ConfigKit/docs/roadmap/implementation.md

# Release Workflow Implementation Plan

Status: implemented on 2026-09-28. All stages are complete.

Implement the [release roadmap](release.md) in the numbered order below. Each
stage is a bounded work unit for a gpt-5.6-terra coding agent; finish its tests
and current documentation before committing it. The completed policy-alignment
record formerly at this path is preserved at
`faf5ff8:docs/roadmap/implementation.md`; use `git show` to read it.

## Scope and settled decisions

Start from completed policy alignment at `faf5ff8`. Follow the
[development guide](../development.md) and applicable adopted policies. Preserve
library behavior, dependencies, the everyday checker, historical release notes,
licensing, and support commitments. This plan authorizes no real release or push.

Use release policy 1.4 and the adoption guide at upstream commit
`d84a38ebab26f048cafbe59a2246700c35d29843`, available from `../policy` or the
existing pinned upstream links. Do not add an upstream-version review. Use shared
wording verbatim unless a substantive ConfigKit decision warrants adaptation.
Preserve exact source metadata; update the local date when revising policy.

The interface is `./scripts/release.sh vMAJOR.MINOR.PATCH`, with the existing
`CONFIGKIT_BASELINE_TOOLCHAIN` prerequisite. Resolve the repository root from the
script location. Retain `scripts/release-check.sh` as its validator and an optional
diagnostic command, never a required prior run. No skip-validation option,
persistent cache, routine confirmation, or automatic commit is needed. Stop using
`.git/configkit-release-candidate`; ignore any old record without deleting it.

Use sequential explicit main-then-tag pushes, not atomic-push negotiation. Keep
Bash 3.2 and existing Git/Go prerequisites; no new release framework is needed.

## Stage 1 — Simplify the release checker and select its gates

**Completed 2026-09-28.** The checker now runs one readonly offline race suite
with test-time vet disabled, one full vet pass, and CGO-disabled Linux amd64/arm64
builds. The disposable module audit and release formatting/whitespace scans were
removed. Focused checker tests, Bash syntax checks, and the real Go 1.25.14
release checker passed on Linux.

Change `scripts/release-check.sh`, checker tests in `release_check_test.go`, and
the affected descriptions in `docs/policy/release.md`, `docs/release.md`, and
`docs/development.md` together. The existing guarded publication blocks remain
operational until Stage 2; do not describe the new command as available yet.

Use this exact gate set with the selected maintained Go 1.25 baseline:

| Gate | Command and rationale |
| --- | --- |
| Full offline race suite | One `go test -mod=readonly -vet=off -race -v ./...`; protects library behavior, ownership, and concurrent use. Includes examples, consumer fixtures, documentation links, and boundary tests. |
| Static analysis | One `go vet -mod=readonly ./...`; full vet provides checks beyond the subset normally run by tests. Disabling test-time vet avoids duplicating that subset. |
| Supported target compilation | `go build -mod=readonly ./...` for Linux amd64 and arm64 with `CGO_ENABLED=0`; these configurations are not covered by native race compilation. |

Keep `GOPROXY=off`, baseline availability/cache diagnostics, fail-fast behavior,
and verbose skip reporting. Dependency/toolchain acquisition is explicit
preparation; there is no automatic upgrade or latest-version survey. Readonly
module commands prevent incidental manifest edits, not impose a new tidy policy.

Remove the disposable-copy module-tidiness audit, release formatting/whitespace
scans, and their unused machinery. Formatting/whitespace remain ordinary
development checks in `scripts/check.sh`; module tidiness is reviewed when
dependencies change. Offline tests/builds establish that the candidate compiles
and works with its declared dependencies. Remove policy wording that makes the
removed checks mandatory release gates. Shell syntax checks belong to development
validation, not a separate publication gate. Do not change the everyday checker.

Split the shared checker regression expectations where their contracts differ.
Retain unusual-filename and whitespace tests for the everyday checker. Controlled
Go subprocesses should verify the release toolchain/offline settings, one suite,
one vet pass, the two target builds, and stopping after failure. Do not recursively
execute the real full suite in a checker-orchestration test.

