The release bot runs every day at 02:00 Asia/Shanghai (18:00 UTC). GitHub schedules can be delayed. If the default branch's source tree is unchanged from the highest successfully published numeric version, it skips the release. Empty commits do not cause a release.
When changed, the bot chooses a version, builds and checks the packages, creates
the tag and publishes the GitHub Release as github-actions[bot]. There is no
manual version bump, tag push or draft publication step. The first release uses
tool.inferrsi.release.initial_version in pyproject.toml (initially 0.1.0);
subsequent releases allocate the next unused numeric patch version. Major/minor
releases need a separately reviewed change to that initial-version floor.
The existing Developer Preview classification is preserved:
tool.inferrsi.release.prerelease = true marks automatic releases as GitHub
prereleases and does not mark them Latest. Changing to a stable channel is a
separately reviewed configuration decision. Published previews count as
successful releases for no-change detection; private drafts do not.
The bot's only repository mutations are creating a new refs/tags/<version>
at an existing reviewed commit and creating/uploading/publishing that version's
Release. It must not create code commits, edit repository files, create/update/
delete code branches, create/approve/merge PRs or enable auto-merge. Configuration
development and review are contributor work; they are not bot capabilities.
Plan/build/notes jobs have Contents read access. Only the publisher receives Contents write; Pull requests permission is explicitly none. The publisher validates its tag/SHA inputs and never checks out or executes repository code. Regression tests execute its actual shell commands against a recording GitHub mock that rejects non-publication writes.
This is an enforced workflow behavior contract, not a tag-only token scope. GitHub's Contents write permission also authorizes file commits and PR merging; setting Pull requests to none does not remove those Contents capabilities. GitHub does not offer a separate tag/Release-only GITHUB_TOKEN permission. Server-side branch restrictions need a separately scoped repository-policy change: the GitHub Actions App identity is shared by other workflows, so this configuration does not silently change their access or existing branch rules.
- Snapshot the default branch and inspect stable tags plus published Releases. Private drafts are not successful-release baselines. Compare exact source trees and reject a published source outside the default branch's history rather than silently release a rollback.
- Create a local tag in a read-only build job.
setuptools-scmderives both distribution and runtime versions from that tag. The build pins SCM's package-specific version override to the validated plan, avoiding ambiguous Git version discovery when failed reservations share a source commit. Tags point directly at the reviewed source commit; the bot does not write version commits tomain. - Run CPU tests/lint, build the sdist and a wheel from the sdist, check metadata
with
twine, and run the complete installed-package checker in a clean wheel environment outside the checkout. Check tag/distribution/runtime parity and generateSHA256SUMS.MANIFEST.inlimits sdist contents to package sources and build metadata; repository evidence/workflows are not release payload. Reusetools/build_codex_plugin.pyandtools/build_workswarm_plugin.pyto buildinferrsi-codex-plugin-vX.Y.Z.zipandinferrsi-workswarm-plugin-vX.Y.Z.zipfrom the same source revision. Each archive's generated host metadata, bundled runtime and tag must have the same version. Check its native entrypoints, verify every bundled runtime hash and exercise the extracted launcher/knowledge reads outside the checkout with isolated Python. These CPU checks do not run WorkSwarm's SDK or UI. - After a successful build, an optional read-only notes job calls the AI API with bounded change metadata for exactly the planned revision range. Its Markdown summary is transferred as a separate artifact. Missing credentials, API/validation errors and unavailable notes artifacts fall back to the existing GitHub-generated notes.
- A separate write job verifies the artifacts, source ancestry and release
state again. It creates the remote tag only after validation, creates an
internal draft, uploads wheel/sdist/both plugin ZIPs/checksums and automatically publishes
after confirming the attachments and tag identity. It does not execute
repository code. These operations stay in the same workflow: tag events
created with
GITHUB_TOKENdo not trigger another workflow.
All source-tree changes, including documentation, are eligible. The
notes configuration groups existing enhancement,
bug and documentation labels and includes other merged PRs. Notes use the
previous published tag rather than incomplete reserved tags. A release describes
software changes; CPU/package checks do not establish campaign PASS, model
quality or service performance. Such claims retain the acceptance and evidence
requirements in contributing.
To enable AI summaries, add the Actions repository secret DS_API_KEY under
Settings → Secrets and variables → Actions. The key is injected only into
the notes step as the helper's internal OPENAI_API_KEY environment variable.
The workflow does not read or fall back to the old OPENAI_API_KEY repository
secret. Do not put the key in workflow source, command arguments, PR text or
logs. API access and usage are billed by the selected provider for that key;
GitHub Copilot access is not required.
The default API model is deepseek-flash, documented by DeepSeek as
DeepSeek-V4.1-Flash-1017
when this configuration was added. This is a provider-managed alias, not a
version pin; DeepSeek may upgrade its target. Set the optional Actions repository variable
OPENAI_RELEASE_NOTES_MODEL to another model available to your API account and
compatible with the Responses API. Set OPENAI_BASE_URL to select the approved
HTTPS provider: https://api.deepseek.com (the fallback),
https://api.openai.com/v1 or https://yibuapi.com/v1. This repository uses
DeepSeek's official API and sends the configured key only to its /responses
endpoint. The existing OPENAI_* setting names also apply to DeepSeek; use a key
issued by the selected provider. The configured API model ID is sent unchanged;
the helper does not substitute a model when it is unavailable. Trailing slashes
are normalized; other hosts,
HTTP, userinfo, query strings and fragments are rejected before sending a key.
Redirects remain disabled. Additional providers require a reviewed allowlist
change. Without the secret, the workflow skips AI generation and continues ordinary
publication.
