stable-release

name	stable-release
description	Promote an already-shipped rolling release (vX.Y.Z) of the DOS kernel to a named stable channel — gated on a green kernel suite + a green third-party CI run on the candidate + a clean truth syscall + a soak window. Writes an evidence file and adds a stable/<codename> git tag on the same commit. Does NOT bump versions or build new artifacts.
disable-model-invocation	false
user-invocable	true
allowed-tools	Read, Edit, Grep, Glob, Bash, Write
argument-hint	<codename> [--from vX.Y.Z] [--window-days N] [--force-promote]
output_root	none

Stable Release — a named channel on top of rolling /release (DOS)

/release ships a rolling tag (vX.Y.Z) whenever DOS lands user-visible change. That tag promises "this commit merged" — it does not promise "the substrate was provably trustworthy here." This skill adds a second, less-frequent channel that does make that promise.

A stable release is a promotion of an already-shipped commit, not a new build. No new version is minted. No new wheel or zip is produced. The gate is read at promote time, frozen into an evidence file, and a second annotated tag (stable/<codename>) is pinned on the same commit.

This is the DOS adaptation of job's stable-release skill. job gates on apply-loop hero metrics (silent-failure share, PSV verified-success rate, funnel-stage regressions) and KEEP-slot baselines from execution-state.yaml + baselines.yaml. None of that exists in DOS — DOS is a domain-free trust substrate, not an apply pipeline; it has no runtime funnel, no hero metric, no baseline ledger. So the gate is re-grounded on the only "known-good" signals DOS actually has: a green kernel suite, a clean truth syscall, and a soak window. See "The gate" below.

Trunk is master. The stable tag is pinned on a commit reachable from master; the push targets master's remote.

The kernel-vs-tooling boundary. The gate context script (scripts/stable_release_context.py) and this skill are dev/release tooling that operates ON the package — never imported BY dos.*. The "kernel imports no host, no scripts/" litmus in CLAUDE.md stays intact.

Relationship to other systems

System	Owns	Stable-release reads	Stable-release writes
/release (scripts/release_bump.py)	rolling vX.Y.Z tag, version markers	the latest vX.Y.Z as promotion candidate	nothing — never re-bumps
the kernel suite (pytest)	substrate correctness	exit code (green == eligible)	nothing
the truth syscall (dos verify)	ground-truth adjudication	exit + verdict (well-formed == eligible)	nothing
git tags	the channel itself	stable/* collision check + previous-stable	one new stable/<codename> tag

The spec deliberately introduces no new state machine, database table, or pipeline event. It reuses git tags + the suite + the syscall.

When to invoke

The user (or a scheduled cron) invokes this skill. Typical triggers:

• After a /release that landed a meaningful kernel change and you want a rollback anchor that says "the substrate was provably good here."
• Quarterly hygiene — pick the highest vX.Y.Z that satisfies the gate.
• After a regression rollback, to mark the recovered state.

Cadence — the two-channel contract. Rolling vX.Y.Z tags ship continuously: /release is deliberately cheap, and a coherent unit that clears the mechanical gates (suite, leak scan, CI, the publish ci-green witness) just ships — several per week is healthy. The stable channel is the deliberate counterweight: an infrequent, gate-driven promotion that answers "which version do I pin in production." There is still no fixed calendar — the gate either passes or it doesn't — but the working rhythm is: promote the highest soaked-green tag (a) after a milestone (a new syscall family, a major/ABI bump, the first PyPI-published era), or (b) when a meaningful run of rolling tags has accumulated since the last stable — every few weeks, not every few days. Rolling moves fast because stable exists to absorb the trust question; stable can afford its soak window because rolling carries the urgency.

Codename convention

Operator picks one short word at promote time, prefixed by year-month: stable/2026-06-aardvark. Memorable, sortable, and rollback becomes meaningful (git checkout stable/2026-06-aardvark says "the substrate was good here" in a way git checkout v0.3.0 never claimed).

Validate: ^\d{4}-\d{2}-[a-z][a-z0-9-]{2,20}$ (the context script enforces this). Reject ambiguous reuses by checking git tag -l 'stable/*-<codename>'.

