[harlan-zw]

🔗 Linked issue

None.

❓ Type of change

• 📖 Documentation
• 🐞 Bug fix
• 👌 Enhancement
• ✨ New feature
• 🧹 Chore
• ⚠️ Breaking change

📚 Description

The search box says "Search skills or repos", but it handled neither repos nor intent well. vercel-labs/agent-skills came back as "Loose matches: No skill contains those words", although the registry holds that Repository. A pasted GitHub URL only offered to index the repo again. A sentence such as "write a postgres migration" was searched as separate words.

The box now calls a new internal GET /api/skills/search, which classifies the query before it searches:

• A GitHub URL or owner/repo answers with a Repository row and its Skills. A Repository the registry does not hold gets the existing index flow.
• @login lists that owner's Skills. A single word runs the existing hybrid search unchanged.
• A sentence goes through @cf/meta/llama-3.1-8b-instruct-fast once. The model returns search terms plus a track, a framework and an author. The hybrid search then runs on the typed words and the terms. Track and framework only lift near ties. They never filter.

Guardrails on the model call:

• The answer is cached per normalised query for 30 days in edge-cache. A reply the model cannot parse is cached as "none" for a day.
• The budget is 800ms. After that, or on any model error, search runs on the typed words. A late reply still lands in the cache.
• Names, refs, URLs and logins skip the model. Each visitor's model calls go through the existing API_GUEST_RATE_LIMIT binding under a search-intent: key. Over the limit, search runs on the typed words.
• Wide events record the fallback reason only, never the query.
• The model is a native Workers AI model billed in neurons, so it needs no AI Gateway credits. A call is about 525 input and 30 output tokens, about $0.036 per 1,000 calls.

The public skills.search v1 answer and the MCP tools are unchanged.

In the panel, each Skill row shows its RunChip compact, and a Repository gets its own row with its Skill count. A name also shows matching Repositories from the typeahead index the browser already holds. The panel is now a grid, because an ARIA option cannot contain a button: Right moves to the run chip and Enter copies it. HomeSearch (the hero from #437) and SkillSearchTrigger share useSkillSearchBox, so Enter means the same in both: open a row selected with the keyboard, otherwise submit. The header field widens while open, the half-empty preview pane is gone, and owner paths truncate instead of crushing Skill names at phone width.

"See all results" for a sentence opens /skills with its meaning search on. That mode was labelled "Ask AI" and "AI search", which COPY.md bans, so it now reads "Search by meaning" and "Results by meaning".

Stage 2, not built

An agent loop could answer what a one-shot rewrite cannot. "stop memory leaks in node" still returns Godot and Tokio Skills, because the embeddings rank them first.

• Loop: Claude Haiku 4.5 through AI Gateway, with the existing mcpTools from layers/mcp/shared/mcp-tools.ts as its tools (search_skills, get_skill, get_collection). Cap it at 4 turns and let it return Skill refs only.
• Cost: about 21k input and 1.2k output tokens per question. At $1 and $5 per million tokens that is about $0.027, or about $0.015 with prompt caching on the system prompt and tools. That is roughly 750 times a stage 1 call.
• Credits: the AI Gateway balance is empty (see send-digests.ts). A $10 top-up covers about 400 to 650 questions.
• Caching: AI Gateway response caching for identical requests, plus an answer cache by normalised question, as stage 1 does.
• UI: opt in only, never per keystroke. A "Find Skills for this task" row at the foot of the panel for sentences, a short streamed answer with Skill rows and run chips, and a daily cap per visitor.

🤖 AI disclosure: Harlan Agent Kit modified this description. My AI open-source policy.

[harlan-github-agent]

🤖 MERGED

Harlan Agent Kit posted this automated review. It is not Harlan's personal review or approval. AI open source policy. Last updated: 2026-10-05 17:31 UTC.

GitHub merged this pull request.

• Logged (42/100): Enter on a word-pair owner/repo query the registry lacks starts a doomed index flow instead of the search it used to run, and hides the loose-match disclosure. View code
  • Ask an agent to verify and fix this finding
• Logged (27/100): A container-query breakpoint keyed to the viewport's md width always matches the capped-width panel, so the sources badge and the run-command key hint never render. View code
  • Ask an agent to verify and fix this finding
• Logged (25/100): A rate limiter error propagates and 500s the search box instead of falling back to the typed words. View code
  • Ask an agent to verify and fix this finding
• Logged (15/100): The pull request's architecture diagram is inaccessible at head, a material documentation gap for the only visual in the description.
  • Ask an agent to verify and fix this finding

Review completed. No finding above 80/100.

Selected findings run after merge and open separate pull requests.

[harlan-zw]

Agent check, run by hand because CI cannot cover it. A local build of this branch served the 20 queries below against production D1, Vectorize and Workers AI, read only. Before is production today. After shows the top three Skills once the understanding is cached.

Query	Before (production)	After (this branch)
vercel-labs/agent-skills	"Loose matches": vercel-agent, benchmark-agents, vercel-sandbox	Repository row (9 Skills), then composition-patterns, deploy-to-vercel, react-best-practices
https://github.com/vercel-labs/agent-skills	Index row only	Same Repository row and Skills
github.com/anthropics/skills	creating-skills, skill-developer, anthropic-architect	Repository row (20 Skills), then its Skills
@antfu	Hybrid search, 58 hits	antfu's 11 Skills
write better commit messages	git-commit-formatter, commit-helper, git-commit-format	writing-commit-messages (agenteq), writing-commit-messages (herdr), commit-messages
make my UI less generic	make-interfaces-feel-better, ui-design, soft-skill	frontend-design (anthropics), ui-design-system, ui-design
reacct hooks	creating-hooks, git-guardrails-claude-code, viral-hook-creator	react-hooks-best-practices, creating-hooks, trigger-realtime
playwrite e2e tests	write-e2e-tests, e2e-testing, playwriter	playwright-e2e-testing, e2e-testing, api-testing
write a postgres migration	postgres-drizzle, neon-postgres-branches, postgres	postgres-database-migration, postgres-drizzle, postgres
set up stripe payments	stripe/ai Skills	stripe/ai Skills, filtered to the stripe owner
review my pull requests	gh-review-requests, skill, pr-review	creating-pull-request, addressing-code-review-comments, pr-review (mixed)
deploy to cloudflare workers	cloudflare-deploy, workers-best-practices, cloudflare	cloudflare-deploy, nextjs-on-cloudflare, cloudflare-vpc-services (mixed)
stop memory leaks in node	godot-debugging-profiling, tokio-troubleshooting, ios-memgraph-analysis	Unchanged miss
test a Vue app, vitest, nuxt, antfu, mattpocock, tailwnd, anthropics/skills/frontend-design	Already right	Same Skills; names also show matching Repository rows

Also checked:

• On a cold local request, 3 of 10 sentences passed the 800ms budget through the local remote-binding proxy and fell back to the typed words. The late replies reached the cache, and the next refresh served the understood results.
• axe-core on the open panel, with the run chip cell active, reported no violations. Enter on that cell copied the command and did not navigate.

Before	After