[gilgtm]

Summary

Offers / skill chips for the addressed local agent (c99d0ce51).

• Listing: in the desktop app, / at a word start lists the skills of the one local Claude or Codex agent that the viewer's message addresses.
  • Claude's skills come from ~/.claude/skills; Codex's from ~/.codex/skills and ~/.agents/skills.
  • They show under a "Skills for " heading. Completion rows gain an optional group.
  • Several addressed agents make the menu ask for one. A remote agent, another person's agent, a person, or a browser build lists nothing.
• The chip: a chosen skill is a chip for that agent. It opens its short description and location in place, as a nested popover. It sends as literal text plus the signed buzz-tokens-v1 tag, and no skill content is copied into the message.
• Native listing: local_skills_list reads only SKILL.md front matter, through the bounded listing helper. It skips front matter over 16 KiB and refuses more than 512 skills.
• Agent grant: a local agent gets one prompt line naming each SKILL.md, only when the message's author is the agent's owner and the chip names that agent.

Compared with the previous version of this PR

The previous version offered session skills through the removed task mode's skills source. Skills now come from the agent's own folders on this machine, for any chat that addresses one local agent.

buzz-review-completed

Testing

Local runs on macOS, on the stack rebased onto main 9596cd3db:

• The pushed heads differ from the tested adbdf424e chain only in tests/browser/solo-rename.spec.mjs. 02f9c0277 makes the new plugin-owned lifecycle test wait for the chat's row before filing it, after the hosted flake check caught it racing the roster load. With that fix, the full solo-rename.spec.mjs passed 180/180 at 02f9c0277 (10 repeats, Chromium and WebKit) and 120/120 at c99d0ce51 (5 repeats).
• At this PR's head adbdf424e: pnpm check (Biome, TypeScript and the design checks) and the icon check passed. Vitest for the files the flake check selects for this PR: 1485 passed, 16 skipped.
• Rust at adbdf424e: cargo fmt --check and cargo clippy -p Buzz --all-targets -- -D warnings passed. local_skills and host_files tests at this head: 4 and 24 passed. cargo-deny check passed at the stack tip.
• At the stack tip adbdf424e: full Vitest, 20556 passed and 23 skipped (includes browser-host/channel-kit.test.mjs).
• Browser journeys at the stack tip, Chromium and WebKit: the full tests/browser/solo-rename.spec.mjs, repeated 5 times, passed 120/120. The 34 other specs that the flake check relates to the stack, run once in each engine: 343 passed, 3 skipped by their own conditions. The full browser suite was not run locally.
• No plugin harness pin includes this commit, so the plugin harness did not exercise it.

Native testing: Gil ran the native desktop app on this stack at adbdf424e, with the matching Nexus plugin build. Gil confirmed that these native-only paths passed:

• ! folder listing through the Tauri repositories_folder_list command (NEXUS-4029: Offer folder chips in the composer #2789);
• / SKILL.md listing through local_skills_list (NEXUS-4033: Offer session skills in the core task composer #2790);
• the Claude --add-dir folder grant (NEXUS-4029: Offer folder chips in the composer #2789);
• relay signing and Outbox retry after a native restart.

Agent testing did not reach these paths: the agent ran no native app, no real relay and no real agent.

Known gaps

• Gil exercised skill listing natively, and it passed. Agent testing did not reach skill listing or the SKILL.md prompt line.

Hosted or local failures that this stack does not cause:

• tests/browser/typing.spec.mjs:290 ("a growing channel/thread composer lifts history with it at the bottom") intermittently fails with React error Add complete reaction toggles to the message menu #185 on main too. Locally it failed 12 of 12 runs at 9596cd3db and 9 of 12 at a later main commit.
• On macOS, cargo test -p Buzz crashes the library test binary (SIGSEGV), and some host_command deadline tests fail. Both happen the same way at base 9596cd3db, before this stack. Hosted Linux CI is the reference for these.

Test changes

Titles are compared only in the test files each range changes. Tests that moved or were renamed are matched by title.

Removed by this PR's diff (0)

• none

Added by this PR's diff (6)

• src/bundled/claude-code/runtime.test.ts: names a skill its owner chose for this agent by chip, never one chosen for another agent
• src/bundled/mentions/SkillCompletion.test.tsx: lists nothing unless one addressed agent is the viewer's own local Claude or Codex agent
• src/bundled/mentions/SkillCompletion.test.tsx: matches / and name characters at a word start, never mid-word or past a space
• src/bundled/mentions/SkillCompletion.test.tsx: offers the one addressed local agent's skills as chips for that agent, listed once per query
• src/bundled/mentions/SkillCompletion.test.tsx: shows a failed listing with Retry, which lists again
• src/features/agent-harness/referenced-folders.test.ts: names only skills the owner chose for this agent, quoted

In the previous version of this PR, not carried over (8)

• src/features/host/service.test.ts: inspects paths with the calling plugin grant and preserves missing/dangling/error results
• src/features/messages/MessageComposer.test.tsx: shows a task chat's skill details through its source, without a failing or malformed report
• src/features/messages/composer-token.test.tsx: opens a skill chip's details in place by pointer or key, and dismisses only them
• src/features/messages/composer-token.test.tsx: returns the caret right after a skill chip when its details close toward it
• src/features/solo/task-skill-completion.test.tsx: matches / only after a boundary, up to the next slash
• src/features/solo/task-skill-completion.test.tsx: offers a task chat's skills for its one tagged agent or primary and rechecks them on selection
• src/features/solo/task-skill-completion.test.tsx: refuses a task source whose skills cannot be called
• src/features/solo/task-skill-completion.test.tsx: reports an empty or failed load, and offers nothing outside a task chat or without a skills source