release-notes

Generate a starting-point GitHub release notes draft for an upcoming Warp release (feature or bugfix, auto-detected from the version). Run from a release…

Release Notes Draft

Generates a markdown draft of the impending Warp release's GitHub release notes, written in GitHub Flavored Markdown. Auto-detects feature vs bugfix release from the version, picks the correct prior tag for diffing, and produces a digestible user-facing summary that the release manager edits and posts to the release page.

Output: a single markdown report, filed according to the destination chosen in Phase 1:

Secret gist (default when gh is available and authenticated): stable filename warp-<version-string>-release-notes.md, stable description Warp <version-string> Release Notes Draft. Later runs against the same version revise the same gist in place; prior versions are preserved in the gist's git history.
Local markdown file (fallback when gh unavailable, or opt-in when gh available): dated path at $(pwd)/warp-<version-string>-release-notes-<YYYY-MM-DD>.md. Not auto-committed; the user reviews, edits, and copy-pastes into the GitHub release form.

Inputs: the positional argument is a branch name (resolved as upstream/<name> → origin/<name> → local), a tag (e.g. v1.13.0), or omitted (defaults to HEAD). Version string comes from VERSION.md at the resolved head, or parsed from the tag name.

Phase 1 — Resolve target and previous release

Resolve the head ref. With no argument, head = HEAD. With an argument, try in order: upstream/<arg>, origin/<arg>, <arg> (local branch or tag). Use the first that resolves via git rev-parse --verify <candidate>. Record both the symbolic ref and the resolved SHA. If nothing resolves, abort with the candidate list shown.
Read the version string. If the argument was a tag matching v<X>.<Y>.<Z>..., parse the version directly from the tag (strip the leading v). Otherwise read VERSION.md at the head:
```
git show <head-ref>:VERSION.md
```
Strip whitespace; this is the raw version string (e.g. 1.13.0rc3, 1.13.0, 1.13.1).
Parse (major, minor, patch) from the leading numeric portion; discard pre-release suffixes (rc1, dev0, .post1). Release notes are written for the eventual release, so 1.13.0rc3 becomes target version 1.13.0.
Classify release type. patch == 0 → feature release (full notes); patch > 0 → bugfix release (slim notes).
Determine the previous release tag (the diff base) using integer math on the parsed tuple:
- Feature release (X.Y.0): highest vX.<Y-1>.* tag. If minor == 0, fall back to highest v<X-1>.*.
```
git tag --list 'v<major>.<minor-1>.*' --sort=-v:refname
```
- Bugfix release (X.Y.Z, Z > 0): highest vX.Y.<Z'> where Z' < Z. Enumerate and filter:
```
git tag --list 'v<major>.<minor>.*' --sort=-v:refname
```
Probe the head for an existing tag. If git tag --points-at <head-ref> returns a tag matching the target version (e.g. v1.13.0), the tag exists but the GitHub release isn't published yet. Note this in the confirmation message; CHANGELOG URLs can use the tag.
Count commits in the range using the same cherry-pick-filtered range the contributor helper uses (so the scope shown to the user matches the work that will be analyzed):
```
git log --no-merges --oneline --cherry-pick --right-only <prev-tag>...<head-ref> | wc -l
```
Note the ... (symmetric difference) and --cherry-pick --right-only flags: these drop commits whose patch-id has an equivalent on the prior release. This matters when the previous tag is on a bugfix branch rather than an ancestor of the head — without filtering, the proposal can overstate scope by hundreds of commits, and the user confirms a misleading count. If zero after filtering, abort.
Probe gh and look up matching gists. Run gh --version && gh auth status. If either fails, force destination to local and skip the rest. Otherwise run gh gist list --limit 1000 and filter rows whose description exactly matches Warp <version-string> Release Notes Draft. Record the matching gist IDs and count (0, 1, N ≥ 2) for Phase 2. Display URLs are https://gist.github.com/<id>.