Acceptance: focused `TestCheckScripts` tests and Bash syntax checks pass. Run the
real revised checker with prepared baseline dependencies on Linux. Do not also
run the everyday suite just to repeat its checks. macOS and newer-toolchain
validation remain optional.

## Stage 2 — Implement single-command publication and recovery

**Completed 2026-09-28.** Added the executable single-command publisher, exact
candidate and destination guards, one checker invocation, explicit main-then-tag
pushes, note-derived annotated tags, and bounded failed-push state checks.
Disposable Git fixtures covered success, refusals, branch-only progress,
conflicting tags, and unknown outcomes without contacting the real remote.
Bash syntax checks, focused tests, the offline default suite, local links, and
whitespace review passed.

Add executable `scripts/release.sh`, calling the existing validator once rather
than duplicating it. Replace `TestReleaseRunbookGuards` with tests of the actual
script: prose headings cease to be an executable interface. Update the release
policy and runbook in this stage to describe implemented behavior and recovery.

### Candidate and publication flow

1. Accept exactly one stable version argument using the existing no-leading-zero
   Semantic Version format. Require `main` and a clean index/worktree including
   nonignored untracked files. Ignored local files alone do not block release.
   Verify the existing authoritative `origin` endpoint and ensure pushes actually
   target it, accounting for separately configured push URLs and multiple push
   destinations. Do not publish to an unnoticed additional destination or print
   credentials in diagnostics.
2. Capture the full candidate commit ID. Read `docs/releases/<version>.md` from
   that commit and require the first line `# ConfigKit <version>`. Meaningful
   summary and applicable compatibility/upgrade content are ordinary preparation,
   not automated prose judgments or four mandatory headings. Retain the existing
   footer and local note-review requirements without a new network check.
3. Require the version unused locally and on the authoritative remote. Query the
   exact remote tag, distinguishing absence from query failure. Read remote main
   and check that publication can advance it without discarding history; main
   already at the candidate is acceptable. No all-remotes or all-tags survey.
4. Run `scripts/release-check.sh` once. Failure stops before any push/tag creation.
   Recheck branch, cleanliness, candidate identity, and unused version state after
   validation. Changed candidates require a fresh invocation after correction and
   commit; do not automatically commit or repair the checkout.
5. Push the captured commit explicitly to `refs/heads/main`, without force or
   incidental follow-tags publication. Override broad configured push defaults
   through explicit destinations/refspecs. A successful push establishes branch
   publication without a separate read-back gate. Recheck local candidate and
   clean state before tagging; ordinary concurrent local edits must not leak
   into publication.
6. Create an annotated tag at the captured commit, using the captured checked-in
   note as its message (`git tag -a -F`). Verify its peeled target and capture its
   tag-object ID. Push only `refs/tags/<version>:refs/tags/<version>`, without
   force or implicit publication of unrelated tags.
7. On successful tag push, report the version/candidate and return success
   immediately. No CI polling, remote read-back, module retrieval, hosted release
   creation, or routine human sign-off gates completion.

Use straightforward private helpers. These are ordinary release guards, not a
guarantee against hostile local processes or concurrent administrators. Do not
add locks, snapshots, signing requirements, or unrelated Git configuration rules.

### Failure outcomes

After a failed or ambiguous push, perform one bounded authoritative-state check;
do not poll or blindly retry:

| State observed | Behavior |
| --- | --- |
| Main push failed, but remote main equals candidate | Continue through remaining guards and tagging. |
| Main push failed, remote differs or cannot be queried | Stop before tagging; distinguish rejection from unknown outcome and explain recovery. |
| Tag push failed, but remote tag-object ID equals the created annotated tag | Report confirmed publication success; main publication was already established. |
| Tag push failed, remote tag confirmed absent | Report branch-only progress and the unpublished local tag; retain it. |
| Remote tag conflicts | Stop; never replace published contents. |
| Remote tag cannot be queried | Report unknown status; retain local state, do not claim success or safe version reuse. |

Document exact-ref recovery commands that distinguish absence from query failure.
Only after confirming a local tag was never published or used by another channel
may a maintainer delete that unpublished tag and rerun with the same version.
A fresh invocation revalidates; this is failure recovery, not a mandatory double
successful-path run. A correctly published tag needs no rerun. Published source
defects need a new version; transient transport failures do not. Never roll back
main or automatically delete tags to recover.

