# Release > Step 0 — Determine target version - Skill: `psd-tools/release` (Agent Skill) - Install (CLI): `npx skillmds@latest add psd-tools/release` - Raw SKILL.md: https://api.skillmd.com/api/skills/psd-tools/release/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: psd-tools (https://skillmd.com/u/psd-tools) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/psd-tools/release --- ## Step 0 — Determine target version **Provided version**: $ARGUMENTS ### If a version was provided (above is non-empty) Validate it as a PEP 440 string (e.g. `1.2.3`, `1.2.3a1`, `1.2.3rc1`, `1.2.3.post1`). Note: PEP 440 versions must **not** start with `v` — that prefix belongs on the git tag, not the version string. Stop and ask the user to correct it if invalid. Store it as **VERSION** for all subsequent steps. ### If no version was provided (above is empty) Analyze the commits listed in Step 1 to recommend the correct next version. The last tag shown in Step 1 uses a `v` prefix (e.g. `v1.14.3`); strip it when computing the next version so the result is a bare PEP 440 string (e.g. `1.14.4`, not `v1.14.4`). Apply these semver rules: - **Major bump** (`X+1.0.0`) — any commit that breaks a public API or documented behaviour - **Minor bump** (`X.Y+1.0`) — any new public feature or API addition, no breaking changes - **Patch bump** (`X.Y.Z+1`) — bug fixes, security patches, chores, docs, or refactoring only Show your reasoning and proposed version to the user, then ask them to confirm or override it. Store the confirmed version as **VERSION** for all subsequent steps. ## Step 1 — Review commits since the last release Fetch tags: !`git fetch --tags -q || echo "Warning: failed to fetch tags — check network/auth and consider retrying."` Last tag: !`git describe --tags --abbrev=0 2>/dev/null || echo "(none)"` Today's date: !`date +%Y-%m-%d` Now list the commits since the last release by running one of these commands: - If "Last tag" above is `(none)`: run `git log --oneline` - Otherwise: run `git log ..HEAD --oneline` (substituting the actual tag) You must run this command and review the output before proceeding to Step 0 or Step 2. ## Step 1b — Prose sweep Rationale docstrings and comments accumulate across a release: each fix PR appends a paragraph explaining what it measured, and none compresses what is already there. Trim them now, while the release is the unit of review — nothing enforces this mechanically (ruff's docstring rules are off, `ruff format` does not reflow comments or docstrings, and no line-length rule is selected), so this step is the only thing that catches it. Measure rather than eyeball — the worst ratios sit on the smallest helpers, which read as fine in a diff: ```bash uv run python tools/prose_density.py ``` Anything above roughly 3x prose-to-code deserves a look. Keep what cost real measurement, and what stops a future reader undoing a fix: why a bound is shaped the way it is, why a guard cannot be tighter, what an experiment ruled out. Cut history the changelog already holds, corpus statistics the tests already assert, and the same explanation repeated at three layers. Trimming is a separate commit from the release commit, and touches comments and docstrings only. If any line of code moves, stop and treat it as a code change. ## Step 2 — Draft changelog entry Read `docs/changelog.rst` to understand the current format, then draft a new entry for **VERSION** using this RST format: ``` VERSION (YYYY-MM-DD) -------------------- - [category] Description (#PR) ``` **Important**: The `-` underline must be at least as long as the title line (RST requirement). Count the exact characters in `VERSION (YYYY-MM-DD)` and use that many dashes. The categories, and the rule for picking between them, are documented in the **Changelog** section of the repo-root `CLAUDE.md`, which loads as project instructions — use that list. It is deliberately not restated here: the second copy is what drifted (#791). Group related changes. Omit purely internal churn that users won't care about. Reference PR numbers where available. Show the draft to the user and ask for approval or edits before continuing. ## Step 3 — Update docs/changelog.rst Prepend the approved changelog entry directly after the `Changelog` header block and the following blank line, leaving a blank line between the header and the new entry. ## Step 4 — Create release branch and commit ```bash git checkout -b release/vVERSION ``` Then update `src/psd_tools/version.py` using the Edit tool — replace the existing `__version__` line with: ```python __version__ = "VERSION" ``` `VERSION` is the bare PEP 440 string **without** the `v` prefix (e.g. `1.15.0`, not `v1.15.0`). Then stage both changed files and commit: ```bash git add docs/changelog.rst src/psd_tools/version.py git commit -m "docs: release vVERSION" git push -u origin release/vVERSION ``` Replace `VERSION` with the actual version string (e.g. `1.15.0`). ## Step 5 — Open a pull request Run `gh pr create` with `--title "Release vVERSION"` and a `--body` containing: - A `## Release vVERSION` heading - A `### Changelog` section with the approved entry from Step 2 pasted in - A `### Release checklist` section with these items: - `[ ] Changelog entry reviewed and accurate` - `[ ] Version follows PEP 440` - `[ ] Prose sweep done (Step 1b), or explicitly skipped` - A closing note: "After this PR is merged, the `auto-tag` workflow will tag the merge commit as `vVERSION` and the `release` workflow will build wheels and publish to PyPI automatically." Replace `VERSION` with the actual version string throughout. ## Step 6 — Done Print the PR URL. Remind the user: > After the PR is approved and merged, the `auto-tag` GitHub Actions workflow tags the merge > commit as `vVERSION` automatically. That tag push triggers the `release` workflow to build > wheels for all platforms and publish to PyPI. No manual tagging or publishing is needed. Replace `VERSION` with the actual version string.