The notes helper sends bounded commit subjects, source links and file-change statistics from the previous successful tag through the planned commit. It omits author metadata, commit bodies and raw patches. Recognizable email/IP literals, URLs, personal paths and common credential forms in selected metadata are replaced with labeled placeholders before transmission; generated canonical commit links remain available. This targeted redaction supplements the repository's privacy rules for source metadata. Inputs are treated as source data, not executable instructions. The prompt asks for a concise English overview of software features, fixes and compatibility changes, with at most six bullets, 250 words of prose and one supplied commit link per bullet. It must not invent performance/quality or campaign PASS claims. The response is labeled AI-generated summary and prepended to the original GitHub changelog, preserving its PR links and source/evidence notice. When source metadata is truncated, the helper appends a fixed coverage notice from the source flags. The model is asked to leave coverage statements to the helper. Model output is read as a file and never executed as shell code. Explicit URLs and Markdown links in model output must be simple inline links to canonical commit URLs present in the supplied source. Images, reference links, bare URL/email autolinks and unsupported or unverified link forms cause fallback. This gate validates generated link syntax; GitHub's own reference and custom autolink rendering remains repository policy.
The helper makes one request using the
Responses API, with
bounded input/output (at most 3,200 output tokens), a timeout, no tools and
store: false. DeepSeek requests
set reasoning effort to none for the short summary. DeepSeek ignores store
and documents stateless Responses compatibility; this setting does
not establish zero data retention. Missing credentials, failed/incomplete/refused
or unusable responses leave no summary. The optional notes job or artifact may
fail without skipping an otherwise valid publication; failed package checks and
workflow cancellation still prevent publication. The AI job has no repository
write permission, and the publisher receives no API key. There is no AI code
commit, merge, account installation or update to an existing public Release.
Each Release contains a Python wheel, source distribution, the versioned Codex
and WorkSwarm plugin ZIPs and SHA256SUMS covering all four packages. Choose
the archive for your host and follow the
Codex guide or
WorkSwarm guide.
Each ZIP contains one standalone plugin, complete bundled runtime and knowledge,
licenses and an installation README. The hosts use different manifests; their
ZIPs are not interchangeable. Either plugin's build failure or a missing/corrupt
installer blocks publication. Publication does not update a plugin branch,
account installation or marketplace listing. No marketplace catalog or
installation wrapper is packaged.
Merge the configuration into the default branch. GitHub Actions and the listed
actions must be allowed by organization policy, and the publisher needs
contents: write. Ordinary releases need no extra secret, PAT, GitHub App or
repository-wide write default. AI summaries additionally use the optional
DS_API_KEY secret described above. The workflow guards the canonical
repository and default branch; copies must deliberately update that repository
guard.
Under Actions → Release bot → Run workflow, select the default branch and
leave dry_run checked. It plans and builds the next release without creating
a remote tag or Release; unchanged trees skip here too. With DS_API_KEY
configured, a changed-source dry run also makes the AI request and uploads the
summary artifact for inspection, consuming API usage. Clearing dry_run
runs the same automatic publication immediately. It takes no version/tag input.
gh workflow run release.yml --repo InferMatrixLab/InferRSI --ref <default-branch> \
-F dry_run=trueActions artifacts are retained for seven days. Package versions come from Git
tags/build-generated _version.py; plain unbuilt source imports report
0+unknown. Build from a full checkout or the supplied sdist. Git archives use
the exported .git_archival.txt; when multiple tags share a commit, use the
uploaded wheel/sdist or explicitly select the SCM build version. Development
builds without a stable tag may
have a development version and are not stable releases.
Runs are serialized across schedule and manual dispatch. A failed build creates
no remote tag or Release. A later tag/upload/publication failure can leave an
unused tag or private incomplete draft. The bot never moves tags, overwrites
notes/assets or uses --clobber. On a later eligible run it allocates the next
unused patch version, so failures may leave gaps in version numbers. Failed
attempts do not advance the successful-release baseline or suppress retry of
unreleased source. Old drafts remain for an authorized maintainer to inspect.
Tag creation and public publication are separate GitHub operations. Client checks cannot lock out concurrent owner actions; manage tag protection/immutable releases separately if needed. Public release corrections use a new version. No PyPI upload, native installer, automatic merge, hardware experiment or new CI lane is added. The existing four lanes remain unchanged.
JiuwenSwarm v0.2.7 shows bot-authored release notes and downloadable packages. Its publisher is a private GitHub App; its pipeline is not public. InferRSI uses its own built-in-token workflow following GitHub's token/trigger rules, generated notes, and publication CLI, plus the setuptools-scm build/archive contract. Remote scheduled execution, repository policy and actual publication remain unverified until observed in an actual release run.
The existing workflow accepts release_v020=true only through manual dispatch.
It plans exactly v0.2.0 with prerelease=false, rejects any already reserved
v0.2.0 or later numeric version, and retains the default-branch, ancestry,
build, package, checksum and no-overwrite guards. Explicit promotion may reuse
the latest preview tree; ordinary runs still skip unchanged trees. Start with
dry_run=true; publication requires a separately authorized dispatch with
dry_run=false after review and validation. Do not push the tag manually.
The reviewed notes in v0.2.0 highlights and generated changelog use v0.1.1 as their comparison baseline. The live publication race check still uses the actual latest release, not that historical notes baseline. This one-off notes path does not call the model API. Wheel, sdist, Codex ZIP, WorkSwarm ZIP and SHA256SUMS remain bound to the planned source commit.
Scheduled dispatch and ordinary manual runs keep the existing pyproject defaults. After v0.2.0 is published, the unchanged semantic patch planner naturally selects v0.2.1 as a preview when source changes; this does not retain the v0.1.x numbering line or turn future releases stable. A failed reserved v0.2.0 is not overwritten or silently replaced: investigate and obtain a new release decision.