phnx-labs
OrganizationAGI CLI ~ a meta-harness for building Agent Factories
Categories
Indexed Skills (61)
browser
Drive a browser via CDP — navigate, click, type, screenshot. Profiles persist login state across sessions, so agents log in once and stay authenticated. Triggers on: 'automate a website', 'browser automation', 'click', 'screenshot', 'scrape', 'log in', 'fill form'.
devices
Register and connect to your machines over Tailscale SSH with agents-cli. Use this skill to sync devices from the tailnet, list them, open a shell on another machine, or see agent sessions running across your whole fleet.
teams
Organize AI coding agents into teams that collaborate on a shared task. Create teams, add teammates, start them, monitor progress, and collect results. Use this skill when you need parallel agent execution. For single-agent dispatch, use `agents run` instead.
routines
Schedule agents to run on a cron schedule or one-shot at a specific time. The scheduler auto-starts on first add. Triggers on: 'schedule an agent', 'recurring job', 'cron', 'daily', 'every weekday', 'run at 2pm'.
secrets
Manage named bundles of environment variables backed by the OS keychain (macOS Keychain, Linux libsecret, Windows Credential Manager). Create bundles, add secrets, generate passwords, and inject them into agent runs. Triggers on: 'API key', 'credentials', 'secrets bundle', 'inject env vars', '--secrets', 'keychain', 'credential manager'.
run
Execute a single agent headlessly or interactively. Supports plan/edit/auto/skip modes, secrets bundle injection, version pinning, fallback chains, balanced rotation, profile dispatch (Kimi/DeepSeek/etc.), and workflow dispatch by name. Triggers on: 'run claude', 'run codex', 'agents run', 'dispatch an agent', 'headless agent', 'one-off agent task'.
sessions
Search, browse, and read agent conversation transcripts across Claude, Codex, Gemini, and OpenCode. Use this skill to find previous sessions, recover context, or inspect what agents have done.
browser
Drive a browser via CDP — navigate, click, type, screenshot. Profiles persist login state across sessions, so agents log in once and stay authenticated. Triggers on: 'automate a website', 'browser automation', 'click', 'screenshot', 'scrape', 'log in', 'fill form'.
devices
Register and connect to your machines over Tailscale SSH with agents-cli. Use this skill to sync devices from the tailnet, list them, open a shell on another machine, or see agent sessions running across your whole fleet.
routines
Schedule agents to run on a cron schedule or one-shot at a specific time. The scheduler auto-starts on first add. Triggers on: 'schedule an agent', 'recurring job', 'cron', 'daily', 'every weekday', 'run at 2pm'.
teams
Organize AI coding agents into teams that collaborate on a shared task. Create teams, add teammates, start them, monitor progress, and collect results. Use this skill when you need parallel agent execution. For single-agent dispatch, use `agents run` instead.
workflows
Create and run reusable multi-agent pipelines. Bundle an orchestrator prompt with optional subagents, skills, and plugins into a named workflow invoked as `agents run <workflow-name>`. Triggers on: 'workflow', 'orchestrator', 'WORKFLOW.md', 'multi-agent pipeline', 'reusable agent task'.
reflect
Recall all feedback, corrections, and constraints from the current conversation before rewriting or iterating. Triggers on: reflect, step back, reconsider, recall feedback, what did I say, incorporate all feedback, or when iterative drafts keep missing the mark.
docs
Write documentation — user-facing, technical, runbooks, onboarding, changelogs. Less is more: only document what code can't tell you.
secrets
Manage named bundles of environment variables backed by the OS keychain (macOS Keychain, Linux libsecret, Windows Credential Manager). Create bundles, add secrets, generate passwords, and inject them into agent runs. Triggers on: 'API key', 'credentials', 'secrets bundle', 'inject env vars', '--secrets', 'keychain', 'credential manager'.
design-drift
Read-only nightly review that scans recently-merged work for DESIGN DRIFT — new primitives added where an existing one should have been reused/extended, producing overlapping surfaces (the `agents feed`/`activity`/`notify`/`message`/`send` smell). Reuses the `quality` skill's engine, ranks findings with a consolidation lens, files (or drafts) Linear tickets, and messages the owner a short summary. Never fixes code. Triggers on: 'design drift', 'overlapping surfaces', 'nightly drift review', 'what merged that should have reused something', 'consolidation opportunities'.
release
Publish packages to registries (npm, CDN, etc.). Discovers repo structure, scaffolds build/release scripts if missing, runs tests, updates changelog, publishes, and tags. Supports monorepos and semver with prereleases. Triggers on: release, publish, ship, npm publish, cut a release.
