feat: Codex skill picker (desktop + mobile) by plutoless · Pull Request #75 · AgoraIO-Community/Newbro

plutoless · 2026-06-08T13:18:29Z

Summary

Adds a per-bro Codex skill picker to the active-session composer on desktop and mobile. Skills are discovered per-executor and a chosen skill rides a turn, activating the agent-native Codex skill via the app-server.

Discovery — skills/list is loaded once at executor start, cached, and carried on the existing register_node capability payload (ExecutorCapabilities → ExecutorNodeExecutor → BroExecutorCapabilitySummary), projected per-bro into the session snapshot. Refreshes on executor reconnect (no live skills/changed wiring in v1).
Activation — the chosen skill rides turn metadata like plan_mode: web skillName → API skill_name → instruction/task/outbound metadata["skill"] → all three CodexExecutor turn-start sites → turn/start with a {type:"skill", name, path} input item + $name marker. Skill and plan-mode are independent/combinable.
Vanished-skill contract — validate-before-write: a skill is only written onto the turn/timeline if present + enabled in the bro's current catalog; otherwise it is dropped with an observable skill_dropped marker and the turn runs plainly (never implies the skill ran).
UI — desktop popover + mobile bottom sheet (ported from the design prototype) read the real per-bro catalog; the chosen skill renders as a pill on the user's timeline bubble.
Lean snapshot — ExecutorSkill carries only name/display_name/short description/hint/path/enabled (no full descriptions, dependencies, or node-local icons), keeping the snapshot small.

Design + protocol contract were verified against a live Codex app-server (codex-cli 0.137.0); a curated fixture is committed at docs/protocol/fixtures/codex-skills-list-sample.json.

Docs: docs/protocol/skills.md, spec docs/superpowers/specs/2026-06-08-skill-picker-design.md, plan docs/superpowers/plans/2026-06-08-skill-picker.md.

Deferred follow-ups (documented in the spec): inline / filter trigger; surfacing the skill_dropped marker as a UI notice.

Test Plan

Backend: .venv/bin/python -m pytest — 710 passed
Web: npx tsc --noEmit clean; vitest targeted suites (SkillPicker, skill-payload, skill-metadata, session-client) — 85 passed
Codex turn-streaming regression (test_codex_multi_message_turn.py) green
Manual smoke: connect a Codex executor, open a bro's active session, pick a skill, send — verify the picker lists real skills, the pill renders, and the turn runs; disable/uninstall a skill and reconnect to confirm it leaves the picker

🤖 Generated with Claude Code

Per-executor Codex skill discovery + chosen-skill-rides-the-turn, mirroring the plan_mode plumbing end to end. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

- model skills across all three capability models (core, wire, summary) + the manual _descriptor copy; discovery otherwise never reaches snapshot - add node_capabilities message for skills/changed (node_status doesn't update registry today) - thread skill explicitly through API → instruction/command metadata → all three codex turn_start sites - vanished-skill: validate-before-write observable contract, no fake fallback - path optional + marker-only activation fallback; dependencies structured; add fixture-first testing and open questions Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Probed codex-cli 0.137.0 skills/list and captured a curated fixture. Verified: path always returned (full skill-item activation is primary); dependencies is structured not list[str]; response grouped per cwd with errors; icons exist but as node-local paths (deferred); defaultPrompt is the picker hint. Snapshot carries a lean ~15KB projection (vs ~40KB raw). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Drop the skills/changed subscription and node_capabilities message. Skills load once when the executor starts, cache on the executor, and ride the existing register_node capability payload — zero protocol additions. Catalog refreshes on executor reconnect; live refresh moves to a documented future enhancement. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

21 TDD tasks across shared models, codex discovery (load-once), runtime/API skill threading, vanished-skill contract, and the web picker + timeline pill. Grounded in the live-app-server fixture. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>