GAIA Release
Run a GAIA release end-to-end against the amd/gaia repo. The skill is a phased checklist with hard gates — each phase produces a concrete artifact, then stops and asks for confirmation before doing anything irreversible (opening a PR, pushing a tag, rerunning a CI job, posting an announcement).
The pre-tag verification phase exists because the v0.17.4 release uncovered two release-blocking bugs that merged-PR CI did not catch (squash-merge silently reverting version.py, and the AppImage smoke test's circular pip install amd-gaia==<unpublished> dependency). Do not skip it.
Argument parsing
Accept one argument in any of these shapes:
0.17.5— bare semverv0.17.5— with thevprefix (the actual tag form)0.17.5.1— hotfix-style four-part version (rare; e.g.v0.15.4.1)- (no argument) — read
__version__from src/gaia/version.py, then suggest the next-patch bump (e.g.0.17.4→ propose0.17.5) and ask the user to confirm or override before continuing.
Normalise to the bare form internally (0.17.5) and the tag form externally (v0.17.5). If the user gives a version that is older than or equal to the current __version__, stop and ask — never silently roll backwards.
Resume detection (run before any phase)
The skill is reinvocation-safe — releases span days, and the user will return mid-flow. Before doing anything else, detect where in the flow we are and skip phases that already landed:
bashgit fetch origin --tags git checkout main && git pull NOTES_ON_MAIN=$(git ls-tree -r origin/main --name-only | grep -c "^docs/releases/v<version>\.mdx$" || true) PR_OPEN=$(gh pr list --repo amd/gaia --search "Release v<version> in:title" --state open --json number --jq '.[0].number' || true) PR_MERGED=$(gh pr list --repo amd/gaia --search "Release v<version> in:title" --state merged --json number --jq '.[0].number' || true) TAG_EXISTS=$(git tag --list "v<version>") RELEASE_EXISTS=$(gh release view "v<version>" --repo amd/gaia --json tagName --jq '.tagName' 2>/dev/null || true)
Resume table:
| State | Action |
|---|---|
RELEASE_EXISTS matches | Release already shipped. Skip to Phase 6 (smoke test + announcement) only. |
TAG_EXISTS, no release | Tag pushed, publish workflow in progress or failed. Resume at Phase 5 (monitor). |
v<version>-rcN tags exist, no final tag | A release candidate is out. Resume at Phase 3.5 step 4 (test pass) for the highest N. List them with git tag --list "v<version>-rc*". |
PR_MERGED, no tag | Notes on main. Resume at Phase 3 (pre-tag verification). Do not re-run Phase 1 — never overwrite merged notes. |
PR_OPEN | PR still open. Tell user "PR #N already open at — waiting for merge." Exit. |
NOTES_ON_MAIN ≥ 1 but no PR | Half-finished prior attempt landed notes without a PR (rare). Stop and ask the user before continuing. |
| Nothing matches | Fresh release. Start at Phase 1. |
Always announce the resume decision before continuing: "Detected v<version> state: PR #831 merged, no tag yet. Resuming at Phase 3 (pre-tag verification)."
A PR_OPEN or PR_MERGED state may have been produced by the nightly automation
rather than a person: .github/workflows/nightly-patch-release.yml
drafts Phase 1-2 on its own when a patch release is due, and stops before merge and
tag. Treat that PR exactly like a human-drafted one — review the notes, then resume
at the phase the table gives.
Hard rules (do not violate)
These map to CLAUDE.md. Re-read them whenever this skill runs.
- Plain language first, in every artifact — the release notes, PR body, Discord post, and your own between-phase status updates all follow CLAUDE.md → How You Communicate: lead with what a user can now do, layer version numbers, SHAs, and pipeline mechanics underneath. Say each point once — don't restate a highlight in the changelog and again in the announcement.
- No Claude attribution anywhere — not in PR titles, PR bodies, commit messages (no
Co-Authored-By: Claude ...trailer), release notes, code comments, or the Discord announcement. - No silent fallbacks — if a validator fails, a step times out, or a workflow run isn't found, stop with an actionable error. Do not retry blindly. Do not "proceed anyway."
