ReleaseKit/docs/roadmap/implementation.md

ReleaseKit staged implementation plan

Status: Proposed. All stages are pending; this document introduces no implemented features or published contracts.

Date: 2026-10-05

Purpose and baseline

This plan gives a Luna coding agent bounded implementation assignments, in execution order, for the target state in the feature roadmap. The roadmap owns product intent and selected release conventions. This document owns implementation steps, dependencies, acceptance evidence, and progress.

The baseline is ReleaseKit commit 6c2695f. It contains documentation, adopted policies, licensing notices, and ADR 0001. It has no Go module, packages, executable, tests, or publication automation. The worktree was clean when this plan was prepared. The available tools were Go 1.26.8 on linux/amd64 and Git 2.53.0; these observations are not minimum support requirements.

The six sibling checkouts provide design evidence; this implementation sequence changes ReleaseKit only. Consumer migrations and measurement of deleted consumer infrastructure belong to later adoption work. Consult code through the knowledge graph, indexing a repository when necessary. Documents, scripts, and CI configuration may be read directly under the discovery instructions.

Scope and completion

Deliver the complete initial library, executable, configuration, shared build and delivery mechanics, Woodpecker examples, and ReleaseKit's own prepared release wiring. Hosted Gitea release pages are part of the selected standard for every project, including source-only projects. Source-only delivery uses the committed release note and zero binary assets; it needs no binary build or checksum manifest.

Proceed through all stages using the local checkout and controlled fixtures. No stage depends on an intermediate release, published tool/runtime pin, consumer pilot, external service action, or human review. Ordinary automated development checks accompany implementation under the testing policy.

After the complete initial implementation, the maintainer will review the architecture, address discovered issues, and then choose the broader testing plan and next steps. Consumer adoption, live integration testing, secret provisioning, CI activation, and actual publication are outside this sequence.

Execution instructions

• Implement stages in integer order. Each stage is one assignment; finish its acceptance checks and local commit before starting the next stage. Record completed work and evidence here. Read the stage's dependencies and relevant decisions before editing.
• Start with the developer guide and applicable policies. Architecture, testing, documentation, and release policies apply to the core implementation. Read the OSS policy when incorporating third-party material.
• Follow the architecture policy and ADR 0001. Add exported declarations as their behavior is implemented, with GoDoc. Implement the selected contract with proportionate internal collaborators.
• The design choices below specify this draft's recommended implementation. Apply the selected scope and decisions throughout the sequence. Later agent-authored code or tests provide evidence of implementation, rather than independent authority to strengthen the contract.
• For each stage, run focused behavioral checks for its changed risks and the established offline default suite. Reuse unchanged evidence where appropriate. Stage acceptance checks are development checks; they add no publication gates beyond the release policy and a repository's selected candidate checks.
• Use temporary repositories, local bare remotes, test programs, and local HTTP servers. Keep fixtures independent of the six consumer checkouts and real services. Temporary fixture module replacements may connect test programs to this checkout; keep them outside committed production module configuration.
• Update canonical documentation with each implemented responsibility. Create consumer, CLI, configuration, integration, and internal documents when they have substantive content. Keep future behavior and stage progress here.
• Keep the implementation self-contained. Build the tool from this checkout; temporary fixture modules may use local replacements. Integration templates accept an operator-selected immutable tool pin for later deployment, without requiring one to complete this plan or claiming that a release already exists.
• Complete stages with local implementation commits and controlled checks. Prepare runnable automation and deployment instructions without activating live workflows. Distinguish implemented integration from verified live use. Finish the whole sequence before the maintainer's architecture review; add no intermediate approval or external-action checkpoint.

Implementation decisions

Module, runtime access, and public ownership

Use one module, gitea.maximumdirect.net/eric/ReleaseKit, preserving the case advertised by the forge's Go import metadata. Set the initial Go language baseline to go 1.25.0, matching the oldest consumer module line. Use APIs available on that baseline. Actual execution uses a prepared local toolchain capable of building the consumer; Narratio already needs Go 1.26. There is no mandatory multi-toolchain release matrix.

The supported packages are the root releasekit facade and buildinfo. Put the executable at cmd/releasekit. Start with standard-library implementations and the installed Git and Go executables. Introduce internal packages as their responsibilities land: runtime information, process execution, release engine, Git operations, builds, Gitea transport, and CLI adaptation. Put command-owned parsing and file-format helpers under cmd/releasekit/internal/cli; keep facade calls in the cmd/releasekit controller. Those helpers return their own settings for explicit conversion by the controller and import neither public facade. The executable therefore consumes the supported APIs while library internal packages retain inward dependencies. There are no public transport clients or assembly APIs.

The initial runtime API is buildinfo.Version() string. It returns an injected release identifier when supplied; otherwise it uses the running executable's main-module version when present and different from (devel), then dev when that metadata is unavailable. This supports source installs as well as explicit release builds. It describes the application version, not the ReleaseKit dependency version. Go build information documents the main-module metadata available to implement this fallback.

