/evo:publish

Canonical release entry point. Owns the full publish lifecycle: pre-flight → bump → changelog → ship → propagate → rollback-on-failure. NOT a synonym for "git push" — see docs/release-protocol.md for vocabulary.

What this skill does

When the user types /evo:publish 18.5.0 (or similar), invoke the go-native release pipeline with the target version. The pipeline owns every step (implementation: go/internal/releasepipeline/, journal under .evolve/release-journal/):

Step	Operation	Failure → action
1	Pre-flight gate (evolve release-preflight)	exit non-zero; abort, no mutations
2	Auto-changelog	abort
3	Version bump (6 markers)	abort
3.5	Rebuild tracked binary go/evolve	abort
4	Consistency check (evolve release-consistency)	abort, files in working tree
5	Atomic ship	abort, nothing pushed
6	Marketplace propagation poll (up to 5 min default)	auto-rollback unless --no-rollback
7	Cache refresh	logged WARN; manual fix

Steps 1–7 are the gh-free pipeline binary. It cannot see GitHub CI — it ends by printing a NOTE: GitHub CI is NOT verified by this pipeline advisory. CI gating is this skill's job (next section); never treat a green pipeline as a green CI.

CI gating (this skill adds what the binary cannot)

evolve release is self-contained and headless-safe, so it does not verify GitHub CI. This skill wraps the pipeline with two gh-based checks. If gh is unavailable, report "cannot verify CI — confirm manually" and stop; never silently skip.

Before invoking the pipeline — base CI must be green (catches "releasing from an already-red main", the v20.1.0 trigger):

gh run list --branch main --limit 1 --json headSha,status,conclusion,url

Require headSha == $(git rev-parse origin/main), status == "completed", conclusion == "success". Anything else → STOP with the run URL:

• in-progress → wait for it.
• failure → fix main green first.
• stale SHA / local main ahead of origin → you'd publish commits CI has never seen; push main and let CI run first, then release. (This is the same gate `/evo:release` runs; it is hoisted here so /evo:publish-direct callers are protected too.)

Also before invoking the pipeline — every LLM CLI must install AND the release binary must answer every core subcommand the installed skills shell out to. This is a deterministic, gh-free local gate (no network, no live LLM): it installs the Claude payload into a throwaway home, renders the Codex/agy projections, checks the Gemini in-repo layout, then smoke-runs each core subcommand on the running binary.

evolve release-verify-clis

Require exit 0 — every CLI row plus the binary:core-subcommands row reports OK. Anything else → STOP; the printed table names the failing target: a CLI whose install/projection broke, or a subcommand the binary no longer answers (the "installed skills silently break" regression). Fix forward before publishing. CI enforces the same matrix via the TestReleaseVerifyCLIMatrix_RealPayload e2e test (go workflow), so a red here predicts a red CI on the release commit.

After the pipeline reports success — the released commit's CI must go green AND the prebuilt binaries must publish. The release (goreleaser) workflow runs on the pushed tag, separately from the gh-free pipeline, so a goreleaser slip ships a binary-less release while evolve release reports success (the v21.1.0 trigger: 0 assets published, only caught by manual check). Watch all three workflows, then confirm the binaries actually landed:

sha=$(git rev-parse origin/main)
for wf in go CI release; do
  rid=$(gh run list --commit "$sha" --workflow "$wf" --json databaseId -q '.[0].databaseId')
  gh run watch "$rid" --exit-status || echo "RED: $wf on $sha"
done
# A green `release` run can still upload nothing on a config slip — confirm the
# prebuilt binaries (evolve_<os>_<arch>.tar.gz) actually landed on the release:
gh release view --json assets -q '.assets[].name' | grep -q '\.tar\.gz$' \
  || echo "MISSING: no prebuilt binaries on the latest release — re-run: gh run rerun <release-run-id>"

• All green + binaries present → done. Report the run URLs + the published asset count (gh release view --json assets -q '.assets | length').
• go/CI red → the release is published but its CI is red. Do not auto-rollback a propagated release — fix forward: land the CI fix on main, then cut the next patch (/evo:publish <x.y.z+1>). Report the failing job + gh run view --log-failed excerpt.
• release red or binaries MISSING → the release shipped without prebuilt binaries (the one-line installer keeps working via build-from-source, just slower). Fix forward: fix .goreleaser.yml/release.yml, then gh run rerun <release-run-id> if the tag is intact, else cut the next patch.

