Finalize release tooling implementation plan

This commit is contained in:
2026-11-03 15:36:23 +00:00
parent 3d396b1a66
commit cb40f1d9f8
2 changed files with 374 additions and 166 deletions

View File

@@ -1,10 +1,11 @@
# Release tooling simplification
Status: implementation roadmap; no implementation has started. Local page-only
Gitea delivery and protection of the supplied token path are selected for
implementation. The recommendations in Open Decisions remain unselected.
Current release behavior stays with its canonical owners until implementation
updates them.
Status: target design finalized for the release-tooling sprint; implementation
has not started. All four generic file checks, registered-tool execution,
page-only hosted delivery, and protection of the supplied token path are selected.
The [implementation plan](implementation.md) owns stage order, tasks, validation,
and completion evidence. Current release behavior stays with its canonical owners
until implementation updates them.
## Purpose and authority
@@ -21,44 +22,6 @@ The [agent policy](../policy/agents.md),
govern implementation. The [release procedure](../release.md) continues to own
the current workflow until implementation updates it.
## Current baseline
Baseline: CommandKit commit `a74fcb20041cbccde93f9b9a1e3b24cb8dd131e7`.
- The root [module](../../go.mod) selects
`go.maximumdirect.net/releasekit v0.8.0` as a Go tool. The release toolchain
remains Go 1.26.8, and the consumer minimum remains Go 1.26.0.
- [Core](../../releasekit.json),
[ckconfig](../../integrations/ckconfig/releasekit.json), and
[ckprompt](../../integrations/ckprompt/releasekit.json) have separate profiles.
They select their physical module directories, the current SSH endpoint, and
an explicit repository-root shared-file list. They enable no optional file
checks or additional product checks, and configure no hosted delivery.
- The 49-line [launcher](../../scripts/release) maps the three selectors, changes
to the repository root, builds the root-selected tool into ignored
`.releasekit/bin/` storage for the host, and delegates publication to ReleaseKit.
- [Launcher tests](../../scripts/release_test.go) occupy 506 lines, including
custom command recording, fake build/publisher processes, and tests of the
assertion helpers themselves.
- The 399-line [publication exercise](../../scripts/release_integration_test.go)
creates disposable Git repositories, release notes, and suite markers. It
checks tag structure, exact remote objects, suite counts, nested tag prefixes,
and endpoint rejection in addition to exercising CommandKit's launcher.
- CommandKit's API, CLI, integration, and maintained consumer behavior is tested
in the owning package suites. There is no separate CommandKit-specific
candidate-validation implementation to replace or preserve.
After the baseline review, the maintainer supplied the local credential file
`.releasekit/gitea-token`. Its contents were not read. The file was confirmed
ignored, untracked, and mode `0600`; authentication was not tested. The existing
directory ignore rule now serves credential storage as well as executable
storage and remains needed after executable-storage removal.
During review, a temporary core profile enabled all four proposed file checks
below. ReleaseKit's local `check` accepted the committed core candidate with
Go 1.26.8. This was diagnostic evidence only: no profiles changed, no publication
occurred, and equivalent integration-module checks remain outstanding.
## Targeted end state
### Ownership
@@ -79,7 +42,7 @@ deliverables and platform support commitments remain unchanged.
### Profiles
Subject to OD-1, select these existing ReleaseKit options in each profile:
Select these existing ReleaseKit options in each profile:
| Selection | Intended protection |
| --- | --- |
@@ -88,7 +51,7 @@ Subject to OD-1, select these existing ReleaseKit options in each profile:
| `validation.module_distribution.workspace_files: true` | Detect committed development workspace files in selected inputs. |
| `validation.module_distribution.local_replacements: true` | Detect filesystem replacements in selected committed module declarations. |
Separately from OD-1, select
Also select
`validation.forbidden_files: [".releasekit/gitea-token"]` in all three profiles.
This delegates protection of the supplied credential path to ReleaseKit's
existing check. The path is repository-relative and the check applies across
@@ -111,27 +74,25 @@ separately for demonstrated needs.
### Launcher and recovery
Subject to OD-2, use the root module's registered `go tool releasekit` through
`exec`, keeping the current selector interface and publication argument mapping.
Use the root module's registered `go tool releasekit` through `exec`, keeping
the current selector interface and publication argument mapping.
