Release crates

Releasing the Halter crates

A release is: pick a version, bump every manifest, write the CHANGELOG entry, land it on master, tag it, let CI publish, verify, then create the GitHub release.

Pushing a v* tag is irreversible. .github/workflows/publish.yml fires on push: tags: ["v*"] and publishes to crates.io, where versions can be yanked but never replaced. Do every reversible step first, verify locally, and get CI green on the release commit before any tag exists.

The release graph

Ten crates publish, in this dependency order — the list lives in bin/crate-release-candidates and that file is the source of truth:

halter-brush-core       vendor/brush-core-vendored/Cargo.toml
halter-brush-builtins   vendor/brush-builtins-vendored/Cargo.toml
halter-protocol         crates/halter-protocol/Cargo.toml
halter-hooks            crates/halter-hooks/Cargo.toml
halter-config           crates/halter-config/Cargo.toml
halter-providers        crates/halter-providers/Cargo.toml
halter-session          crates/halter-session/Cargo.toml
halter-tools            crates/halter-tools/Cargo.toml
halter-runtime          crates/halter-runtime/Cargo.toml
halter                  crates/halter/Cargo.toml

Facts that shape every step:

• The eight crates/halter-* crates plus halter move as one semver-coupled unit. They always share a version number. halter-cli (publish = false) is bumped in lockstep so the workspace reads consistently; examples/software-factory (publish = false) is not versioned.
• The two vendored brush crates version independently. They only move when vendor/ changes, and their numbers (currently halter-brush-core 0.5.x, halter-brush-builtins 0.2.x) have nothing to do with the halter line. Do not sweep them with a blanket find-and-replace, and do not assume they are unchanged — check.
• Versions are literal strings, not version.workspace. Each manifest has its own version = "X.Y.Z" under [package], and inter-crate dependency lines carry { path = "...", version = "X.Y.Z" }. Both must move. crates/halter/Cargo.toml alone has seven dependency lines to update.
• Cargo.lock is gitignored, so there is no lock churn to commit. docs/ is gitignored too.

1. Establish the baseline

LAST=$(git describe --tags --match 'v*' --abbrev=0)
git log --oneline "$LAST"..HEAD
git diff --stat "$LAST"..HEAD -- crates/ vendor/ examples/

Read the commits, not just the stat. You are looking for the story the release notes have to tell and for anything that breaks a downstream consumer.

Then let the release tooling tell you what would actually publish:

CRATE_RELEASE_BASE_REF="$LAST" bin/crate-release-candidates

Run this before bumping (to see what the last release left behind — the vendored crates in particular can carry an unpublished bump from a prior merge) and again after bumping to confirm the final publish set. Its stderr narrates candidate: / unchanged: per crate; its stdout is the name|version|manifest rows the workflow consumes.

2. Choose the version

Pre-1.0, so under Cargo's semver rules the minor position is the breaking position. Bump minor when anything in the diff breaks a downstream compile or silently changes a documented value:

• a new public struct field (breaks struct literals)
• a new variant on a public enum that is not #[non_exhaustive] (breaks exhaustive matches)
• any change to a public function signature
• a trait method signature change (breaks external implementors, e.g. custom SessionStore)
• a semantic change to an existing public or persisted field

Patch is for genuinely additive, compile-compatible work. When it is close, take the minor — the cost of an extra minor on a 0.x line is nil, and a wrong patch bump breaks builds silently.

To find which public items the diff touched, read the hunk headers rather than the changed lines — git names the enclosing item, so a change inside a pub enum surfaces even though the variant line itself carries no pub:

git diff -U0 "$LAST"..HEAD -- 'crates/**/*.rs' \
  | grep -E '^@@.*@@ *(pub |impl )' | sed 's/^.*@@ //' | sort -u

Every pub enum / pub struct / pub trait in that list needs eyeballing against the list above; read the actual hunk before deciding. impl blocks are included because git attributes an indented pub fn to its enclosing impl line, so a method signature change shows up only as the impl.

Do not grep the changed lines for ^[-+]\s*pub — enum variants, struct fields, and methods inside impl blocks are not pub-prefixed, so it silently reports nothing for exactly the change class that forces a minor here. It returned zero matches for both v0.3.0..v0.4.0 (added ReasoningEffort variants) and v0.4.0..v0.5.0 (added StreamEvent / SessionEventPayload variants) — the last two breaking releases.

cargo semver-checks --baseline-rev "$LAST" decides this mechanically and is worth installing if you cut releases often; the hunk-header scan is the zero-dependency fallback.

3. Bump the manifests

Rewrite version = "<old>" → version = "<new>" across the nine crates/*/Cargo.toml files, covering both [package] and inter-crate dependency lines. Confirm the vendored manifests and the brush-* entries in the root [workspace.dependencies] were not caught by the sweep unless you intend to move them.

grep -rn 'version = "' crates/*/Cargo.toml
grep -n 'brush-' Cargo.toml
grep -n '^version' vendor/*/Cargo.toml

4. Write the release notes

CHANGELOG.md is the canonical write-up; the GitHub release body is derived from it. Add a section directly under ## [Unreleased], following the shape of the 0.3.0 and 0.4.0 entries:

## [X.Y.Z] - YYYY-MM-DD

One paragraph: what this release is, and *why this version position* — name the
specific break that forces a minor, or state that the work is additive.

