Stack misdetected as 'angular' in pure-Python repo; 5 retries all failure_class 'unknown', no strategy rotation

[wesleysimplicio]

Stack misdetected as "angular" in a pure-Python repo — 5 consecutive pipeline failures, all classified failure_class: "unknown"

Problem

Running against wesleysimplicio/simplicio-loop (a stdlib-Python repo: pyproject.toml, scripts/*.py, hooks/*.py, zero JS/TS beyond none at all), simplicio-dev-cli recorded the target stack as "angular" and burned 5 attempts on the same target, every one failing with failure_class: "unknown".

Evidence

.simplicio/runs.jsonl in the simplicio-loop repo (2026-06-23):

{"ts":"2026-06-23T21:10:22Z","model":"","provider":"claude","prompt_variant":"default","mode":"pipeline","attempt":1,"ok":false,"failure_class":"unknown","tokens_estimated":1110,"target":"hooks/headroom_dashboard.py","stack":"angular"}
{"ts":"2026-06-23T21:10:28Z","model":"","provider":"claude","prompt_variant":"default","mode":"pipeline","attempt":2,"ok":false,"failure_class":"unknown","tokens_estimated":959,"target":"hooks/headroom_dashboard.py","stack":"angular"}
{"ts":"2026-06-23T21:10:35Z","model":"","provider":"claude","prompt_variant":"default","mode":"pipeline","attempt":3,"ok":false,"failure_class":"unknown","tokens_estimated":982,"target":"hooks/headroom_dashboard.py","stack":"angular"}
{"ts":"2026-06-23T21:10:41Z","model":"","provider":"claude","prompt_variant":"default","mode":"pipeline","attempt":4,"ok":false,"failure_class":"unknown","tokens_estimated":1006,"target":"hooks/headroom_dashboard.py","stack":"angular"}
{"ts":"2026-06-23T21:10:46Z","model":"","provider":"claude","prompt_variant":"default","mode":"pipeline","attempt":5,"ok":false,"failure_class":"unknown","tokens_estimated":1002,"target":"hooks/headroom_dashboard.py","stack":"angular"}

Three independent smells in one log:

1. Stack detection wrong: target is hooks/headroom_dashboard.py (a .py file) in a repo with pyproject.toml at root, yet stack: "angular". Whatever heuristic ran (marker files? cached global default? previous-repo bleed-through?) it ignored both the file extension of the target and the repo manifest. Wrong stack ⇒ wrong skill-router prompt, wrong test harness ⇒ the 6-layer contract can't pass.
2. failure_class: "unknown" × 5: the failure classifier extracted nothing actionable from 5 identical failures. Even "stack/test-harness mismatch" would have been diagnosable from attempt 1 (running an Angular verify against a Python file).
3. Retry loop without strategy change: 5 attempts, same prompt_variant (default), same mode, ~same token cost, ~6s apart — no escalation, no variant rotation, no early abort on repeated identical failure.

Also: model: "" is empty in every row — either a logging gap or the model was never resolved; either way it makes the ledger less useful for debugging.

Proposed fix

1. Stack detection: when the edit target is *.py or the repo root has pyproject.toml/setup.py and no angular.json/package.json, resolve stack python. If detection comes from a cache, key it by repo path and invalidate on manifest change. Log the detection source (stack_source: "angular.json" | "cache" | "default") in runs.jsonl so future misdetections are one-glance diagnosable.
2. Failure classifier: add a class for harness/stack mismatch (verify command not applicable to target language) so it can't hide under unknown.
3. Retry policy: on N (e.g. 2) consecutive failures with identical failure_class AND identical prompt_variant, rotate variant or abort with a diagnostic instead of burning the full attempt budget.
4. Ledger: populate model (or log why it's unresolved).

Acceptance criteria

• Running against a repo with pyproject.toml and a .py target logs stack: "python" (test with a fixture repo).
• stack_source recorded in runs.jsonl.
• Harness/stack mismatch produces a named failure_class, not unknown.
• Identical-failure retries rotate strategy or abort early (test).