v1.18.16 server/web mode: exactly 2 project skills silently dropped when project is a git repo (works in CLI, works without .git)

Summary

opencode v1.18.16 (server/web mode, --port): exactly 2 project skills are silently dropped when the project is inside a git repo — regardless of skill name, content, or mtimes. Without a .git directory, all skills load.

Same binary loads all skills in TUI/CLI mode, and in server mode all skills load if the project has no .git directory. The dropped skills are absent from /api/skill and from the skill tool — no error is surfaced.

Related: #41213 (external skill discovery silently finds nothing), #39871 (v2 skill discovery + git worktree).

Environment

• opencode v1.18.16 (latest release, installed via curl installer) — also latest on npm (opencode-ai@1.18.16)
• Linux x86_64
• Server started with opencode --port <port> (same code path the VSCode extension uses)
• Project: git repo, skills configured via skills.paths: ["./.opencode/skills"], 8 skills in .opencode/skills/<name>/SKILL.md

Steps to reproduce

1. Create a project with N skills in .opencode/skills/ and a git init.
2. Start the server: opencode --port 4129 (or via VSCode extension).
3. Query skills: curl http://localhost:4129/skill (or run /skill in a session attached to the server).

Result: N-2 skills are returned. The same 2 are always missing (in our case backend-firebase and frontend-design-system).

Controls that rule out obvious causes:

Systematic count across fresh copies (isolated, --port mode)

The pattern is always exactly 2 dropped when git is present, never an N-relative threshold — and it is deterministic (same skills dropped every run, reproducible across restarts).

Expected behavior

Server/web mode should load the same set of skills as the CLI/TUI. Nothing is dropped silently; at minimum a warning should be emitted.

Actual behavior

The skill tool description lists N-2 skills; the 2 missing ones return "not found". No error or warning appears in server logs.

Analysis / hypothesis

The drop appears tied to git-repo/worktree detection (issue #39871 describes the fs.up({ targets, start, stop }) discovery that stops at the git worktree boundary in v1 — #41213 confirms skill discovery silently returning nothing without surfacing an error). In server mode the project-root/worktree resolution seems to interfere with skills.paths, dropping a fixed subset. Only 2 are dropped (rather than all) even with a single skills.paths entry, which suggests per-entry dedup/last-wins behavior (like the Map keyed by skill id in v2's SkillV2.list()).

Happy to run any additional repro, capture logs, or bisect to a specific version if maintainers point me to the relevant code path.