Keep the injected string private in the internal runtime implementation. Keep the linker symbol identifier in a small internal contract used by the builder. The builder and runtime implementation change together. Consumers call the accessor; their configuration supplies no linker symbol. The runtime import graph uses the standard library and small runtime implementation, independent of orchestration. A tool/runtime version mismatch must be exposed by the native version smoke check; do not introduce an exact dependency-version equality gate.

Orchestration API

Implement these root operations as the corresponding stages land:

Operation	Request and result responsibilities
Check(ctx, ReleaseRequest)	Capture and validate a local committed candidate; return its commit and version. This diagnostic operation does not query publication services, create a tag, or claim remote readiness.
Publish(ctx, ReleaseRequest)	Validate once, check authoritative state, create and publish the selected annotated tag, and return publication state and retained recovery facts.
Build(ctx, BuildRequest)	Build the selected targets with the requested version, verify a native version invocation, and return source identity and generated artifacts.
Deliver(ctx, DeliveryRequest)	Verify tagged source and supplied artifact identity, create or find a hosted release, compare content, and add missing assets with explicit partial or uncertain outcomes.

Each operation returns a facade-owned result and an error. Requests carry the repository directory explicitly. ReleaseRequest also carries the display product name, version, selected remote, optional validation callback, optional build specification, and optional output writer. The remote defaults to origin; publication uses main. The callback receives a facade-owned Candidate containing directory, captured commit, and selected version. A nil callback selects the default Go test invocation. A supplied callback replaces that invocation with the repository's complete selected validation; it is called once. The CLI adapts configured commands into this callback.

A build specification carries main package, binary name, and targets. Each facade-owned Target contains Go OS and architecture strings. A build request adds version and output directory. A delivery request identifies version, product, repository directory, Gitea destination, and optional build result. The Gitea destination carries base URL, owner, repository name, and caller-supplied token. Library code obtains no credentials from ambient configuration. Before the first remote write, an inspection failure reports that this invocation has not attempted mutation. Reserve uncertain-mutation classification and exit 3 for writes whose authoritative outcome remains unresolved.

Publication results identify version, candidate commit, captured endpoint, phase, and tag object when created. Use distinct states for not_attempted, not_published, published, conflict, and unknown. These describe this invocation; a pre-existing remote release is not described as unpublished. Delivery results identify release URL and the matching, uploaded, pending, and uncertain assets, with not_attempted, complete, partial, conflict, or unknown state. Ordinary errors preserve useful causes through wrapping. Document exactly when nonzero results accompany errors. Add only the named types needed to represent these responsibilities; field spelling and internal package layout can follow idiomatic Go without changing the decisions here.

CLI and configuration

Use standard-library flag parsing and JSON configuration. The default profile is <repository>/releasekit.json; --config selects another file. Configuration paths are resolved from the selected repository, whose default is the current directory. A profile supplies product identity, remote, candidate commands, optional build specification, and optional Gitea delivery declaration. The command supplies the version rather than persisting it in the profile.

Use this eventual profile shape; add each optional responsibility only when its stage implements it:

{
  "product": "Scriptorium",
  "remote": "origin",
  "checks": [{"args": ["sh", "scripts/check-release-candidate.sh"]}],
  "build": {
    "main_package": "./cmd/scriptorium",
    "binary_name": "scriptorium",
    "targets": ["linux/amd64", "linux/arm64"]
  },
  "delivery": {
    "base_url": "https://gitea.maximumdirect.net",
    "owner": "eric",
    "repository": "scriptorium",
    "token_env": "GITEA_RELEASE_TOKEN"
  }
}

This is a planned example, not an existing file or working invocation. Command entries also accept an env map. Run argument vectors directly, without implicit shell parsing. A repository that needs shell syntax selects sh or bash explicitly. Run commands in the captured repository directory and supply RELEASEKIT_VERSION and RELEASEKIT_CANDIDATE for candidate-specific checks. Those captured values take precedence over conflicting command environment entries. Inherit the remaining environment and overlay explicitly owned values without mutating the caller process.

An absent or empty check list selects go test -count=1 ./...; a nonempty list is the repository's complete validation sequence. Decode profile fields strictly so unsupported build or delivery settings cannot silently disappear from the selected checks. Validate required values for the operation being run; source-only checks and publication need neither a main package nor credentials. Hosted page delivery requires a destination and token for source-only profiles as well as binary-producing profiles. The CLI loads the token named by delivery.token_env, defaulting to RELEASEKIT_GITEA_TOKEN, only for delivery. Build receipt identity covers resolved build inputs; delivery credentials, destination, and unrelated check settings are excluded from that identity.

The eventual command forms are:

releasekit version
releasekit check [--repo DIR] [--config FILE] VERSION
releasekit publish [--repo DIR] [--config FILE] VERSION
releasekit build [--repo DIR] [--config FILE] [--output DIR] VERSION
releasekit deliver [--repo DIR] [--config FILE] [--assets DIR] VERSION

Flags precede the positional version. version prints releasekit <version>. Use exit 0 for completed success, 1 for a known failure or conflict, 2 for usage or invalid configuration, and 3 for an uncertain remote mutation. Library outcomes retain more detail. Send validation progress and diagnostics to stderr; stdout presents the command's final result. Signals cancel the operation through its context. There is no interactive confirmation loop.

