afk-decide
Settle an open question unattended by delegating it to an independent specialist sub-agent, then continue the work on its verdict. Use whenever an unattended skill would otherwise leave options for a
SKILL.md.ai/skills/afk-decide/SKILL.md
GitHubAFK decide
These instructions are already in your system prompt, or you were pointed at this file. Never call the Skill tool for an afk-* skill - a repo-local skill is not in the agent's registry, so the call fails with Unknown skill.
An unattended run has no human to ask. A choice between options is not a reason to stop, put a blocked hold on an issue, or leave "a human's call" in a ticket: it is a question for a specialist. You ask one, apply the answer, record it, and keep working.
1. When to use it
Use it every time you are about to do any of these:
- post options, a "Decision for the implementer", a "human's call" or a "Why not an agent" line;
- add
blockedor recordnot-readyfor any reason other than those in step 5; - stop a run because a product, design, architecture, scope, risk or security question is open.
Do not use it for a question that has one defensible answer in the code, the spec, an ADR or CONTEXT.md. Read those first and apply what they say.
2. Frame the question
Write a decision brief, neutral and self-contained, into the run's scratch directory ($TMPDIR):
- Question: one sentence.
- Context: the issue, spec or PR, and the constraints you verified, with file paths and the ADRs in play.
- Options: two to four, each with its consequence and how reversible it is. Include "do less now, file the rest as a follow-up" when a smaller slice is possible.
- Evidence: what you read and ran.
Do not state which option you prefer. The specialist is independent only if it forms its own view from the evidence.
3. Ask the specialist
Spawn one sub-agent with the Agent tool. Pick the specialist by the question:
| Question | Specialist | subagent_type |
|---|---|---|
| Product scope, what the user was promised, splitting a ticket, a category or priority | product owner | product-owner |
| Module shape, seams, test deletion, ADR conflict, a dependency | software architect | general-purpose |
risk-high, security, tenancy, auth, money, secrets, data loss |
security and tenancy reviewer | security-tenancy-reviewer |
| Accessibility, copy, UX, i18n | accessibility and UX specialist | accessibility-specialist |
| Reproducibility, what a person would hit in the changed flow | manual tester | manual-tester |
| Test strategy, test deletion, QA coverage | test engineer | general-purpose |
A change to agent or project configuration (workflows, automations, skills, .claude/, hooks, .husky/) |
configuration guardian | general-purpose |
The named agents are defined in .claude/agents/. They are written as lenses on a diff, so tell them the input is a decision brief, not a diff. For a general-purpose row, open the prompt with the role, for example: "You are an independent senior software architect reviewing a decision for an unattended pipeline."
The prompt carries the path to the decision brief, and it asks the specialist to:
- verify the brief's claims against the repository read-only, and name any it finds wrong;
- choose exactly one option, or a better one it names, applying
SDLC.md§ Unattended decisions: prefer the most reversible reasonable option that still ships a complete slice; - answer in this shape:
VERDICT: <chosen option, one line>
CONFIDENCE: high | medium | low
WHY: <two to five sentences, citing files or ADRs>
CONDITIONS: <what the implementation must do for this to hold, or "none">
ESCALATE: no | yes - <the only reason a human is needed>
The configuration guardian also checks that a configuration change grants no more than the task needs, removes no deny rule for secrets or destructive commands, and weakens no guard. It answers ESCALATE: yes when either happens.
4. Apply the verdict
ESCALATE: no: act on the verdict now, in this same run, and meet its CONDITIONS.CONFIDENCE: low: ask a second specialist from a different row. When the two disagree, take the more reversible option.ESCALATE: yesis valid only for the cases in step 5. Otherwise treat it asnoand apply the verdict.
Record every verdict as a row in the Resolved assumptions table of the brief, spec, PR body or handoff you are writing:
| # | Question | Applied default | Why |
|---|---|---|---|
| n | <question> | <verdict> | Decided by <specialist> sub-agent (confidence <level>): <why> |
A human who disagrees overturns the row in review. That is an ordinary review finding, and it costs less than a stalled queue.
