Orchestrate the opendatahub-io release for all llm-d components in one cycle. Collects upstream(llm-d) versions, auto-discovers the release tracker issue, then spawns parallel sub-agents — one per component (release branch, Konflux onboarder workflow, PR validation, approve+merge, Quay image verify, GitHub draft release) plus one KServe metadata PR sub-agent — and posts the final #Release# tracker comment. Use when the release manager runs on/before ODH code-freeze date for the llm-d team.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Orchestrate the opendatahub-io release for all llm-d components in one cycle. Collects upstream(llm-d) versions, auto-discovers the release tracker issue, then spawns parallel sub-agents — one per component (release branch, Konflux onboarder workflow, PR validation, approve+merge, Quay image verify, GitHub draft release) plus one KServe metadata PR sub-agent — and posts the final #Release# tracker comment. Use when the release manager runs on/before ODH code-freeze date for the llm-d team.
Runs once per ODH release cycle by the release manager. Drives release
branch creation, the Konflux onboarder workflow, onboarder-PR validation
and merge, Quay image verification, KServe metadata PR, and the tracker
comment — for every llm-d component in one shot.
Single input:--version v3.5-ea2. Three values derive from it:
Use
Value
ODH release branch on each component repo
release-v3.5-ea2
Konflux onboarder version input + Quay image tag
v3.5-ea2
GitHub release tag on each component repo
odh-v3.5-ea2
Value flow (for --version v3.5-ea2)
This diagram shows where the version goes, what name it takes at each stop,
and which script flag receives it. Three derived values are colored below:
B = release branch, T = image tag (unchanged from VERSION),
R = GitHub release tag.
release manager
└─ /odh-llm-d-release --version v3.5-ea2
orchestrator (SKILL.md)
VERSION = v3.5-ea2
RELEASE_BRANCH= release-v3.5-ea2 (B = release_branch_prefix + VERSION)
IMAGE_TAG = v3.5-ea2 (T = VERSION, unchanged)
RELEASE_TAG = odh-v3.5-ea2 (R = github_release_tag_prefix + VERSION)
│
├── per-component sub-agent (one per component, in parallel)
│ step 1 create-release-branch.sh --branch B → branch release-v3.5-ea2
│ step 2 onboarder-trigger-gha.sh --version T → Konflux builds quay.io/.../<image>:v3.5-ea2
│ step 6 release-image-checker.sh --tag T → verifies quay.io/.../<image>:v3.5-ea2
│ step 7 create-github-release.sh --tag R → draft GH release tagged odh-v3.5-ea2
│
└── KServe metadata sub-agent
open-kserve-metadata-pr.sh --version VERSION → PR title/branch use v3.5-ea2
(the upstream component versions in --updates
are SEPARATE — collected from the release
manager via AskUserQuestion in step 1)
Step 0 — Parse arguments and load config
Parse $ARGUMENTS:
--version <ver> (required) — the ODH release version. Accepted forms:
GA: v3.5,
v3.5.0
Early Access: v3.5-ea1, v3.5-ea.1, v3.5.0-ea.1 (also v3.5.0-ea1)
Validate that --version matches ^v[0-9]+\.[0-9]+(\.[0-9]+)?(-[A-Za-z0-9.]+)?$.
If missing or malformed, prompt the user via AskUserQuestion and stop on cancel.
For the KServe metadata PR, gather the upstream version for every component
and every entry in kserve_extra_entries (vLLM today). Use a singleAskUserQuestion call so the release manager fills them all in at once
rather than answering one prompt at a time.
tracker-issue-finder.sh exited non-zero (GitHub API/search failure):
tell the user the search failed (show stderr), and ask whether to retry,
to provide an issue number directly, or to skip the final comment step.
count == 0 — tell the user no tracker issue matches ${VERSION}, ask
whether to skip the final comment step or to provide an issue number.
count == 1 — show the single match (title + URL) and confirm.
count > 1 — show the matches in an AskUserQuestion and let the user
pick.
Save TRACKER_ISSUE_NUMBER (or "" if skipped).