Candidate checks and Git publication

Parse stable versions using the roadmap's exact syntax. Validate decimal components as strings without introducing a machine-integer size limit. Read the release note from the captured commit at docs/releases/<version>.md. Its first line must be a Markdown H1 containing the configured product name and exact version as complete title elements. Preserve the note content for the annotated tag. The shared tool imposes no additional note section inventory.

Check checks a clean worktree and index on main, captures HEAD, checks the committed note and unused local version, runs validation and selected builds, and rechecks local mutable state. Publish adds authoritative endpoint and remote branch/tag checks before validation and immediately before publication. Cleanliness includes untracked files; ordinary Git ignore rules still apply. Snapshot resolved request values and CLI profile content for the invocation so configuration is not reloaded into different checks midway through the run. Source changes remain subject to the ordinary candidate rechecks. The Go API accepts caller-supplied values without requiring a particular profile file.

Resolve the selected remote's effective push destination using Git's own remote configuration semantics. A missing destination or more than one distinct effective publication destination is an actionable configuration error because this operation needs one authoritative endpoint. A separate fetch URL is supported. Inspect the publication endpoint, rather than assuming the selected remote's normal fetch URL is authoritative.

Use a transient, command-scoped named remote with explicit read and push URLs and mirror mode disabled. Verify that both effective URLs equal the captured publication destination after Git URL rewriting. Reuse that scoped remote for queries and push; persist no new remote configuration. Retain normal Git credential and transport settings, carrying remote-specific transport settings into the scoped remote as needed. If rewriting prevents establishing this equality, report the unresolved destination before mutation. Recheck the original remote resolution before pushing. Test ordinary insteadOf, pushInsteadOf, and explicit push-URL cases using local remotes. Git's remote documentation describes the effective URL lookup on which this relies.

Use exact ref names in remote queries. Require authoritative main to equal the captured candidate and the selected tag to be absent. Create the annotated tag from the committed note with signing disabled for this invocation, inspect its object type and peeled commit, then freeze its object ID. Push <tag-object-id>:refs/tags/<version> with automatic following of tags and mirror defaults disabled. Publish no branch updates, unrelated tags, or force updates.

A successful push returns immediately as published. After a failed or ambiguous push, make one authoritative tag query: an exact object match confirms success; absence confirms not published at observation; a different object is conflict; an unavailable or uninterpretable answer is unknown. Retain the local tag in all failed or uncertain cases. Caller cancellation after push starts also needs this bounded reconciliation attempt using a separate finite recovery context. Use a 30-second recovery deadline as an operational default, not a guarantee that the endpoint will respond. Cancellation before push does not initiate it.

The last remote branch check is an observation before the tag push. This design provides no transaction locking remote main against concurrent writers. Do not add branch pushes or assert stronger concurrency guarantees in tests.

Go execution, builds, and assets

For ReleaseKit-owned candidate Go commands and builds, use prepared dependencies and a local toolchain. Set child environment values GOWORK=off, GOPROXY=off, GONOPROXY=none, GOSUMDB=off, and GOTOOLCHAIN=local. Set module mode to readonly, preserving a repository's deliberately selected testing flags. ReleaseKit-owned builds replace ambient GOFLAGS with -mod=readonly and use their explicit build arguments; custom validation commands retain their deliberately configured flags. Preparation of caches and tools occurs separately. These settings do not sandbox arbitrary repository callbacks or establish a general network firewall.

Build with CGO_ENABLED=0, -mod=readonly, -trimpath, -buildvcs=false, and linker flags -s -w plus the shared version injection. This selects the pure-Go build convention already used by the current asset builders. Additional cgo toolchains remain outside this implementation. The injected string must use a linker-supported initializer. Go's linker documentation defines the -X mechanism.

Initially verify linux/amd64, linux/arm64, and darwin/arm64. Represent targets as values rather than a closed enum; other explicit Go OS/architecture pairs can be passed to the prepared compiler and may fail according to its support. This establishes no native-execution promise for other platforms.

Asset names are <binary-name>-<version>-<goos>-<goarch>, with .exe when the requested Go OS needs it. Binary names and target components must be nonempty filename components without path separators, backslashes, CR, or LF, so names stay in the output directory and have unambiguous manifest entries. Normalize repeated target pairs to one target, retaining their first position. Distinguish the display product name used by notes from this binary name. Preserve selected target order for progress; generate manifests in filename order.

For nonempty targets, invoke the built application with version and require <binary-name> <version> plus a final newline. Reuse a target artifact when it matches GOHOSTOS/GOHOSTARCH; otherwise build a temporary native executable with the same flags and injection. Determine the host independently of ambient GOOS/GOARCH. Native smoke binaries are temporary and are not extra assets. An empty target set performs no Go build or version invocation. Source-only projects may retain an explicitly selected application check; no binary check is silently imposed on them.

Candidate build checks use a temporary output directory outside the worktree and discard it. Artifact builds record the commit, version, main package, targets, compiler identity, and generated filename/hash pairs. Return these facts in BuildResult. The CLI persists them in an internal local receipt, .releasekit-build.json, alongside the generated files after a successful complete build and smoke check. A partial build returns progress without a receipt authorizing delivery. This receipt is for tool coordination, with no public format or attestation promise. It contains no credentials and is not a hosted asset. Build output defaults to dist; create the directory as needed, regenerate owned local filenames, and preserve unrelated files. Report generated files on partial local failure and clean up only tool-owned temporary paths.

