fix(cli): remove duplicate skill catalog

[marius-kilocode]

The skill catalog was being emitted twice in ordinary agent requests: once in the system prompt and again in the dynamically generated skill tool description. This repeated the same available-skill context on every request and increased prompt size without improving discoverability.

This change keeps the system-prompt catalog as the single effective catalog. The skill tool retains its static usage instructions and input schema, and invoking it still lazy-loads the complete skill body. The registry deletion is limited to the Kilo-specific dynamic description path, so permission filtering, project/global/builtin discovery, trusted skill handling, plugin definition hooks, and subagent behavior remain unchanged.

Historical provenance: the removed describeSkill block was introduced by commit bfcb911648d in PR #12204, the OpenCode v1.17.4 compatibility merge. That compatibility work added the concise skill inventory to the tool description, while the system-prompt catalog remained present. This PR removes only that later duplicate representation and preserves the surrounding upstream-compatible tool behavior.

Validation included 50 skill, registry, discovery, permission, built-in, trust, worktree, system-prompt, and lazy-loading tests, plus CLI typecheck, lint, formatting, changeset, promise-facade, and OpenCode annotation guards.

The generated /experimental/tool endpoint was measured before and after using deterministic JSON character counts. The endpoint continued to expose 17 tools, while total tool-definition JSON decreased from 41,583 to 35,706 characters. The skill description decreased from 6,247 to 399 characters, and its catalog-marker count changed from 1 to 0. A live backend check returned the same corrected 399-character static description and unchanged schema for both a normal worktree and an Agent Manager-managed worktree.

The persisted local session for this worktree also showed lower uncached model input after the fix. Comparing recorded step-finish data around the registry change, average uncached input fell from 7,268 to 2,502 tokens per step, a 65.6% reduction. Cache-read tokens naturally increased as the conversation grew, so they were not used as the prompt-size comparison.

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Files Reviewed (4 files)

• .changeset/fix-duplicate-skill-catalog.md
• packages/opencode/src/tool/registry.ts
• packages/opencode/test/kilocode/skill-catalog.test.ts
• packages/opencode/test/tool/skill.test.ts

---

_{Reviewed by grok-4.6 · Input: 53.3K · Output: 10K · Cached: 424.8K}

_{Review guidance: REVIEW.md from base branch main}