Retain host normalization, workspace-off execution, the local toolchain, and
read-only module metadata. Preserve other effective Go settings rather than
silently replacing them while arranging read-only selection. Go 1.26.8's
`go tool` does not accept `-mod=readonly` directly; use its supported Go settings
mechanism and verify the intended behavior.
read-only module metadata. Preserve other effective Go settings while arranging
read-only selection.
The implementation plan specifies the wrapper interface and invocation details.
Remove explicit executable storage when the launcher no longer creates it.
Retain `/.releasekit/` in `.gitignore` for the credential file, and update its
comment to describe that purpose. Update recovery instructions to use the
registered tool or the entry point printed for the saved attempt. Preserve
ReleaseKit's distinction between publication outcome and record-storage errors.
Go's cache supplies executable storage; the launcher creates no persistent
binary directory. `/.releasekit/` remains ignored for the credential file, with
an explanatory comment. Recovery instructions use the registered tool or the
entry point printed for the saved attempt, preserving ReleaseKit's distinction
between publication outcome and record-storage errors.
Share module selection and host-safe tool invocation between publication,
delivery, and recovery where applicable. Delivery selects the same module
profile as publication; recovery uses its recorded attempt rather than inferring
identity from a current profile. Keep the existing publication interface and
provide an explicit delivery invocation without automatically chaining the two
operations. Choose a small wrapper arrangement during implementation rather
than adding a coordination framework or duplicating selector tables.
operations. Keep one selector mapping and shared invocation plumbing.
A pending upstream request concerns host-safe consumer launcher examples and
regression coverage. It is not a dependency of this work: v0.8.0 already supports
@@ -204,115 +165,34 @@ Exercise module routing separately from exit statuses rather than multiplying
every combination. Assert behavior and meaningful selections, not helper
implementation details or exact usage prose.
Remove the disposable publication exercise and its Git fixtures, synthetic
notes, markers, tag inspection, and environment-isolation helpers when no longer
used. Remove assertion-helper self-tests and the custom build-call log/parser
with their obsolete consumers. Trust upstream publication and delivery coverage;
do not move a duplicate suite into a different local package or export a new
test-helper API. There is no line-count target for the replacement tests.
The resulting local suite omits the disposable publication exercise, generic
tag inspection, assertion-helper self-tests, and custom build-call log/parser.
Upstream publication and delivery coverage owns those generic behaviors. No
replacement publication engine or exported test-helper API is needed. There is
no line-count target for the replacement tests.
## Work from baseline to target
## Selected choices and rationale
### 1. Configure validation and hosted delivery
All four generic file checks are selected as release-time checks, in addition
to the agreed forbidden token path. Formatting and local-link validation automate
existing contributor review work. Workspace-file and filesystem-replacement
checks address accidental inclusion of development metadata that a successful
local test run can overlook. This adds specific release checks; it does not
make every optional ReleaseKit validator mandatory or prohibit development
workspaces and replacements outside the selected committed candidate inputs.
Selecting only formatting and Markdown would change less release behavior, but
would leave these distribution mistakes to manual review. The selected upstream
checks avoid a second local implementation.
- [ ] Resolve OD-1 and record the chosen scope here.
- [ ] Update the three profiles for the selected options while preserving their
module identities, endpoint, and shared-file selections.
- [ ] Add the agreed delivery destination and forbidden token path to each
profile. Preserve the ignored token file and the directory ignore rule.
- [ ] Run the selected checks against each module's committed candidate inputs.
Review findings against existing contracts; fix relevant defects without
introducing broad exclusions or unrelated remediation by default.
- [ ] Update the release policy's delivery scope and the release procedure's
selected-check and hosted-delivery descriptions with the profile changes.
Host-normalized registered-tool execution is selected because the root module
already registers ReleaseKit and Go caches the executable. Retaining a persistent
native executable would provide a recovery entry point independent of a later
Go-driver invocation, but would retain executable-storage and bootstrap test
maintenance. Recovery instead uses the registered tool or ReleaseKit's reported
entry point. The credential directory remains ignored under either approach.
### 2. Simplify the launcher and its tests
- [ ] Resolve OD-2 and implement the selected invocation path.
- [ ] Add or adapt focused wiring tests before deleting the coverage they replace.
- [ ] Remove `scripts/release_integration_test.go`, assertion-helper self-tests,
and unused build/publisher scaffolding after reviewing their coverage ownership.