Delivery uses a clean checkout whose HEAD equals the selected local tag's peeled commit; detached HEAD is valid. Read the note from that commit and snapshot the selected settings. Repository-local tracked profiles come from that source; explicit caller-supplied settings are identified in the build result. Verify the hosted repository already has the same tag targeting that commit before creating a release. Existing lightweight historical tags may be inspected and delivered; new publication follows the annotated-tag standard. Do not require current remote main to remain at an old release commit during delivery retries.

The CLI's deliver --assets DIR consumes a verified build receipt and checks file hashes, selected filenames, and source/build-input identity before remote mutation. When no directory is supplied, build into a temporary directory from the tagged checkout and then deliver. With zero targets, hosted delivery creates the required note page without a binary manifest or build receipt. An absent build specification is equivalent to zero targets for source-only delivery. Neither mode reruns the candidate test suite or changes source-publication status.

Hosted delivery and Woodpecker

Use a small internal standard-library Gitea REST client. On 2026-10-05 the host's version endpoint reported Gitea 1.25.4; its served API schema confirmed release-by-tag lookup, JSON release creation, asset listing, and multipart attachment upload. Implement the operations described here against local fixtures and record them in the integration guide. Verification against the deployed host belongs to later live testing and is not needed to complete this sequence. This observation is not an all-version Gitea compatibility promise.

Read the tag through /api/v1/repos/{owner}/{repo}/tags/{tag} and verify its commit.sha. Use the repository API prefix for each following operation. Find the release through /releases/tags/{tag}. Create it through POST /releases with the existing tag, captured commit, title from the note's H1, complete committed note as body, and stable, non-draft settings. A found release must match the note body and tag and be stable/non-draft; retain its existing title, including a historical title chosen by other tooling. A content conflict is reported; its body, title, assets, and tag are not edited. Existing historical pages remain intact.

List attachments through /releases/{id}/assets. Identify expected filenames and stream their downloads to compare SHA-256 content; size and name alone do not prove a match. Use multipart field attachment and the filename query parameter for missing uploads. The client obtains no automatic tag-creation responsibility. Keep authorization on the configured service origin when following attachment URLs or redirects; never include token values in receipts or diagnostics.

For nonempty binary sets, generate SHA256SUMS with lowercase hashes, two spaces, filenames, and LF-terminated lines, sorted by filename. It describes the complete declared binary set and excludes itself and the private receipt. An existing manifest must match that expected content. Fill missing assets covered by that manifest only with matching bytes. A manifest describing a different or smaller set cannot be replaced to extend that version. Leave unrelated attachments untouched.

Inspect all known declared asset and manifest conflicts before uploading. Upload missing binaries in filename order, then a missing manifest. HTTP writes are separate operations. After an ambiguous creation or upload, re-read the affected resource once; classify matching content as completed, conflict as conflict, and inaccessible or still-indeterminate content as unknown. Handle concurrent creation by finding and verifying the created resource. Do not blindly retry writes or delete attachments. Record partial progress so an operator can rerun against the same tag and retained assets.

Woodpecker uses tag-triggered repository-local YAML with a pinned Go image and an explicit tool preparation step. Provide source-only and binary-producing examples. Consumer templates accept an operator-selected immutable ReleaseKit source pin for deployment; template preparation and local execution use the implementation checkout and require no published pin. Prepare consumer module caches separately, then run builds and delivery with offline Go execution. This honors ReleaseKit's source-only distribution; a dedicated ReleaseKit container image would require a separate deliverable decision. Use a Go image that satisfies the repository's module requirements.

ReleaseKit's own prepared workflow builds cmd/releasekit from its tagged checkout using the local toolchain and delivers its hosted note page. It needs no previously released executable and produces no distributed binary assets. Prepare this workflow locally; activation belongs to the later next-step plan.

Use CI_COMMIT_TAG and CI_COMMIT_SHA to bind the workflow to tagged source, with an explicit tag fetch when the checkout omits local tag refs. For binary profiles, the build step writes to ignored dist and the delivery step shares the workspace. Source-only workflows call delivery directly without an artifact build step. Supply the Gitea token only to delivery from a Woodpecker secret. Document operator substitution of image digests and source pins for eventual consumer deployment. Pipeline preparation may access dependency services; candidate validation and builds use the prepared caches. CI does not call publish, push refs, or gate source completion. The supported retry is a rerun from the same tagged source with content comparison; reproducible bytes across arbitrary toolchains or elapsed time are not promised.

Stage 1: Establish the Go module and development baseline

Status: Complete. Depends on the baseline only.

1. Add the module with the selected path and language baseline. Add root package documentation for the orchestration facade and establish ordinary Go layout.
2. Keep committed module state independent of sibling checkouts and workspaces. Start without external runtime dependencies or a generic plugin framework.
3. Add the working offline developer test command and initial implemented-package inventory. Update README to its actual implemented boundary.

Acceptance: module loading and go test ./... work with the prepared local toolchain and network disabled. Review module identity and licensing notices. Do not invent tests merely to assert that an empty package exists.