- Release notes are bulleted, plain, and short — see Notes format in Phase 1. Every entry is one bullet, never a prose block; no narrative overview paragraph; no emoji; a hard word budget that is checked, not eyeballed. Sections in order:
## Breaking Changes(only if any),## What's New,## Bug Fixes,## Known Issues(only if any),## Contributors,## Full Changelog. Patch releases do not include apip installblock. - Match the previous release PR body shape exactly — read the most recent merged
Release vX.Y.ZPR (e.g.gh pr list --repo amd/gaia --state merged --search "Release v in:title" --limit 3). Open with# GAIA vX.Y.Z Release Notes(no MDX frontmatter in the PR body), end with aRelease checklistsection. Style drift here costs review cycles. - The
version.pybump and its notes land in the same commit.docs.ymlandtests/unit/test_docs_json_release.pyboth resolvedocs/releases/v<version>.mdxand thedocs.jsonnavbar label from the declared__version__, so bumping ahead of the notes (the old0.17.3→ "bump to 0.17.4 for development" pattern) turns every later docs and unit run red, not just the release PR. Never bump on its own. - Bulletproof commits only — every change made by this skill must satisfy the four criteria in CLAUDE.md (validated, critiqued, scope-clean, no half-finished work) before being committed.
- Pushing tags is irreversible. Always confirm the SHA the tag will point to and the green status of the pre-tag verification run before
git push origin v<version>. - Manual approval gate at the publish step is human-only — Claude cannot click the GitHub environment "approve" button. Surface the run URL and stop.
Phase 1 — Draft release notes (PR-ready)
Goal: produce docs/releases/v<version>.mdx matching house style; update navigation and the UI package.json.
Steps
-
Survey commits since previous tag, then sanity-check the version request against scope.
bashPREV=$(git tag --sort=-v:refname | grep -E '^v[0-9]+\.[0-9]+\.[0-9]+(\.[0-9]+)?$' | head -1) echo "Previous tag: $PREV" git log "$PREV..HEAD" --pretty=format:'%h %s' echo echo "Total: $(git log $PREV..HEAD --oneline | wc -l) commits" echo "Feat: $(git log $PREV..HEAD --oneline | grep -cE '^[a-f0-9]+ feat') feat commits" echo "Fix: $(git log $PREV..HEAD --oneline | grep -cE '^[a-f0-9]+ fix') fix commits"Group commits by theme (features, fixes, infra, docs). Extract the PR number from each subject (
(#NNN)). For each non-trivial entry, open the linked PR or commit body to get the why — release notes need motivation, not just titles.Scope-vs-version sanity check (do not skip): apply this rubric against the requested target version:
Commit shape Suggested release shape Only fix/docs/chore/ci, nofeatPatch ( vX.Y.Z+1) ✓1–2 featcommits, small scopePatch acceptable, mention them as "What's New" 3+ featcommits, or any commit titled... vX.Y.Z milestone, default-model swap, breaking change, package layout changeMinor ( vX.Y+1.0) — push backRemoval of a public API, CLI flag deletion, config schema break, version-pin floor raised in a non-additive way Major ( vX+1.0.0) — push backIf the requested version doesn't match the rubric, stop and surface the mismatch: "You asked for
v<requested>(patch). I see N feat commits since<prev>including<one or two examples>— this looks minor-shaped. Continue as patch, or bump tov<suggested>?" Do not silently proceed. -
Read recent release notes for structure only — frontmatter shape, section headings, PR-link format. Take length and tone from Notes format below, not from the files: v0.21.1 is the model at 267 narrative words, while v0.22.0 (1705) and v0.23.0 (1119) were flagged in review as too much text in too-blocky a form — they are the regression this section exists to prevent.