- [ ] If registered-tool execution is selected, remove executable-storage setup
while retaining `/.releasekit/` for credentials and updating its comment.
Do not delete user-local files as part of the refactor.
- [ ] Expose explicit module-aware delivery using shared invocation plumbing.
Keep publication and delivery independently invocable and retryable.
- [ ] Update recovery commands, delivery prerequisites and retry instructions,
and token environment/file precedence. Remove obsolete opt-in publication
exercise instructions from the release and development guides with the
corresponding behavior change.
### 3. Validate and finish
- [ ] Validate with Go 1.26.8 and prepared dependencies. Exercise the real
registered tool's host execution and read-only behavior once where the local
wiring tests use a controlled process; use nonpublishing operations.
- [ ] Run the owning module suites and selected ReleaseKit checks against the
final candidates. ReleaseKit `check` already runs the selected module's suite;
avoid repeating the same successful suite solely for another checklist item.
- [ ] Distinguish workspace behavior evidence from each module's independent
dependency graph. For disposable committed-candidate checks, use matching
notes and existing module metadata; do not invent future dependency versions
or checksums to make validation pass.
- [ ] Review the documentation, local links, and final diff. Confirm that
domain-specific package tests remain and that removed helpers have no users.
Check future-note guidance for module-specific full tags and immutable forge
links; do not rewrite published notes to adopt the new workflow.
- [ ] Verify delivery profile routing and operation forwarding with controlled
processes. Rely on upstream tests for Gitea authentication, HTTP writes, and
content reconciliation rather than adding a duplicate local delivery engine
suite. No real token or hosted writes are needed for these migration checks.
- [ ] Record validation results and completed work here. Retire the completed
roadmap through the usual documentation lifecycle when appropriate.
These are migration tasks, not additional permanent release gates. Verification
needs no production tag push or hosted delivery. Ordinary releases continue to
use one publication command, with no required preliminary `check` run, followed
by a separate hosted-delivery operation. Confirmed source publication remains
the source-release completion point; hosted-delivery failure is reported and
handled separately.
## Open Decisions
### OD-1: Which optional checks become release gates?
**Recommendation:** Select all four options in the profiles table. Formatting
and local link checks automate existing contributor review work. The two module
distribution checks address accidental inclusion of development-only metadata
that a successful local test run can overlook. Reusing upstream mechanisms keeps
this protection declarative and avoids maintaining local checker code.
This is a deliberate addition to release-time validation, not merely deletion
of duplicate code. The observed core pass does not establish every module's
compatibility with the selected input scope. Resolve integration findings and
the applicability of vendored or fixture inputs before treating the selection
as adopted. Record the chosen release scope in the release procedure.
The specific forbidden token path and page-only delivery are already selected
for implementation and are not part of this open decision.
**Viable alternative:** Select formatting and Markdown now and defer the module
distribution checks. This automates familiar review work with less change to
accepted release contents, and preserves flexibility for legitimate committed
module fixtures. It leaves accidental workspace files or filesystem replacements
to ordinary review. Choose this alternative if the broader distribution checks
would reject justified inputs or their added protection does not warrant the
restriction. Do not compensate by writing equivalent local validators.
### OD-2: Registered tool or persistent native executable?
**Recommendation:** Use host-normalized `go tool releasekit`. The tool is already
registered in the root module, the Go cache supplies the executable, and the
launcher can forward its exit status without managing `.releasekit/bin/`.
This also removes executable-storage test machinery and documentation, while
the directory ignore rule remains for the token. Preserve read-only metadata
selection and test the effective environment before replacing the current
bootstrap.
**Viable alternative:** Retain the current native build and persistent executable
while simplifying profiles and tests. It is already proven to handle inherited
cross-build targets, accepts an explicit build-time read-only flag, and provides
a durable recovery entry point independent of a later Go-driver invocation.
It retains storage, bootstrap, and recovery-documentation maintenance, so it
removes less local infrastructure. Prefer it only if verification identifies a
meaningful recovery or tool-resolution need that the registered-tool path does
not meet; the pending upstream example change alone is not such a blocker.
The implementation plan records the existing integration dependency-version
mismatch separately from validation of the current source. This infrastructure
sprint does not publish releases or select new library dependency versions.
Historical hosted-page backfilling and CI delivery remain separate future
choices, not prerequisites for the selected local page-only workflow.