| name | zero-touch |
| description | Canonical definition of zero-touch-by-operator for PR autonomy measurement. Use when evaluating, reporting, or labeling PR autonomy status. |
Zero-Touch-by-Operator
Definition
A PR is zero-touch when AO (Agent Orchestrator) produced a PR that completed
the draft-first quality sequence and reached /green without manual
intervention from a Claude Code terminal session or human operator. Canonical
definitions live in draft-first-pr and pr-green-definition.
This repo requires explicit human MERGE APPROVED authorization before any
merge (see this repo's merge-safety policy) — that authorization step, and the
gh pr merge command it triggers, are not by themselves a zero-touch
violation. The metric measures AO's autonomy in producing a mergeable-quality
PR, not who executes the final merge click. (Prior to
PR #8217, condition
#3 required skeptic-cron.yml to have auto-merged the PR; that gate and its
auto-merge mechanism no longer exist in this repo, so this condition is retired
rather than left permanently unsatisfiable.)
Specifically:
- AO spawned the worker session autonomously (via
ao spawn, lifecycle-worker, or poller)
- The worker completed
/es, /er, and /advice, then reached /green, without external nudges
- No human/terminal session pushed commits, resolved comments, fixed CI, or posted reviews on the PR branch before it reached that state
Quality-complete criteria
Use ~/.claude/skills/draft-first-pr/SKILL.md for the draft-phase sequence and
~/.claude/skills/pr-green-definition/SKILL.md for the two /green gates.
CodeRabbit and Bugbot are advisory and do not determine zero-touch status.
Repo evidence exceptions still apply where the evidence standards permit them.
What breaks zero-touch
- A Claude Code terminal session (human-operated) pushed fixes to the PR branch
- A human or terminal session posted review comments or approvals
- A human dismissed a bot review to unblock merge
- A terminal session ran
@coderabbitai approve or similar to bypass a gate
- A human manually triggered CI re-runs to get past flaky tests
Not counted as breaking zero-touch: the mandatory MERGE APPROVED human
authorization and the merge command itself — every PR in this repo requires
this regardless of how autonomously it was produced.
What does NOT break zero-touch
- Jeffrey asking questions about the PR in Slack (observation, not intervention)
- Bot-to-bot interactions (CR, Bugbot, Skeptic, evidence-review-bot)
- AO workers posting
@coderabbitai approve or @coderabbitai all good? (worker doing its job)
- Jeffrey approving the AO spawn itself (dispatch is expected; execution must be autonomous)
Measurement — GitHub actor audit
A PR is zero-touch if, up to the point of the merge action, the only GitHub
actors (commits, reviews, comments) are:
jleechan2015 (AO agent GitHub identity)github-actions[bot] (CI)coderabbitai[bot] (code review)cursor[bot] (Bugbot)
The merge action itself is excluded from this audit. Since PR #8217 removed
the auto-merge mechanism, every merge in this repo is executed by a human-facing
session under explicit MERGE APPROVED authorization — the merge event is not a
zero-touch signal one way or the other. Only audit actions that happened before
the merge command.
If $USER (Jeffrey's personal account) or any other human/terminal identity
appears as a commit author, reviewer, or commenter before the merge, it's
operator-assisted.
KPI
zero_touch_rate = (zero-touch merged PRs) / (total merged PRs) * 100
Measured weekly. Target: increasing trend.
Labeling (planned)
Previously planned to run via skeptic-cron.yml; that workflow no longer exists
(removed by PR #8217). Labeling mechanism TBD — not yet implemented, tracked in beads.
Reference
(No external doc currently exists — this SKILL.md is the sole source of truth.
A prior reference to ~/.openclaw/docs/ZERO_TOUCH.md/.html pointed to files
that do not exist on disk; removed rather than left as a dead link.)