Patch releases are short. Minor/major releases carry more bullets and a
pip installblock — the per-bullet shape is identical, only the count grows. -
Create docs/releases/v.mdx with this skeleton (adapt to whether it's patch / minor / major):
mdx--- title: "v<version>" description: "<one plain line: what shipped. Not a pitch.>" --- # GAIA v<version> Release Notes <Optional single sentence naming what this release is about. Drop it entirely if the bullets already say it — that is the common case. Never a paragraph.> ## Breaking Changes - **<what changed>** — <what to do instead>. (PR [#NNN](https://github.com/amd/gaia/pull/NNN)) ## What's New - **<what the user can now do>** — `<gaia command>`. <At most one more sentence, only if the title genuinely needs it.> (PRs [#NNN](https://github.com/amd/gaia/pull/NNN), [#NNN](https://github.com/amd/gaia/pull/NNN)) ## Bug Fixes - **<what was broken>** — <what now happens>. (PR [#NNN](https://github.com/amd/gaia/pull/NNN)) ## Known Issues - **<what still does not work>** — <workaround, or "tracked in [#NNN](https://github.com/amd/gaia/issues/NNN)">. ## Contributors - [@handle](https://github.com/handle) — <what they contributed>. (PR [#NNN](https://github.com/amd/gaia/pull/NNN)) ## Full Changelog **N commits** since v<previous>: - `<sha>` — <subject> - ... Full Changelog: [v<previous>...v<version>](https://github.com/amd/gaia/compare/v<previous>...v<version>)Omit
## Breaking Changesand## Known Issueswhen empty — no "None" placeholder. Drop the---rules between sections; the headings already separate them.Generate the changelog by introspecting git, and escape it for MDX. Do not hand-transcribe subjects, and do not pipe
git logoutput in raw — a subject containing<or{is valid git and invalid MDX, which fails CI'smintlify validate. v0.22.0 hit this on#1791's subject ("fix Mintlify MDX validation (unescaped '<' breaks the docs check)") — the one commit whose subject describes the bug it causes.bash# Generate the `- `<sha>` — <subject>` lines, escaping the two characters MDX # treats as syntax. `\<` and `\{` render as literal `<` / `{` (verified against # `mintlify validate`), and shas never contain them, so escaping the whole line # is safe. Subjects like "AMD <> SpecificAI" survive correctly. git log v<previous>..HEAD --pretty=format:'- `%h` — %s' \ | sed -e 's/</\\</g' -e 's/{/\\{/g' > /tmp/changelog.txt # Guard: nothing hazardous may survive. grep -nE '(^|[^\\])[<{]' /tmp/changelog.txt && echo "MDX HAZARD — fix before continuing" || echo "changelog MDX-safe"Sanity-check the count against
git log --oneline | wc -lafter writing the file (wc -lunder-counts by one when the last line has no trailing newline — v0.22.0 claimed 197 when the real count was 198). The claimed count, the listed lines, andgit logmust all agree.Notes format (apply to every entry — this is the point of the skill). Tone is already governed by CLAUDE.md → How You Communicate — plain language, outcome first, each point made exactly once. Do not restate or soften it here. What is release-notes-specific is the shape:
- Bullets, not paragraphs. Every entry is one bullet: a bold clause naming what the
user can now do, then at most one sentence, then the PR links. No
###prose blocks insideWhat's New. - Say it once. A highlight appears in the bullet list or in the opening sentence,
never both. The frontmatter
descriptionis not a third copy. - Factual, not promotional. State the capability and stop. Banned: emoji, "we're excited to announce", "finally", "blazing(-fast)", "game-changer", "seamless", "powerful", "unlock", "Here's the good stuff", "makes X effortless", "no more crashes", invented benchmarks.
- Order by what users run. Agents and commands first (
gaia hub,gaia email, …); SDK, CI, and refactor work last, or omitted when it has no user-visible effect. - Word budget — a cap, not a target. ≤ 350 words for a patch, ≤ 600 for a
minor/major, measured from the top down to the first of
## Bug Fixes/## Known Issues/## Contributors/## Full Changelog. Those four are reference lists — their length is set by how many fixes actually shipped, and squeezing them hides work. The cap is on the narrative part, which is what bloats. Over budget means cut entries, not reflow them. Step 8 checks it. (Calibration, same measure: v0.21.1 shipped at 267, v0.22.0 at 1705, v0.23.0 at 1119.)
