feat: / Skill Invocation in Group Channels - #4259

#4233 | hub #4231 | stacked on #4252 | Agent tasks: claude-code/local/b7955cb7-65c2-4b41-bafd-b18256e4a38b (UI, integration, review), codex/local/01a1e0df-afbe-7283-bd17-0b534bc78f98 (ranking, routing, digests)

Closes #4233

Why the change

S1 made / Skill Invocation work in a Bot DM. A Group Channel holds several PersonaBots, so its / menu must list every member's Skills, let a Human find a Bot's Skills by its name, and keep routing driven by mentions only.

Special things to note

• Stacked on feat: invoke a Bot's Skills from / in a Human DM #4252 (S1). Retarget to main after feat: invoke a Bot's Skills from / in a Human DM #4252 merges so Closes #4233 takes effect.
• Merge risk: two-way door, revert the PR; no schema change. Small blast radius: the Group composer menu, skills-list ranking, and the Group digest / pending-context text a Bot reads.
• A Skill Invocation still addresses no one (ADR-0346): picking Alice's Skill and mentioning Bob wakes only Bob, and Bob reads Alice's snapshot with owner bot:alice.
• Ordinary Group digests and pending context now carry the <skill_invocation> snapshot too; before, a Bot woken by digest instead of a mention saw only /name.
• External IM mirrors (Discord/Lark) already receive the message body, so a chip shows as plain /name. A unit test covers the outbound payload. The real-IM browser check (AC-4) needs the signed-in IM QA browser and is tracked in QA: Skill chip falls back to /name in Discord/Lark mirrors (Skills S2 AC-4) #4260.
• The sent chip in the timeline now shows its owning Bot's avatar, as the composer chip and the menu already did, since a Group message can carry another Bot's Skill.

Change outline

/ menu in a Group Channel: Before (S1 head), a Group's / offered only commands.

After: commands, then Skills. Alice's /qa-echo and Bob's /triage-bugs carry their avatars; each Host Skill is listed once.

Bot name ranking: /bob finds Bob's Skill through its owner.

Alice's Skill, addressed to Bob: the composer holds an @Bob mention and Alice's /qa-echo chip.

Only Bob replies (real model, both wake policies mentions-only). The sent chip shows Alice's avatar.

Isolated Host, viewport 1440×900 cropped to the conversation pane, light theme, English, AX model from ax-model.json. Canonical read (channelTimeline): the Human message stores skillInvocations[0] = {name: qa-echo, owner: {kind: bot, botSlug: <Alice>}, path: <Alice memory>/.agents/skills/qa-echo/SKILL.md, content} with a single mention of Bob; the only Bot entry is Bob's SKILL_QA_7731. channelSkillCatalog returned both Project Skills and each Host Skill once.

Ranking: rankSkillEntries(entries, query, ownerNames?) gains a Bot-name tier: name prefix > name subsequence > owning Bot display name or slug > description. The composer passes each Bot Skill's display name and slug; skills-list --query does the same through the list RPC, so CLI and menu agree.

Group catalog: the Group composer loads channelSkillCatalog and reloads when the Bot roster changes. Host Skills appear once; each Bot Skill row shows its Bot's avatar.

Same name: a bare /name that matches a command still runs the command; the same-named Skill row stays selectable with the arrow keys.

/alice  →  /alice-intro · Bob   (name prefix)
           /triage · Alice      (owner name)
           /notes · Host        (description)

Tests: packages/core/test/skill-invocation.test.ts (Group routing independence, digest snapshots, plain IM payload), packages/core/test/bot-live-cli.test.ts (CLI ranking), packages/client/test/skill-composer.test.ts (owner-name ranking, command precedence). Codex ran the ci-plan affected suites (4,040 passed, 16 skipped) before integration; pnpm lint and tsc are clean on the integrated head.

🤖 Generated with Claude Code