release-headless
Gotchas for the agents-cli headless release. The macOS home base signs, notarizes, and npm-publishes over SSH with no GUI and no Touch ID — three things must be primed or the release stalls. Triggers on: release stuck at sign/publish, errSecInternalComponent, "no npmjs.com bundle", "Missing embedded.provisionprofile", making the release autonomous.
agents-cli
Run and manage many AI coding agents (Claude Code, Codex, Cursor, Gemini/Antigravity, Grok, Kimi, OpenCode, Droid, …) from one CLI — the `agents` command. Use this to run multiple coding agents in parallel, manage multiple Claude Code accounts, keep working when I hit my usage limit, resume a session on another machine, or pin the agent CLI version. Triggers on: run multiple coding agents in parallel, manage multiple Claude Code accounts, I hit my usage limit, resume a session on another machine, pin the agent CLI version.
run
Execute a single agent headlessly or interactively. Supports plan/edit/auto/skip modes, secrets bundle injection, version pinning, fallback chains, balanced rotation, profile dispatch (Kimi/DeepSeek/etc.), and workflow dispatch by name. Triggers on: 'run claude', 'run codex', 'agents run', 'dispatch an agent', 'headless agent', 'one-off agent task'.
workflows
Create and run reusable multi-agent pipelines. Bundle an orchestrator prompt with optional subagents, skills, and plugins into a named workflow invoked as `agents run <workflow-name>`. Triggers on: 'workflow', 'orchestrator', 'WORKFLOW.md', 'multi-agent pipeline', 'reusable agent task'.
sessions
Search, browse, and read agent conversation transcripts across Claude, Codex, Gemini, and OpenCode. Use this skill to find previous sessions, recover context, or inspect what agents have done.
myskill
Project-scoped skill fixture. Loads when the agent is launched from inside this project tree.
cgraph
Use FIRST for any question about THIS codebase's structure, symbols, or relationships — where a function/class/file is defined, what calls or imports it, what breaks if you change it, how two parts connect, or to load focused source context before editing or reviewing. Routes to the cgraph MCP tools (graph_query / graph_explain / graph_impact / graph_path / graph_context), which serve ranked file:line results, node neighborhoods, transitive blast radius, and token-budgeted source bundles from a resident per-project graph daemon in ~10ms — far cheaper than grepping and reading files. Prefer over blind grep/read for code navigation, dependency tracing, and impact analysis.
scripts
Scripts directory contract — canonical build/test/install/release.sh patterns for deployable projects. Triggers on: scripts/, release.sh, build.sh, deploy, publish.
security
Security audit of a codebase via parallel agents — one per vulnerability class. Reads code fast with Explore agents, cross-checks against current advisories via web search, filters false positives hard. Triggers on 'security scan', 'security audit', 'vulnerability scan', 'check for leaked secrets', 'scan for injection', or scheduled security checks.
run
Execute a single agent headlessly or interactively. Supports plan/edit/auto/skip modes, secrets bundle injection, version pinning, fallback chains, balanced rotation, profile dispatch (Kimi/DeepSeek/etc.), and workflow dispatch by name. Triggers on: 'run claude', 'run codex', 'agents run', 'dispatch an agent', 'headless agent', 'one-off agent task'.
browser
Drive a browser to automate websites — fill forms, click buttons, take screenshots, scrape pages. Uses the built-in `browser` command (or `agents browser`).
routines
Schedule agents to run on a cron schedule or one-shot at a specific time. The scheduler auto-starts on first add. Triggers on: 'schedule an agent', 'recurring job', 'cron', 'daily', 'every weekday', 'run at 2pm'.
secrets
Manage named bundles of environment variables backed by macOS Keychain. Create bundles, add secrets, and inject them into agent runs. Use this skill when working with credentials, API keys, or sensitive configuration.
sessions
Search, browse, read, and move agent conversation transcripts across Claude, Codex, Gemini, and OpenCode. Use this skill to find previous sessions, recover context, inspect what agents have done, or export/import sessions as portable bundles over the SSH fleet.
teams
Organize AI coding agents into teams that collaborate on a shared task. Create teams, add teammates, start them, monitor progress, and collect results. Use this skill when you need parallel agent execution. For single-agent dispatch, use `agents run` instead.