Published crates: <list>. `halter-cli` also moves to `X.Y.Z` but remains
`publish = false`.

### Added / ### Changed / ### Fixed

- Prose entries. Say what a consumer sees, not what the commit touched.

### Upgrading from <prev-minor>

- One bullet per thing that breaks a build, with the mechanical fix.

The Upgrading from section is the part that earns its keep. Every break identified in step 2 gets a bullet with the concrete remedy ("estimate_context_tokens takes a trailing usage_anchor_floor: usize; pass SessionState::usage_anchor_floor, or 0 for the previous behavior").

Also update prose docs the release invalidates — the root README.md and any affected crates/*/README.md. These carry no version strings, so the edits are content-only: new config keys, new enum values, changed provider behavior.

5. Verify locally

cargo fmt --all -- --check
cargo clippy --workspace --all-features --all-targets -- -D warnings
cargo test --workspace --all-features
CRATE_RELEASE_BASE_REF="$LAST" bin/crate-release-candidates | cut -d'|' -f1,2

That last line is the publish set. Reconcile it against the CHANGELOG's "Published crates" list — a mismatch means one of the two is wrong, and it is usually the changelog.

6. Land on master and get CI green

Commit as a dedicated release commit (Bump crate versions for vX.Y.Z), or fold the bump into the feature commit when the release exists to ship that one change — both patterns are in the history. Push to master and wait for CI:

gh run list --branch master --limit 5 \
  --json displayTitle,status,conclusion,headSha,workflowName \
  --jq '.[] | "\(.headSha[0:7])  \(.workflowName)  \(.status)/\(.conclusion // "-")  \(.displayTitle)"'

Do not tag until CI is green on the exact release commit. Publishing from a red master ships a broken version permanently.

7. Tag — hand this to the user

The tag commands are gated by the permission classifier, correctly: pushing the tag is the irreversible publish. Do not route around it. Give the user the two commands to run with the ! prefix:

! git tag -a vX.Y.Z -m "vX.Y.Z" -m "See CHANGELOG.md for release notes."

! git push origin vX.Y.Z

Two things that have bitten this repo before:

• Use two -m flags, never an embedded newline. A multi-line -m "..." string leaves zsh in an unmatched-quote continuation. Git joins repeated -m with a blank line, producing the same message the previous tags carry.
• A repository ruleset restricts tag creation. The push succeeds for the repo owner but prints Bypassed rule violations for refs/tags/vX.Y.Z ... creations being restricted. That line is expected, not a failure — check for * [new tag] in the same output.

8. Watch the publish

gh run list --workflow publish.yml --limit 3 \
  --json databaseId,status,conclusion,headBranch,event,createdAt \
  --jq '.[] | "\(.databaseId)  \(.status)/\(.conclusion // "-")  ref=\(.headBranch)  \(.createdAt)"'
gh run watch <run-id> --exit-status

The workflow runs bin/crate-release-candidates against the previous tag, then bin/publish-crate-release-candidates, which publishes in dependency order and blocks on the sparse index between crates (30 attempts × 10s) so each dependency is resolvable before its dependents publish. It skips versions already present, so it is safe to re-run.

Typical duration is under 5 minutes. Then verify against crates.io directly rather than trusting the workflow's exit code:

for spec in halter-protocol halter-hooks halter-config halter-providers \
            halter-session halter-tools halter-runtime halter; do
  curl -fsSL "https://index.crates.io/ha/lt/$spec" \
    | grep -q "\"vers\":\"X.Y.Z\"" && echo "  ok $spec X.Y.Z" || echo "  MISSING $spec X.Y.Z"
done

The index path is {name[0:2]}/{name[2:4]}/{name} for names of four or more characters — so all halter* crates live under ha/lt/. Check the vendored crates at their own versions when they are part of the release.

If the publish fails partway

Crates already on the index stay there. Fix the cause, then re-run gh workflow run publish.yml (or -f publish_all=true to sweep every listed crate whose current version is missing from the index). The publisher's skip-if-published check makes re-runs idempotent — do not bump versions to work around a partial publish.

9. Create the GitHub release

Only after crates.io is confirmed. Use --verify-tag so the command can never create a ref:

gh release create vX.Y.Z --verify-tag \
  --title "vX.Y.Z - <short theme>" \
  --notes-file <notes.md>

Body format, following v0.1.2:

Compared against <prev-tag> (<sha>).

Published crate bumps:
- <crate> <old> -> <new>
  ...

Changes:
- <consumer-facing bullet>
  ...

Keep it a condensation of the CHANGELOG section, not a second independent write-up. Link the CHANGELOG for the full Upgrading from detail.

Tags v0.2.0 through v0.4.0 were published to crates.io but never got a GitHub release. If you are asked to backfill, gh release create <tag> --verify-tag works on any existing tag and does not re-trigger the publish workflow.

Report

Close out with: version and previous version, the exact crates published and at what versions, the run ID and link, index verification result, and the GitHub release URL. State plainly if any crate failed or was skipped.