# Outstanding MMA Test Failures — Track Proposal

**Date:** 2026-06-27
**Branch:** `tier2/post_module_taxonomy_de_cruft_20260627`
**Latest commit:** `635ca552` (partial fix)

---

## Status: 1 critical test still failing in tier-3-live_gui

```
tests/test_mma_concurrent_tracks_sim.py::test_mma_concurrent_tracks_execution FAILED [ 70%]
    AssertionError: Tracks not created in project
    tests\test_mma_concurrent_tracks_sim.py:66: AssertionError
```

After plan-epic succeeds (2 proposed tracks), the test clicks `btn_mma_accept_tracks`. The bg_task logs "Starting 2 tracks..." but only 1 sprint-ticket mock call is observed (for track-a). The 2nd sprint call for track-b never happens. Test polls `tracks` for 30 seconds and times out.

**Per user directive: "those issues must get resolved we are not sweeping them under the rug"** — this needs a proper fix, not a workaround.

---

## Root Cause Analysis

The failure is the result of a **chain of cruft_elimination_20260627 changes that propagated incompletely through the production code and the test mock**:

### 1. `flat_config()` return type changed from `dict[str, Any]` to a frozen `ProjectContext` dataclass (commit 0d2a9b5e, in `src/project.py`)

**Impact:** 3 production sites in `src/app_controller.py` mutated the returned object via dict-style assignment:
- `_do_generate` (line 4027): `flat["files"] = ...` and `flat["files"]["paths"] = ...`
- `_cb_plan_epic` (line 4604): `flat.setdefault("files", {})["paths"] = ...`
- `_start_track_logic_result` (line 4793): `flat.setdefault("files", {})["paths"] = ...`

Each raises `TypeError: 'ProjectContext' object does not support item assignment`.

**Status:** ✅ **FIXED** in commits `a4901fa2` and `635ca552` (call `flat.to_dict()` to get a mutable dict).

### 2. `conductor_tech_lead.topological_sort()` return type changed from `list[str]` to `list[Ticket]` (likely also in 0d2a9b5e or related)

**Impact:** `_start_track_logic_result` in `src/app_controller.py` iterated over `sorted_tickets_data` and used `t_data["id"]`, `t_data.get("description")`, etc. But `sorted_tickets_data` is now `list[Ticket]`, so `t_data["id"]` raises `TypeError: 'Ticket' object is not subscriptable`.

**Status:** ✅ **FIXED** in commit `635ca552` (use Ticket attribute access: `t_data.id`, `t_data.description`, etc.).

### 3. `gemini_cli_adapter` uses session persistence via `--resume` flag (commit 0d2a9b5e or related)

**Impact:** The mock `tests/mock_concurrent_mma.py` was written when each LLM call was stateless. Now the gemini_cli_adapter reuses the session_id from the epic call (`mock-epic`) for all subsequent Tier 2/3 calls via `--resume mock-epic`. The mock's response routing (based on prompt substrings) broke because:
- Epic init: `if 'PATH: Epic Initialization' in prompt` (prompt is real)
- Sprint: `if 'generate the implementation tickets' in prompt` (prompt is empty in resume mode!)
- Worker: `if 'You are assigned to Ticket' in prompt` (prompt is empty)

So all resume calls fell to the default case, which returns a generic mock response that doesn't parse as JSON.

**Status:** ✅ **PARTIALLY FIXED** in commit `635ca552` (mock now parses `--resume` from sys.argv and uses a persistent call counter to route to per-track responses).

### 4. ✅ **RESOLVED** — Production bug: NameError on `models.Metadata` call site

After all 3 prior fixes in commit `635ca552`, only 1 sprint-ticket call was observed (for track-a). The for loop in `_cb_accept_tracks._bg_task` was reached but track-a's `_start_track_logic` raised a `NameError` that was NOT caught by the EXCEPT block (which only catches 7 specific exception types). The io_pool worker died, the for loop never reached track-b.

**Root cause:** The de-cruft migration in commit `ee763eea` removed `from src import models` from `src/app_controller.py` but did not update the call site `models.Metadata(...)` at line 4830. The line is:
```python
meta = models.Metadata(id=track_id, name=title, status="todo", created_at=datetime.now(), updated_at=datetime.now())
```

`models` is no longer in scope, so this raises `NameError: name 'models' is not defined`.

**Status:** ✅ **FIXED** in commit `e9919059` (added `TrackMetadata` to the `from src.mma import` line; changed `models.Metadata(...)` to `TrackMetadata(...)`).

**Verification:** 5 consecutive PASS runs of `test_mma_concurrent_tracks_execution` (7.49s, 7.54s, 7.97s, 8.02s, 8.45s). The full diag log shows both tracks are created:
```
[DIAG] _start_track_logic_result self.tracks.append OK title='Track A' track_id=track_ef3ff66ba50c
[DIAG] _start_track_logic_result ENTER title='Track B' goal='Track B Goal' skeletons_len=0
[DIAG] _start_track_logic_result AFTER generate_tickets title='Track B' raw_tickets_count=1
...
[DIAG] _start_track_logic_result self.tracks.append OK title='Track B' track_id=track_52e6741b0748
```

### 5. ✅ **RESOLVED** — Mock bug: session_id-based routing for sprints is fragile

The session_id-based routing added in commit `635ca552` had two sub-bugs:
- `call_n` literal matching (`== 2`, `== 3`) is fragile to test ordering: the file-based counter persists across tests in the same session, so `call_n != 2` for the 1st sprint if a prior test ran.
- `session_id="mock-sprint-A"` means "this is a follow-up call after the 1st sprint returned mock-sprint-A", so the response should be sprint-B (2nd track tickets), not sprint-A. The prior code routed this to sprint-A, causing track-b's worker to have stream id `ticket-A-1` (not `ticket-B-1`).