learn
Post-session reflection that writes durable improvements forward. Two modes. Default: reflect on the session you just finished — recall what was used, distill only the lessons that generalize, route each to its right home (skill / rule / memory / nothing). Target mode (`/learn <skill|plugin|command|workflow>`): audit every past session that used that target, surface the recurring problems as an HTML triage report (each framed expectation → what happened → why, anchored to the session that surfaced it), then apply only the fixes you approve. Built to NOT downgrade existing workflows and NOT overfit to one session. Triggers on: 'learn from this', 'reflect and improve', 'update your skills', 'what did we learn', 'audit my rush:design skill', 'where does <skill> keep going wrong', 'retro', 'post-mortem this session'.
loop
General-purpose unattended work drain across every project and kind — code, browser, outreach, design, portal tasks. Spreads load with agents teams + balanced account rotation so one logout or rate-limit does not kill the night. Drives each item to landed — engineering merges on green behind a non-author review, never waiting on the user to click merge. Its `triage` mode forces every open item to keep-and-schedule-this-cycle or cancel instead of draining. Triggers on: 'work loop', '/work:loop', 'overnight drain', 'drain the board', 'finish everything unattended', 'keep moving on all projects', 'triage the board', 'clean up the backlog', 'decide keep/cancel on every ticket'.
refactor
Restructure a codebase the way a principal engineer does as a product grows — merge redundant concepts, extract the horizontal layer four modules reimplemented, draw real module boundaries, lift a cohesive core out into its own package/SDK so it is testable and reusable, reorganize the tree so it matches the architecture, shrink an overgrown public surface, and give a concept the contract its job calls for — the provider pattern (Go interface, TS interface, Python Protocol, Rust trait) with one implementation per variant in a registry, instead of the same if/else-by-name chain repeated across twenty files. Evidence-first: a module dependency graph (god modules, cycles, extraction candidates, upward imports), measured agent traffic per file, and a surface census — then before/after architecture figures rendered with artifacts-cli, then behavior-preserving PRs. Triggers on: 'refactor this codebase', 'clean up the architecture', 'the codebase has gotten messy', 'agents keep getting lost in this repo', 'merge the
review
Review PRs and repo health with one skill, three modes. Default: recap the session's goal, discover every PR it opened, and review/merge each in dependency order. Given PR number(s): a deep sub-agent review (not the author) with file:line grounding, an architecture rubric (reuse, cross-cutting-at-source, no duplicate surfaces, doc-asserted invariants), and a security pass on risk-touching diffs. Given `repo` / a path / `--since <date>`: a read-only whole-repo architecture-and-quality diagnostic (four orthogonal categories, HTML report, no merge verdict). Triggers on: 'review the PR', 'code review', 'review #N', 'before I merge', 'block this PR', 'security review', 'what's wrong with this branch', 'health check', 'parallel implementations', 'audit drift'.
continue
Resume previous agent work in this session — load a prior transcript (or a group of them), verify what actually landed, and finish the remaining work here. Reattach only when the session is genuinely still live and interactive. Also covers post-crash multi-session recovery (prefer finishing headlessly over resurrecting terminals). Triggers on: /continue, 'pick up where I left off', 'resume that session', 'continue the auth work', crash recovery that finishes work rather than reopening windows.
finish
Drive the current task to fully delivered — recover the goal, finish what remains, verify end-to-end, docs, commit, PR, release checklist, close the ticket. Never stops at a recap, blocker, or partial handoff. Triggers on: /finish, 'finish this', 'drive it to done', 'ship it', 'don't stop at a recap', 'see it through'.
insights
Analyze how you and your agents work across harnesses — orchestrates agents insights, insights mix, perf, and sessions stats into one evidence-backed action list. No separate trends/perf slash commands. Triggers on: /insights, /sessions:insights, 'session insights', 'how have I been working', 'where do agents stall', 'skill dead weight', usage friction, harness mix.
search
Pull ALL relevant prior-session context on a topic — fast, ranked, snippet-level, without loading full transcripts. `agents sessions "<q>"` alone misses assistant answers (not indexed) and returns whole-session refs, not snippets. This skill layers CLI discovery with a bundled fallback script that recovers what the index misses. Triggers on: /sessions:search, /recall, 'find what I said about X before', 'did we already solve this', 'pull context from a past session', 'search my sessions for'.
share
Publish an agent-generated HTML artifact (a plan, viz, or report) to a shareable link via the managed Phoenix endpoint by default, or the user's own Cloudflare R2 when BYO is configured. Public gets an auto Open Graph cover so the link unfurls into a preview card in Slack/iMessage/Twitter/Discord; pass --private for an unlisted auto-expiring link with no card. Use when an agent has produced HTML worth handing to a human, or when a plan/viz should outlive /tmp. Triggers on: 'share this', 'publish the plan', 'make a link', 'shareable link', 'send me the plan', 'og image / preview card for this', '/share --private'.