Invocation

/evo:publish 18.5.0                       # full publish, default 5-min poll, auto-rollback on
/evo:publish 18.5.0 --dry-run             # simulate every step, mutate nothing
/evo:publish 18.5.0 --skip-tests          # hot-fix path: skip preflight gate-test execution
/evo:publish 18.5.0 --no-rollback         # post-push failure → exit 3, no auto-revert
/evo:publish 18.5.0 --max-poll-wait-s 600 # tolerate slower marketplace propagation

The slash command translates to:

"$CLAUDE_PROJECT_DIR/go/evolve" release <args>

Optional hardening: --require-preflight runs the full-dry-run harness before any step; --strict-pass rejects WARN preflight verdicts.

When to use this skill

• Always for any version bump (patch, minor, major). The pipeline guarantees consistency, propagation, and rollback.
• Gate readiness first with `/evo:release` (read-only checks: preflight, consistency, CI-green-on-main, no-WIP-commits) — it delegates here when green.

When NOT to use this skill

• Not for non-release commits. If you're committing a feature without bumping the version, use `/evo:commit` (gated attestation → evolve ship --class manual). The release pipeline assumes a version bump and will fail-fast in preflight if <target> equals current.
• Not as a substitute for testing. --skip-tests is for hot-fix scenarios where CI already verified. Routine releases must run the full preflight test suite.

Checking what /evo:publish would do (dry-run)

Operators new to the pipeline: always start with --dry-run to see the proposed changelog block, version bump diff, and lifecycle plan before mutating:

/evo:publish 18.5.0 --dry-run

The pipeline emits each step's proposed output without writing or committing.

Common failure modes

Symptom	Likely cause	Fix
preflight: target X not greater than current Y	Version already bumped, OR you typo'd the arg	Run cat .claude-plugin/plugin.json and confirm; pick a higher target
preflight: most recent audit-report.md does not declare 'Verdict: PASS'	Last audit was WARN/FAIL or stale	Run a fresh audit cycle (evolve loop) or evolve subagent run auditor
marketplace-poll: TIMEOUT after git push	Marketplace checkout didn't pull within deadline	Pipeline auto-rolls-back. Investigate: git -C ~/.claude/plugins/marketplaces/evo log --oneline | head
SELF_SHA_TAMPERED on the next ship	Rebuilt binary pinned but not committed in the release	Known structural residue — see runtime-reference.md binary-rebuild procedure; fix tracked in the release-rebuild-binary-not-committed work package
Hand-curated CHANGELOG entry overwritten	(Won't happen)	Changelog step is idempotent — if ## [<version>] exists it skips

Vocabulary refresher

If unsure what "publish" means in this project: open docs/release-protocol.md. The short version:

• push ≠ publish. git push only moves a remote ref; publish runs the full lifecycle.
• ship is the per-commit primitive (evolve ship --class manual|cycle); publish/release is the version lifecycle on top of it.

Implementation note

This skill is a thin wrapper over the go-native pipeline (evolve release, go/internal/releasepipeline/) plus the gh-based CI gating the binary can't do (the pipeline is intentionally gh-free so it stays headless/cron-safe). All evolve release flag semantics match the underlying command verbatim; the pre/post CI checks above are the skill's own and live only here. (History: through v8.13.x this wrapped legacy/scripts/release-pipeline.sh; the bash pipeline was deleted in the go-only consolidation and this skill now fronts its go-native successor.)