Deployment Skill (Agent Sessions)

This skill is an agent-facing entrypoint that avoids duplicating the deployment runbook.

Canonical Sources (Single Source of Truth)

• Runbook: docs/deployment.md
• Unified tool: tools/release/deploy (see tools/release/deploy --help)
• Recommended pre-release QA checklist: docs/release/pre-release-qa.md

If anything here disagrees with the runbook, follow docs/deployment.md.

Workspace Policy (Hard Rule)

• Always run deployment from the user's current local repository checkout.
• Do not clone to temporary directories and do not switch to alternate worktrees as a deployment workaround.
• If the local worktree is dirty, stop and tell the user to clean the tree first (commit, stash, or discard), then continue in the same local repo.

Resource Preflight (Before QA, Bump, and Release)

Heavy steps (QA test runs, the Release build, notarization) crawl when the Mac is out of memory. During 6.0 the Release build ran 20+ minutes at 15% CPU because swap was nearly full. The cause was fseventsd, a macOS daemon holding 21 GB (20 GB compressed). The machine reported "31% free" and ps RSS showed every process small, so both looked healthy. Run this check before each heavy step, and again if a Release build passes 12 minutes:

sysctl -n vm.swapusage                                    # swap used / total
top -l 1 -o mem -n 8 -stats pid,command,mem,cmprs         # MEM includes compressed memory
df -h / | tail -1                                         # free disk
du -sh ~/Library/Developer/Xcode/DerivedData .deriveddata-* 2>/dev/null

Use top, not ps: RSS hides compressed and swapped memory, which is where the problem was. Do not trust memory_pressure's "free percentage" either.

Act on what you find:

• Swap above half of its total, or any process above 8 GB in top MEM — fix before the heavy step. If the process is fseventsd, restart it: launchd respawns it at once, and nothing is lost. Try sudo -n killall fseventsd first. When that fails (no cached sudo), stop and give the user that command without -n to run in Terminal, then re-check. Never kill other system daemons. For user apps (Chrome, other Claude sessions, test runners), name the biggest ones and ask the user to quit them; do not kill them yourself.
• Free disk under 20 GB, or the DerivedData folders above 10 GB together — recommend deleting the repo's .deriveddata-* folders and ~/Library/Developer/Xcode/DerivedData. They are build caches and regenerate. Delete them only when the user agrees, and never while a build is running.
• Otherwise continue. Report the swap and top-memory numbers in one line.

QA Gate (Mandatory — Run Automatically Before Deploy)

• Always run QA automatically before any bump/release/verify step, unless the user explicitly says to skip it (e.g. "skip QA", "no QA").
• Do not ask whether to run QA — just run it.
• QA execution order:
  1. Scope — git log --oneline --decorate -n 30 and git diff --name-only <LAST_TAG>..HEAD; identify high-risk areas.
  2. Build — xcodebuild -project AgentSessions.xcodeproj -scheme AgentSessions -configuration Debug build
  3. Full test suite — ./scripts/xcode_test_stable.sh
  4. Targeted tests — run suites for touched high-risk areas (session parsing, usage tracking, onboarding, etc.)
  5. Warnings sweep — flag any new actionable warnings in build output.
  6. Manual smoke reminder — list the manual steps from docs/release/pre-release-qa.md §3–4 and ask the user to confirm GO/NO-GO after completing them.
• If automated gates fail → stop, report failure, do not proceed to bump/release.
• If user says "skip QA" or "no QA" → proceed without running, note it was skipped.

Test count. QA reads the authoritative pass count from the .xcresult bundle (stdout reports per-bundle totals and this scheme has two — it under-reports) and compares it to tools/release/test-count-baseline.txt. A drop warns and continues, deliberately: it does not block a release. So the warning has to actually be read — a deleted suite still exits 0 from xcodebuild, and this line is the only thing that says so. When the count changes for a real reason, update the baseline file in the same release.

Before Starting (Ask the User)

1. Target version (X.Y for major/minor releases, X.Y.Z only for patch releases; never ship X.Y.0)
2. Any headline changes (new agents, major features) that must be reflected in docs/CHANGELOG.md
3. Whether this is a major release that requires onboarding updates
4. Public copy updates needed for README/GitHub Pages (major changes to highlight, renamed features, or outdated wording to fix)

Do NOT ask about QA status — QA always runs automatically as part of pre-deploy (see QA Gate above).

