Bump Image

Bump Image Tag in deployment-charts

Update the Docker image tag for one or more LiftWing inference services (isvcs) in operations/deployment-charts after Jenkins publishes a new image.

Scope note: This skill is intentionally limited to bumping image tags. If it ever expands to handle other values.yaml changes (env vars, replica counts, etc.), it should be renamed to reflect that broader scope.

Input

$ARGUMENTS — <cl-number> [isvc-name ...]

• cl-number: the inference-services Gerrit CL number where the patch merged and PipelineBot posted the published image (e.g. 1279278). Required.
• isvc-name: zero or more isvc names to pre-filter candidates (e.g. revertrisk-language-agnostic revertrisk-multilingual). Optional. When omitted, all isvcs whose image matches the published image are shown for selection.

Note: isvc names and image names do not map 1:1. For example, both revertrisk-language-agnostic and revertrisk-language-agnostic-pre-save share the image machinelearning-liftwing-inference-services-revertrisk. Always confirm the exact set of isvcs to update in the interactive step below.

Steps

1 + 1b + 2. Parallel data fetch (do all three in one batch)

Make the following three calls in parallel — they are independent and should never be sequential:

1. PipelineBot images — WebFetch the Gerrit messages endpoint:

URL:    https://gerrit.wikimedia.org/r/changes/<cl-number>/messages
Prompt: Extract only Docker image references published by PipelineBot matching
        machinelearning-liftwing-inference-services-*:*-publish.
        Return as a compact list, one entry per line: <image-name>:<tag>.
        Omit all other message content, timestamps, and metadata.

Note: mcp__gerrit__list_change_comments cannot be used here — it does not return full message bodies. Replace with that tool once the gerrit-mcp-server exposes /changes/<id>/messages.

1b. Changed files for auto-pre-filter — call mcp__gerrit__list_change_files on the same CL. Look for paths under src/models/<model-name>/. The model directory name is a strong signal for which image was the primary target. Cross-reference against the image list from step 1 — images whose name contains the model directory name are pre-selected. If isvc-name args were provided in $ARGUMENTS, use those instead and skip inference.

2. Phabricator task — call mcp__gerrit__get_commit_message on the same CL and extract the Bug: TXXXXX trailer. Reuse it in the deployment-charts commit message.

If no PipelineBot messages are found, ask the user to paste the image reference(s) directly.

3. Locate the deployment-charts repo

Check the DEPLOYMENT_CHARTS_PATH environment variable. If it is set and points to a valid directory, use it as the repo root.

If it is not set, ask the user:

"What is the path to your local operations/deployment-charts checkout?"

Then tell them they can avoid this prompt in future by adding the path to .claude/settings.local.json (gitignored, so safe for local paths):

{
  "env": {
    "DEPLOYMENT_CHARTS_PATH": "/path/to/operations/deployment-charts"
  }
}

4. Find candidate isvcs

Use Grep (not Read) to search helmfile.d/ml-services/*/values.yaml and helmfile.d/ml-services/*/values-ml-staging-codfw.yaml for lines containing the image name from step 1. Grep returns only matching lines — do not read entire files at this stage. Also grep for the surrounding context lines (use -C 3 or similar) to capture the isvc name and current tag alongside the image reference.

If isvc-name args were provided, further filter candidates to only those matching the given names.

Show the user a preview table. Mark pre-selected entries with *:

  Namespace / file                                  isvc name                       Current tag                New tag
* outlink/values.yaml                               outlink-topic-model             2024-04-01-100000-publish  2024-05-07-143022-publish
  revertrisk/values.yaml                            revertrisk-language-agnostic    2024-04-01-100000-publish  2024-05-07-143022-publish
  ...

Ask the user to confirm the selection (pre-selected entries are the default) or adjust it, then choose scope: both prod + staging, staging only, or prod only.

5. Edit the values files

For each confirmed isvc + file combination, use Edit to replace the old tag string with the new tag. Only the tag string is replaced — no other YAML keys are touched.

6. Draft the commit message

Do not run git commit. Instead, draft the commit message and display it for the user to review, modify if needed, and commit themselves.

Commit message format (Wikimedia conventions):

• Subject: ml-services: update <primary-isvc-name> image tag to <new-tag>
  • Use the first/most significant updated isvc name as the primary one.
  • If multiple isvcs were updated, list the rest in the What: body.
• Body:

ml-services: update <primary-isvc-name> image tag to <new-tag>

Why:

- New Docker image published after patch merge in inference-services

What:

- Bump <primary-isvc-name> version to <new-tag>
- Bump <other-isvc-name> version to <new-tag>  (if applicable)

Assisted-by: Claude Sonnet 4.6
Bug: TXXXXX

7. Remind about next steps

Next steps:
  git add <modified files>
  git commit            # paste the drafted message above
  git review            # submit for review

  # After merge — on deployment.eqiad.wmnet, cd to the namespace dir:
  cd /srv/deployment-charts/helmfile.d/ml-services/<namespace>

  # Staging:
  helmfile -e ml-staging-codfw diff
  helmfile -e ml-staging-codfw sync

  # Production (both DCs):
  helmfile -e ml-serve-eqiad diff
  helmfile -e ml-serve-eqiad sync
  helmfile -e ml-serve-codfw diff
  helmfile -e ml-serve-codfw sync