[Bug]: OpenCode worktree snapshot caches a builtins-only /skill: list mid-thread and never refreshes (recovered by "Restart agent session") #16912

Before submitting

• I searched existing issues and did not find a duplicate.
• I included enough detail to reproduce or investigate the problem.

Area

apps/server

Steps to reproduce

Observed twice on the same thread; not yet deterministic, but the conditions below reproduce the shape:

1. Provider: OpenCode v2 (opencode serve per workspace, spawned by T3), desktop app, local environment, project
   dotfiles opened in worktree mode (thread cwd ~/.t3/worktrees/<repo>/<id>).
2. Progress a thread past skill/config churn that touches the OpenCode skill registry: e.g. commits editing skills and
   opencode.json, a stow --adopt -R . restow, or post-merge publish steps (setup-agent-skills.sh + stow).
3. Mid-thread, OpenCode's per-directory skill registry goes through a rebuild cycle (logged as
   watcher stopped → watcher subscribe → watcher started for <worktree>/.agents/skills, twice 15 s apart in our
   log at 14:08:53Z / 14:09:08Z on 2026-10-07, interleaved with the periodic provider.updated/model.updated
   refresh pair for the same cwd at 14:09:58Z).
4. If T3's per-cwd workspace-snapshot probe resolves inside that rebuild window, the composer's
   workspaceSnapshots[...matching cwd].skills list is cached as OpenCode's built-in skills only
   (OpenCode, Report) — all discovered personal/global/project skills drop out.
5. Open / in that thread afterwards: only /model, /compact, /init, /review, /usage-limits and the two
   built-in skills remain. It never heals on its own — still broken after ~40 min, a PR merge, and a machine-wide
   skill republish.

Expected behavior

The palette keeps listing all registered skills (probed or retained from the previous healthy snapshot), and a
mid-thread registry rebuild never leaves the thread with a partial, non-empty skill list.

Actual behavior

The per-cwd snapshot caches the partial (built-ins-only) skill list. Two gaps make it sticky:

1. The auto-refresh effect only fires when no workspaceSnapshots entry exists for the cwd; a bad-but-present
   entry suppresses refresh indefinitely (see _chat-CBE06zzM.js:
   if (t) { Lr.current = e; Rr.current = null; return } before the refreshProviders call).
2. The empty-list retention guard only rescues skills.length === 0
   (skills: nextProvider.skills.length === 0 ? previousProvider.skills : nextProvider.skills,
   shouldRetainMissingOpenCodeMetadata); a 2-item builtins-only list passes through untouched.

On the OpenCode side the narrowed list is real state, not just UI caching: the session's available-skills list is
pushed as deltas ("The following skill IDs are no longer available and must not be used: …"), so the agent's
model-visible skill set narrows the same way until recovery.

Recovery confirmed on the live thread: command palette (Ctrl+K) → "Restart agent session" (runs switchRef, then
refreshProviders with fresh: true for {instanceId, cwd: worktreePath}), then one short message → full
/skill: list returns.

Impact

Minor bug or occasional failure — the agent's invocable-skill set silently shrinks mid-thread (registry deltas, not
just UI caching), and the $-picker/skill invocations stop working on that thread until a manual "Restart agent
session"; threads recover cleanly with that workaround at the cost of agent-session continuity.

Version or commit

T3 Code 0.0.46-nightly.20261004.2644 (pacman t3code-nightly-bin 0.0.46_nightly.20261004.2644-1),
OpenCode v2 provider CLI 2.0.23.

Environment

Arch Linux (x64), T3 Code desktop app (/usr/lib/t3code-nightly), local environment + Tailscale-serve web client
(both clients share the same t3 serve snapshot, so the shrink should be visible on phone/web too), per-thread
OpenCode server 127.0.0.1:34323, provider opencode, model glm-5.3-flash.

Logs or stack traces

~/.local/share/opencode/log/opencode.log, 2026-10-07 (OpenCode v2.0.23, run ids cbd46a36 / e0b1da69):

14:08:53.455Z watcher stopped   path=~/.t3/worktrees/dotfiles/t3code-0bc72be4/.agents/skills type=directory
14:08:53.462Z watcher subscribe path=~/.claude/skills type=directory
14:08:53.495Z watcher subscribe path=~/.agents/skills type=directory
14:08:53.508Z watcher subscribe path=~/.t3/worktrees/.../t3code-0bc72be4/.agents/skills type=directory
14:08:53.512Z watcher started   path=~/.t3/worktrees/.../t3code-0bc72be4/.agents/skills backend=inotify
14:09:08.003Z watcher stopped   (same skills path again — second rebuild cycle 15 s later)
14:09:58.890Z event type=provider.updated location.directory=~/.t3/worktrees/.../t3code-0bc72be4
14:09:58.906Z event type=model.updated  location.directory=~/.t3/worktrees/.../t3code-0bc72be4

Palette state after this window: only built-in skills (/skill:OpenCode, /skill:Report) remained; all 61 discovered
dotfiles skills missing. Same persisted cache also holds a machine snapshot with skills:0 and slashCommands
[compact] (~/.t3/caches/opencode.json), so partial snapshots reach disk.

The per-cwd prober resolves only after a fresh skill.updated + command.updated event pair for the directory
(listOpenCode2Workspace: pending = Set("command.updated","skill.updated"), then skill.list/command.list are
taken as-is) — a skill.updated emitted at the start of a rebuild (before the personal/project/compat sources are
re-registered) satisfies the wait while skill.list still returns only built-ins.

Screenshots, recordings, or supporting files

(1) Shrunk state — unfiltered / palette, only the two OpenCode built-in skills:

(2) Same thread after Restart agent session + one message — full list restored:

Workaround

"Restart agent session" on the affected thread (via command palette Ctrl+K → Actions), then send a message. It
costs the thread's agent-session continuity. Creating a fresh thread also shows the full list. No non-destructive
refresh path found for an existing thread.

Related but not duplicates

• [Bug]: Workspace snapshot taken before the first Claude probe finishes leaves the / menu without provider commands for the whole session #11575 — Claude driver: snapshot taken before first capability probe; ours is the OpenCode v2 ACP driver, skills
  surface (not provider commands), mid-thread onset after a healthy start.
• [Bug]: One failed Claude capability probe empties the slash-command menu for 5 minutes and persists the empty list #7111 — one failed Claude probe empties the slash menu; ours is a partial (non-empty) skills list, OpenCode driver,
  and the retention guard explicitly does not catch it.
• Provider CLI updates leave discovery caches and sibling instance catalogs stale #14316 — stale discovery caches across provider updates; same family but different driver and refresh-suppression
  mechanism in the composer per-cwd effect.
• [Bug]: Project skills that a new worktree receives after creation never appear in the $ picker #14801/fix(server): rescan workspace skills after session start hooks run #14877 — worktree skills never reach the $ picker (Claude, gitignored skills dir + SessionStart hook);
  different root cause.