Phase 2 — Confirm scope and destination with user

Print a single proposal block and wait for explicit user confirmation. Mandatory pause; do not proceed to Phase 3 until the user replies.

Lead with the scope line:

Drafting <feature|bugfix> release notes for Warp .

Head: <head-ref> @ <short-sha> <(tag <tag> exists at this SHA)>

Previous release: <prev-tag> @ <short-sha>

Commits in range: <N>

Append the output block matching the current gh / match state:

gh unavailable:

Output: local markdown file at <cwd>/warp-<version-string>-release-notes-<today>.md (gh not available). Confirm, or specify different refs?

gh available, 0 matches:

Output: (a) new secret gist [default], (b) local markdown file in cwd. Confirm refs + pick.

gh available, 1 match:

Output: (a) revise existing gist <url> [default], (b) new secret gist, (c) local markdown file in cwd. Confirm refs + pick.

gh available, N matches (N ≥ 2):

Multiple existing gists share the stable title:

<url-1> — updated <time-1>

<url-2> — updated <time-2> ...

Output: (a) revise gist by number, (b) new secret gist, (c) local markdown file in cwd. Confirm refs + pick.

Translate the reply into one destination token (local, new-gist, revise-gist:<id>). Letters (a)/(b)/(c) are positional within the prompt shown; for the N-match branch a numeric reply (e.g. "revise 2") selects from the listed gists. New refs in the reply re-run Phase 1.

Phase 3 — Gather inputs

3a — CHANGELOG section

Read CHANGELOG.md at the head ref:

git show <head-ref>:CHANGELOG.md

Locate the section for this release. Try ## [<version-string>] first (e.g. ## [1.13.0]), then fall back to ## [Unreleased] if the release branch hasn't promoted yet.

Collect bullets per subsection (### Added, ### Removed, ### Deprecated, ### Changed, ### Fixed, ### Documentation). For each bullet, capture: full raw text, section name, GH refs (regex \bGH-(\d+), deduped), experimental flag (**Experimental** present), breaking flag (**Breaking:** present).

3b — Commit list and contributor attribution

Resolve <skill-dir> to the directory containing the currently loaded SKILL.md for this skill (for example, <repo>/.claude/skills/release-notes or <repo>/.codex/skills/release-notes). Do not hardcode .claude or .codex; the skill content must work from either tree.

Run the contributor enumeration script via uv run so it picks up the repo's pinned Python (the script uses 3.10+ syntax and the repo policy in AGENTS.md requires uv run over the system interpreter). Stays inside Bash(uv run *list_contributors.py*). The temp changelog uses the warp- prefix so the existing cleanup glob catches it:

git show <head-ref>:CHANGELOG.md > /tmp/warp-<version-string>-changelog.md
uv run "<skill-dir>/scripts/list_contributors.py" \
  --base <prev-tag> \
  --head <head-ref> \
  --changelog /tmp/warp-<version-string>-changelog.md \
  --target-version <target-version-string>

Delete /tmp/warp-<version-string>-changelog.md after the script returns.

The script emits JSON with commits[] (sha, subject, author name+email, GH refs from subject+body+CHANGELOG diff, files touched) and contributors[] aggregated per (name, email). Each contributor record carries: commit_count, gh_login (best-effort: noreply email pattern, then gh api), classification (nvidia | external — see references/contributor-attribution.md), pr_summary (PR number + subject pairs from squash commits), changelog_sections (CHANGELOG section labels their GH refs land in), already_shipped (true when all attributed work falls in a prior CHANGELOG section, meaning they were already thanked), prior_commit_count (all-time commits authored by this email reachable from <base>, inclusive), is_first_time_contributor (true when prior_commit_count == 0), and nvidia_low_commit_count (true when classification == "nvidia" AND commit_count <= 3 AND prior_commit_count <= 3, flagging likely non-Warp-team NVIDIA contributors for manual review; thresholds live in the script and are the source of truth).

