docs(autoheal): roadmap — sequencing slices A–E + apply-path boundary #75

Merged

pdurlej merged 1 commit from claude/autoheal-roadmap into main

2026-06-08 19:28:58 +02:00

claude commented

2026-06-08 18:38:37 +02:00

Collaborator

What

A living roadmap doc (docs/operations/autoheal-roadmap.md) sequencing the supervised auto-heal gate from what's on main today to a fully-wired, evidence-driven gate. PLAN only — no code, no execution. Sibling of pyfallow-integration-plan.md.

Why now

Auto-heal is the operator's current priority. #66 (PR gate) + #70 (evidence evaluator) shipped, but the remaining work was only in scattered issues (#65, #68) + my head. This captures the sequence so the next session/cousin picks it up cold, and each slice stays bounded + fail-closed.

Where we are on main: #66 PR gate, #70 evidence evaluator.
The boundary invariant every slice keeps: Iskra owns evidence+execution, Patchwarden owns the gate, operator owns approval. A runtime mutation never reaches an "eligible/go" verdict (D20/D21).
Slice sequence A–E with scope, acceptance, deps, suggested executor, M2/D21 status:
- A — deploy_drift_probe plan-object (closes substance of #65) — AGY-able
- B — policy-driven repair classes (TOML) — AGY-able, claude-reviewed
- C — wire evaluator into iskra-openclaw (cross-repo, codex; makes it live)
- D — schema contract test for runtime-repair-verdict — trivial
- E — real Iskra evidence model — blocked on live fixtures
Sequencing diagram (A/B/D parallel now → C → E).
The apply-path question (parked): Patchwarden never applies; at most gates an Iskra-executed, operator-approved, deterministic-executor apply per the clawsweeper pattern (#69).
Decision triggers, wrongness signals, M3 open questions.

Scope

Docs only — one new file. No src//tests//spec/ changes.

Next move after merge

Slices A + B + D are AGY-gemini-able now (operator's Antigravity budget); C is a scoped codex issue (cross-repo). This doc is the spec source for all of them.

Refs #68, #65, #69. D20/D21.

## What A living roadmap doc (`docs/operations/autoheal-roadmap.md`) sequencing the supervised auto-heal gate from what's on `main` today to a fully-wired, evidence-driven gate. **PLAN only — no code, no execution.** Sibling of `pyfallow-integration-plan.md`. ## Why now Auto-heal is the operator's current priority. #66 (PR gate) + #70 (evidence evaluator) shipped, but the remaining work was only in scattered issues (#65, #68) + my head. This captures the sequence so the next session/cousin picks it up cold, and each slice stays bounded + fail-closed. ## Contents - **Where we are** on `main`: #66 PR gate, #70 evidence evaluator. - **The boundary invariant** every slice keeps: Iskra owns evidence+execution, Patchwarden owns the gate, operator owns approval. A runtime mutation never reaches an "eligible/go" verdict (D20/D21). - **Slice sequence A–E** with scope, acceptance, deps, suggested executor, M2/D21 status: - **A** — `deploy_drift_probe` plan-object (closes substance of #65) — AGY-able - **B** — policy-driven repair classes (TOML) — AGY-able, claude-reviewed - **C** — wire evaluator into iskra-openclaw (cross-repo, codex; makes it *live*) - **D** — schema contract test for `runtime-repair-verdict` — trivial - **E** — real Iskra evidence model — blocked on live fixtures - **Sequencing diagram** (A/B/D parallel now → C → E). - **The apply-path question** (parked): Patchwarden never applies; at most gates an Iskra-executed, operator-approved, deterministic-executor apply per the clawsweeper pattern (#69). - **Decision triggers**, **wrongness signals**, **M3 open questions**. ## Scope Docs only — one new file. No `src/`/`tests/`/`spec/` changes. ## Next move after merge Slices A + B + D are AGY-gemini-able now (operator's Antigravity budget); C is a scoped codex issue (cross-repo). This doc is the spec source for all of them. Refs #68, #65, #69. D20/D21.

Rows
Columns