happycatlabs/bradley-skills

Define, validate, snapshot, and inspect Bradley's privacy-safe provider-neutral agent organization directory. Use when changing exact task-to-role bindings, primary reportsTo edges, reparenting or host-placement revisions, observer subscriptions, or advertised contact capabilities. This skill validates organization metadata; it does not own task lifecycle or message delivery.

Maintain a machine-configured Notion Agent Workspace as durable shared context for projects and agent teams. Use when creating, updating, organizing, retrieving, or summarizing architecture, decisions, plans, research, role contracts, project notebooks, or privacy-safe handoffs that should survive across machines and sessions. The destination comes from a local profile; never infer it from a page title or repository copy.

Use when working with Apple's container CLI on macOS, especially to run isolated smoke tests, disposable Linux labs, port-published services, or container-machine experiments. Trigger when the user mentions apple/container, Apple Container, container run, container machine, isolated test harnesses, snapshots/checkpoints for container work, or asks whether Apple Container should be used instead of Docker/worktrees.

Cross-agent brainstorming and idea critique for early-stage decisions, architecture options, product/workflow design, debugging hypotheses, implementation strategies, or "what are we missing?" questions. Use when the user asks to brainstorm, workshop, explore options, stress-test an idea before implementation, compare possible approaches, or get Claude/Codex perspectives on a hypothetical plan. Prefer critique for reviewing existing diffs, completed work, push readiness, or concrete artifacts.

Dispatch Claude Code (claude -p) as a subagent — one-shot consults for an independent second opinion, scoped worker lanes (implementation in a worktree, read-only research, structured extraction), and resumed follow-ups. Use when the user invokes /claude, says "ask Claude" / "use claude" / "delegate to claude", when an orchestration plan routes a lane to Claude, or when a lane benefits from Claude's skills/CLAUDE.md/subagent ecosystem. Supports resume for follow-up turns in the same session.

Dispatch the Codex CLI (gpt-5.5) as a subagent — one-shot consults on hard questions, scoped worker lanes (implementation, research, browser/computer-use tasks), and resumed follow-ups. Formerly ask-codex. Use when the user invokes /codex or /ask-codex, says "ask codex" / "use codex" / "delegate to codex", when an orchestration plan routes a lane to Codex, or when a browser-driving task suits Codex's native browser_use. Supports resume for follow-up turns in the same session.

Create, update, inspect, pause, resume, or delete Codex automations on the local machine or a connected remote Codex machine. Use when the user asks to start, schedule, install, change, test, or remove a recurring Codex job, especially when they name another machine or host such as Eve. Also use when an originating Codex task needs one bounded, task-owned follow-up monitor that returns actionable or terminal results to that same task. Resolve the target host and project explicitly; remote automations are installed by a short-lived Codex thread running on the owning host because automation state is host-local.

One-shot cross-agent consultation for an independent second opinion, design call, debugging hypothesis, architecture tradeoff, implementation plan check, or "am I thinking about this right?" question. Use when the user asks to consult another agent, get Claude/Codex's opinion, validate an approach, or compare model perspectives without running a full brainstorm or critique. Prefer brainstorm for multi-lane idea generation and critique for reviewing concrete work.

Run a bounded multi-agent critique of existing work, concrete design choices, uncommitted code, committed diffs, or another agent/session's output. Use when the user says /critique, asks for a review/stress-test/readiness check, wants Claude and Codex perspectives before committing or pushing, or asks to review dispatched/orchestrated work without low-signal findings. Prefer brainstorm for early-stage idea generation or hypothetical options. Routes differently by host agent: Codex should use Claude plus focused Codex subagents; Claude should use consult or the codex skill plus focused Claude subagents. Do not require codex-review or REVIEW.md.

Ask or delegate to Cursor CLI Agent — Composer 2.5 for fast advisory/review, Grok 4.5 for long-running coding lanes — read-only advice, plan review, code review, or scoped implementation. Use when the user invokes /cursor, says ask Cursor, check with Cursor, use Composer or Grok, delegate to Cursor, or wants a Cursor CLI lane inside an orchestrated workflow.