Two robustness filters run by default: a cherry-pick filter (drops commits whose patch-id appears in <prev-tag>, common when fixes land on bugfix branches before main) and CHANGELOG-section attribution (sets already_shipped). See references/contributor-attribution.md for the rationale.

Capture the JSON. Phase 5d auto-renders only classification == "external" AND already_shipped == False contributors; the nvidia_low_commit_count and is_first_time_contributor flags surface candidates for manual decision in the Phase 6 chat summary, not in the rendered Acknowledgments section.

3c — Past-release calibration (optional)

For tone calibration, fetch one recent release body of matching type via gh release view <tag> --repo NVIDIA/warp --json body --jq .body. Mirror structure and register; do not copy content.

Phase 4 — Plan grouping and inspect implementations (feature releases only)

Bugfix releases skip Phase 4 and proceed to Phase 5 (include a one-sentence intro in Phase 5a). Read references/style-rules.md, references/feature-investigation.md, and references/feature-release-template.md before writing.

Identify the lead feature — the single most exciting capability in the release. Gets its own ### … block at the top of ## New features with a code example. Prefer items that unlock previously-impossible workflows (new artifact, file format, cross-language boundary) over new parameters. Honor the lead the user named at invocation.
Plan theme groups. Bucket remaining ### Added / ### Changed entries into 3–6 themes that reflect what users do with the features. Common shapes: "Tile programming enhancements", "JAX integration", "warp.fem enhancements", "Compilation and tooling", "Performance improvements", "Language enhancements", "New examples".
Order by impact. Lead feature first, then themes by user-facing impact. ## Bug fixes near the end, omitted if no fix warrants a highlight.
Tag experimental items. Any bullet flagged **Experimental** in 3a gets a > [!IMPORTANT] admonition under its ### … heading.
Inspect implementations for caveats, artifacts, and in-tree examples. Apply the protocol in references/feature-investigation.md to the lead feature and to every ### … block you plan to write. The protocol surfaces (a) "not yet supported" / TODO limits buried in the implementation, (b) on-disk or wire artifacts beyond the obvious file (sidecar directories, compute-arch pinning, version metadata), (c) existing in-tree examples worth linking (warp/examples/**, warp/tests/**), and (d) cross-language requirements (any feature crossing Python ↔ C++ ↔ JAX needs a snippet on each side, not just Python). Skipping this step is the difference between "looks like a release announcement" and "lets me start using the feature."

Phase 5 — Draft sections

5a — Intro paragraph

Open the body with a 2-3 sentence paragraph that names the shape of the release. Lead with the unlock, not API names. Feature-release intros name the lead feature first, then 1-2 supporting themes. Bugfix-release intros are one sentence: Warp v<version> is a bugfix release following v<prev>. plus a changelog pointer.

5b — Sections

Fill in references/feature-release-template.md (feature release) or references/bugfix-release-template.md (bugfix release). Both templates carry per-section instructions inline. Code examples are the default: every ### … block describing a new public API or behavioral change ships with a runnable Python snippet unless the change is a single-parameter tidy-up. Snippets must NOT call wp.init() (Warp initializes implicitly). Breaking changes in ## Bug fixes or ## Changes ship with before/after snippets, not prose. When a snippet has a print(...), run it via uv run /tmp/release_notes_example_<name>.py and embed the real captured output. Render single-line output as a trailing # comment; render multi-line or structured output as a separate fenced ```text block following the Python block (see references/style-rules.md "Rendering output"). Required when the output IS the feature (memory tracker, diagnostics, dtype promotion). Use > [!NOTE] / > [!IMPORTANT] / > [!CAUTION] admonitions, never ad-hoc **Note:** or quoted-bold forms. See references/style-rules.md.

5c — Announcements

This section appears in both feature and bugfix release notes (deprecations carry forward across releases until they actually land). It announces CHANGES in this release: removals, deprecations, platform-support shifts, release-cadence shifts. It does NOT restate unchanged policy.

If a sub-section has no actual change to announce in this release, drop the sub-section. If no sub-section applies, drop the entire ## Announcements section.

