feat: remove full eval from PR builds, add daily scheduled cron eval

[hadarzaguri]

Summary

• PR builds: removed the 30-min EvalForge eval poll entirely. PRs now run lightweight validation only — YAML parse, file path existence, and tag name check against EvalForge. Fast (~seconds), cheap, still catches structural errors before merge.
• Daily cron (skill-eval-scheduled.yml): new workflow runs at 8am UTC every day. Finds yaml/wix-manage/**/documentation.yaml files changed in the last 25 hours, collects their tags, and runs the full eval against those tags. If nothing changed → skips. On failure → opens a GitHub Issue labeled skill-eval-regression. On recovery → auto-closes the issue.
• Action refactored: mode input added (required: pr | scheduled). runPr and runScheduled are separate code paths. Removed unused formatEvalPassed/Failed/Timeout/NoScenarios. Unified scheduled formatters into single formatScheduledResult(status, sha, opts).

One-time setup (before first cron fires)

gh label create skill-eval-regression --color e11d48 --description "Skill eval regression detected by scheduled run"

Note

dist/index.js needs to be rebuilt locally before merging:

cd .github/actions/skill-eval && yarn install && yarn build

The skill-eval-ci.yml workflow will verify this on the PR.

Test plan

• Open a PR touching yaml/wix-manage/ — confirm workflow completes in seconds, PR comment shows validation passed/failed (no EvalForge eval run)
• Trigger skill-eval-scheduled via workflow_dispatch — confirm it runs the full eval and logs correctly
• Confirm a regression creates a skill-eval-regression issue and a passing run closes it

🤖 Generated with Claude Code