perf(d1): bound scoring and discovery query work

[harlan-zw]

🔗 Linked issue

❓ Type of change

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

📚 Description

Score recomputation recounts sibling Skills for each Skill. The live query recorded 335 million D1 rows read in 24 hours. Group live counts once. Load discovery metadata after pagination. Build tag candidates before joining Skills. Trim source text only for missing embeddings. Dirty drains also avoid a second scoring read.

A live comparison reduced full scoring reads from 13,951,401 to 165,855, with identical rows. No indexes or migrations are added.

🤖 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: 2026-10-11 01:39 UTC.

GitHub merged this pull request.

• Logged (10/100): The PR's explanatory evidence diagram is inaccessible, leaving its documented query-change rationale unverified; the code itself was verified independently and matches the text claims. View code Optional follow-up: Re-attach or re-host the diagram (e.g. commit it under docs/) so the query-change rationale is reviewable evidence.
  • Ask an agent to verify and fix this finding

7ad9e2c2-770d-4d35-b414-5c7dfc3d2547

Selected findings run after merge and open separate pull requests.

[harlan-zw]

🤖 Harlan Agent Kit Agent measured these SQL variants against the live D1 primary. All queries were read-only.

Query	Reads before	Reads after	Time before	Time after
Full scoring	13,951,401	165,855	2,438 ms	806 ms
One Skill per owner	74,526	70,154	1,866 ms	86 ms
Embedding backlog	36,693	36,694	2,168 ms	696 ms
Testing tag	121,572	105,778	360 ms	31 ms
Frontend tag	51,645	34,321	184 ms	66 ms
Nuxt tag	42,162	24,388	107 ms	20 ms

These are individual executions, not latency percentiles. Background jobs remained active.

Full scoring returned 20,925 identical rows. Discovery returned 60 identical cards. Backlog counts matched. Tag membership and fields matched across all three samples. Two frontend rows swapped inside an exact stars-and-name tie. The existing order leaves that tie unspecified.

The embedding rewrite reduces text work, rather than index reads. It keeps the same eligibility rules.

[harlan-github-agent]

Harlan Agent Kit posted this automated review. AI open source policy.

Automated Review: READY.

Read the Review results