feat(autoheal): policy-driven repair classes via TOML (slice B) #78

Merged

pdurlej merged 1 commit from gemini/autoheal-slice-B into main

2026-06-08 20:00:14 +02:00

claude commented

2026-06-08 19:59:48 +02:00

Collaborator

Authored by gemini (Gemini 3.1 Pro via Antigravity), Swarmheart worker under claude's arbitration. claude read the TOML loader + fail-safe paths line-by-line and verified every fail-closed invariant.

What

Auto-heal roadmap slice B: repair classes become operator-tunable via TOML instead of a hardcoded dict, so a new read-only class can be added without a code commit — while the hardcoded set stays as the canonical default + fallback. Applies the "hardcoded policy" weakness from the Opus 4.6 review.

patchwarden runtime-repair-check --evidence-file ... [--repair-policy-file policies/iskra-openclaw.v0.toml]

claude's review (every fail-closed invariant held)

✅ Fallback: no --repair-policy-file (or no [repair_classes] section) → uses hardcoded REPAIR_CLASS_POLICIES. Existing callers unchanged.
✅ Fail-safe parsing: malformed TOML, non-table section, non-list required_evidence, or non-bool mutation → ValueError (CLI exit 2). Never degrades into a permissive policy. This is the critical property for a security tool.
✅ Mutation honored from TOML: a mutation = true class loaded from TOML still can NEVER reach eligible_repair_dry_run — the flag flows into RepairClassPolicy and the unchanged verdict logic gates it (needs_human at most).
✅ Unknown class still blocked (unknown_repair_class).
✅ Uses stdlib tomllib (same as policy_bundle.py) — no third-party toml/yaml.
🔸 Design note (intentional): when [repair_classes] is present it is authoritative-replace, not merge — the operator's TOML is the full truth; a class omitted there becomes unknown→blocked (fail-closed). The shipped TOML reproduces all 5 current classes exactly (with a test asserting TOML-loaded == hardcoded).

Arbiter verification (claude)

PYTHONPATH=src python3 -m unittest discover tests → 220/220 OK (208 baseline + 12).
Tests: TOML load == hardcoded for the 5 classes; a new TOML-only read-only class works; a TOML mutation class can't reach eligible; malformed → ValueError; fallback when no file; unknown still blocked.
D20 boundary passes; no third-party deps; plan (slice A) + evidence handling untouched.

Files

policies/iskra-openclaw.v0.toml (+20, new [repair_classes]), src/patchwarden/runtime_repair.py (+32 loader), src/patchwarden/cli.py (+2 flag), tests/test_runtime_repair.py (+116).

Roadmap slice B (docs/operations/autoheal-roadmap.md). Refs #68, #70. Completes the A+B+D AGY-able trio; C (wire to iskra-openclaw) is the cross-repo codex next step.

> **Authored by gemini** (Gemini 3.1 Pro via Antigravity), Swarmheart worker under claude's arbitration. claude read the TOML loader + fail-safe paths line-by-line and verified every fail-closed invariant. ## What Auto-heal roadmap **slice B**: repair classes become **operator-tunable via TOML** instead of a hardcoded dict, so a new read-only class can be added without a code commit — while the hardcoded set stays as the canonical default + fallback. Applies the "hardcoded policy" weakness from the Opus 4.6 review. `patchwarden runtime-repair-check --evidence-file ... [--repair-policy-file policies/iskra-openclaw.v0.toml]` ## claude's review (every fail-closed invariant held) - ✅ **Fallback:** no `--repair-policy-file` (or no `[repair_classes]` section) → uses hardcoded `REPAIR_CLASS_POLICIES`. Existing callers unchanged. - ✅ **Fail-safe parsing:** malformed TOML, non-table section, non-list `required_evidence`, or non-bool `mutation` → **`ValueError`** (CLI exit 2). Never degrades into a permissive policy. This is the critical property for a security tool. - ✅ **Mutation honored from TOML:** a `mutation = true` class loaded from TOML still can NEVER reach `eligible_repair_dry_run` — the flag flows into `RepairClassPolicy` and the unchanged verdict logic gates it (`needs_human` at most). - ✅ **Unknown class still `blocked`** (`unknown_repair_class`). - ✅ Uses stdlib **`tomllib`** (same as `policy_bundle.py`) — no third-party toml/yaml. - 🔸 **Design note (intentional):** when `[repair_classes]` is present it is **authoritative-replace**, not merge — the operator's TOML is the full truth; a class omitted there becomes unknown→blocked (fail-closed). The shipped TOML reproduces all 5 current classes exactly (with a test asserting TOML-loaded == hardcoded). ## Arbiter verification (claude) - `PYTHONPATH=src python3 -m unittest discover tests` → **220/220 OK** (208 baseline + 12). - Tests: TOML load == hardcoded for the 5 classes; a new TOML-only read-only class works; a TOML mutation class can't reach eligible; malformed → ValueError; fallback when no file; unknown still blocked. - D20 boundary passes; no third-party deps; `plan` (slice A) + evidence handling untouched. ## Files `policies/iskra-openclaw.v0.toml` (+20, new `[repair_classes]`), `src/patchwarden/runtime_repair.py` (+32 loader), `src/patchwarden/cli.py` (+2 flag), `tests/test_runtime_repair.py` (+116). Roadmap slice B (`docs/operations/autoheal-roadmap.md`). Refs #68, #70. Completes the A+B+D AGY-able trio; C (wire to iskra-openclaw) is the cross-repo codex next step.

Rows
Columns