Why. Once the dojo grades a seat on an SOP, "may this seat do X unsupervised?" is a test result, not an opinion. Belts stop being decoration and become the permission system; seats domesticate themselves by passing katas, and nobody hand-assigns rights.
Ledger.time/belts.json, committed, written only by the dojo runner (#85) from a passing ladder: {seat: {skill: {belt, run, at, lapsed}}}. Belts are soulbound and monotonic, as in orgchain v1.14: a promotion is a new entry over the old one, never a removal. A failed re-run sets lapsed: true — the belt stays on the record, the permission switches off until the seat passes again. Re-runs every TIME_BELT_REVIEW_DAYS (default 28) or whenever the SOP changes. The operator may grant a belt by a decision logged in docs/decisions.md (the founder's prerogative), and it is marked granted.
review.engine: an Approve on a PR under .github/time/ counts in verdict() only from a seat with ≥ brown in review; a lower belt's review is recorded as a COMMENT.
review.product: Approve counts from ≥ green in review.
plan.queue: queuing cards (Astra) needs ≥ brown in management (the sizing kata).
steward.repair: the steward tick (#80) acts only with black in management; otherwise it runs in report-only mode.
escalate.tier: the automatic one-tier retry (#78) targets only seats with ≥ green in build.
Deterministic jobs (housekeeping, Time's own merges, the governor) need no belt: code, not judgment.
Mechanics.belts.py: belt(seat, skill), allowed(seat, action); verdict(), plan_spec/Astra, the steward and the retry path call allowed(). Enforcement is behind TIME_BELTS=on (default off), so the first ladder run (#85) can grant the initial belts before anything is gated; with it off, allowed() logs what it would have refused. cmd_doctor prints the belt table and which gates are live.
Audit. Every kata run commits its transcript next to the result, so a belt can be contested by re-running the kata; graders are deterministic where possible and never the seat under test. Later: mirror each promotion as an orgchain inscription on the seat's sat (lineage/orgchain v1.14 belts), so the chain certifies what the ledger says.
Tests. A brown-in-review seat's Approve on an engine PR merges; a green one's is recorded as a comment and does not; a lapsed black steward runs report-only; with TIME_BELTS off nothing is refused and the would-have-refused lines appear; a granted belt counts like an earned one.
CHANGELOG line under "### Added". Docs: .agents/skills/belts/SKILL.md, linked from AGENTS.md.
DONE: PR open with the code above, python3 .github/time/test_time.py green including the new checks, CI green
Depends on #85
**Why.** Once the dojo grades a seat on an SOP, "may this seat do X unsupervised?" is a test result, not an opinion. Belts stop being decoration and become the permission system; seats domesticate themselves by passing katas, and nobody hand-assigns rights.
**Ledger.** `time/belts.json`, committed, written only by the dojo runner (#85) from a passing ladder: `{seat: {skill: {belt, run, at, lapsed}}}`. Belts are soulbound and monotonic, as in orgchain v1.14: a promotion is a new entry over the old one, never a removal. A failed re-run sets `lapsed: true` — the belt stays on the record, the permission switches off until the seat passes again. Re-runs every `TIME_BELT_REVIEW_DAYS` (default 28) or whenever the SOP changes. The operator may grant a belt by a decision logged in `docs/decisions.md` (the founder's prerogative), and it is marked `granted`.
**Permissions table.** `time/permissions.json`: action → (skill, minimum belt). Starting rules:
- `review.engine`: an Approve on a PR under `.github/time/` counts in `verdict()` only from a seat with ≥ brown in `review`; a lower belt's review is recorded as a COMMENT.
- `review.product`: Approve counts from ≥ green in `review`.
- `plan.queue`: queuing cards (Astra) needs ≥ brown in `management` (the sizing kata).
- `steward.repair`: the steward tick (#80) acts only with black in `management`; otherwise it runs in report-only mode.