Pull from:

CHANGELOG ### Removed at the head ref. Each entry is a candidate ### Removals in this release bullet. Apply the tone rules in references/style-rules.md "Tone for removals and deprecations": neutral migration step tied to the user's outcome, no scolding about prior deprecation warnings.
CHANGELOG ### Deprecated at the head ref. Each entry is a candidate ### Upcoming removals bullet. Only state a specific removal version when a primary source commits to it (CHANGELOG text, runtime DeprecationWarning emit string, docstring, design doc, or maintainer statement). Without a primary source, use the neutral framing "will be removed in a future feature release per the standard deprecation timeline." See references/style-rules.md "Forward-looking claims".
CHANGELOG ### Deprecated from previous releases that explicitly named this release as the removal version. These deprecations have now landed; promote them under ### Removals in this release.
Platform-support changes, ONLY when they actually changed in this release. Read directly from the source at the head ref AND at the previous-release tag, then diff:
- Python: pyproject.toml (requires-python and Programming Language :: Python :: 3.X classifiers) and .python-version.
- CUDA Toolkit: warp/_src/build_dll.py for MIN_CTK_VERSION and the CTK_* constants in build_lib.py.
- OS / arch: the cibuildwheel config in pyproject.toml and any manylinux_* / macosx_* constraints in build scripts.
If nothing changed between the two tags, drop the ### Platform support sub-section. If something changed, state ONLY the change ("Python 3.9 is no longer supported"), not the unchanged baseline.
Release-cadence changes, ONLY when the cadence actually changed. Use aspirational language ("Warp aims to publish a feature release every month"), not declarative.
Out-of-CHANGELOG facts the user is responsible for (allocator policy changes, security advisories, planned support changes the build doesn't yet enforce). Surface candidates in the chat summary so the user can confirm or drop.

5d — Acknowledgments

Use contributors[] from Phase 3b filtered by already_shipped == False (those marked True were thanked in a prior release; surface in the chat summary instead).

Auto-rendered acknowledgments are limited to classification == "external" contributors. For each:

If gh_login is non-null, render - @<gh-login> for <one-line summary> (#<NNNN>)..
If gh_login is null (no gh available, or no GitHub user could be associated with the commit email), do NOT render @None / @null. Render - <name> for <one-line summary> (#<NNNN>). and surface the entry in the chat summary so the release manager can resolve the handle or drop the line before posting. The trailing (#<NNNN>) is omitted when no CHANGELOG bullet matches the contributor's commit (per references/contributor-attribution.md).

If the auto-rendered list is empty after filtering, render No external contributions in this release. under the section heading.

Surface for manual decision in the chat summary (do NOT auto-render acknowledgment lines for these):

nvidia_low_commit_count == True contributors: NVIDIA-classified, low commit count in this range, low all-time commit history. Likely outside the Warp team. The release manager may want to acknowledge them like an external contributor. Show name, GitHub login (if known), commit count, prior commit count, and the contribution subjects from pr_summary.
is_first_time_contributor == True contributors (regardless of classification): no commits in the repo before the diff base. The release manager may want a "Welcome our first-time contributors:" callout regardless of email domain. Show the same fields.

These two flag sets can overlap (a first-time NVIDIA contributor is flagged on both). Surface each contributor once; note both flags when both apply. See references/contributor-attribution.md "Two reasons to acknowledge a contributor classified as nvidia" for the rationale.

5e — Self-audit before sending to user

Before handing the draft to the release manager, re-read references/style-rules.md (and references/feature-investigation.md for feature releases; bugfix releases skipped Phase 4 and can skip its rules) and verify each rule was applied. Every unapplied rule is an iteration the release manager would otherwise spend.

Phase 6 — Render and write to chosen destination

Fill in every {{PLACEHOLDER}} in the selected template, applying references/style-rules.md (chiefly: #NNNN not [GH-NNNN](...), experimental → > [!IMPORTANT], no em dashes, no skill-internal terminology, proper-noun capitalization). The destination was chosen in Phase 2: one of local, new-gist, or revise-gist:<id>.