**Status:** ✅ **FIXED** in commit `913aa48c` (replaced session_id-based sprint routing with prompt-content-based routing; the original pre-`635ca552` design).

**Verification:** 3 consecutive PASS runs after the fix.

The test counter is at 2 after the test runs (one epic + one sprint). This proves the mock was called twice. The third call (sprint-B) never happens.

**Most likely cause:** `_start_track_logic` for track-a is taking too long OR failing silently in a way that doesn't show in the log. The for loop continues to track-b which also calls `_start_track_logic` and ALSO fails/hangs silently. The 30-second test poll times out before either track completes.

---

## What's Needed

### Option A: Continue investigation in this iteration (Tier 2 autonomous track)

1. **Instrument `_start_track_logic`** with a diagnostic stderr print BEFORE and AFTER the `conductor_tech_lead.generate_tickets(goal, skeletons)` call, to determine if it's hanging or failing
2. **Run the test in isolation** with the instrumentation
3. **If hanging:** check `aggregate.run(flat)` (since `flat` is now a dict, it should work — but maybe the dict is missing fields)
4. **If failing:** the except block in `_start_track_logic_result` catches it; add a print before the `return Result(data=None, errors=[err])` to see the error

### Option B: Open a new Tier 2 track

Create `conductor/tracks/fix_mma_concurrent_tracks_sim_20260627/spec.md` with:
- **Goal:** Make `test_mma_concurrent_tracks_sim::test_mma_concurrent_tracks_execution` pass in the batched test suite
- **Scope:** Investigate the second-track-not-firing issue, fix the root cause (production OR mock), verify
- **Owner:** Tier 2 autonomous (this session) or Tier 1 manual review
- **Estimated scope:** 3-5 files changed (production in `src/app_controller.py` and/or mock in `tests/mock_concurrent_mma.py`), 1-2 hour investigation + fix + verify

---

## Files Currently Modified (uncommitted in working tree)

| File | Change |
|------|--------|
| `src/app_controller.py` | `flat.setdefault(...)["paths"] = ...` → `flat = flat.to_dict() if hasattr...; flat.setdefault(...)["paths"] = ...` (2 sites); `t_data["id"]` → `t_data.id` (1 site) |
| `tests/mock_concurrent_mma.py` | Parse `--resume` arg from sys.argv; use persistent call counter for per-call response routing |

**Not committed yet** — staged for the next tier2 autonomous run.

---

## Recommendation

**Open a dedicated track** for this work. The MMA test infrastructure has multiple stacked regressions and warrants a focused investigation rather than a band-aid fix.

If the user wants me to **continue in this session**, I can:
1. Add stderr instrumentation to `_start_track_logic` to diagnose
2. Run the test in isolation
3. Fix the root cause based on the diagnosis
4. Verify the test passes
5. Commit the fix

Per user direction, no sweeping under the rug — this needs a real fix.

### 6. ✅ **RESOLVED** — Mock bug: epic branch only matches one literal prompt

**Date:** 2026-06-27 (discovered after the fix_mma_concurrent_tracks_sim_20260627 track SHIPPED)

The stress test (`tests/test_mma_concurrent_tracks_stress_sim.py::test_mma_concurrent_tracks_stress`) uses `mma_epic_input='STRESS TEST: TRACK A AND TRACK B'`, which the mock's epic branch did NOT match (it only matched `'PATH: Epic Initialization'`). The stress prompt fell to the Default branch which returns text (not JSON), and the production's `orchestrator_pm.generate_tracks` failed to parse it, returning 0 tracks.

**Root cause:** The mock's epic branch was a literal-substring check for a single test-specific prompt. It was not robust to other test prompts.

**Status:** ✅ **FIXED** in commit `fad1755b` (restructured routing so sprint and worker are checked first, and any non-empty prompt that doesn't match those patterns is treated as an epic request returning 2 tracks).

**Verification:** 3 consecutive PASS runs of both `test_mma_concurrent_tracks_execution` AND `test_mma_concurrent_tracks_stress` (13.94s, 14.81s, 14.13s).


### 7. ✅ **RESOLVED** — Production bug: 'refresh_from_project' task overwrites self.tracks

**Date:** 2026-06-27 (discovered after the second batched test run)

After the epic catch-all fix, the batched test still failed. Diagnostic logging revealed that `self.tracks` was being replaced between track appends (different `id(self.tracks)` values in the log). Root cause:

`_start_track_logic_result` (and `_cb_accept_tracks._bg_task`) appended a `'refresh_from_project'` task to `_pending_gui_tasks` at the end. The main thread processed this task by calling `_refresh_from_project`, which does:

    self.tracks = project_manager.get_all_tracks(self.active_project_root)

This REPLACED `self.tracks` with a fresh disk read. In batched test environments, the disk read returned 0 tracks (due to timing or path issues), losing the in-memory tracks that were just appended by `self.tracks.append(...)`.

**Fix:** Remove the `'refresh_from_project'` task appends from both `_start_track_logic_result` and `_cb_accept_tracks._bg_task`. The bg_task already updates `self.tracks` directly via `self.tracks.append(...)`. The refresh is unnecessary for the accept flow because the other state (files, disc_entries, etc.) doesn't change during the accept.

**Status:** ✅ **FIXED** in commit `55dae159`.

**Verification:** 3 consecutive PASS runs of the failing test combination (test_context_sim_live + test_mma_concurrent_tracks_execution + test_mma_concurrent_tracks_stress) at 100.57s, 100.29s, 100.18s. Also passes 15 wider tests (237.63s) with no regressions.