Example — one highlight:
Bad (prose block, promotional, three sentences where one works):
Triage your inbox from the terminal —
gaia emailPoint GAIA at your inbox and it sorts the noise from what needs you: drafts replies to routine mail, flags what's urgent, leaves the rest. Runs locally, so your mail never leaves your machine. Try it:
gaia email.Good (one bullet, plain, factual):
- Triage your inbox from the terminal —
gaia emailsorts mail, drafts replies to routine messages, and flags what needs you. Runs locally. (PR #1234)
- Bullets, not paragraphs. Every entry is one bullet: a bold clause naming what the
user can now do, then at most one sentence, then the PR links. No
-
Update docs/docs.json:
- Add
releases/v<version>to the Releases tab. - Bump the navbar label (e.g.
v<previous-version> · Lemonade <previous-lemonade>→v<version> · Lemonade <current-lemonade>). Read src/gaia/version.py for theLEMONADE_VERSIONconstant — it is the source of truth, and the navbar may be drifted from it (Lemonade bumps land outside release PRs).
- Add
-
Sync the UI package version.
bashnode installer/version/bump-ui-version.mjsConfirm src/gaia/apps/webui/package.json now reads the new version.
-
Bump the hub component manifests. The terminal hub and Agent UI publish to the Agent Hub R2 catalog on a version tag, and
release_components.ymlgates the tag against each manifest — R2 paths are immutable, so publishing under a stale version cannot be undone. A mismatch fails the release, by design.bashsed -i '' "s/^version: .*/version: <version>/" \ hub/components/terminal-hub/gaia-agent.yaml \ hub/components/agent-ui/gaia-agent.yaml # GNU sed: drop the '' grep -H '^version:' hub/components/*/gaia-agent.yaml # both must read <version>Then bump the terminal hub's
min_gaia_versionto<version>too. It needs daemon host API v1.1, which no release before this one ships — so the release it publishes alongside is its minimum, andrelease_components.ymlrefuses to publish it under an older core. Leaveagent-ui's alone: it talks togaia.ui.server, not the daemon control plane.bashsed -i '' "s/^min_gaia_version: .*/min_gaia_version: \"<version>\"/" \ hub/components/terminal-hub/gaia-agent.yaml # GNU sed: drop the '' python util/check_component_core_api.py --release-version <version>Finally, record what the previous release shipped in
RELEASED_DAEMON_APIin util/check_component_core_api.py — the guard resolves an already-published version from that table, and a missing row makes it fall back to trusting this tree, which is the drift it exists to catch. -
Confirm
__version__is correct.bashgrep -E '^__version__' src/gaia/version.pyThe post-prior-release bump usually handles this, but a squash-merge can revert it silently. If it's wrong, edit it. If it's right but reverted later (see Phase 3), the validator will catch it.
-
Validate — both checks. Run from the repo's activated venv (
source .venv/bin/activateon Linux/macOS,.venv\Scripts\activateon Windows; the bare-pythonMicrosoft Store stub will fail). If you're working from a git worktree without its own venv, run from the parent checkout's venv.bashpython util/validate_release_notes.py docs/releases/v<version>.mdx --tag v<version> (cd docs && npx -y mintlify@latest validate) # the docs.yml `validate` job — MUST also pass # Word budget from *Notes format* — the narrative part only, not the reference lists. # Cap: 350 words for a patch, 600 for a minor/major. Over budget means cut entries. awk '/^## (Bug Fixes|Known Issues|Contributors|Full Changelog)/{exit} {print}' \ docs/releases/v<version>.mdx | wc -wBoth must exit 0. Fix any errors before continuing. If either fails for reasons unrelated to your changes (missing dep, broken import), stop and surface that — do not silently bypass.
validate_release_notes.pyprints the first failing check (missing/renamed section, absentcompare/link, tag mismatch) — read that line to localise the fix; it has no--verboseflag.validate_release_notes.pypassing is not sufficient — it is not an MDX parser. CI'svalidatejob additionally runsmintlify validatefromdocs/, and v0.22.0 failed it after the notes passed the Python validator (see the escaping rule in step 3). Its error is misleading: an unparseable.mdxsurfaces as"releases/v<version>" is referenced in the docs.json navigation but the file does not exist— the file exists, it just never parsed. Pre-existing parse errors underdocs/plans/anddocs/superpowers/are reported but non-fatal; leave them alone.
Gate 1 — show the user the draft
Show the diff (git diff --stat plus the new .mdx file inline). Ask: "Approve these release notes and continue to Phase 2 (open PR)?" Wait for explicit yes. Iterate on tone/wording before continuing — much cheaper than fixing on main.
Phase 2 — Open the release PR
Goal: branch, commit, push, open PR, request review.