Completion evidence: added the Go 1.25.0 module and documented root facade, updated the README and developer guide to describe the implemented boundary and offline test command, and confirmed module loading and go test ./... with network access disabled. The module has no external dependencies; existing LGPL licensing and attribution notices remain consistent.

Stage 2: Implement the runtime version facade

Status: Complete. Depends on Stage 1.

1. Implement buildinfo.Version and its internal injected-string and main-module fallback logic. Keep runtime state private and the public accessor small.
2. Add the internal injection-symbol contract without importing orchestration into the runtime path. Test injected precedence and missing/development metadata through stable internal seams and a small executable fixture.
3. Add cmd/releasekit with a working version command using this accessor. Record its implemented command contract and the runtime consumer guidance.

Acceptance: a source-built fixture reports dev; a fixture linked with a stable version reports that version. Fallback metadata uses the executable main module. CLI version output is correct and no runtime orchestration initialization occurs.

Completion evidence: implemented buildinfo.Version, its private injected string and main-module fallback, and the internal linker-symbol contract. Unit tests cover injection precedence and missing, development, and stable main module metadata. Built and ran a separate consumer executable from source and with an injected stable version using network-disabled Go commands; both reported the selected value. CLI tests cover its version line, and the CLI's dependency closure contains no orchestration package. Added runtime consumer guidance and documented the command contract.

Stage 3: Protect public contracts and dependency direction

Status: Complete. Depends on Stage 2.

1. Add a compiled-export boundary checker for both facades using go list export data and go/types. Inspect exported signatures, fields, callbacks, aliases, generic constraints, and pointer/value method sets recursively.
2. Assert facade ownership of public named types. Permit deliberately selected standard-library contracts. Protect the private-state forwarding approach against leaked internal types and promoted implementation methods.
3. Add import-direction checks: internal production packages use internal contracts; the runtime facade's dependency closure excludes orchestration, process, build, Gitea, and CLI implementation.
4. Use small negative fixtures proving that the checker catches a leaked type and promoted internal method, and accepts a private implementation field.

Acceptance: actual exports pass and meaningful negative fixtures fail. Tests inspect contracts and imports, with no fixed inventory of future internal package names. Extend these checks as subsequent exports land.

Completion evidence: added internal/apicheck, which loads compiled exports from go list -deps -export and inspects signatures, fields, aliases, generic constraints, and pointer and value method sets through go/types. The checks pass for both public facades and enforce internal import direction and the runtime dependency boundary. Negative fixtures detect leaked types, aliases, callbacks, generic constraints, and promoted implementation methods; a private implementation field and selected standard-library contracts pass. Added the checker command and internal package inventory to the canonical developer docs.

Stage 4: Implement profile loading and command adaptation

Status: Complete. Depends on Stage 3.

1. Implement JSON product, remote, and command settings under an internal CLI adapter. Add strict field decoding, directory resolution, required-value diagnostics, direct argument execution, and per-child environment overlays.
2. Implement an internal cancellable process runner with directory, argument, environment, stdout, stderr, and exit-error support. Use os/exec directly; expose no public runner abstraction solely for test assembly.
3. Add CLI parsing infrastructure and selected usage/exit behavior. Advertise commands only when their following stages implement them.
4. Add small profile examples and docs/config.md for implemented fields.

Acceptance: paths and arguments containing spaces remain separate arguments; selected environment values reach the child without changing process globals; missing executables, invalid configuration, and cancellation report useful errors. Tests use controlled local child commands and temporary directories.

Completion evidence: added strict profile decoding, repository-relative and absolute profile resolution, required product and command diagnostics, and direct check argument adaptation. Added a cancellable os/exec runner with working directory, inherited environment plus child-only overlays, stdout, stderr, and preserved exit errors. Tests cover spaced paths and arguments, environment isolation, missing executables, exit failures, cancellation, and invalid profiles. Added a profile example and canonical CLI, configuration, and internal package documentation.

Stage 5: Capture and validate one local candidate

Status: Complete. Depends on Stage 4.

1. Implement stable-version syntax, clean main and index/worktree inspection, captured HEAD, profile snapshot/note loading, title identity, and unused local tag checks. Preserve immutable candidate facts in internal values.
2. Implement default offline Go validation and callback adaptation. Bind the callback to the captured candidate; pass captured version/commit to CLI commands. Call the complete validation callback once, stopping on failure.
3. Recheck branch, HEAD, and cleanliness after validation. Retain the captured settings throughout the operation. Implement the public Check operation and local diagnostic check command.
4. Document that local check success provides local candidate evidence and makes no authoritative remote availability claim.

Acceptance: temporary repositories cover invalid versions, missing or mismatched committed notes, dirty tracked/untracked state, existing local tags, detached HEAD, callback failure, and a callback changing HEAD or the worktree. Count a successful callback once. No test needs a publication service, and Check creates no tag or remote ref.

Evidence: internal/candidate validates stable versions and committed note identity, captures a clean main commit, runs either the offline default or the configured callback once, and rechecks local mutable state. The public releasekit.Check API and releasekit check command provide the local diagnostic. Temporary Git repositories cover the acceptance cases; offline go test -count=1 ./... and go run ./internal/apicheck pass.