debug
Debug end to end with swarm verification — clarify intent vs observed, check the feature's spec for the gap that let it slip, trace the data path, attribute regressions to the agent/session that caused them, have independent agents (different providers, blind) confirm the root cause, then close the loop with a viewable artifact, a ticket, and a dispatched fix. Use for any 'here's what I saw, here's what I expected' problem. Triggers on: 'debug', 'swarm debug', 'root cause', 'why is this happening', 'what was my intent vs what happened', 'confirm the bug', 'independent debug'.
orchestrate
Fan a task out across a swarm of parallel coding agents via `agents teams` — the engine behind every /swarm:* command. Build a distribution plan with boundary contracts, spawn a mixed team (claude/codex/antigravity), monitor, and synthesize. Use whenever a task is wide enough that one agent would serialize 4+ independent edits, or when you want independent agents to verify a conclusion instead of trusting one. Triggers on: 'swarm', 'fan out', 'spin up a team', 'parallel agents', 'distribute this', 'independent verification'.
plan
Plan a feature with swarm verification — research hard, produce mock-ups for any UI/flow, draft a behavior-first change proposal, then have independent agents plan the same thing blind and reconcile. Use before building anything non-trivial. Triggers on: 'swarm plan', '/swarm plan', 'plan with verification', 'change proposal', 'plan and check the approach'.
spec
Write the durable source-of-truth specification for a capability so other agents and humans do not invent wrong behavior — describe how it actually behaves in plain language: what it is for, what it does, the sharp cases, and what must not change, grounded in the real code. Includes mock-ups for any UI/flow, blind-verified by a swarm. Not a change plan. Triggers on: 'swarm spec', '/swarm spec', 'specify this capability', 'source-of-truth spec', 'how does this behave', 'what does this guarantee'.
demo
Demonstrate landed work instead of just claiming it shipped. Recover the ORIGINAL intent (not the diff), exercise the shipped thing in its REAL environment — the installed/deployed artifact, never the dev build — drive it on the real surface with `agents browser`/`agents computer` signed in as the owner on REAL representative inputs (never toy examples), put before/after side by side with a measured delta, then deliver an analyzed HTML report on the owner's screen and attach it to the PR. Triggers on: /demo, /work:demo, 'show me a demo', 'prove it works', 'test it in prod', 'how does it look side by side', 'did we ship all of it', 'demo what we just shipped', after any landing/merge/release.
research
Answer a hard research question by fanning it out across DISTINCT search modalities, each blind to the others, then synthesizing one sourced artifact. Codex agents run aggressive web search; Grok agents mine X/Twitter community chatter (its privileged data); Antigravity works Google; Perplexity runs a broad browser-driven Deep Research report; the Claude fleet deep-reads and synthesizes. Multi-modal because no single engine sees everything — a company's funding is in a press release, its reputation is in X replies, its moat is in a founder's blog. Triggers on: /research, /research:research, 'research X across sources', 'deep research on', 'market/competitive/landscape research', 'what's the real story on <company/topic>', 'pull everything on', 'multi-source research'. For a hands-on PRODUCT exploration (drive it, screenshot each journey, prove claims visually) use /research:product.
resume
Pick a whole PROJECT's work back up. Best-effort auto-detects the project from the CWD (GitHub repo + subdirectory → Linear project), reconstructs its in-flight work (live + interrupted sessions, open PRs, worktrees, open/doing tickets), presents it not-progressing-first, then OFFLOADS each item to a role=worker device — never the interactive/personal box. Not sessions:continue (one transcript) and not work:loop (the whole board). Triggers on: /work:resume, /resume, 'pick this project back up', 'resume the work on <project>', 'what was in flight here', 'get this project moving again'.
cloud
Dispatch and manage agent tasks in the cloud across providers — Rush Cloud (GitHub repo + branch, auto-opens a PR), Codex Cloud (pre-built env), and Factory pods (Droid + computer-use). Covers run, list, status, logs, message, cancel, providers. Triggers on: 'cloud run', 'dispatch to the cloud', 'rush cloud', 'codex cloud', 'run an agent in the cloud', 'cloud task', 'agents cloud'.
computer
Drive native macOS apps — screenshot windows, click, type, drag, read text. Uses the built-in `agents computer` command (Accessibility + ScreenCaptureKit daemon). Triggers on automating desktop apps (Photoshop, Parallels VMs, Finder, any non-browser GUI), "computer use", clicking/typing in a Mac app, or capturing an app window.