---

Step 1 — Pre-digest the gate context

One helper collapses every gate read into one JSON payload:

python scripts/stable_release_context.py --codename <codename>
#   --from vX.Y.Z       pin the candidate (default: latest semver tag)
#   --window-days N     min soak age for the candidate tag (default 3)
#   --skip-pytest       smoke / dry inspection only (do NOT use for a real promote)
#   --force-promote     report blockers but mark all_green-eligible (records rationale)

The script returns:

{
  "candidate_tag": "v0.3.0",
  "candidate_sha": "abc123…",
  "codename": "2026-06-aardvark",
  "window_days": 3,
  "previous_stable": {"tag": "stable/2026-05-zephyr", "sha": "…"},
  "tag_collision": false,
  "idempotency": {
    "tag_exists": false, "tag_sha": null, "tag_matches_candidate": null,
    "evidence_file_exists": false, "evidence_path": "docs/stable-releases/2026-06-aardvark.md"
  },
  "gate": {
    "pytest_suite_green": {"name": "...", "pass": true, "exit_code": 0, "summary_tail": "..."},
    "ci_green":           {"name": "...", "pass": true, "verdict": "GREEN", "advisory": false, "reason": "..."},
    "dos_verify_clean":   {"name": "...", "pass": true, "exit_code": 1, "verdict": {...}},
    "tag_age":            {"name": "...", "pass": true, "age_days": 5.2, "window_days": 3}
  },
  "summary": {"all_green": true, "blockers": [], "forced": false}
}

If summary.all_green is false, the skill stops and reports the failing rows — exactly which gate row blocked promotion and its reading. This is the entire point of the skill; never paper over a red gate.

dos_verify_clean exits 1 on a healthy no-plan repo — that is a PASS. The truth syscall's exit code carries the ship verdict (0 shipped / 1 not), not execution health. The gate probes with a sentinel (plan, phase) that never shipped, so a healthy syscall returns shipped=false, source="none" and exit 1; the script treats a well-formed verdict dict + exit ∈ {0,1} as the pass. A crash (non-{0,1} exit, traceback, no JSON) is the only failure. Don't "fix" this by demanding exit 0.

Step 2 — Decide the promotion candidate

By default, promote the most recent vX.Y.Z tag that has soaked ≥ window_days (so the gate had time to observe it). If --from vX.Y.Z is passed, use that.

Refuse to promote (the context script flags these as blockers):

• A tag already promoted (a stable/* tag points at the same SHA → tag_collision).
• A tag younger than window_days unless --force-promote is set with a written rationale in the evidence file.

If candidate_tag is null, there is no vX.Y.Z tag yet — DOS has not cut a rolling release. Stop and tell the operator to run /release first; a stable promotion has nothing to promote.

Advisory check (PyPI live since 2026-06-10): did the candidate actually publish? The stable promise includes "pin this in production," and the pin is pip install dos-kernel==<underlying> — which only resolves if the candidate tag's publish.yml run was approved and uploaded (the tag push triggers the pipeline, but the upload waits on the ci-green witness + the pypi-environment approval, so a tag can exist unpublished). Check pip index versions dos-kernel (or the publish run on the tag). If unpublished: prefer approving the pending publish first; promoting anyway is allowed (the git tag still anchors the rollback) but the evidence file must note the pin does not resolve. Hand-checked for now — not a script gate row yet.

Step 2.5 — Idempotent re-run (a partial-failure replay is SAFE)

A promotion writes two durable side-effects in sequence — the stable/<codename> tag (Step 5) and the docs/stable-releases/<codename>.md evidence file (Step 4 + its commit in Step 6). If a prior run was interrupted between them (operator Ctrl-C, a git push that failed), re-running with the same codename must be safe — detect what already exists and skip it, never error or write a duplicate. This is the idempotency property the userland app bakes into its promote orchestrator; DOS reads it from the context script's idempotency block and the skill body acts on it (DOS keeps the promotion script-light, so this is data the skill obeys, not a wrapper).

Read idempotency from Step 1 and branch:

• tag_matches_candidate: true (the stable/<codename> tag already exists and points at the candidate SHA) → the tag step already succeeded. Skip Step 5 (do not re-tag — a stable tag is never re-pointed); record it on an idempotent_skips list for the final summary. This is the clean replay case, NOT a tag_collision blocker (the context script already excludes a same-codename tag from the collision check).
• evidence_file_exists: true → the evidence file is already written. Re-read it; if its gate snapshot matches today's, skip re-writing Step 4 and its Step 6 commit (note the skip). If it's stale/partial, finish writing it and commit — the file is the artifact, completing it is the point.
• tag_exists: true AND tag_matches_candidate: false → the codename already tags a different commit (the word was reused). This is a hard blocker the context script already raised — stop; a stable tag is immutable, so pick a fresh codename and add a ## Supersedes note in the new evidence file.
• All false → a normal first run; proceed through Steps 4–6 as written.