Stage 6: Resolve and inspect the authoritative endpoint

Status: Complete. Depends on Stage 5.

1. Implement effective push-URL capture, command-scoped remote configuration, matching read/write resolution, and endpoint rechecks described above.
2. Implement exact remote main/tag queries with distinct present, absent, conflicting, and unreadable results. Query failures never establish absence.
3. Compose the pre-validation and pre-publication remote guards around captured local state. Require main equality, rather than ancestor-only publication.

Acceptance: local remotes cover a separate fetch endpoint, explicit push URL, ordinary URL rewrites, multiple push destinations, missing main, conflicting tags, main ahead/behind the candidate, endpoint changes, and remote failure. Queries leave local tags, persistent remote configuration, and remote refs intact.

Evidence: internal/git captures effective push destinations through Git, rejects multiple distinct destinations, verifies scoped URL rewriting, and queries exact main and version-tag refs. A candidate check requires main to equal the captured commit and the version tag to be absent. Local bare remote tests cover separate fetch/push URLs, both rewrite modes, missing main, an existing tag, main ahead and behind, endpoint changes, unreadable remote state, and unchanged persistent config and refs.

Stage 7: Create and push the exact annotated tag

Status: Complete. Depends on Stage 6.

1. Implement public publication request/result values and explicit conversion into the engine. Validate once between the complete preflight guards.
2. Create an annotated tag from captured note bytes, inspect type and peeled commit, capture its object ID, and push only that object to the selected ref.
3. Recheck mutable local state and endpoint before the push, including the local tag still identifying the captured object. A changed local tag prevents this invocation from proceeding; cleanup remains deliberate recovery.
4. On successful push, return published state without another remote read-back. On push error, retain the tag and return unknown until Stage 8 adds resolution.

Acceptance: a local bare remote receives the expected tag object and commit; its main and unrelated refs do not change. Follow-tags and mirror configuration cannot expand publication. Note content, tag identity, and one-validation behavior are checked through the public facade. Restrict stage trials to test remotes while failure handling is incomplete.

Evidence: the public releasekit.Publish API captures the committed note, validates once between exact remote guards, creates an unsigned annotated tag, rechecks local and remote state, and pushes only the captured tag object. Bare remote tests verify the note bytes, peeled commit, tag object, unchanged branch and unrelated refs, disabled follow-tags and mirror expansion, validation count, remote conflicts, and retained local tags after push error; Stage 8 adds outcome classification through remote reconciliation.

Stage 8: Resolve rejected and ambiguous publication

Status: Complete. Depends on Stage 7.

1. Add the one-query reconciliation path and public error/result classification. Preserve the original push cause when resolution does not confirm success.
2. Give post-push cancellation a separate bounded recovery context. Retain local tag/object facts when the read fails or recovery is cancelled/times out.
3. Add real rejected-push tests using bare-remote hooks. Simulate lost acknowledgments at the subprocess boundary after a real successful push. Separate known absence, conflicting object, and unavailable-query cases.
4. Cover a remote branch change during validation and local tag mutation before push. Keep concurrency promises at the observed guard boundaries.

Acceptance: exact remote object match succeeds despite a reported push error; missing, conflicting, and unavailable remote tags have distinct outcomes. Failed pushes issue one reconciliation query, successful pushes issue none, and no recovery force-update or automatic tag deletion occurs.

Evidence: failed tag pushes receive one bounded exact-tag query. Matching object identity returns published; absence returns not_published with the original push error; a different object returns conflict; query failure returns unknown while preserving the push cause. Tests cover rejection by a bare remote's receive hook, an installed conflicting tag, unavailable reconciliation, and a lost acknowledgment after the bare remote accepted the tag while the caller context was canceled. Query counts confirm one recovery read after errors and none after a successful push. Tests also verify that a remote branch change during validation and a changed local tag before push prevent publication.

Stage 9: Expose publication and document ReleaseKit bootstrap

Status: Pending. Depends on Stage 8.

1. Wire the complete public publisher into releasekit publish, with selected progress output, final result, exit classification, and signal cancellation.
2. Add ReleaseKit's source-only profile and one short maintainer entry script. The profile selects the offline default suite; no inherited race, lint, documentation, image, or platform gates are added.
3. Create docs/release.md now that publication works. Name authoritative origin at the ReleaseKit forge URL and main; document preparation, successful-push completion, exact-object recovery, local tag retention, and deliberate cleanup of proven unpublished tags.
4. Document first-release execution from the captured candidate source using the local Go toolchain, before a released ReleaseKit executable exists. Include immutable release-note licensing footer construction and dependency preparation separate from validation.
5. Update current development documentation and the policy-adoption roadmap's bootstrap status to the implemented boundary. Preserve pending deliverables.

Acceptance: an end-to-end CLI fixture publishes to a local remote through a committed profile and returns the expected exit/output behavior. Runbook commands are exercised using a temporary repository configured like ReleaseKit. This stage prepares self-publication and does not publish a real ReleaseKit release.

Stage 10: Build selected targets with shared version injection

Status: Pending. Depends on Stage 9.