Public Copy Update (Required for All Releases)

Always update (every release)

• README.md download link: v{VERSION}/AgentSessions-{VERSION}.dmg and label Download Agent Sessions {VERSION} (DMG)
• README.md Option A download link (second occurrence under Install section)
• docs/index.html download button URL and label
• docs/index.html version meta-line (Version {VERSION} · Free & open source · No telemetry)

Do NOT version-pin the site meta descriptions

docs/index.html's description, og:description and twitter:description are deliberately version-agnostic as of the 2026-08-29 SEO work. This checklist used to demand "mention current version + key change" there, which is how the homepage description reached 541 chars reading "Version 5.0 makes every agent a plug-in adapter…" — past the ~155-char SERP budget, and stale the day after every release.

Leave them alone. If the product's positioning genuinely changes, rewrite them on their own merits and keep the budget: title ≤60, description ≤155. Adding a version string is the one edit that is always wrong. See the reference-docs-site-seo-conventions memory.

Update for minor/major or user-visible feature releases

• README.md "What's New in X.Y" section: update heading to new version, rewrite TL;DR and Highlights to reflect this release's key changes (do not keep old version's copy)
• docs/index.html hero/feature copy if features were renamed or new agents added

Never add

• Versioned "What's New in X.Y" section to docs/index.html
• Detailed release notes to README or website (those live in docs/CHANGELOG.md)

Never Announce a Bug the User Never Had (Hard Rule)

A fix to a feature that ships in this same release is not a Bug Fix. It is development.

When a feature is new in X.Y, every defect found and fixed in it before X.Y shipped was never in anyone's hands. Listing those under "Bug Fixes" invents a history of breakage users never experienced, and buries the actual feature under a list of things that sound broken. Ship the feature; the fixes are part of it.

Before writing any Bug Fix entry, ask: which released version had this bug? If the answer is "none — the code is new in this release", the entry does not exist. Fold anything user-visible into the feature's own Highlight instead.

Worked example (4.8, Grok CLI's first release):

• ❌ "A Grok session shows the title its own sidecar records" — Grok shipped in 4.8; no user ever saw the wrong title.
• ❌ "Grok transcripts open with content in them" — same.
• ❌ "A Grok session's message count matches its transcript" — same.
• ✅ "Grok CLI is the eleventh current agent source…" — the Highlight, which already says transcripts, images, Analytics and resume work.

Mixed entries need splitting, not deleting: a fix spanning shipped and new sources is real for the shipped ones. Describe it in terms of those, and drop the new source from the list. In 4.8 the CLI PATH-masking fix covered Cursor, Kimi and Pi (all shipped) plus Grok (new) — it stayed, naming only the three.

This applies identically to docs/CHANGELOG.md, the Sparkle notes, the GitHub release body, and the README "What's New". The changelog is the source all of them derive from, so fix it there first.

After pushing

• Verify GitHub Pages reflects updated docs/index.html (check the download button and version meta-line; the meta descriptions should read the same as before the release)

Pre-Deploy Checklist (Run Before Bump)

• docs/CHANGELOG.md [Unreleased] section has full, accurate content for this release (this is the only changelog — root CHANGELOG.md is a pointer to it, not a copy to sync)
• Every Bug Fix entry names a defect that existed in a released version — no fixes to features shipping in this same release (see "Never Announce a Bug the User Never Had")
• In-app What's New has an entry for this version in AgentSessions/Onboarding/Models/WhatsNewCatalog.swift — both a teasers line and a bundled array. hasContent goes true on the auto-generated new-provider row alone, so a forgotten release still shows the card, just with one generic line and no teaser. Do not author a row for a new source by hand: providerHighlights(for:) generates it from versionIntroduced, and authoring it again shows it twice.
• README.md download links updated to new version (both occurrences)
• README.md "What's New" section updated to new version heading + rewritten highlights
• docs/index.html download button URL, label, and version meta-line updated
• docs/index.html meta descriptions left unchanged — they are version-agnostic on purpose (see "Do NOT version-pin the site meta descriptions")
• All above files committed before running deploy bump (or bump will overwrite)

Sparkle Notes Are Short and Fun (Hard Rule)

Sparkle notes appear in a small update window. Nobody reads prose there. The 5.0 notes shipped as full changelog paragraphs and had to be republished after the fact — don't repeat that.

