conductor(gui_ux): Add Phase 6 — live streaming, per-tier model config, parallel DAG, auto-retry

Addresses three gaps where Claude Code and Gemini CLI outperform Manual Slop's
MMA during actual execution:

1. Live worker streaming: Wire comms_log_callback to per-ticket streams so
   users see real-time output instead of waiting for worker completion.
2. Per-tier model config: Replace hardcoded get_model_for_role with GUI
   dropdowns persisted to project TOML.
3. Parallel DAG execution: asyncio.gather for independent tickets (exploratory
   — _send_lock may block, needs investigation).
4. Auto-retry with escalation: flash-lite -> flash -> pro on BLOCKED, up to
   2 retries (wires existing --failure-count mechanism into ConductorEngine).

7 new tasks across Phase 6, bringing total to 30 tasks across 6 phases.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

conductor/tracks/comprehensive_gui_ux_20260228/plan.md

@@ -45,3 +45,14 @@ Focus: Dense, responsive dashboard with arcade aesthetics and end-to-end verific
 - [ ] Task 5.4: Write an end-to-end integration test (extending `tests/visual_sim_mma_v2.py` or creating `tests/visual_sim_gui_ux.py`) that verifies via `ApiHookClient`: (a) track creation form produces correct directory structure; (b) tier streams are populated during MMA execution; (c) approval indicators appear when expected; (d) cost tracking shows non-zero values after execution.
 - [ ] Task 5.5: Verify all new UI elements maintain >30 FPS via `get_ui_performance` during a full MMA simulation run.
 - [ ] Task 5.6: Conductor - User Manual Verification 'Phase 5: Visual Polish & Integration Testing' (Protocol in workflow.md)
+
+## Phase 6: Live Worker Streaming & Engine Enhancements
+Focus: Make MMA execution observable in real-time and configurable from the GUI. Currently workers are black boxes until completion.
+
+- [ ] Task 6.1: Wire `ai_client.comms_log_callback` to per-ticket streams during `run_worker_lifecycle` (multi_agent_conductor.py:207-300). Before calling `ai_client.send()`, set `ai_client.comms_log_callback` to a closure that pushes intermediate text chunks to the GUI via `_queue_put(event_queue, loop, "response", {"text": chunk, "stream_id": f"Tier 3 (Worker): {ticket.id}", "status": "streaming..."})`. After `send()` returns, restore the original callback. This gives real-time output streaming to the Tier 3 stream panels from Phase 1.
+- [ ] Task 6.2: Add per-tier model configuration to the MMA dashboard. Below the token usage table in `_render_mma_dashboard`, add a collapsible "Tier Model Config" section with 4 rows (Tier 1-4). Each row: tier label + `imgui.combo` dropdown populated from `ai_client.list_models()` (cached). Store selections in `self.mma_tier_models: dict[str, str]` with defaults from `mma_exec.get_model_for_role()`. On change, write to `self.project["mma"]["tier_models"]` for persistence.
+- [ ] Task 6.3: Wire per-tier model config into the execution pipeline. In `ConductorEngine.run` (multi_agent_conductor.py:105-135), when creating `WorkerContext`, read the model name from the GUI's `mma_tier_models` dict (passed via the event queue or stored on the engine). Pass it through to `run_worker_lifecycle` which should use it in `ai_client.set_provider`/`ai_client.set_model_params` before calling `send()`. Also update `mma_exec.py:get_model_for_role` to accept an override parameter.
+- [ ] Task 6.4: Add parallel DAG execution. In `ConductorEngine.run` (multi_agent_conductor.py:100-135), replace the sequential `for ticket in ready_tasks` loop with `asyncio.gather(*[loop.run_in_executor(None, run_worker_lifecycle, ...) for ticket in ready_tasks])`. Each worker already gets its own `ai_client.reset_session()` so they're isolated. Guard with `ai_client._send_lock` awareness — if the lock serializes all sends, parallel execution won't help. In that case, create per-worker provider instances or use separate session IDs. Mark this task as exploratory — if `_send_lock` blocks parallelism, document the constraint and defer.
+- [ ] Task 6.5: Add automatic retry with model escalation. In `ConductorEngine.run`, after `run_worker_lifecycle` returns, check if `ticket.status == "blocked"`. If so, and `retry_count < max_retries` (default 2), increment retry count, escalate the model (e.g., flash-lite → flash → pro), and re-execute. Store `retry_count` as a field on the ticket dict. After max retries, leave as blocked.
+- [ ] Task 6.6: Write tests for: (a) streaming callback pushes intermediate content to event queue; (b) per-tier model config persists to project TOML; (c) retry escalation increments model tier.
+- [ ] Task 6.7: Conductor - User Manual Verification 'Phase 6: Live Worker Streaming & Engine Enhancements' (Protocol in workflow.md)

conductor/tracks/comprehensive_gui_ux_20260228/spec.md

@@ -92,8 +92,21 @@ This track enhances the existing MMA orchestration GUI from its current function
 - Hook API for testing: [docs/guide_tools.md](../../docs/guide_tools.md)
 - Simulation patterns: [docs/guide_simulations.md](../../docs/guide_simulations.md)
 
+## Functional Requirements (Engine Enhancements)
+
+### Live Worker Streaming
+- During `run_worker_lifecycle`, set `ai_client.comms_log_callback` to push intermediate text chunks to the per-ticket stream via the event queue. Currently workers are black boxes until completion — both Claude Code and Gemini CLI stream in real-time. The callback should push `{"text": chunk, "stream_id": "Tier 3 (Worker): {ticket.id}", "status": "streaming..."}` events.
+
+### Per-Tier Model Configuration
+- `mma_exec.py:get_model_for_role` is hardcoded. Add a GUI section with `imgui.combo` dropdowns for each tier's model. Persist to `project["mma"]["tier_models"]`. Wire into `ConductorEngine` and `run_worker_lifecycle`.
+
+### Parallel DAG Execution
+- `ConductorEngine.run()` executes ready tickets sequentially. DAG-independent tickets should run in parallel via `asyncio.gather`. Constraint: `ai_client._send_lock` serializes all API calls — parallel workers may need separate provider instances or the lock needs to be per-session rather than global. Mark as exploratory.
+
+### Automatic Retry with Model Escalation
+- `mma_exec.py` has `--failure-count` for escalation but `ConductorEngine` doesn't use it. When a worker produces BLOCKED, auto-retry with a more capable model (up to 2 retries).
+
 ## Out of Scope
-- Automated "Auto-Fix" loops without user intervention.
 - Remote management via web browser.
 - Visual diagram generation (Dear PyGui node editor for DAG — future track).
 - Docking/floating multi-viewport layout (requires imgui docking branch investigation — future track).