[harlan-zw]

🔗 Linked issue

Follow-up to #418.

❓ Type of change

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

📚 Description

Agents spent 103–178 seconds forking Skills while reading extra instructions and retrying unsupported install JSON output. Fork links now request a short fork-only guide, with a shallow fetch at one commit, batched checks, and plain installation output.

Two fresh preview trials took 96s and 115s, compared with 103s and 178s in earlier live trials. These are single-model smoke trials using preloaded HTML fixtures.

🤖 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 13:24 UTC.

GitHub merged this pull request.

• Logged (30/100): The negotiation redirect drops ?action=fork, so the documented fallback never reaches the fork-only Markdown. View code
  • Ask an agent to verify and fix this finding
• Logged (15/100): The PR's screenshot evidence is inaccessible from this session, so the measured fork-time improvement is unverified.
  • Ask an agent to verify and fix this finding

86d4d52d-5a48-498d-b9c7-a5d000dd6db7

Selected findings run after merge and open separate pull requests.

[harlan-zw]

Production workflow and timing checked by Codex, because CI cannot perform fresh-Agent trials:

Fork URL	Earlier live trial	Candidate preview trial
Specific Skill	103s	96s
Short repository URL	178s	115s

Both fresh Agents received a normal URL and an isolated Codex project. They received no implementation history. The preview used production HTML fetched before timing and candidate Markdown from the exported composer. Metadata and Git remained live, at the same source commits as the earlier trials. These are single-model smoke trials, not a general speed benchmark.

Neither Agent tried install JSON output or fetched full Git history. Independent checks verified original source bytes and modes, licence evidence, provenance, local lock records, and Agent copies. Edited copies survived reinstall. Upstream checks identified local sources.

The built Cloudflare application also served the fork-only query response from /gh/vercel-labs/skills.md?action=fork. Canonical and explicit Skill query routes pass boundary tests. Normal Markdown responses still contain the original Skill body.

The generated candidate guide measured 2,928 bytes for generate-package-skill and 2,863 bytes for find-skills. Earlier live Markdown measured 12,071 and 8,115 bytes. The fixture responses excluded module-added frontmatter and sitemap text.

[harlan-github-agent]

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

Automated Review: READY.

Read the Review results

[harlan-zw]

Codex Agent verified production after merge. Deploy run 37316365035 passed at descendant 389db91. Both normal Skill pages link to the fork-only guide. Both guides return HTTP 200 and include shallow fetch and plain install. Normal Markdown still includes the full Skill. Evidence: https://notes.localhost/notes/skilld-fork-speed