Build recommendation quality feedback loop using skip history and mute preferences

[Shubb07]

Summary

Improves recommendation quality by incorporating contributor feedback signals into the recommendation ranking pipeline.

The recommendation engine now:

• tracks optional skip reasons
• down-ranks repeatedly skipped repositories
• down-ranks repeatedly skipped languages
• respects contributor mute preferences for repositories and languages
• preserves recommendation diversity through soft penalties instead of hard filtering

Type of Change

• Bug fix
• New feature
• UI / UX improvement
• Refactor
• Documentation
• Other

Related Issue

Closes #91

What was changed?

• Added optional skip_reason persistence for skipped recommendations
• Added centralized recommendation ranking penalties/constants
• Implemented repository skip down-ranking
• Implemented language skip down-ranking
• Added muted repository/language preference persistence
• Added muted repository/language ranking penalties
• Added efficient bulk skip-history aggregation with bounded history windows
• Avoided N+1 queries using shared in-memory lookup maps
• Preserved recommendation diversity through subtractive scoring only
• Added runtime-safe fallback handling for missing language/skip data

Screenshots

N/A (backend/recommendation pipeline enhancement)

Checklist

• My code follows the project structure and conventions
• I tested this locally (npm run dev)
• No hardcoded secrets or credentials
• I have updated documentation if needed

[vercel]

@Shubb07 is attempting to deploy a commit to the codersogs-3057's projects Team on Vercel.

A member of the Team first needs to authorize it.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
mergeship	Ready	Preview, Comment	May 20, 2026 2:04pm

[Siddhartha-singh01]

The penalty design here is solid subtractive soft penalties with documented
constants in constants.ts is exactly the right approach. But I ran CI's steps
locally and the PR isn't mergeable yet:

1. Typecheck fails. You added repoLanguage: string | null as a required field on
   ScoredIssue, but didn't update recommend.test.ts its issue() factory
   builds ScoredIssue objects without repoLanguage. Exact error:
   recommend.test.ts(4,61): error TS2322: ... not assignable to type 'ScoredIssue'.
   Fix: add repoLanguage (e.g. repoLanguage: null) to the factory's base object.

2. Format check fails on src/app/actions/profile.ts. Run npm run format and
   commit.

3. Migration numbering: 0010_recommendation_feedback.sql is the third 0010_
   migration main already has 0010_add_failed_webhook_events and
   0010_add_profile_fields. Please rename it to the next free number.

4. Most important for a "recommendation quality" PR: the new penalty logic (repo
   skip, language skip, muted repo/language) has no test coverage at all —
   recommend.test.ts wasn't touched. Please add tests for the penalties in
   rankScore so the ranking behavior is actually verified.

Run npm run typecheck, npm run format, and npm test locally all three need
to be green before this can merge. The design is good; it just needs the build
fixed and the new logic tested.

[Ayush-Patel-56]

Please see this #67 (comment)