monitors
Durable event-triggered watchers: watch a source, detect a change, and fire an agent, a routine, or a notification. The cross-agent layer — agents watching sources (the fleet, other agents) and reacting. Triggers on: 'watch for', 'monitor', 'when X changes run Y', 'notify me when', 'poll until', 'fire an agent on', 'watch the fleet', 'watch CI'.
product
Explore a PRODUCT hands-on and prove every claim visually — don't just read about it. Install or sign up, DRIVE the real product through each user journey, screenshot every meaningful step, record a short clip of the headline flows, then put landing-page CLAIMS next to what the product ACTUALLY did. Produces a highly-visual artifact: favicon/logo-tagged product cards, per-journey flow diagrams, screenshot strips, an embedded clip, and a claims-vs-reality table — never one idle screenshot and a wall of text. Composes research:research for the public intel, browser + secrets to drive the surface, and artifacts to render. Triggers on: /research:product, 'explore <product> hands-on', 'how does <product> actually work', 'test out <product> and show me', 'does <product> really do X', 'walk through <product>'s features', 'compare these products by actually using them', 'evaluate <product>'s user flows'.
reflect
Recall all feedback, corrections, and constraints from the current conversation before rewriting or iterating. Triggers on: reflect, step back, reconsider, recall feedback, what did I say, incorporate all feedback, or when iterative drafts keep missing the mark.
docs
Write documentation — user-facing, technical, runbooks, onboarding, changelogs. Less is more: only document what code can't tell you.
myskill
Project-scoped skill fixture. Loads when the agent is launched from inside this project tree.
design
One keyless, offline-first front door for design. Routes a design intent to a mode and renders it as self-contained HTML/SVG (no CDN, no paid keys): UI screens and flows, clickable prototypes, design systems and tokens (including brand identity, BRAND.md), architecture/flow/ER diagrams, infographics and data-stories, slide decks, vector assets (OG cards, SVG logos, icon sets, posters), critique of an existing screen, and anticipating a flow's dead-ends. True raster (photo, illustration, painterly cover) is an optional layer that degrades to a spec plus an editable placeholder rather than hard-failing. Every mode loads design-core first (hierarchy, WCAG AA contrast, colorblind-safe palettes, brand-probe, the anti-tells catalog of what makes a design look AI-generated, precise non-marketing copy, render/screenshot/critique verification). Triggers on: design a screen/page/UI, mock up, prototype, design system, tokens, brand, BRAND.md, wireframe, diagram this, infographic, dataviz, slide deck, logo, OG image, soc
fork
Fork a session into a NEW, independent same-harness sibling — the 'git branch' of sessions. Resolved across the fleet and seeded with a recap of the source (label, cwd, ticket, last state, changed files), so it works for a session on ANY device in ANY harness and picks up where the original left off. Where continue resumes the SAME thread, fork launches a new one. Triggers on: /sessions:fork, /fork, 'fork this session', 'branch this conversation', 'copy this session into a new one', 'explore an alternative direction without touching this thread', 'fork <id>'.
workweave
Render a fast, private WorkWeave-style engineering-intelligence report from every locally indexed agent session for the current repository. Reuses agents sessions, insights, output, cost, resource stats, hook/command latency, and guard-friction engines; writes Markdown + HTML and opens the result. Triggers on: /yc:workweave, workweave report, project agent analytics, engineering intelligence dashboard, chart our agent sessions.
artifacts
Author plans, reports, and visual explanations as Markdown, then render them with artifacts-cli into self-contained branded light/dark HTML. Use for implementation plans, plan mode, architecture diagrams, infographics, dashboards, comparisons, data stories, or any request to render or present an artifact visually.
mq
Structure-aware context query for large files — probe structure, then extract only the section you need, instead of reading whole files into context. Works on Markdown, HTML, PDF, JSON, YAML, CSV, XLSX, DOCX, PPTX AND source code (Go/Python/TS/Rust/...). Triggers on: reading any file 200+ lines, exploring a directory you haven't seen, extracting one section/function/endpoint from a large file, searching a topic across many files, 'use mq', or when reading a full file would waste context.
tickets
Work with the project's issue tracker (Linear, GitHub Issues, Jira, GitLab, etc.) — auto-detect whichever tracker is available (a loaded tracker skill, an installed CLI, or a repo signal), then list / claim / comment / close / create / search issues, always closing with proof. Also drives the check-first / open-if-missing / close-on-delivery ticket lifecycle the conventions rule asks for. Triggers on: 'tickets', 'issues', 'issue tracker', 'my queue', "what's on my plate", 'claim/close/comment/create/search a ticket or issue', 'open an issue', 'move it to In Progress', 'linear', 'gh issue', 'jira', 'the board'.
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.