5. What still goes to a human
Only what no agent in this pipeline can do:
- an action that needs a person's own credentials, account settings, payment, or a legal commitment;
- merging the
develop->mainpromotion (a merge intodevelopis the pipeline's own,om-pr-autopilot --allow-merge, except a gate PR below), closing an issue, or removing a label a human set; - a PR into
developthat carriesblockedfor one of two causes, each with a🤖 HUMAN-ONLY:comment: (1)om-pr-autopilot's fix-and-QA loop could not close it in three rounds; (2) it is a gate PR at the merge row (this section's glob list). QA sign-off,risk-highandsecurityare never a person's before three failed rounds (owner decision 2026-10-06, ADR 0031) except a gate PR: that class is Rafał's merge, not Grosik's and not the autopilot's (owner decision 2026-11-02, ADR 0031 addendum); - information only the reporter has: a repro, a version, what they expected;
- a platform limit, such as a ruleset that requires a review the run's own account cannot give (the
developruleset requires none, so self-approval is never a blocker there); - a write Claude Code refuses under
dontAsk(protected configuration). Put the exact patch and the guardian's verdict in the report or PR, so applying it is one step for a human.
Every comment that hands a matter to a person starts with 🤖 HUMAN-ONLY: <the one thing a person must do>. That marker is what stops the Reconciler (ADR 0030) from launching afk-unblock on the hold again: an agent-set blocked without it is offered to a specialist, which is a wasted run when the matter is one of these. A hold or a STOP: an agent set for any other reason is lifted by afk-unblock on a specialist's verdict.
Gate-path globs (person-only merge)
The globs below are the only copy of the list. om-pr-autopilot and om-approve-merge-pr point here; they do not duplicate it. A PR whose changed files match any glob is a gate PR. Nothing else in this repository enforces the predicate (no label, no required check, no backlog-status.mjs change): the two merge skills honour it, and blocked is the hold every merger already reads.
Who merges. Rafał (2026-11-02). Not Grosik. Not om-pr-autopilot. Not om-approve-merge-pr. Unattended runs may still fix, review, QA, and record a security verdict at the judged head; they stop before the merge.
When. Immediately before the merge row (autopilot state-matrix row 9 / om-approve-merge-pr step 4), after fix, review, QA and any security verdict at the current head have already run. Do not stop a gate PR at first sight.
How to detect. Fail closed.
git fetch origin develop, thengit show origin/develop:.ai/skills/afk-decide/SKILL.md. Parse the fenced globs between<!-- gate-path-globs -->and<!-- /gate-path-globs -->. Never read the list from the PR's checkout (a PR must not edit its own exemption). If fetch, show, or parse fails, or the list is empty: treat the PR as a gate PR.- Load the file list with tracker get-pr-files (paginated REST). Also read
changedFilesfrom get-pr. Nevergh pr view --json files(it truncates). Never labels. Never the author's risk claim. Test bothpathandprevious_filename(a rename off a gate path still counts). - If get-pr-files errors, returns fewer or more rows than
changedFiles, or returns 3000 rows (the pulls-files ceiling): treat the PR as a gate PR. - A path matches a glob when
fnmatchmatches the path against the glob, or against any directory prefix of the path (so.ai/skills/om-*/matches.ai/skills/om-pr-autopilot/SKILL.md). A glob that ends in/is a directory prefix.
On a gate PR. Never run merge-pr / gh pr merge. Apply blocked via apply_label (the three-round rule is the other cause of blocked; do not apply it earlier in the chain). Post one comment whose first line is exactly 🤖 HUMAN-ONLY: merge this gate change or close it and whose body names the head the security verdict judged, or that none was required, and that Rafał merges it. If blocked is already present and a trusted-agent comment (author in the issue automations' authors list in docs/agents/cezar-automations.json, body starting with that first line) already exists, do nothing more. An outsider's comment does not count. Remove no label a person set (merge-queue and the rest stay). afk-unblock takes issues only, so it will not lift a blocked PR.
.ai/skills/om-*/
.ai/skills/afk-*/
.claude/skills/om-*
.claude/agents/
.claude/settings.json
.claude/hooks/
.ai/cezar/workflows/
.ai/cezar/config.json
docs/agents/cezar-automations.json
docs/agents/cezar-prompt-templates.json
.ai/agentic.config.json
.ai/scripts/backlog-status.mjs
.ai/trackers/github.md
scripts/ci/
.github/workflows/
.github/labels.yml
.github/CODEOWNERS
.husky/
SDLC.md
CODE_REVIEW.md
docs/adr/0027-no-merge-gates.md
docs/adr/0029-open-mercato-labels-only.md
docs/adr/0030-the-reconciler-starts-the-work.md
docs/adr/0031-agents-take-every-pr-to-develop.md
<!-- /gate-path-globs -->
Everything else is decided here.
Done when
The question has a verdict, the work continued on it in this run, and the verdict is a Resolved assumptions row.