Step 3 — Spawn sub-agents in parallel
You will issue one message containing multiple Agent tool calls so they
all run in parallel. Build the prompts by reading the templates and
substituting variables.
3a — Per-component sub-agents
For each key in .components (iterate via yq):
Read references/per-component-task.md.
Substitute only SKILL_DIR, COMPONENT_KEY, VERSION (and optionally
QUAY_TIMEOUT). The sub-agent reads everything else (ODH repo,
default branch, workflow component, konflux files, Quay org, etc.)
directly from components.yaml.
Issue one Agent tool call with the substituted prompt and
subagent_type=general-purpose. Description: Release <COMPONENT_KEY>.
3b — KServe metadata PR sub-agent
Read references/kserve-pr-task.md.
Substitute SKILL_DIR, VERSION, UPDATES. The sub-agent
reads kserve_metadata_file and kserve_keep_entries from
components.yaml directly (upstream repo and target branch are
hardcoded in the script).
Issue one Agent tool call.
All sub-agent calls go in the same response message so the harness
launches them concurrently. Each sub-agent returns a JSON payload (see the
template files for the schema).
Step 4 — Wait for results and surface failures
When all sub-agent tool results return, parse each JSON payload.
For each sub-agent:
Show a one-line summary: component name + verdict (ok, partial,
failed) + key URLs (workflow run, PR, release, Quay images).
If any per-component sub-agent reports validation.status=fail, show the
problems array verbatim — the release manager will diagnose those
manually.
If any quay_images[].status is missing, surface a clear warning
block with each missing ref and browse_url. Say explicitly that the
release was completed without that image landing in the Quay-poll
window; the release manager needs to confirm whether the Konflux build is
still in progress or actually failed.
If any sub-agent reports errors, surface them.
Step 5 — Compose the #Release# tracker comment
If TRACKER_ISSUE_NUMBER was set (step 2), build the comment body. Format
per the doc:
component_name uses the registry's tracker_name (the short form used by
humans in the tracker, e.g. llm-d-router, wva, batch-gateway) — not
the full kserve_entry_name.
Image-pending marker. A component is still included in the tracker
comment when any of its quay_images reports missing — the release
branch was created, the onboarder PR was merged, and a draft GH release
exists, so the line carries genuine value. Append the literal text
(image pending — verify) (with a single space before the backtick) at
the end of the line in that case, so the
release manager (and the tracker readers) can tell at a glance which
components need a manual Quay check.
Write the body to a temp file. Show it to the user via AskUserQuestion —
options: "Post as-is", "Edit first", "Skip". On "Edit first", invite the
user to edit the temp file and re-confirm.
The KServe metadata sub-agent (per references/kserve-pr-task.md) calls:
scripts/open-kserve-metadata-pr.sh.
All scripts support --help.
Gotchas
Sub-agent parallelism is at-most-N. The harness may serialize if it
hits rate limits. Sub-agents are independent and idempotent (each
resource creation script checks-then-creates), so a re-run picks up where
it left off — never overwrites past work.
Konflux runs are asynchronous. Step 2 inside each per-component
sub-agent only triggers the workflow; step 3 polls for the resulting PR
(15 min budget, 1 min interval) and step 6 polls Quay for the image tag
(15 min budget, 1 min interval). The Quay budget is tunable via the
sub-agent's QUAY_TIMEOUT.
PR validation can fail safely. If the onboarder PR's tekton files
don't match the expected output-image, path-context, or dockerfile
for a component, the sub-agent reports the diff and refuses to merge.
The release manager fixes the registry (or the onboarder), then re-runs.
KServe PR uses a fork. The release manager normally doesn't have
write access to opendatahub-io/kserve; the KServe sub-agent forks
under the release manager's gh user, pushes there, and opens a
cross-fork PR against master.
The skill never publishes the GitHub release. Releases are created
as drafts; the release manager publishes manually after reviewing the
auto-generated notes.
The skill never auto-posts the tracker comment. It composes the
comment, shows the body, and asks before posting.
