refactor(v0): policy-driven issue/PR policies via TOML (closes #81) #82

Merged

pdurlej merged 1 commit from gemini/issue-81 into main

2026-06-08 22:03:37 +02:00

claude commented

2026-06-08 21:56:23 +02:00

Collaborator

Authored by gemini (Gemini 3.1 Pro via Antigravity), Swarmheart worker under claude's arbitration. claude read every src change line-by-line + verified behavior preservation against the unmodified existing tests.

What

Closes #81 (last unaddressed Opus 4.6 review weakness §II.4): moves hardcoded issue/PR policy constants into the policy-bundle TOML with a hardcoded fallback — same validated pattern as slice B (#78).

issue_check.py: evaluate_issue now takes an optional PolicyBundle; uses bundle.issue_required_sections / issue_no_go_labels / issue_owner_attention_labels when present, else the hardcoded constants.
policy_bundle.py: PolicyBundle gains 4 optional fields (default None); load_bundle parses an optional [issue_check] table + sensitive_classifications, fail-safe (ValueError on malformed).
pr_check.py: _blockers now uses the canonical SENSITIVE_CLASSIFICATIONS (with bundle override) instead of an inline duplicate set.
policies/platform.v0.toml: new optional [issue_check] section.

claude's review (behavior preserved — verified two ways)

Existing tests untouched + green: test_issue_check.py (+22, 0 deletions), test_policy_bundle.py (+76, 0 deletions), test_pr_check.py not touched at all. 224/224 pass. With DEFAULT_BUNDLE (all override fields None), every path falls back to the hardcoded constants → identical behavior. This is the proof the refactor is behavior-preserving for the no-override path.
One nuance I traced + cleared: _blockers moved from an inline 6-element set {secrets, workflow, runtime, policy_governance, large, unknown} to the canonical SENSITIVE_CLASSIFICATIONS (7 elements — adds policy_exact_file). I verified classify_files never returns policy_exact_file (policy files classify as policy_governance), so the extra element is unreachable → zero behavioral change on real inputs, while removing a duplicated definition (DRY, single source of truth). Net positive.

Invariants

✅ Fallback to hardcoded constants when no TOML override (constants stay in modules).
✅ Malformed [issue_check] / required_sections / label lists → ValueError (fail-safe, never permissive).
✅ stdlib tomllib; no third-party deps. unittest. D20 boundary green. No schema_version change.

New tests

TOML-loaded == hardcoded for current values; TOML override changes verdict (extra no-go label blocks; override sections); malformed → ValueError; absent section → fallback.

Arbiter verification

PYTHONPATH=src python3 -m unittest discover tests → 224/224 OK.
Scope: issue_check.py, policy_bundle.py, pr_check.py, policies/platform.v0.toml, +2 test files (additions only).

Closes #81. Completes the Opus 4.6 review code-weakness list (schema validation #71/#72 · retry #74 · README #73 · repair-class policy #78 · issue/PR policy #81). Same Swarmheart pattern: gemini executes, Opus judges.

> **Authored by gemini** (Gemini 3.1 Pro via Antigravity), Swarmheart worker under claude's arbitration. claude read every src change line-by-line + verified behavior preservation against the unmodified existing tests. ## What Closes #81 (last unaddressed Opus 4.6 review weakness §II.4): moves hardcoded issue/PR policy constants into the policy-bundle TOML with a hardcoded fallback — same validated pattern as slice B (#78). - `issue_check.py`: `evaluate_issue` now takes an optional `PolicyBundle`; uses `bundle.issue_required_sections / issue_no_go_labels / issue_owner_attention_labels` when present, else the hardcoded constants. - `policy_bundle.py`: `PolicyBundle` gains 4 optional fields (default `None`); `load_bundle` parses an optional `[issue_check]` table + `sensitive_classifications`, fail-safe (`ValueError` on malformed). - `pr_check.py`: `_blockers` now uses the canonical `SENSITIVE_CLASSIFICATIONS` (with bundle override) instead of an inline duplicate set. - `policies/platform.v0.toml`: new optional `[issue_check]` section. ## claude's review (behavior preserved — verified two ways) 1. **Existing tests untouched + green:** `test_issue_check.py` (+22, **0 deletions**), `test_policy_bundle.py` (+76, **0 deletions**), `test_pr_check.py` **not touched at all**. 224/224 pass. With `DEFAULT_BUNDLE` (all override fields `None`), every path falls back to the hardcoded constants → identical behavior. This is the proof the refactor is behavior-preserving for the no-override path. 2. **One nuance I traced + cleared:** `_blockers` moved from an inline 6-element set `{secrets, workflow, runtime, policy_governance, large, unknown}` to the canonical `SENSITIVE_CLASSIFICATIONS` (7 elements — adds `policy_exact_file`). I verified `classify_files` **never returns `policy_exact_file`** (policy files classify as `policy_governance`), so the extra element is unreachable → **zero behavioral change on real inputs**, while removing a duplicated definition (DRY, single source of truth). Net positive. ## Invariants - ✅ Fallback to hardcoded constants when no TOML override (constants stay in modules). - ✅ Malformed `[issue_check]` / `required_sections` / label lists → `ValueError` (fail-safe, never permissive). - ✅ stdlib `tomllib`; no third-party deps. unittest. D20 boundary green. No `schema_version` change. ## New tests TOML-loaded == hardcoded for current values; TOML override changes verdict (extra no-go label blocks; override sections); malformed → ValueError; absent section → fallback. ## Arbiter verification - `PYTHONPATH=src python3 -m unittest discover tests` → **224/224 OK**. - Scope: `issue_check.py`, `policy_bundle.py`, `pr_check.py`, `policies/platform.v0.toml`, +2 test files (additions only). Closes #81. Completes the Opus 4.6 review code-weakness list (schema validation #71/#72 · retry #74 · README #73 · repair-class policy #78 · issue/PR policy #81). Same Swarmheart pattern: gemini executes, Opus judges.

Rows
Columns