roster

Use when infrastructure work needs the current Estate map, resource state, declarations, provider readback, an exact plan, or the provider-native management loop. Trigger: /estate-infrastructure, /estate.

Discover and dispatch the agent fleet. Any Roster agent may commission any other agent; there is no declared delegates list. Use when: "delegate this", "dispatch an agent", "who can do this", "hand this to a builder/reviewer/ verifier", "launch a focused agent", or any time work should run in a fresh, independently composed context instead of this one. Trigger: /roster, /dispatch.

Enumerates the peer AI agent CLIs installed on this machine (codex, pi, goose, opencode, claude, cursor-agent, grok, agy, hermes, oracle) and how to invoke each headlessly. A capability map, not a quota: useful for fresh-context adversarial review on a different model family, second opinions, competing attempts, and wide benches. Use when: "ask codex", "ask another model", "second opinion", "cross-model review", "what AI tools do I have", "other agents", "different model family", "adversarial critique from another provider". Trigger: /peer-harnesses.

Prepare credential-free public repository handoffs on dedicated Fly Sprites: remote, isolated, scale-to-zero sandboxes for heavy work. A clean owned checkpoint makes every handoff reproducible without copying workstation, Harness, or Git credentials. Use when: "prepare this on a sprite", "remote sandbox", "offload a public checkout", "bake a sprite", or heavy work that an external launch owner will supervise. Trigger: /sprites, /sprite-lane.

Use when an agent needs to inspect, claim, update, request input for, or complete work cards in a Powder instance. Powder is the self-hostable, agent-first work board: a durable card store with run sessions, activity, audit events, relations, optional proof, and human-in-loop states.

Shape an unresolved idea into a decision and executable context packet. Use when: "shape this", "write a spec", "design this feature", "plan this", "spec out", "context packet", "technical design". Trigger: /shape, /spec, /plan, /cp.

Create or update root VISION.md as a first-class project north-star artifact: conversational interrogation, repo research, philosophy distillation, and wiring repo-local primitives to read it. Use when: "vision", "vision.md", "project vision", "north star", "what is this project", "clarify product direction", "write/update VISION.md", "project philosophy", "why does this repo exist". Trigger: /vision, /north-star.

Grill the operator relentlessly about a plan, decision, or idea by batching every open question into one round instead of asking one at a time. Use when: the operator wants to stress-test their thinking, or uses any 'grill' trigger phrases. Trigger: /grill, /grilling.

Produce a consistently-styled, self-contained HTML report served from a local static directory. One house template (Silver Age comic-ops palette, dark/light toggle, and a mandatory copy-page button) so every report an agent hands the operator looks and behaves the same. Use when: "make an HTML artifact/report", "write up a brief/report/dashboard as a page", or any time you'd otherwise dump a long analysis into chat. Trigger: /artifact.

Interface and design-system design by parallel bench: every /design invocation fans out blind subagent lanes — each obeying one vendored design philosophy end to end — and composes their rendered explorations into one clickable lab catalog the operator browses and gives verdicts on. Needs a screenshot, URL, rendered artifact, or file plus intent. Use when: "improve the design", "redesign this", "critique this screen", "show me options", "prototype this", "deslop this", "polish the UI", "is this accessible", "scaffold DESIGN.md", or any product-facing visual artifact — docs layout, dashboards, charts, diagrams. Do not use when: a mechanical style fix with no design decision (just make the edit), or behavior verification (/qa). Trigger: /design, /prototype.

Route Misty Step factory application capabilities. Use when choosing, auditing, integrating, or operating Canary, Powder, Landmark, Aesthetic, Bitterblossom, or mint: production observability, incidents, health checks, error logging, backlog/work-card state, release intelligence, UI/UX system adoption, event-triggered agent workflows, or routing a credentialed outbound API call through the agent credential broker. Trigger: /factory-apps, /factory-stack.

Use when an agent needs to make an outbound HTTP call to a vendor API (OpenAI, GitHub, Stripe, or any third-party service) and would otherwise need an API key, token, or other secret. mint is the fleet's agent credential broker: route the call through mint's egress proxy with a credential placeholder — your tailnet identity IS the auth; never request, embed, or expect real credential bytes in agent context. Trigger phrases: "mint", "credential broker", "egress proxy", "call the API through mint", "I need an API key", "OpenRouter key", "secret placeholder", "__mint.", "proxy call".

Web research, multi-AI delegation, and multi-perspective validation. /research [query], /research delegate [task]. Use when: "search for", "look up", "research", "delegate", "get perspectives", "web search", "find out", "investigate", "introspect", "check readwise", "saved articles", "reading list", "what are people saying", "X search", "trending", "which model", "compare models", "best model for", "model selection". Trigger: /research.