### Regression coverage and acceptance

Use disposable real Git working repositories and bare remotes. Reuse the existing
fixture-only authoritative-URL substitution or Git URL rewriting, with local-file
transports and isolated identity/signing configuration. No production test-only
bypass flags. Stub the expensive validator with a loggable result; use real
rejection hooks and narrow Git-process wrappers for lost responses/unavailable
queries. Tests must never contact the actual remote or recurse into the suite.

Cover success with main initially behind the candidate and already equal;
single validation; note-derived annotated tag/commit identity; ignored files;
and unrelated local tags remaining unpublished. Cover malformed version, wrong
branch/destination, dirty/changed candidate, missing/mismatched note, local/remote
tag collision, failed validation, non-fast-forward main, and the meaningful
failure outcomes above. A note without empty boilerplate headings must pass.
Assert resulting refs and exit status, not helper structure or exact diagnostics.
Verify completion has no CI/consumer-tool dependency.

Acceptance: Bash syntax checks, focused checker/publication tests, and the offline
default suite pass. The script is tested only against fixtures. Current docs must
no longer require a manually recorded candidate before single-command publication.

## Stage 3 — Complete documentation alignment and integrated validation

**Completed 2026-09-28.** Reconciled the adopted release policy with its pinned
upstream source and ConfigKit decisions; aligned the runbook, development guide,
and roadmap records with the implemented command. Bash syntax, focused release
tests, local documentation links, whitespace review, and the final real baseline
release checker passed on Linux. No publication or push was performed.

Reconcile `docs/policy/release.md` with the pinned upstream release policy. Adopt
shared paragraphs verbatim except for substantive local decisions. Remove the
temporary deferral, all-remotes freshness rule, and stale prospective v0.3.0 note
instruction. Preserve known-published-version nonreuse, ConfigKit compatibility,
source-only deliverables, annotated tags, selected gate rationale, Linux support,
best-effort macOS, no-CI completion, and historical exceptions. No compensating
controls are required merely to use successful-tag-push completion.

Keep base version 1.4 and its existing pinned source; update the local date.
Existing attribution and license scope already cover these edits. Do not change
other policies, copyright/license grants, or historical release notes unnecessarily.

Review `docs/release.md` as the operational owner: preparation, one publication
command, completion, partial/unknown-state recovery, and distinctly optional
diagnostics. Retain exact version-specific licensing footer URLs and optional
consumer retrieval; neither adds a network gate. Remove duplicate manual publish
instructions. Update `docs/development.md` navigation, gate summary, and test map.
Search active documentation/scripts for stale candidate-file, section-template,
module-tidiness gate, CI-wait, and manual publication instructions. Fix affected
links without rewriting historical records.

Correct the completed policy roadmap's implementation-record reference to
`faf5ff8:docs/roadmap/implementation.md`, since this path now owns the new plan.
Mark release-roadmap implementation status accurately after acceptance. Preserve
unfinished/deferred work; retirement belongs to actual release preparation.

Run Bash syntax checks, focused orchestration tests, documentation-link tests,
and `git diff --check`. Run the revised real release checker once on the final
implementation with prepared baseline dependencies. Reuse that successful suite
result instead of separately repeating examples, boundaries, vet, builds, or the
everyday suite. Report missing prerequisites rather than claiming success. Never
invoke publication against origin as an acceptance test.

Acceptance: scripts, policy, and runbook agree; tests protect selected behavior
without extra contracts; current docs describe only implemented behavior. Commit
each completed unit with a plain-English message and report validation/workspace
state without publishing. No product decisions remain open in this plan.