[oyi77]

Summary

This PR completes and validates the 3.6.9 skills+memory investigation track by shipping provider-aware skills behavior end-to-end (backend + dashboard), improving AUTO injection quality, and fixing a runtime/dev blocker discovered during live validation.

Why this change

• Skills source/provider handling was not explicit enough in settings and install endpoints (skillsmp vs skillssh paths were not fully aligned with provider intent).
• AUTO mode relied on simplistic matching and could miss relevant skills or pick weak candidates.
• Skill enable/mode transitions had an inconsistency when toggling enabled state without an explicit mode.
• A stale tracked build artifact (src/lib/dataPaths.js) could shadow TS module behavior and break dev startup in certain flows.

What changed

1) Provider-aware skills settings and install routing

• Added persistent provider setting plumbing for skills provider selection.
• Enforced provider-aware install routing in skills marketplace endpoints.
• Enriched skill metadata to support better UI and decisioning:
  • mode
  • sourceProvider
  • tags
  • installCount
• Added migration:
  • src/lib/db/migrations/027_skill_mode_and_metadata.sql

2) Dashboard skills UX improvements

• Added provider selector to Memory & Skills settings.
• Improved skills page with:
  • search/filter refinements
  • provider-aware defaults/popular behavior
  • explicit per-skill controls: on / off / auto
  • clearer uninstall/install interactions

3) Smarter AUTO skill injection

• Expanded injection context passed into injectSkills(...) with model, formats, background reason, messages, and provider hints.
• Replaced naive matching with a scored contextual approach and guardrails:
  • thresholded selection (AUTO_MIN_SCORE)
  • top-N cap (AUTO_MAX_SKILLS)
  • bias for provider-relevant matches
  • preserved explicit on/off semantics

4) API consistency fix for enable/mode transitions

• PUT /api/skills/[id] now keeps enabled and mode coherent when clients toggle enabled without explicitly sending mode.

5) Runtime hardening and test stabilization

• Deleted stale tracked artifact src/lib/dataPaths.js (prevent JS shadowing TS export behavior at runtime).
• Stabilized translation-path assertions to match actual provider-preference behavior.
• Forwarded clientHeaders through BaseExecutor retry path to preserve request behavior across retries.

Commit breakdown

1. 85b6cabf feat: add provider-aware skills metadata and install routing
2. e275fa30 feat: improve skills dashboard filtering and provider-aware marketplace UX
3. ff012d0d feat: add scored AUTO skill injection with contextual signals
4. f1f3e8d2 fix: stabilize translation-path assertions and remove stale dataPaths build artifact
5. 32382598 fix: forward client headers through base executor retries

Files touched (high-level)

• Skills injection/runtime:
  • src/lib/skills/injection.ts
  • open-sse/handlers/chatCore.ts
  • src/app/api/skills/[id]/route.ts
  • open-sse/executors/base.ts
• Skills provider/backend:
  • src/lib/skills/providerSettings.ts (new)
  • src/shared/validation/settingsSchemas.ts
  • src/lib/skills/registry.ts
  • src/lib/skills/types.ts
  • src/app/api/skills/route.ts
  • src/app/api/skills/marketplace/install/route.ts
  • src/app/api/skills/skillssh/install/route.ts
  • src/lib/db/migrations/027_skill_mode_and_metadata.sql (new)
• Dashboard/UI:
  • src/app/(dashboard)/dashboard/settings/components/MemorySkillsTab.tsx
  • src/app/(dashboard)/dashboard/skills/page.tsx
• Test/runtime cleanup:
  • tests/unit/chatcore-translation-paths.test.ts
  • tests/unit/db-core-init.test.ts
  • tests/unit/skills-injection.test.ts
  • tests/unit/skills-provider-setting.test.ts
  • tests/unit/skills-skillssh.test.ts
  • src/lib/dataPaths.js (deleted)

Verification

All key checks passed on this branch:

npm run build
npm run typecheck:core
node --import tsx/esm --test tests/unit/skills-injection.test.ts
node --import tsx/esm --test tests/integration/skills-pipeline.test.ts

Additional run during investigation:

npm run test:coverage

No failing tests remained in the validated set after fixes.

Notes

• Push permissions for origin were unavailable (403), so this PR is submitted from fork branch oyi77:investigation/skills-provider-memory-3.6.9 into upstream release/v3.6.9.
• Local workspace contains unrelated untracked investigation artifacts not included in this PR.

[gemini-code-assist]

Code Review

This pull request introduces a more sophisticated skill management system, including per-skill activation modes (on, off, auto), a semantic scoring mechanism for automatic skill injection, and support for multiple skill providers (SkillsMP and skills.sh). The changes span the database schema, API routes, and the dashboard UI. Feedback focuses on optimizing the semantic scoring loop by pre-calculating constant tokens and improving context extraction for the Responses API format to ensure accurate automatic skill selection.

[gemini-code-assist]

The contextTokens and backgroundTokens are re-extracted from the context text and background reason for every candidate skill in the loop. Since these values are constant for the entire injection process, they should be computed once in injectSkills and passed to scoreAutoSkill to avoid redundant processing, especially as the number of installed skills increases.

[oyi77]

Addressed in 52345ef.

✅ Optimized AUTO scoring by precomputing contextTokens and backgroundTokens once per injectSkills(...) call, then passing them into scoreAutoSkill(...) for each candidate.

This removes repeated token extraction in the candidate loop while keeping scoring behavior unchanged.

[gemini-code-assist]

The skill injection context currently only includes body.messages. For requests using the Responses API format, the conversation history is stored in body.input. This means contextual AUTO skill selection will be less accurate for these requests. Consider including body.input when body.messages is missing.

Suggested change

	messages: Array.isArray(body.messages) ? body.messages : undefined,
	messages: Array.isArray(body.messages) ? body.messages : (Array.isArray(body.input) ? body.input : undefined),

[oyi77]

Addressed in 52345ef.

✅ chatCore now uses Responses API input as fallback context for skills injection:

• body.messages when present
• otherwise body.input when it is an array

Also added integration coverage in tests/integration/skills-pipeline.test.ts (responses input context participates in AUTO skill injection) to verify AUTO selection works for /v1/responses payloads.

[diegosouzapw]

Thanks @oyi77 for this great contribution! 🎉 We've resolved the merge conflict locally and integrated it directly into the release/v3.6.9 branch. It will be part of the next release. We appreciate your effort!