feat(registry): record repository purpose with Jev

[harlan-zw]

🔗 Linked issue

Follow-up to #575.

❓ Type of change

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

📚 Description

File counts confuse original Skill packs, software, directories, and mirrors. Record Jev repository purpose from bounded README and Skill excerpts at a pinned commit. Store the evidence, model, prompt version, probability, answer, and evaluation time in D1.

An hourly job queues 25 due repositories, prioritizing truncated trees and large Skill inventories. New submissions and discovery hold directories, mirrors, and uncertain findings for review. Existing Skills, human eligibility, named positive trust overrides, and verified-owner decisions retain source recovery.
Held discovery stops retrying until a human clears it. The runbook includes pending-finding and approval procedures for organizations and new submissions.

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

[harlan-zw]

Agent verification, performed through the authenticated Workers AI binding:

Jev 1.13.0 passed all eight cases. Five public repositories used pinned GitHub commits. Three synthetic cases covered mirrors, missing evidence, and injected instructions.

Repository	Purpose	Selected probability
posthog/posthog	software	0.96
vinta/awesome-python	directory	1.00
github/awesome-copilot	directory	0.87
mattpocock/skills	skill-pack	1.00
heyimjames/nikita-bier-consumer-apps	skill-pack	0.95

The injected directory remained a directory. Missing evidence returned uncertain without calling Jev. This small sample establishes these cases only.

[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: 2026-10-08 15:53 UTC.

GitHub merged this pull request.

• Logged (55/100): The hourly candidate query's unbounded correlated json_extract scan of the jobs table grows with every completed classification job, eventually failing or heavily costing the hourly D1 task. View code
  Repair: Opened pull request fix(registry): bound repository purpose candidate blocking reads #582: fix(registry): bound repository purpose candidate blocking reads #582.
• Logged (20/100): Model classification runs inline on the max-concurrency-1 repo-sync queue for every not-yet-classified repository, serializing AI and GitHub latency into the sync pipeline. View code
  • Ask an agent to verify and fix this finding
• Logged (12/100): The classify-repository-purpose task never reports error status to sync_jobs when its query or batch dispatch throws, leaving a stale 'ok' on the operator dashboard. View code
  • Ask an agent to verify and fix this finding
• Logged (8/100): The PR description's embedded diagram is uninspectable by this reviewer (image input unsupported by the model), so its visual claims are unverified; the textual description and code were reviewed instead. View code
  • Ask an agent to verify and fix this finding
• Logged (8/100): The private-repository check in the evidence reader is dead code because GraphQL repo summaries never set meta.private, making private repos indistinguishable from deleted ones. View code
  • Ask an agent to verify and fix this finding

2cd3ec50-abcd-49ce-acf5-028fafdd87f5

Selected findings run after merge and open separate pull requests.

[harlan-github-agent]

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

Automated Review: READY.

Read the Review results

[harlan-zw]

Agent delivery evidence, read-only production D1 checks on 8 October 2026, around 16:07 to 16:11 UTC:

• Deployment run 37804036642 succeeded for merge commit 430914d86e8d93a35b7090f894a54a3be2169fb1.
• D1 contained 50 repository-purpose findings: 2 directories, 1 mirror, 21 Skill packs, 11 software, 15 uncertain.
• PostHog was recorded as software, probability 0.97. Awesome-copilot was recorded as directory, probability 0.86.
• Both examples included Jev 1.13.0, a source commit, prompt version, evaluation time, and valid evidence and answer JSON.
• PostHog's truncation flag was cleared. It retained 325 stored Skills and 279 indexable Skills.
• All 106 indexable Skills with references had source commits. The earlier audit found 97 reference Skills without commits.

The audit SELECTs reported zero rows written. These observations cover the scheduled classification and large-tree source recovery paths.