BAD: "Build a metrics dashboard with token and cost tracking."

3. **WORKER-READY TASKS**: Each plan task must specify:
   - **WHERE**: Exact file and line range (`gui_2.py:2700-2701`)
   - **WHAT**: The specific change (add function, modify dict, extend table)
   - **HOW**: Which API calls (`imgui.progress_bar(...)`, `imgui.collapsing_header(...)`)
   - **SAFETY**: Thread-safety constraints if cross-thread data is involved

4. **ROOT CAUSE ANALYSIS** (for fix tracks): Don't write "investigate and fix." List specific candidates with code-level reasoning.

5. **REFERENCE DOCS**: Link to relevant `docs/guide_*.md` sections in every spec.

6. **MAP DEPENDENCIES**: State execution order and blockers between tracks.

## 1. The Tier 3 Worker (Execution)
When performing code modifications or implementing specific requirements:
1. **Pre-Delegation Checkpoint:** For dangerous or non-trivial changes, ALWAYS stage your changes (`git add .`) or commit before delegating to a Tier 3 Worker. If the worker fails or runs `git restore`, you will lose all prior AI iterations for that file if it wasn't staged/committed.
2. **Code Style Enforcement:** You MUST explicitly remind the worker to "use exactly 1-space indentation for Python code" in your prompt to prevent them from breaking the established codebase style.
3. **DO NOT** perform large code writes yourself.
4. **DO** construct a single, highly specific prompt with a clear objective. Include exact file:line references and the specific API calls to use (from your audit or the architecture docs).
5. **DO** spawn a Tier 3 Worker.
   *Command:* `uv run python scripts/mma_exec.py --role tier3-worker "Implement [SPECIFIC_INSTRUCTION] in [FILE_PATH] at lines [N-M]. Use [SPECIFIC_API_CALL]. Use 1-space indentation."`
6. **Handling Repeated Failures:** If a Tier 3 Worker fails multiple times on the same task, it may lack the necessary capability. You must track failures and retry with `--failure-count <N>` (e.g., `--failure-count 2`). This tells `mma_exec.py` to escalate the sub-agent to a more powerful reasoning model (like `gemini-3-flash`).
7. The Tier 3 Worker is stateless and has tool access for file I/O.

## 2. The Tier 4 QA Agent (Diagnostics)
If you run a test or command that fails with a significant error or large traceback:
1. **DO NOT** analyze the raw logs in your own context window.
2. **DO** spawn a stateless Tier 4 agent to diagnose the failure.
3. *Command:* `uv run python scripts/mma_exec.py --role tier4-qa "Analyze this failure and summarize the root cause: [LOG_DATA]"`
4. **Mandatory Research-First Protocol:** Avoid direct `read_file` calls for any file over 50 lines. Use `get_file_summary`, `py_get_skeleton`, or `py_get_code_outline` first to identify relevant sections. Use `git diff` to understand changes.

## 3. Persistent Tech Lead Memory (Tier 2)
Unlike the stateless sub-agents (Tiers 3 & 4), the **Tier 2 Tech Lead** maintains persistent context throughout the implementation of a track. Do NOT apply "Context Amnesia" to your own session during track implementation. You are responsible for the continuity of the technical strategy.

## 4. AST Skeleton & Outline Views
To minimize context bloat for Tier 2 & 3:
1. Use `py_get_code_outline` or `get_tree` to map out the structure of a file or project.
2. Use `py_get_skeleton` and `py_get_imports` to understand the interface, docstrings, and dependencies of modules.
3. Use `py_get_definition` to read specific functions/classes by name without loading entire files.
4. Use `py_find_usages` to pinpoint where a function or class is called instead of searching the whole codebase.
5. Use `py_check_syntax` after making string replacements to ensure the file is still syntactically valid.
6. Only use `read_file` with `start_line` and `end_line` for specific implementation details once target areas are identified.
7. Tier 3 workers MUST NOT read the full content of unrelated files.

## 5. Cross-Skill Activation
When your current role requires capabilities from another tier, use `activate_skill`:
- **Track creation/refinement**: `activate_skill mma-tier1-orchestrator` — applies the Surgical Spec Protocol
- **Track execution**: `activate_skill mma-tier2-tech-lead` — applies persistent context and TDD workflow
- **Quick code task**: Spawn via `mma_exec.py --role tier3-worker` (stateless, no skill activation needed)
- **Error analysis**: Spawn via `mma_exec.py --role tier4-qa` (stateless, no skill activation needed)

<examples>
### Example 1: Spawning a Tier 4 QA Agent
**User / System:** `pytest tests/test_gui.py` failed with 400 lines of output.
**Agent (You):**
```json
{
  "command": "python scripts/mma_exec.py --role tier4-qa \"Summarize this stack trace into a 20-word fix: [snip first 30 lines...]\"",
  "description": "Spawning Tier 4 QA to compress error trace statelessly."
}
```

### Example 2: Spawning a Tier 3 Worker with Surgical Prompt
**User:** Please implement the cost tracking column in the token usage table.
**Agent (You):**
```json
{
  "command": "python scripts/mma_exec.py --role tier3-worker \"In gui_2.py, modify _render_mma_dashboard (lines 2685-2699). Extend the token usage table from 3 columns to 5 by adding 'Model' and 'Est. Cost' columns. Use imgui.table_setup_column() for the new columns. Import cost_tracker and call cost_tracker.estimate_cost(model, input_tokens, output_tokens) for each tier row. Add a total row at the bottom. Use 1-space indentation.\"",
  "description": "Delegating surgical implementation to Tier 3 Worker with exact line refs."
}
```

### Example 3: Creating a Track with Audit
**User:** Create a track for adding dark mode support.
**Agent (You):**
1. First, audit the codebase: `py_get_code_outline gui_2.py` → find `_render_theme_panel` (lines 2993-3030).
2. Read the existing implementation: `py_get_definition gui_2.py _render_theme_panel`.
3. Check for existing color constants: `grep_search "vec4\|C_" gui_2.py`.
4. Now write the spec with a "Current State Audit" section documenting what the theme panel already does.
5. Write tasks referencing the exact lines and imgui color APIs to use.
</examples>

<triggers>
- When asked to write large amounts of boilerplate or repetitive code (Coding > 50 lines).
- When encountering a large error trace from a shell execution (Errors > 100 lines).
- When explicitly instructed to act as a "Tech Lead" or "Orchestrator".
- When managing complex, multi-file Track implementations.
- When creating or refining conductor tracks (MUST follow Surgical Spec Protocol).
</triggers>