Steps
-
Branch and commit.
bashgit checkout -b v<version>-release git add docs/releases/v<version>.mdx docs/docs.json src/gaia/version.py \ src/gaia/apps/webui/package.json src/gaia/apps/webui/package-lock.json \ hub/components/terminal-hub/gaia-agent.yaml \ hub/components/agent-ui/gaia-agent.yaml git status # confirm scope-clean — no drive-by edits git diff --cached --stat git commit -m "Release v<version>" git push -u origin v<version>-releaseIf
git statusshows anything outside those five files, stop and ask (bump-ui-version.mjsrewritespackage-lock.jsontoo — the prior release commit carries all five). The release PR must be scope-clean. -
Read the most recent release PR body to match shape.
bashgh pr list --repo amd/gaia --state merged --search "Release v in:title" --limit 3 \ --json number,title,body | jq -r '.[0]'Use that as the structural template for this PR body — same opening, same checklist, same section order. Do not invent a new shape.
-
Prepare the PR body file. Build it by stripping the MDX frontmatter from the release notes and appending the
Release checklistsection copied from the previous release PR.bash# Strip the YAML frontmatter (everything between the first two '---' lines) awk '/^---$/{c++; next} c>=2' docs/releases/v<version>.mdx > /tmp/release-body.md # Append the Release checklist section from the previous release PR PREV_PR=$(gh pr list --repo amd/gaia --state merged --search "Release v in:title" --limit 1 --json number --jq '.[0].number') gh pr view "$PREV_PR" --repo amd/gaia --json body --jq '.body' \ | awk '/^## Release checklist/{found=1} found' \ >> /tmp/release-body.md # Sanity-check the body before opening the PR head -3 /tmp/release-body.md # should start with "# GAIA v<version> Release Notes" tail -10 /tmp/release-body.md # should end with the checklistIf the awk pipeline produces an empty file, the previous PR didn't have a
## Release checklistheading — fall back to copying the entire body and editing it manually before continuing. -
Open the PR. Title:
Release v<version>. Body: the file you just built.bashgh pr create --repo amd/gaia \ --title "Release v<version>" \ --body-file /tmp/release-body.md \ --reviewer kovtcharov-amdSurface the resulting PR URL.
Gate 2 — wait for merge
Stop. Tell the user: "PR opened at . Waiting for review/merge before pre-tag verification. Re-invoke this skill (or /loop) to continue once merged."
Do not poll, do not auto-merge. Do not push the tag from the un-merged branch.
Phase 3 — Pre-tag verification (the hard-won gate)
Goal: prove the merged commit on main actually builds clean before the tag locks it in. Do not skip this. Two release-blocking bugs in v0.17.4 would have shipped without this step.
Steps
-
Sync local main to the merged commit, and capture the SHA from the release PR (not just the local HEAD).
bashgit checkout main && git pull # Re-derive from the release PR — survives gate pauses across sessions/shells. RELEASE_PR=$(gh pr list --repo amd/gaia --state merged --search "Release v<version> in:title" --json number --jq '.[0].number') MERGED_SHA=$(gh pr view "$RELEASE_PR" --repo amd/gaia --json mergeCommit --jq '.mergeCommit.oid') echo "Will tag: $MERGED_SHA (from PR #$RELEASE_PR)" test "$(git rev-parse HEAD)" = "$MERGED_SHA" || { echo "Local main ($( git rev-parse HEAD)) does not match merged release PR SHA ($MERGED_SHA) — pull, or main has moved past the release commit"; exit 1; }When you reach Gate 3 below, carry both
$RELEASE_PRand$MERGED_SHAinto the gate question text so Phase 4 can re-derive them from the answer rather than depending on shell variables that don't survive the pause. -
Re-verify
__version__post-merge. Squash-merges have silently reverted this. If it doesn't match the target version, stop — open a follow-up PR to fix it before tagging. Never tag a wrong version.bashgrep -E '^__version__' src/gaia/version.py -
Re-run the release notes validator on the merged tree.
bashpython util/validate_release_notes.py docs/releases/v<version>.mdx --tag v<version> -
For a final: check the flagship agent was released from this commit.