• Highlight: 2–4 sentences, hard max. Lead with the change, keep one concrete number if there is one, cut everything a curious reader can find in the changelog.
• Feature / Bug Fix bullets: 1–2 sentences. One for what changed, at most one for why it was wrong before. A good bug-fix line can be a single sentence ("Five agents used to skip it silently.").
• Whole notes: aim under ~250 words. If the preview scrolls, it's too long.
• Tone: dry fun is welcome — "takes no for an answer", "Bring the agent you use". Personality yes, marketing-speak no, emoji never.
• The full detail lives in docs/CHANGELOG.md; the notes may compress it freely. If an entry can't be compressed without losing the point, the changelog entry is overweight — tighten it there first (changelog entries also don't need to be essays).
• Links in changelog entries that feed the notes must be absolute URLs — relative links break in the appcast and the GitHub release body.
• To republish notes after a release: edit docs/CHANGELOG.md, then python3 tools/release/sparkle_release_notes.py --version <V> --changelog docs/CHANGELOG.md --appcast docs/appcast.xml --github-url <release-url> --lint --out-text /tmp/notes.txt, then gh release edit v<V> --notes-file /tmp/notes.txt, commit and push the appcast.

Sparkle Release Notes (Approval Gate)

• The release pipeline generates structured Sparkle notes from docs/CHANGELOG.md:
  • Highlights or grouped current-release changes
  • Other changes (summary)
  • Reminder from the baseline release (for patch releases: A.B)
• During tools/release/deploy release <VERSION>, the deploy script prints a Sparkle release notes preview after build/sign/notarization and appcast validation.
• Treat the preview as user-facing product copy, not raw commit history:
  • Lead with the headline change the user should care about.
  • Do not include internal cleanup, validation fixes, or pre-release stabilization as “Bug Fixes” if users never received that broken behavior.
  • If the preview is misleading, stop and edit docs/CHANGELOG.md before publishing.
• If SKIP_CONFIRM is not 1, it will pause and ask for approval before publishing (pushing appcast, updating Homebrew, updating the GitHub release).
• SKIP_CONFIRM=1 requires RELEASE_NOTES_REVIEWED=1 at the appcast publish gate; set it only after manually inspecting the Sparkle preview or for a rerun whose notes were already reviewed.
• The notes generator fails before publishing if notes contain obvious internal/process wording or put Bug Fixes ahead of a headline section.
• GitHub Release notes must use the same curated/linted notes as Sparkle, not raw commit history or raw changelog extraction.
• If the current release has no structured bullets, the generator adds a fallback highlight: Small bug fixes and stability improvements.

Standard Workflow (Use the Unified Tool)

tools/release/deploy changelog [FROM_TAG]
tools/release/deploy bump [patch|minor|major]
git push origin main
tools/release/deploy release <VERSION> [--dry-run]
tools/release/deploy verify <VERSION>

After the Release Is Published

• Set sequence and fingerprint in docs/prices-v2.released.json to the values the released build bundles (RunwayPriceTable.makeForTesting().sequence and .manifestFingerprint), in a commit of its own. Ordinary price edits never touch that file. testSequenceIsRaisedWheneverTheManifestDiffersFromTheLastRelease then forces a higher sequence on the next edit of any byte of docs/prices-v2.json.
• Make sure docs/prices-v2.json is live on the docs site before or with the app release.

Failure Handling

• First stop: docs/deployment.md → Troubleshooting, logs, and rollback guidance.
• Rollback only after reviewing logs: tools/release/rollback-release.sh <VERSION>.

Read the real exit code, and check what actually published

Two traps, both hit during 5.1.1:

• Do not wrap the release in ; echo "EXIT: $?". The wrapper's own status is what the caller sees, so a pipeline that correctly exited 4 was reported as 0 and briefly looked like a tooling bug. Run tools/release/deploy release <V> on its own and read its status.

• A completed run is not a published release. Confirm against the world, not the log:

  gh release view v<VERSION> --json tagName,isDraft,assets
  git ls-remote --tags origin v<VERSION>
  grep -n "sparkle:shortVersionString" docs/appcast.xml

  Then tools/release/deploy verify <VERSION> for the full check. A failure before the publish steps leaves nothing public — no release, no tag, appcast unchanged — which is the state to expect after an aborted build.

Watching a long run

The release log embeds the entire QA test output, so a bare keyword grep matches test names, not events — "Aborted", "Homebrew" and "rollback" all produced false alarms during 5.1.1. Anchor on the script's own line markers (^==> , ^✅ , ^ERROR: ) and run the pattern against the live log before trusting it. Note also that ERROR: Release QA stamp is not valid is expected and self-heals whenever HEAD moved after deploy qa — it re-runs QA rather than failing. When matching newest-file-by-glob (/tmp/notarization-<V>-*.log), guard on mtime or you will read the previous run's artifacts.