1. Implement public build values and internal build execution using the fixed flags, version symbol, filename convention, and prepared Go environment.
2. Build explicit OS/architecture pairs and verify the application version invocation using a host artifact or temporary native build. Preserve caller-owned output files and report partial generated output on failure.
3. Return source identity, artifact paths, filenames, targets, hashes, and compiler identity. Keep byte reproducibility and cross-platform execution promises outside the result contract.
4. Add a temporary consumer module importing only buildinfo and exposing the canonical version command. Its local replacement is fixture-only.

Acceptance: the fixture compiles for the three initial targets and a native execution reports the injected version. An ambient cross-target environment does not select the smoke host. Empty targets perform no build or smoke test; missing main packages, build failures, incorrect version output, unsafe binary names, and cancellation fail with useful partial results.

Stage 11: Integrate candidate builds and artifact command

Status: Pending. Depends on Stage 10.

1. Add build profile fields, CLI build, and the private artifact receipt. Bind receipts to source/build-input identity and generated hashes. Validate data on reload; leave the receipt outside the public API and hosted asset list.
2. Integrate selected builds into Check and Publish after the single candidate callback. Use disposable output outside the worktree for candidate builds. Recheck mutable state after the entire validation/build sequence.
3. Add ignored dist output and source-only and binary-producing examples. Document the API, profile target selection, version invocation, and build output ownership.
4. Refresh facade and runtime dependency checks with the new values.

Acceptance: source-only publication invokes no builder; selected build failure prevents tag creation; selected builds inject the exact candidate version; callback validation runs once. CLI artifacts and receipts agree on filenames, hashes, commit, and resolved build settings. Generated validation output does not dirty the candidate. Continue directly to hosted delivery implementation; no publication or consumer pilot is needed to complete this stage.

Stage 12: Find and create Gitea release pages

Status: Pending. Depends on Stage 11.

1. Implement the small internal REST transport using the served API operations, explicit destination/token inputs, context cancellation, and structured failure mapping. Use standard HTTP and JSON packages.
2. Verify the hosted tag targets the captured source commit. Find a release by tag, create a missing stable release from the committed note, and verify existing note body/tag and stable, non-draft state before accepting it. Preserve the existing title.
3. Handle create conflicts and lost acknowledgments by one lookup and content comparison. Never edit the existing page to make it match.
4. Add docs/integrations/gitea.md describing implemented API interactions, credential supply, source binding, and failure semantics.

Acceptance: httptest.Server covers found, missing, created, conflicting, unauthorized, malformed, cancelled, and ambiguous-create results. A missing or wrong hosted tag prevents page creation. Existing historical titles are retained when the body, tag, and release state match. Fixture assertions cover actual request paths, payloads, and authentication without using real credentials.

Stage 13: Inspect asset and manifest content

Status: Pending. Depends on Stage 12.

1. Generate the expected binary manifest from the declared build result. Add attachment discovery and streaming hash comparison of existing assets.
2. Compare existing SHA256SUMS with the complete declared manifest. Identify matching, missing, conflicting, and unverifiable content separately.
3. Inspect all declared conflicts before any upload and preserve unrelated attachments. Keep download authorization confined to the configured origin.

Acceptance: equal name/size with different bytes is conflict; matching content is accepted; a full matching manifest can cover missing binaries; a subset or different manifest is conflict; unreadable content is not treated as missing. Source-only delivery generates no binary manifest. No write endpoint is used by inspection tests.

Stage 14: Upload missing assets and report partial delivery

Status: Pending. Depends on Stage 13.

1. Add multipart uploads for missing binaries in filename order, followed by a missing manifest. Record completed work and pending asset identities.
2. Resolve ambiguous uploads by one resource read and content comparison. Handle concurrent matching uploads without creating replacement content; detect duplicate/conflicting attachment names as conflicts.
3. Implement rerun behavior that accepts existing matching files, adds missing ones, and stops on conflicts. Preserve every existing page and attachment.
4. Add tests where the first binary succeeds, a later write fails, and a second invocation completes only the missing assets; also cover a lost manifest acknowledgment and failure of the reconciliation read.

Acceptance: partial and unknown outcomes retain exact completed/pending facts; retries do not overwrite or delete content; no blind repeat POST is made after an ambiguous acknowledgment. Published manifest constraints hold across reruns.

Stage 15: Expose tagged-source delivery through the facade and CLI

Status: Pending. Depends on Stage 14.

1. Implement Deliver, its facade-owned request/result values, and conversions. Bind source note, resolved build settings, tag commit, and hosted repository as specified.
2. Add delivery profile fields, environment-name token loading in the CLI, and deliver with an optional verified assets directory. With no directory, build declared tagged-source artifacts in temporary storage before uploading. An absent build specification or empty target set selects page-only delivery: skip the builder, native smoke, receipt requirement, and manifest generation.
3. Accept detached tagged checkouts and historical tags without requiring old remote main equality. Verify receipt identity/hashes before supplied assets are used. Keep the library credential input explicit.
4. Add current consumer/CLI guidance for source status, delivery status, retained assets, and retry commands. Extend public boundary checks.