gaia init,gaia hub install gaiaand the Windows setup install the flagship from the Agent Hub andhub/agents/gaia/npm/binaries.lock.json, which onlyrelease_agent_gaia.yml(tagagent-pkg-gaia-vX.Y.Z) fills.publish.ymlandrelease_components.ymlrun this same check on a final tag and fail it; run it first so a missing flagship release is found before the tag exists. (An RC skips it: it freezes its own flagship.)bashgit fetch --tags origin python util/check_flagship_release.py --ref "$MERGED_SHA"If it fails, release the flagship in this order. Skipping step 2 deadlocks:
release_agent_gaia.ymlrefuses to run until the terminal-hub version the lock pins is on the hub, and terminal-hub normally publishes only on the final tag, which this gate blocks.- Bump the flagship in a PR: edit
__version__inhub/agents/gaia/python/gaia_agent/__init__.py, runpython hub/agents/gaia/python/packaging/stamp_version.py, add ahub/agents/gaia/npm/CHANGELOG.mdentry. - Publish terminal-hub first. Dispatch Publish Hub Components (R2) (
release_components.yml) on main withdry_rununchecked; a dispatch skips this gate. Its agent-ui and Windows-setup legs fail because there is no GitHub release or new flagship yet; they publish nothing. Skip this ifhub/components/terminal-hub/gaia-agent.yaml's version is already on the hub. - Tag
agent-pkg-gaia-v<flagship-version>on the merged commit and approverelease_agent_gaia.yml. - Land the
binaries.lock.jsonit regenerated (its log prints it) and re-run Phase 3 on that commit. Code the binary is frozen from must not change between the agent tag and the core tag.
- Bump the flagship in a PR: edit
-
Trigger Build Installers via
workflow_dispatchagainst the merged SHA. This builds the same artifacts the tag push will build, on the same code, before the tag exists.bashgh workflow run "Build Installers" --repo amd/gaia --ref main sleep 5 RUN_ID=$(gh run list --repo amd/gaia --workflow "Build Installers" --limit 1 --json databaseId -q '.[0].databaseId') echo "Watching run $RUN_ID" gh run watch "$RUN_ID" --repo amd/gaiaAppImage smoke jobs (
appimage-distro-matrix,appimage-userns-restricted) are the most flaky —appimage-userns-restrictedhas a 300sstate: readypoll (raised from 90s to cover a first-run model download). On real failure, stop and fix root cause; on transient flake (timeout-only, no logic error),gh run rerun $RUN_ID --failedis acceptable.
Gate 3 — green run on the merged commit
Required before continuing:
__version__matches the target.validate_release_notes.pypasses.- For a final,
check_flagship_release.pypasses. - Build Installers run on the merged commit is green (not yellow, not "mostly green except smoke tests").
Show the user the run URL, the release PR number (#$RELEASE_PR), and the merged SHA ($MERGED_SHA). Ask: "Pre-tag verification green on <sha> (release PR #N). Push tag v<version>?" — the SHA and PR number being in the question text means Phase 4 can re-derive them even if the user resumes in a fresh shell.
Phase 3.5 — Release candidate (default for minor/major; ask for a patch)
Goal: put the exact build users will get in front of a tester before it becomes the default. An RC tag runs the same publish.yml, behind the same approval gate, but stays off every stable channel:
Final v<version> | RC v<version>-rcN | |
|---|---|---|
| GitHub Release | normal, becomes latest | pre-release, never latest |
| PyPI | <version> | <version>rcN — pip install amd-gaia ignores it without --pre |
npm @amd-gaia/agent-ui | dist-tag latest | <version>-rc.N under dist-tag next |
Agent Hub R2 (release_components.yml) | publishes | does not run — the catalog has no pre-release channel |
| Website | stable downloads | a collapsed "Try the release candidate" entry, hidden once the final ships |
Flagship agent (gaia-agent) | the hub's, from release_agent_gaia.yml | frozen from the RC's commit by rc-flagship, attached to the pre-release; the RC core installs it from there and never from the hub |
Context7 refresh, release branch, release-notes bot | run | skipped |
| Release notes | docs/releases/v<version>.mdx | the same file — no RC-specific notes |
version.py stays <version>; the pipeline stamps the RC suffix from the tag (util/release_tag.py), so the final can be tagged on the RC's commit with no further version bump. The terminal's .pkg/.deb/.rpm packages are not built for an RC (they need a numeric-only version); the RC carries the one-step Windows setup (gaia-X.Y.Z-rc.N-win-x64-setup.exe, built by publish.yml's rc-windows-setup around that run's own desktop app build — never published to R2), the desktop app installers, the raw terminal binaries, and the wheel/sdist. A failed setup leaves it off the pre-release and turns the run red via rc-windows-setup-missing.
