* fix(ci): store metadata suggestions in the review instead of actions/cache
pull_request_target runs can't write to the Actions cache ("cache write
denied: token has no writable scopes"), so the cache was never saved and
the model was asked again on every push. That produced new suggestions
anchored on the same lines as earlier ones.
Store the model's suggestions per icon name as hidden JSON in the review
body and read them back from the bot's earlier reviews, so the API is only
called for icons without stored suggestions. This drops the gh api and
actions/cache steps.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TSdUq9uFpjuQsoobCBqz4j
* fix(ci): handle earlier metadata suggestion comments on older PRs
- Read values from the oldest comment format too, which doesn't start with
"Suggested `field`" and was posted without a closing fence; the field
is taken from the suggestion's `"field": [` line.
- Skip a field while an earlier suggestion for it is still applicable (not
outdated), so two suggestions never replace overlapping lines.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TSdUq9uFpjuQsoobCBqz4j
* fix(ci): read icon metadata at the reviewed commit
The review comments are anchored to the commit that triggered the run,
but the icon JSON was read through the PR files' raw_url, which always
points at the latest head. When a commit was pushed while the action
was running, suggestions were anchored to the wrong lines, e.g. a
use-cases block replacing the "tags": [ line, producing invalid JSON.
Read each file at the reviewed commit instead, default that commit to the
PR's head SHA, and skip files that don't exist at it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TSdUq9uFpjuQsoobCBqz4j
---------
Co-authored-by: Claude <noreply@anthropic.com>
* feat(ci): run metadata suggestions on every PR update with cached results
- Run on opened/reopened/synchronize and queue runs per PR.
- Cache the model's suggestions per icon name with actions/cache, so the
OpenAI API is only called again for icons that are added or renamed.
- Skip suggestions that were already posted instead of skipping the whole
PR once it has a review.
- Suggest only the values to append after the last array item instead of
the whole array; empty and inline arrays still get the full block.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TSdUq9uFpjuQsoobCBqz4j
* refactor(ci): key the suggestions cache on the PR's icon names
Replace the separate restore/save steps and the `cache-updated` output
with a single actions/cache step keyed on a hash of the changed icon
files, so the API is only called again when icon names change.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TSdUq9uFpjuQsoobCBqz4j
* style(ci): drop trailing blank line in suggest-metadata action
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TSdUq9uFpjuQsoobCBqz4j
* fix(ci): dedupe metadata suggestions by value instead of comment body
The append-style suggestion includes the array's current last item, so
comparing whole comment bodies reposted the same values after a
contributor added a tag or applied part of a suggestion. Read the values
back from the bot's earlier suggestion blocks per file and field, and
exclude them like the current values.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TSdUq9uFpjuQsoobCBqz4j
---------
Co-authored-by: Claude <noreply@anthropic.com>
* fix(scripts): keep contributors out of metadata suggestions
The shared metadata instructions include a Contributors section asking to
add the PR author, but the suggestion schema has no contributors field, so
the model leaked it into use-cases. Strip that section from the prompt and
explicitly restrict suggestions to tags, categories and use-cases.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MThCx7cVXV3XWfo5C327Ci
* refactor(ci): move metadata suggestions into a dedicated composite action
Move scripts/suggestMetaData.mts to .github/actions/suggest-metadata and
wrap it in a composite action with explicit inputs. The workflow now uses
the local action; package.json scripts point to the new location.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MThCx7cVXV3XWfo5C327Ci
* chore(ci): add .env.example for testing the suggest-metadata action locally
Document the env vars the script reads, point the debug/watch scripts at
an action-local .env, and fix the USE_FILE_SYSYEM typo.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MThCx7cVXV3XWfo5C327Ci
* chore(ci): make .github/actions a pnpm workspace
Add .github/actions/* to the pnpm workspace and give the suggest-metadata
action its own package.json with its dependencies (openai, zod, dotenv,
@octokit/rest). openai, zod and dotenv move out of the root package.json;
@octokit/rest stays there for updateContributors. The workflow now only
installs the action's dependencies, and the root suggest:metadata scripts
delegate to the action package.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MThCx7cVXV3XWfo5C327Ci
* Upgrade to gpt-6-luna
---------
Co-authored-by: Claude <noreply@anthropic.com>