The load-bearing property: every side-effect detects its own already-done state, so the skill can be re-run after any partial failure and converges to one tag + one evidence file. Report every skip under idempotent_skips in the final summary so the operator sees what the replay reused versus wrote.

Step 3 — The gate (the load-bearing step)

The gate is four rows, all from the context script. Each is a real check, not a remembered number:

Row	Source	Pass condition
pytest_suite_green	python -m pytest -q	exit 0 — the kernel suite is green on the promoting machine. The substrate's entire value is deterministic, correct verdicts; a red suite means it can't anchor a stable channel.
ci_green	the CI/Checks oracle (dos.drivers.ci_status) on the candidate commit	not RED — the third-party CI record, not a local self-report. GREEN passes; RED blocks (a real, accountable failure, e.g. a Windows/py3.11 matrix leg the promoter's box never ran); PENDING/NO_SIGNAL are advisory passes (the local suite row stays load-bearing). See the asymmetry note below.
dos_verify_clean	dos verify (sentinel probe)	well-formed verdict dict + exit ∈ {0,1} — the truth syscall executes cleanly (see the exit-code note above).
tag_age	committer date of the candidate tag	age ≥ window_days — the candidate soaked, not a just-cut tag promoted in the same breath.

ci_green is asymmetric — it can only ADD a blocker, never remove one. It is the substrate eating its own distrust (docs/93): pytest_suite_green is a self-report from the very machine cutting the promotion, so ci_green reads the verdict the CI system recorded on infrastructure the promoter doesn't control. Only RED blocks (CI proves the build is broken — the row's whole reason to exist, catching matrix-leg failures invisible locally). PENDING (CI not finished) and NO_SIGNAL (gh unwired/unauthed, or the candidate was never pushed) are advisory passes — hard-blocking on them would make the gate unrunnable offline, contradicting the oracle's fail-safe "never fabricate a verdict" discipline. The local pytest row remains the load-bearing suite gate; CI is purely additive signal. Use --skip-ci for an offline smoke. If ci_green is PENDING and you want CI-verified promotion, wait for the run to finish and re-run the gate rather than --force-promote.

This row is itself the "richer known-good history across consumers" the gate was always designed to grow into: as more signals appear (an N-day clean-dos verify soak, a consumer-side loop reporting back), additional rows thread into stable_release_context.py the same way — without touching the kernel. That is the package's own modularity rule: policy (the gate) is data + a driver-shaped script; mechanism (the kernel) stays untouched.

Step 4 — Write the evidence file

Single markdown file at docs/stable-releases/<codename>.md (create the dir on the first promotion). Frontmatter is the gate JSON snapshot; body is human-readable.

---
codename: 2026-06-aardvark
underlying_version: v0.3.0
candidate_sha: abc123…
promoted_at: 2026-06-15
window_days: 3
previous_stable: stable/2026-05-zephyr
gate:
  pytest_suite_green: {pass: true, exit_code: 0}
  ci_green: {pass: true, verdict: GREEN, advisory: false, repo: anthony-chaudhary/dos-kernel}
  dos_verify_clean: {pass: true, exit_code: 1, source: none}
  tag_age: {pass: true, age_days: 5.2, window_days: 3}
---

# Stable promotion — 2026-06-aardvark

Promotes `v0.3.0` (commit `abc123…`).

## What's known-good here

- Kernel suite green (`pytest -q`, exit 0) at this commit.
- Truth syscall executes clean (`dos verify` → well-formed verdict).
- Candidate soaked 5.2 days (≥ 3-day window).
- Pinned install: `pip install dos-kernel==0.3.0` (published on PyPI — or note
  here if the candidate's publish run was never approved).

## Rollback target

`git checkout stable/2026-06-aardvark` returns to this exact commit + a
provably-green substrate.

## Gate evidence (mechanically generated; do not edit)

[full gate table from the context JSON]

This file is the artifact /stable-release exists to produce. Everything else is bookkeeping. Use the date from the operator / the system clock provided in context — do not invent a timestamp.