Filename conventions:

Local file (in $(pwd)): warp-<version-string>-release-notes-<today>.md — dated, user-facing.
Gist file (inside the gist): warp-<version-string>-release-notes.md — no date. Stable name so later runs revise the same gist in place.

Stable gist description (used when creating a new gist; also the matching key for Phase 1.8):

Warp <version-string> Release Notes Draft

If destination is local: write the rendered markdown to $(pwd)/<local-filename> using the Write tool.

If destination is new-gist:

Write to /tmp/<gist-filename> using the Write tool.
gh gist create --desc "<stable-desc>" /tmp/<gist-filename> — capture the URL from stdout.
Delete /tmp/<gist-filename>.

If destination is revise-gist:<id>:

Write to /tmp/<gist-filename> (same stable name the existing gist already uses).
gh gist edit <id> --filename <gist-filename> /tmp/<gist-filename>. The --filename flag selects which file inside the gist to replace; the trailing local path supplies the new content. Without --filename, gh treats the local path as a gist-side filename selector, fails to find it, and falls through to its interactive editor. Do NOT pass --desc: keeping the description stable is what lets the next run match this gist again. Prior versions are preserved automatically in the gist's git history.
Delete /tmp/<gist-filename>.

Never pass --public. Never file a destination the user did not choose.

Print a one-line chat summary with output path/URL and headline counts (features, announcements, external contributors). Append as warranted:

"revised in place; prior versions kept in gist git history" for revise-gist.
"Already-shipped contributors skipped: @user1 (sections=[1.12.1]), ..." for already_shipped == True entries the user can spot-check.
"Unresolved GitHub handles (rendered with name only): (), ..." for any external contributors whose gh_login is null in the rendered acks, so the release manager can resolve or drop those lines before posting.
"First-time contributors (consider a shoutout): @user1 (1 commit, ''), ..." for contributors with is_first_time_contributor == True. Include classification in the listing so the release manager sees both NVIDIA and external first-timers.
"NVIDIA contributors with low commit count (likely outside Warp team, consider acknowledging): @user1 (2 in range, 2 all-time, ''), ..." for contributors with nvidia_low_commit_count == True (excluding any already covered as first-time above).

For revising an existing gist (the revise-gist:<id> destination), remind the release manager that prior versions are preserved in gist git history and can be diffed across iterations.

Failure modes

Argument doesn't resolve. Print the candidate list (upstream/<arg>, origin/<arg>, <arg>) and abort. Do not silently fall back to HEAD.
VERSION.md missing or unparsable at the head. Abort with the raw file contents shown.
No previous tag found. Feature release: no prior minor exists (only at v1.0.0-equivalents). Bugfix release: no prior vX.Y.<Z'> with Z' < Z. Abort and ask the user to specify the base manually.
CHANGELOG section not found. Neither ## [<version-string>] nor ## [Unreleased] matches. Abort; the user likely passed the wrong head ref.
gh unavailable. list_contributors.py falls back to email-only classification (only @nvidia.com registers as internal); private-membership NVIDIA staff committing from noreply / personal addresses will be misclassified as external. Note in the chat summary.

Regexes and parsing rules (inline reference)

GH ref: \bGH-(\d+).
Tag pattern: ^v(\d+)\.(\d+)\.(\d+)([a-z0-9.+-]*)$. The trailing group captures rc1, dev0, .post1, etc.
CHANGELOG section header: ^## \[([^\]]+)\] — the bracketed contents are the version or Unreleased.
CHANGELOG subsection header: ^### (Added|Removed|Deprecated|Changed|Fixed|Documentation).
Experimental marker: \*\*Experimental\*\*:? (bold, optional trailing colon).
Breaking marker: literal **Breaking:**.
noreply.github.com email login: ^\d+\+([^@]+)@users\.noreply\.github\.com$ → group 1 is the GitHub username.