- `escalate.tier`: the automatic one-tier retry (#78) targets only seats with ≥ green in `build`.
- Deterministic jobs (housekeeping, Time's own merges, the governor) need no belt: code, not judgment.
**Mechanics.** `belts.py`: `belt(seat, skill)`, `allowed(seat, action)`; `verdict()`, `plan_spec`/Astra, the steward and the retry path call `allowed()`. Enforcement is behind `TIME_BELTS=on` (default off), so the first ladder run (#85) can grant the initial belts before anything is gated; with it off, `allowed()` logs what it would have refused. `cmd_doctor` prints the belt table and which gates are live.
**Audit.** Every kata run commits its transcript next to the result, so a belt can be contested by re-running the kata; graders are deterministic where possible and never the seat under test. Later: mirror each promotion as an orgchain inscription on the seat's sat (`lineage/orgchain` v1.14 belts), so the chain certifies what the ledger says.
**Tests.** A brown-in-review seat's Approve on an engine PR merges; a green one's is recorded as a comment and does not; a lapsed black steward runs report-only; with `TIME_BELTS` off nothing is refused and the would-have-refused lines appear; a granted belt counts like an earned one.
CHANGELOG line under "### Added". Docs: `.agents/skills/belts/SKILL.md`, linked from AGENTS.md.
DONE: PR open with the code above, `python3 .github/time/test_time.py` green including the new checks, CI green
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Depends on #85
Why. Once the dojo grades a seat on an SOP, "may this seat do X unsupervised?" is a test result, not an opinion. Belts stop being decoration and become the permission system; seats domesticate themselves by passing katas, and nobody hand-assigns rights.
Ledger.
time/belts.json, committed, written only by the dojo runner (#85) from a passing ladder:{seat: {skill: {belt, run, at, lapsed}}}. Belts are soulbound and monotonic, as in orgchain v1.14: a promotion is a new entry over the old one, never a removal. A failed re-run setslapsed: true— the belt stays on the record, the permission switches off until the seat passes again. Re-runs everyTIME_BELT_REVIEW_DAYS(default 28) or whenever the SOP changes. The operator may grant a belt by a decision logged indocs/decisions.md(the founder's prerogative), and it is markedgranted.Permissions table.
time/permissions.json: action → (skill, minimum belt). Starting rules:review.engine: an Approve on a PR under.github/time/counts inverdict()only from a seat with ≥ brown inreview; a lower belt's review is recorded as a COMMENT.review.product: Approve counts from ≥ green inreview.plan.queue: queuing cards (Astra) needs ≥ brown inmanagement(the sizing kata).steward.repair: the steward tick (#80) acts only with black inmanagement; otherwise it runs in report-only mode.escalate.tier: the automatic one-tier retry (#78) targets only seats with ≥ green inbuild.Mechanics.
belts.py:belt(seat, skill),allowed(seat, action);verdict(),plan_spec/Astra, the steward and the retry path callallowed(). Enforcement is behindTIME_BELTS=on(default off), so the first ladder run (#85) can grant the initial belts before anything is gated; with it off,allowed()logs what it would have refused.cmd_doctorprints the belt table and which gates are live.Audit. Every kata run commits its transcript next to the result, so a belt can be contested by re-running the kata; graders are deterministic where possible and never the seat under test. Later: mirror each promotion as an orgchain inscription on the seat's sat (
lineage/orgchainv1.14 belts), so the chain certifies what the ledger says.Tests. A brown-in-review seat's Approve on an engine PR merges; a green one's is recorded as a comment and does not; a lapsed black steward runs report-only; with
TIME_BELTSoff nothing is refused and the would-have-refused lines appear; a granted belt counts like an earned one.CHANGELOG line under "### Added". Docs:
.agents/skills/belts/SKILL.md, linked from AGENTS.md.DONE: PR open with the code above,
python3 .github/time/test_time.pygreen including the new checks, CI green