Manage the operator's Todoist as the system of record for tasks, reminders, and follow-ups — capture, triage, complete, and organize. Agent-native via the Todoist MCP; scriptable via the `td` CLI. Use when: "add this to my todoist", "add a task", "remind me to", "capture this", "what's on my todoist", "what's due today", "what's in my inbox", "mark this done", "create a project/label", "log a follow-up". Trigger: /todoist, /task, /todo.

Shared vocabulary for designing deep modules. Use when the user wants to design or improve a module's interface, find deepening opportunities, decide where a seam goes, make code more testable or AI-navigable, or when another skill needs the deep-module vocabulary.

Grill the user relentlessly about a plan, decision, or idea. Use when the user wants to stress-test their thinking, or uses any 'grill' trigger phrases.

Reference for writing and editing skills well — the vocabulary and principles that make a skill predictable.

Dispatch-shaped code review: fan the diff out to fresh-context reviewers across diverse providers and model families, synthesize, fix blockers, re-review until clean. Use when: "review this", "code review", "is this ready to ship", "second-model review". Trigger: /code-review, /review.

Convene a council of DISTINCT OpenRouter model families (via opencode/pi), each in a generative persona, to think divergently about one question, then synthesize as chair. The generative sibling of /peer-harnesses' adversarial bench: councils generate and reframe, roster critiques. Use when: "convene a council", "thinktank", "brainstorm with different models", "get diverse perspectives", "panel of AIs", "what would different experts think", "divergence pass", "ideate broadly", "stress-test this direction with other models". Trigger: /council, /thinktank.

Operate Bitterblossom's `bb` event-plane CLI for agent workloads. Use when Codex needs to inspect or run `bb`, configure `plane.toml`, `agents/`, or `tasks/`, dispatch or audit tasks, handle runs, dead letters, recovery, parked tasks, submissions, gates, review-factory workflows, or help another repo consume Bitterblossom. Trigger phrases: "Bitterblossom", "bb", "event plane", "agent workload", "run a task", "inspect runs", "DLQ", "parked task", "submission loop", "review factory".

Engineer Roster primitives: skills, shared doctrine, provider roster, harness configs, gates, evals, bootstrap, sync. Use for "improve the harness", "harness engineering", "bootstrap is wrong", "AGENTS.md is stale", "skill health", "undertriggering skill", "description tax", "eval skill", "sync primitives", "stack/infrastructure defaults", "one core many faces", "API/CLI/MCP/SDK skill template", "factory product template", "generate repo-local skill", "domain agent skill". Trigger: /harness-engineering, /harness, /skill.

Investigate, audit, triage, and fix. Systematic debugging, incident lifecycle, domain auditing, and issue logging. Feedback-loop-first protocol: reproduce or replay before root cause, pattern analysis, hypothesis test, and fix. Use for: any bug, test failure, production incident, error spikes, audit, triage, postmortem, "diagnose", "why is this broken", "debug this", "production down", "is production ok", "audit stripe", "log issues". Trigger: /diagnose.

Verify the running thing works. Browser walks for web, request replay for APIs, local API emulation for supported third-party services, shell smoke for CLIs, consumer builds for libraries, tool-call replay for MCP. "Tests pass" is not QA. Use when: "run QA", "verify the feature", "test this", "check the app", "smoke test", "exploratory test", "capture evidence". Trigger: /qa.

Backlog grooming as one always-on loop: tidy the board, then sweep — investigate, research, critique, and shape ambitious work. Tidy every run. Use when: "groom", "backlog", "what should we build", "prioritize", "rethink this", "biggest opportunity", "moonshot", "audit skills", "skill quality audit". Trigger: /groom, /groom audit, /backlog, /rethink, /moonshot, /scaffold.

Audit, design, and run repo-owned CI gates, host-agnostic by default — one repo-owned contract that local, GitHub Actions, Azure, or any runner calls. Roster's own gate is fmt/clippy/test plus `cargo run --locked -p roster-cli -- check`; consumer repos keep their native gate. Use when: "run ci", "fix ci", "ci is red", "is ci passing", "audit/design/strengthen CI", "host-agnostic CI", "Dagger", "gates are slow". Trigger: /ci, /gates.

Capture one compounding repo-technical learning while a solved problem is still fresh. Use when: after a bug fix, diagnosis, delivery, review, or incident reveals a reusable pattern worth adding to `docs/solutions/`. Trigger: /compound, /capture-learning, /learning.