Create and hand off Codex worktree sessions for Linear-ticket or repo work. Use when the user asks to fork, dispatch, spin up, start background sessions, create worktree-backed Codex threads, seed child sessions with orchestrate, prepare multiple ticket-specific agents, or hand off work from a coordinating chat. Pair with monitor only when the user asks to watch, check, steer, or wrap child sessions. For scheduled or recurring Codex jobs, including jobs owned by another machine, use codex-automation instead.

Operate as Bradley's interruptible engineering lead, project lead, or domain lead across multiple agent tasks. Use when the user assigns a lead role, asks what is being worked on across a project, wants several existing lanes advanced, or wants durable project context that survives individual chats. Reconcile the Agent Workspace, tracker, live task providers, repositories, GitHub, CI, and shipping authorities; route substantive work to durable owners and observation to watchers instead of implementing in the lead task.

Inspect local Claude Code session transcripts when the user wants to find, summarize, monitor, or wait on another Claude session. Use for requests like "inspect Claude session", "what is this Claude agent doing?", "find the latest Claude session", "is Claude blocked?", "Codex is waiting on a Claude agent", or when coordinating work by reading Claude session JSONL files under ~/.claude. Read-only by default.

Inspect local or connected-remote Codex Desktop or CLI tasks, threads, sessions, and rollout transcripts to find current work, reconstruct conversation, determine lifecycle state, summarize blockers, monitor progress, or recover handoff context. Use when a user gives a Codex task/thread/session id, asks what another Codex agent is doing, wants the latest or final message, asks whether a task is running or blocked, or needs read-only evidence from Codex app tools or host-local JSONL/log storage. Supports Codex-specific lifecycle semantics plus conservative fallbacks for other agent harnesses; interact with a task only when explicitly requested.

Inspect Cursor composer/session IDs directly or resolve Cursor UI request IDs, then read local agent transcripts to summarize, monitor, or wait on another session. Use whenever the user provides a Cursor session, composer, conversation, request, generation, or invocation ID; asks "inspect Cursor session", "what is this Cursor agent doing?", "find the latest Cursor composer", or "is Cursor blocked?". Reads Cursor request traces, hook logs, JSONL transcripts, and composer metadata. Read-only by default.

Plan and persist privacy-safe material progress, blocker, user-input, and terminal event delivery to the exact current primary reportsTo binding. Use for quiet inspection, deterministic restart dedupe, bounded provider retry/readback, observer copies, or Linear fallback. This skill does not own provider transport, task lifecycle, or routine heartbeats.

Monitor existing Codex child threads with a lead monitor and optional watcher subagents. Use when the user asks to monitor, check progress, get status, wait on child sessions, inspect dispatched work, decide whether a child is clean, send steering or wrap-up instructions, push/deploy/close after a child finishes, supervise multiple Codex worktree sessions, or use a dispatch ledger. Do not use for creating new sessions; use dispatch for that.

Coordinate a small number of genuinely parallel subagents for multi-slice implementation, codebase-wide audits or investigations, architecture follow-up, or review feedback. Also use its pull-request-ownership-only mode whenever a Codex or Claude task opens or adopts a PR, even if no delegation is needed. Use when the user asks to orchestrate, dispatch subagents inside the current work, parallelize implementation, audit, investigation, or review slices, integrate agent work, or act as a lead orchestrator for an active task. First confirm that delegation will reduce elapsed time or materially lower risk; keep bounded or serial work in the main agent even when orchestration is requested. For monitoring already-created Codex child threads, use monitor instead.

Rewrite or review human-facing prose so it reads like a thoughtful person wrote it. Use for internal documentation, onboarding guides, technical explanations, knowledge-base pages, tickets, pull request descriptions, handoffs, status updates, and other writing that feels robotic, jargon-heavy, over-structured, ambiguous, or obviously AI-generated. Preserve the writer's meaning and tone while improving first-read comprehension, specificity, rhythm, and natural voice.

Reconcile active agent work across a configured Agent Workspace, Worklog, live provider tasks, trackers, Git, GitHub, CI, and project shipping gates. Use when auditing open work, answering what is active across a project or domain, deciding whether an existing owner should continue, advancing several authoritative owners, or wrapping abandoned but still relevant work.