docs(autoheal): iskra-openclaw evidence-bundle integration spec (Slice C / #452) #94

Merged

pdurlej merged 1 commit from claude/autoheal-evidence-bundle-spec into main

2026-06-17 23:30:10 +02:00

claude commented

2026-06-17 23:25:12 +02:00

Collaborator

What

A self-contained contract doc — docs/operations/iskra-openclaw-evidence-bundle-spec.md — for wiring the Patchwarden auto-heal gate live in iskra-openclaw (roadmap Slice C, iskra-openclaw#452).

Why

Slice C is codex / cross-repo territory and currently unstarted. Codex shouldn't have to read all of Patchwarden to learn the interface, and the contract shouldn't live only in an issue body where it can drift. This doc pins, in one place inside the Patchwarden repo:

exact bundle JSON shape (field-by-field) + the evidence-counting rule (status==success AND sha==target_sha);
the runtime-repair-check CLI invocation + exit-code contract (0 = eligible_repair_dry_run, 1 = needs_human/blocked);
the 5 repair classes + required evidence + best-possible verdict (mutation classes cap at needs_human);
verdict + plan shapes (plan only for read-only eligible);
surface rules (artifact floor; optional patchwarden-bot comment) and the fail-closed boundaries;
a nudge to capture 1–2 real bundles as Slice E fixtures during the first dry-run.

Safety / discipline

Docs-only; 296 tests green.
The doc explicitly defers to spec/schemas/runtime-repair-verdict.schema.json + policies/iskra-openclaw.v0.toml as source of truth (cross-links throughout) so it can't become a second, drifting contract.
Reinforces D20/D21: gate not executor, mutation never "go", operator-approval only from an explicit action.

From the 2026-06-16 opportunity audit (auto-heal finder). Unblocks codex on iskra-openclaw#452. Do not merge — operator merge only.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

## What A **self-contained contract doc** — `docs/operations/iskra-openclaw-evidence-bundle-spec.md` — for wiring the Patchwarden auto-heal gate live in `iskra-openclaw` (roadmap **Slice C**, [`iskra-openclaw#452`](https://git.pdurlej.com/pdurlej/iskra-openclaw/issues/452)). ## Why Slice C is `codex` / cross-repo territory and currently unstarted. Codex shouldn't have to read all of Patchwarden to learn the interface, and the contract shouldn't live only in an issue body where it can drift. This doc pins, in one place inside the Patchwarden repo: - exact **bundle JSON shape** (field-by-field) + the **evidence-counting rule** (`status==success` AND `sha==target_sha`); - the **`runtime-repair-check` CLI invocation** + **exit-code contract** (0 = `eligible_repair_dry_run`, 1 = `needs_human`/`blocked`); - the **5 repair classes** + required evidence + best-possible verdict (mutation classes cap at `needs_human`); - **verdict + `plan`** shapes (plan only for read-only eligible); - **surface rules** (artifact floor; optional `patchwarden`-bot comment) and the **fail-closed boundaries**; - a nudge to **capture 1–2 real bundles as Slice E fixtures** during the first dry-run. ## Safety / discipline - **Docs-only**; 296 tests green. - The doc explicitly **defers to `spec/schemas/runtime-repair-verdict.schema.json` + `policies/iskra-openclaw.v0.toml` as source of truth** (cross-links throughout) so it can't become a second, drifting contract. - Reinforces D20/D21: gate not executor, mutation never "go", operator-approval only from an explicit action. From the 2026-06-16 opportunity audit (auto-heal finder). Unblocks codex on `iskra-openclaw#452`. **Do not merge — operator merge only.** Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Rows
Columns