Deliver one routed ticket end to end: make the contract executable, build the smallest coherent change, prove the live outcome, review it, and leave it merge-ready or ship when explicitly asked. Use for "deliver this", "build this ticket", "make it merge-ready", "take this end to end". Trigger: /deliver.

Generate world-class, source-verified reference documentation for a codebase: a multi-agent loop that surveys the repo, plans the information architecture, writes facet-scoped pages, and adversarially verifies every claim against live source before committing markdown + HTML + diagrams to docs/. Always runs the full verify loop; scope is incremental by provenance. Use when: "document this codebase", "generate the docs", "build a codebase wiki", "write architecture docs", "onboarding docs", "documentation site", "keep the docs in sync", "world-class docs". Trigger: /document, /docs, /wiki.

Design an LLM/agentic eval that can change a decision — a task + a model or agent under test producing fresh output + a grader — and separate real capability evals from instrumentation (linters, CI gates, KPIs). For proving a harness skill earns its keep, use /skill-eval instead. Use when: designing or critiquing an eval, building an eval corpus, designing or calibrating an LLM-as-judge, or comparing models/harnesses on a measured capability. Trigger: /eval-design.

Edit, audit, or rewrite prose so it sounds like a specific human wrote it, not a generic AI draft — cutting AI tells, filler, and detector-bait while preserving truth, voice, and audience fit. Use when: "humanize this", "make this sound less AI", "remove AI slop", "de-slop this", "edit this prose", "make this sound natural", "fix the writing voice", "rewrite this copy". Trigger: /human-writing, /deslop.

Recommend the best next move from live thread and repo state. Use when: "what's next", "what next", "now what", "what should I do next", "what should we do next", "anything else to do", "where are we now, what's next", "what next in the backlog". Trigger: /next, /what-next, /now-what.

Browser-mode-only Oracle consults: bundle a prompt plus selected files and ask a signed-in ChatGPT GPT-5.5 Pro browser session for a second opinion. Use when stuck, debugging hard bugs, reviewing an architecture plan, or cross-checking a substantive diff with large file context. Never use Oracle API mode from Roster. Trigger: /oracle, /consult.

Fast orientation from live local evidence — at session start, or picking up work a dead session left unfinished. Use when: "orient yourself", "new session", "where are we", "catch me up before acting", after compaction, after switching worktrees, before choosing a workflow; or "pick up where it left off", "the session died / hit its usage limit", "continue that agent's work". Trigger: /orient, /ground, /session-start, /pickup.

Architecture refactor mode: set a concrete improvement goal, refactor until the architecture is simpler and coherent, live-test after each significant step, autoreview, commit green milestones, and track progress in /tmp/refactor-{project}.md. Use when: "refactor this", "clean up the architecture", "make the design better", "refactor until you're happy", "pay down design debt", "simplify this subsystem". Trigger: /refactor.

Turn a working repo into credible external-facing proof: demoability audit, deterministic demo path, marketing site, case study, screenshots, demo video, launch copy, and consulting portfolio assets. Use when: "productize this", "make this demoable", "make this polished", "make a marketing site", "show this off", "demo video", "case study", "portfolio piece", "consulting asset", "launch page", "sales demo". Trigger: /showcase, /productize, /demoability.

Prove a harness skill beats no-skill with a falsifiable A/B eval, or retire it: name the one claim the skill must earn, run it skill-on vs raw same-model, grade blind, return a keep/adapt/cut verdict. For designing an eval for a product or model, use /eval-design instead. Use when: "eval this skill", "does this skill help", "prove the skill beats no skill", "write an eval for a skill", "skill A/B", "skill regression test". Trigger: /skill-eval, /eval-skill, /prove-skill.

Reverse-lookup glossary that turns a vague description of a web animation or motion effect into its exact term ("the bouncy thing when a popover opens" → Pop in; "the iOS rubber-band scroll" → Rubber-banding). Use when the user asks "what's it called when…", or describes a motion effect without knowing its name and wants the right word to prompt an AI or designer with. For naming an effect, not designing or building one.

Apple's approach to interface design and fluid, physical motion, translated for the web. Use when building or reviewing gesture-driven UI, spring animations, drag/swipe/sheet interactions, momentum and interruptible transitions, translucent materials and depth, typography (optical sizing, tracking, leading), reduced-motion, or the design foundations (feedback, spatial consistency, restraint) behind Apple-style interfaces.

This skill encodes Emil Kowalski's philosophy on UI polish, component design, animation decisions, and the invisible details that make software feel great.

Reviews animation and motion code against a high craft bar derived from Emil Kowalski's design engineering philosophy. Default to flagging; approval is earned.