Step 5 — Tag

git tag -a stable/<codename> <candidate_sha> -m "stable/<codename> — promoted from v0.3.0"
git push origin stable/<codename>

Idempotent skip (Step 2.5): if idempotency.tag_matches_candidate was true, the tag already exists on the right commit — do not run these commands; record idempotent_skips: stable/<codename> tag (already present) and move on. If only the local tag exists but the push didn't land (a rare interrupt), re-running git push origin stable/<codename> is safe and idempotent.

Never delete or re-tag a stable tag. Mistakes get a new codename + a ## Supersedes section in the next evidence file.

Step 6 — Commit the evidence file

The evidence file is a real doc; commit it (it is not the promoted commit — the tag points at the candidate, the evidence lands as a fresh commit on master):

Stage and commit by explicit pathspec (-- <path>) — master here is a concurrently-written tree, and a bare git commit would sweep another session's staged content under this message (the same hot-tree discipline /release Step 5 follows):

git add docs/stable-releases/<codename>.md
git commit -m "docs/stable: promote v0.3.0 to stable/<codename>" -- docs/stable-releases/<codename>.md
git push origin master

Idempotent skip (Step 2.5): if idempotency.evidence_file_exists was true and the file's gate snapshot already matches today's, the evidence commit likely already landed — check git log for the docs/stable: promote … <codename> commit; if present, skip this step and record idempotent_skips: evidence file + commit (already present). Only write + commit if the file is missing or stale.

Step 7 — Final summary

Print:

• Codename + underlying tag + commit SHA (short)
• Gate result table (one row per check, reading vs pass/fail)
• Evidence file path
• Stable tag + push status
• idempotent_skips (if any) — each side-effect a re-run reused rather than re-created (the tag, the evidence file, the evidence commit), so the operator sees this was a replay, not a fresh promotion.
• Previous stable, for the rollback hint

---

Supporting script

scripts/stable_release_context.py produces the Step 1 JSON. It runs the suite, probes the truth syscall, computes tag age, checks for a stable/* collision, and reports the idempotency block (Step 2.5 — whether this codename's tag + evidence file already exist) so a partial-failure re-run is safe. It is dev/release tooling — never imported by dos.*. There is no heavier orchestrator (job's stable_release_promote.py / agents.sr.*): DOS's gate is small enough that the skill body drives the tag + evidence-file steps directly, and keeping it script-light avoids adding surface the kernel ethos discourages. The idempotency is therefore data the skill obeys, not a wrapper that enforces it — the same mechanism/policy split the kernel keeps everywhere.

Anti-patterns (don't do these)

• Don't bump versions. This promotes an existing tag. If you're editing pyproject.toml or src/dos/__init__.py, you're in the wrong skill (use /release).
• Don't build new artifacts. No new wheel, no new zip. The git tag IS the artifact in DOS.
• Don't --skip-pytest for a real promotion. That flag is for dry inspection only; the green suite is the load-bearing gate row.
• Don't --skip-ci to dodge a RED CI run. The flag is for offline smokes. If ci_green is RED, the build is broken on a leg your machine didn't run — fix it, don't skip it. (PENDING/NO_SIGNAL already don't block; you only reach for --skip-ci to silence them, which is fine for a dry run and wrong for a real one.)
• Don't override a red gate without writing why. --force-promote is allowed, but the evidence file must contain a ## Force-promote rationale section explaining which row failed, why it was overridden, and who approved it.
• Don't promote two stable tags in one window without a --force-promote rationale.
• Don't fold the rolling channel into the stable channel. Two tags, two audiences: rolling = "latest code," stable = "last provably-good substrate."

What this skill consciously punts

• A stable install pointer. job flips stable_current.txt via install_versioned.py; DOS has no versioned-install snapshot at all (it's a pip package), so there is no pointer to flip. pip install dos-kernel==<underlying> pins the promoted version directly (live on PyPI since 2026-06-10) — note the dist name is dos-kernel, not the bare dos (an unrelated PyPI package).
• GitHub Releases entry for stable tags. The evidence file + annotated tag are the canonical artifacts. Revisit if operators ask.
• Auto-promotion via cron. Easy to add later (/loop + this skill), but the first few promotions should be operator-driven so the gate's calibration gets validated by hand.
• Richer gate rows (multi-day soak, consumer feedback). Threaded into the context script when DOS has the history to support them — not before.