Steps
-
Dry-run the classification — a malformed tag fails
validatebefore anything builds, but catching it locally is cheaper:bashpython util/release_tag.py classify v<version>-rc1 python util/validate_release_notes.py docs/releases/v<version>.mdx --tag v<version>-rc1IS_RC=true,PEP440_VERSION=<version>rc1,NPM_VERSION=<version>-rc.1. Anything else (-rc.1,rc1without the dash,-rc0,-beta1) is rejected. -
Tag the verified SHA from Gate 3 and push — same confirmation rule as Phase 4: the SHA and the green pre-tag run go in the question.
bashgit tag -a v<version>-rc1 <merged-sha> -m "Release candidate v<version>-rc1" git rev-list -n1 v<version>-rc1 # MUST equal <merged-sha> git push origin v<version>-rc1 -
Monitor and approve exactly as in Phase 5. The approval gate is the same
publishenvironment. There is norefresh-context7job to wait for;redeploy-website-rcasks the website to rebuild so the RC entry appears. -
Personal test pass on a machine that does not already have the release installed (CLAUDE.md: test from the user's real initial state):
bashpip install amd-gaia==<version>rc1 && gaia -v # must print <version>rc1 npm view @amd-gaia/agent-ui dist-tags # latest unchanged, next = <version>-rc.1 gh release view v<version>-rc1 --repo amd/gaia # Pre-release, installers + gaia-agent-* attached gaia init && gaia hub list # gaia agent version ends in +v<version>-rc1Then install the desktop app from the pre-release and walk the golden paths the release notes advertise.
-
Problems → fix forward, then
-rc2. Land the fix onmainthrough a normal PR, re-run Phase 3 on the new merged SHA, and tagv<version>-rc2from step 2. Never move or delete an RC tag — each one is a published version on PyPI and npm.
Gate 3.5 — RC passed
Ask: "v<version>-rcN passed testing on <sha>. Tag the final v<version> on that same SHA?" Phase 4 then tags that SHA (re-derive it with git rev-list -n1 v<version>-rcN), not necessarily the release PR's merge commit.
Phase 4 — Tag and trigger the publish pipeline
Goal: push the annotated tag; do nothing else.
Steps
-
Confirm you are on
mainat the verified SHA. Re-derive$MERGED_SHAfrom the release PR (the SHA from Gate 3) — do not trust shell state across the gate pause.bashgit checkout main git pull # Re-derive from the release PR — the PR number was in the Gate 3 question. MERGED_SHA=$(gh pr view <release-pr-number> --repo amd/gaia --json mergeCommit --jq '.mergeCommit.oid') test "$(git rev-parse HEAD)" = "$MERGED_SHA" || { echo "main moved past verified SHA $MERGED_SHA — re-verify (Phase 3) before tagging"; exit 1; } -
Tag and push.
bash# Annotated, on the verified SHA — every prior release tag is annotated ("Release vX.Y.Z"). # A bare `git tag v<version>` creates a lightweight tag and breaks convention. git tag -a v<version> <merged-sha> -m "Release v<version>" git rev-list -n1 v<version> # MUST equal <merged-sha> before pushing git push origin v<version> -
Confirm the publish workflow picked it up.
bashsleep 10 gh run list --repo amd/gaia --workflow publish.yml --limit 3The flow is:
validate → build (build-pypi + build-npm + build-desktop-installers) → approve-publish (manual gate) → publish (publish-pypi + publish-npm) → post-publish-smoke → github-release → refresh-context7. Surface the run URL.
Phase 5 — Monitor and approve
Goal: watch the run, distinguish flake from real failure, surface the manual approval gate to the human.
