Skip to content
Back to skills

Autoevo Review

ASecurity

Triage actionable Autoevo findings; empty cycles remain machine receipts.

  • 4 stars
  • 0 votes
  • 0 copies
  • 1 view
  • Added October 9, 2026
ai-agentsrustgogit

Security analysis

A100/100

Scanned October 9, 2026

npx -y skills add mhxie/Atelier --skill autoevo-review --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Autoevo Review?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Autoevo Review
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/mhxie-autoevo-review-atelier/badge)](https://www.skillsdirectory.com/skills/mhxie-autoevo-review-atelier)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: autoevo-review
description: Triage actionable Autoevo findings; empty cycles remain machine receipts.
---
# /autoevo-review

Triage pending Autoevo proposals with the user: apply, skip, defer, or explain.
Follow `protocols/autoevo.md`. This is an interactive approval workflow, not a
continuation of the nightly model run.

Never edit canonical nightly JSON or its derived reports during review; doing so
cannot prove an incomplete cycle. Optional human history belongs in a separate
`<paths.agent_findings>/autoevo-review-<YYYY-MM-DD>.md`.

## 1. Load and housekeep

Resolve `<paths.meta>/autoevo_pending.toml` and the current local date once.
Load through `scripts/autoevo_pending.py list --status pending`; a missing queue
or zero returned entries means there is nothing to review. A malformed queue is
a blocker: report it and do not rewrite it.

Before presentation, stage `autoevo_pending.py auto-dismiss --today <date>` on
private state for entries at three surfaces or over the default 30-day age.

Never mutate live state directly. After approval, the orchestrator writes
explicit note, queue, ledger, and optional review-log changes as plain files
after rechecking their hashes; Reflect commits and pushes them.

## 2. Present the queue

Group remaining entries by `redundant`, `time-stale-A`, `time-stale-B`,
`contradicted`, and `low-signal`. Show a compact category count followed by each
entry's id, action, evidence, peers, proposal date, surface count, and any
`default_action` / `default_at`.

Explain veto semantics exactly:

- Skip vetoes a `stale-banner` default.
- Apply vetoes a `dismiss` default; skip confirms it.
- Defer restarts the 14-day window.
- Silence leaves an eligible default to the later trusted nightly parent.

Let the user choose a category or all entries. If unclear, walk all entries in
order. For each item ask: apply, skip, defer, explain, or quit. Apply and skip
require one user-authored sentence of reason; never invent or accept an empty
reason. Defer may carry an optional reason.

## 3. Resolve one item

### Apply

Dispatch the relevant normal, user-approved workflow. Curator drafts note
changes and retains its full preservation/approval gates. Autoevo's
`auto_apply_safe` flag does not bypass approval here.

A queued contradiction is not necessarily confirmed: the nightly queues both
genuine and unproven contradictions. Unless the entry carries a complete,
gap-free Challenger verdict tied to the exact claim and peer, dispatch
Challenger first. A genuine verdict may proceed to Curator's normal wiki-update
proposal; a rhetorical verdict performs no wiki write and returns to the user
for a dismissal decision.

After approval, stage `autoevo_pending.py resolve --id <id> --status applied
--reason <reason> --today <date>` on private state. Write it with the approved
note change. Archive means byte-preserving addition plus deletion, not a model
write or promised rename.

If the user rejects the proposal, stage a `dismissed` resolution with their
reason instead. No content change is written.

### Skip

Stage `autoevo_pending.py resolve --id <id> --status dismissed --reason
<reason> --today <date>` against private state. Tell the user whether this
vetoes `stale-banner` or confirms `dismiss`; then include it in the next queue
write.

### Defer

Stage `autoevo_pending.py defer --id <id> --today <date>` with the optional
reason. It increments surfacing state and, when applicable, restarts the default
deadline. Optionally ask for a cue snooze (default seven days); cue snoozing is
group-wide, not per entry. Batch compatible defer-only state changes into one
write.

### Explain

Show the complete stored evidence without adding authority. For redundant QMD
findings, show peer paths, QMD ordering scores, and source-reviewed overlap; do
not describe a score floor. For contradictions, show the claim, peer, signal,
probe verdict, and any unresolved gaps. Then ask again.

### Quit

Stop. Write only already confirmed staged changes; leave every other entry
pending. If writing has not begun, discarding scratch state has no live effect.

## 4. Write and summary

Before writing, recheck live hashes and limit the write to approved
note/state/review-log paths. Never commit or push. Rollback is plain
`git restore`.

Report applied, dismissed, deferred, remaining, and any failed write.
Name the queue path and separate review log if one was created. A review log is
ordinary human-triage history only; it is never an Autoevo verifier input.

## Edge cases

- Missing peers or changed evidence make an entry stale. Explain the mismatch
  and ask the user whether to dismiss; do not silently resolve it.
- A wiki change always uses Curator's normal approval path and appends the
  required Revision Log.
- Any daily-note source is a nightly bug. Refuse the content operation and ask
  the user whether to dismiss the queue entry; daily notes remain read-only.
- On ambiguous or partly written effects, stop and report `needs_review`.
  Do not replay the model, automatically roll back, or overwrite state.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.