Acceptance: facade/CLI fixtures cover both page-only delivery and binary delivery from tagged source to a local server, with content-aware reruns. Page-only delivery preserves the complete note, needs no main package or receipt, and performs no Go build or asset upload. Wrong commit, stale build settings or receipt, tampered assets, and missing credentials prevent relevant writes. Candidate tests and Git publication are not run by delivery. A delivery failure leaves the existing source tag intact and reports only delivery failure.

Stage 16: Provide the common Woodpecker integration

Status: Pending. Depends on Stage 15.

1. Add maintained source-only and binary-producing tag-pipeline templates. Parameterize the Go image and immutable ReleaseKit source pin for eventual deployment, with explicit installation/cache preparation and a delivery secret. Document required template inputs rather than inventing a published pin. Create no new ReleaseKit image-distribution requirement.
2. Bind tag and commit variables to the checkout, fetch the selected tag when necessary, share generated assets between steps, and keep output ignored.
3. Document preparation versus offline execution, asynchronous status, and rerunning the same immutable tag. The source-only template delivers the note page directly; the binary template first builds its declared assets. Neither workflow executes the publisher.
4. Exercise both command sequences locally with a tool built from this checkout, fixture repositories, and a local server. Substitute the local executable for the templates' installation step during development checks. Check YAML using locally available tooling or static inspection; no live runner or new external validator is needed to complete the stage.

Acceptance: the maintained commands reach common delivery for both profile types and common builds for declared binary targets, with the selected version and commit. Secrets enter only the delivery step. Agent-performed static YAML checks and local command evidence are recorded accurately. No published tool pin, container pull, configured secret, or live runner is a stage prerequisite. CI success is not a source release gate. The Woodpecker environment reference owns the runner-provided tag and commit variables.

Stage 17: Prepare ReleaseKit's own hosted-page delivery

Status: Pending. Depends on Stage 16.

1. Extend ReleaseKit's source-only profile with its Gitea destination and token environment name. Keep the build target set empty and retain its selected offline candidate suite. No distributed executable or binary manifest is introduced.
2. Prepare a repository-specific Woodpecker workflow under the maintained examples, using the tagged checkout to build the local release tool and call page-only delivery. Keep it outside active CI discovery during this sequence. Document where to place it for later activation and how to supply its secret.
3. Update the release policy's deliverable declaration and docs/release.md together to identify the hosted note page, automation responsibility, and page-delivery retry command. Source completion remains the exact tag push.
4. Exercise the prepared command sequence against a temporary tagged repository and local HTTP server. Retain the note's complete body, including licensing footer, on the hosted page.

Acceptance: ReleaseKit's source-only example creates a hosted page and accepts an identical rerun without building consumer binaries, generating a manifest, requiring a receipt, or consulting a previously released tool. Delivery failure leaves source status unchanged. Workflow preparation requires no real tag, credentials, remote push, CI activation, or maintainer checkpoint.

Stage 18: Complete the initial implementation and handoff documentation

Status: Pending. Depends on Stage 17.

1. Map each initial capability in the feature roadmap to its implemented owner and existing behavioral evidence. Record the concrete package and dependency inventory, both supported facades, and their implemented responsibilities in the canonical internal and consumer documentation.
2. Finish current API, CLI, configuration, integration, example, development, and release documentation. Check links and consistency with actual behavior; keep licensing and attribution accurate for incorporated material.
3. Run the established offline default suite and relevant existing boundary checks for the assembled implementation. Resolve implementation defects within this scope without adding a new permanent release matrix or live testing requirement.
4. Record the completed initial implementation and evidence in this plan. Identify prepared integrations whose live operation remains unverified and hand off the implementation for the maintainer's subsequent architecture review. Keep consumer migrations and their anticipated code savings marked as future adoption work.

Acceptance: the initial library and executable provide check, publication, build, runtime version access, page-only and binary delivery, partial/uncertain recovery, and prepared Woodpecker integration. The documentation describes the implemented boundary accurately. Completion depends on local implementation and development evidence; architectural review, a broader testing plan, consumer migrations, and release/deployment decisions follow this sequence.

Coverage of the feature roadmap

Initial target responsibility	Implementing stages
Two public facades and small runtime dependencies	1–3, extended in 10–11 and 15
Stable version and committed note identity	5
Candidate capture, one validation, mutable guards	5–8, 11
Authoritative endpoint and remote ref state	6–8
Annotated object publication and failure recovery	7–9
ReleaseKit source-only bootstrap/runbook	9, 17
Shared version injection and selected cross-builds	2, 10–11
Native smoke, empty target sets, consistent names	10–11
Hosted page/note creation, including source-only projects	12, 15, 17
Full checksums, comparison, partial/additive retries	13–15
Prepared Woodpecker workflows for both delivery profiles	16–17
Complete initial implementation and canonical documentation	18

Consumer infrastructure reduction and convergence on the shared standard remain feature-roadmap adoption goals. Their migration sequence, real-world testing, and measured savings will be determined after the initial architecture review.

Open Questions

None block this initial implementation sequence. The maintainer has selected hosted pages for source-only projects and deferred pilots, releases, consumer adoption, and broader testing decisions until after implementation and the subsequent architecture review. Actual release pins and deployment credentials are inputs to later adoption; no stage requires them.