Steps
-
Watch the run.
bashgh run watch <run-id> --repo amd/gaia -
On flake (timeout-only, no logic error in the failing step's logs):
gh run rerun <run-id> --failedis non-destructive — re-runs only failed jobs and their downstream. Do not rerun the whole workflow; that wastes the publish budget. -
On real failure (logic error, missing artifact, validator failure that wasn't there before): stop. Do not push past a red build. Do not delete and re-tag — that path is messy. Surface the failing step + log to the user.
-
Manual approval gate — when the run reaches the
approve-publishjob (gated on thepublishGitHub environment), surface the URL to the user. Tell them: "Manual approval required at . Claude cannot click this — please approve in browser when ready." Then wait. -
After approval, PyPI + npm + GitHub Release jobs run in parallel. Watch to completion.
-
refresh-context7is the terminal job and may legitimately fail — that is not a release failure. This job runs after PyPI, npm, GitHub Release, and the desktop installers are already published. Context7's API rejects refresh requests inside a short cooldown window (observed: ~3–6 days between releases) with HTTP 429 (rate-limited); the workflow tolerates 429 and treats any other status as a hard failure. If the job is red, the release is still live. Open the job log, read theResponse body:block between the::stop-commands::markers, and either accept the cooldown reason or file a follow-up about a new rejection cause. Do not delete or re-tag — see the recovery guidance in the Notes section below. -
Never add
set -xorcurl -vto therefresh-context7step to "debug" a failure. GHA only masks the verbatim secret value;curl -vprints theAuthorization: Bearer <token>header, which is a transformed form that GHA's masking does not catch. Read the captured response body instead — it carries the same diagnostic signal without the leak risk.
Phase 6 — Post-release verification + announcement
Goal: confirm artifacts are live, draft the Discord announcement.
Steps
-
Smoke test the published wheel.
bashpip install --upgrade amd-gaia==<version> gaia -vMust report
<version>. If it reports the previous version, the squash-mergeversion.pyregression slipped through — escalate immediately. -
Verify the GitHub release page.
bashgh release view v<version> --repo amd/gaiaRequired artifacts:
.whl,.tar.gz,.deb,.AppImage,.dmg,.exe, and thelatest*.ymlfiles for the Electron auto-updater. If any are missing, the corresponding build job didn't run or didn't upload — investigate. -
Draft the Discord announcement. Read the just-shipped release notes (
docs/releases/v<version>.mdx) to populate the highlight list — reuse theWhat's Newbullets near-verbatim plus any Bug Fix worth surfacing. Same Notes format rules apply: bullets, plain, factual, no emoji.The template and format rules live in
reference/discord-announcement.md— read it and copy the template verbatim. Every element in it is load-bearing (the@gaiarole mention, the backticked version, the fenced install block, the fixed boilerplate sentences); do not improvise the shape.
Gate 6 — present, do not auto-post
Show the user the smoke-test output, the artifact list, and the drafted Discord announcement in a fenced markdown block. Ask: "Post the announcement to Discord?" Wait for explicit yes — Discord posting is human-only (Claude does not have Discord access here, and announcements are visible to all users).
Output between phases
After every phase, output:
- Phase N complete.
- What changed / what was verified (1–3 bullets).
- Concrete artifact (URL, SHA, file path, or fenced draft).
- Gate question (always ends with
?, names the next destructive action).
Do not bundle two phases into one user prompt. The gates exist for review.
Notes
- The argument-passing convention is the target tag, not the previous tag. If the user says "release v0.17.5", that is what gets created — the previous tag is the newest final tag (Phase 1 step 1's
PREV) — never an-rcNtag, or the notes would cover only the changes since the candidate. - Hotfix releases (
v0.15.4.1) follow the same flow; thevalidate_release_notes.pycheck accepts the four-part form. - Release candidates are
v<version>-rcNonly (Phase 3.5). There is no RC form of a four-part hotfix tag. - Minor/major releases (
v0.18.0,v1.0.0) need a richer notes structure — a "Highlights" block, thepip installinstructions, and migration notes if breaking. The skeleton above is patch-shaped; expand for non-patch releases by mirroring the prior minor/major release notes. - If the publish run fails partway through (e.g. PyPI publishes but npm doesn't), do not delete the tag and start over. Resolve the failing job, rerun only that job (
gh run rerun <run-id> --failed), and let the rest of the pipeline complete idempotently. The tag is the source of truth — preserve it.


