8000 GitHub - laminko/anyteam · GitHub
[go: up one dir, main page]

Skip to content
laminkoPublic

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

anyteam — one AI dev team, any agent harness

anyteam

One AI dev team, any agent harness. A portable hub-and-spoke specialist team you can install into any of several AI coding harnesses with one command. The team is a small roster of role-specialised subagents — Product Manager, Architect, Backend, Frontend, two UI/UX roles, and two security roles — plus an operating manual that tells the lead agent how to delegate to them and synthesise their work.

One canonical source, rendered into each harness's native subagent format. Define the team once; install it everywhere.

Supported harnesses

Only harnesses with a file-definable subagent roster (where a lead can delegate to predefined named specialists) are supported.

Harness Roster location Manual file Model tiering
Claude Code .claude/agents/*.md CLAUDE.md opus / sonnet
OpenCode .opencode/agents/*.md AGENTS.md anthropic/claude-*
Gemini CLI .gemini/agents/*.md GEMINI.md gemini-2.5-pro / flash
Codex CLI ~/.codex/agents/*.toml (user-level) AGENTS.md gpt-5-codex + reasoning effort
Pi .pi/agents/*.md AGENTS.md anthropic/claude-*

The lead role runs on the heavy tier; pm, architect, and the two security roles are heavy, the rest are light. Model ids are sensible defaults — tune them in src/harnesses.json and re-run the generator (see Contributing).

Install

git clone https://github.com/laminko/anyteam.git
cd anyteam
bash install.sh

Run in a terminal, the installer prompts you to pick the harness(es) to install (space-separated numbers/names, or all; default: claude-code) and to confirm the target project directory. Any flags you pass pre-fill those prompts, so pressing Enter accepts them.

Or skip the prompts entirely with flags (also how piped / CI runs are driven — with no terminal the installer doesn't prompt and falls back to flags + auto-detection):

bash install.sh --harness claude-code   # one harness
bash install.sh --harness opencode,pi   # several
bash install.sh --all                   # every supported harness
bash install.sh --dir /path/to/project  # target a different project
bash install.sh --dry-run               # show what would happen, copy nothing

The installer is no-clobber by default (it skips files that already exist; pass --force to overwrite) and is plain bash (3.2-compatible — works with macOS system bash).

Per-harness notes

  • Codex installs the roster user-level at ~/.codex/agents/ (honors $CODEX_HOME) to avoid a project-scope spawn bug; the manual still lands in the project as AGENTS.md.
  • Pi installs only the (inert) roster files by default. Making its subagents work requires a separate, third-party community extension — see Enabling Pi subagents below. The installer asks for your consent before installing it.

Enabling Pi subagents (opt-in)

The Pi roster does nothing until you install @tintinweb/pi-subagents — a third-party community package (MIT, maintained outside anyteam). This is the only step in anyteam that runs code from outside this repo on your machine, so it's opt-in — the installer asks (default: no) before installing it, and you can also run it yourself, pinned to a version you've reviewed:

pi install npm:@tintinweb/pi-subagents@0.10.0

Then trust the project folder when Pi prompts. (Without the extension, the .pi/agents/ files are simply ignored — nothing breaks.)

After install: ground the team

Installing copies files; it does not teach the team your codebase. Run the one-time intake (INTAKE.md): the lead scans the repo and fills the ## Project brief section of the installed manual (stack, frameworks, run/build/test, conventions, …) so every specialist inherits the project's real context. In Claude Code, the bundled SKILL.md wraps install + intake as a /bootstrap-team skill.

The team

Agent Role Owns
pm Product Manager Spec, scope, acceptance criteria (WHAT / WHY)
architect Architect Data model, API contract, component boundaries (HOW)
be Backend Engineer Server-side implementation against the contract
fe Frontend Engineer Client / UI against the contract
uiux-audit UI/UX Auditor Read-only usability / accessibility review
uiux-research UI/UX Researcher Patterns / evidence before building
security-research Security Researcher Threat model + security requirements before building
security-audit Security Auditor Read-only review of code / diffs for vulnerabilities

The human talks to the lead (the main agent thread); the lead delegates to specialists and synthesises their results. Specialists do not talk to each other.

How the team works

A feature flows through a pipeline the lead adapts per task — pm (spec) → architect (contract) → security-research (threat model) → be/fe (build) → uiux-audit + security-audit (review) — and the lead synthesises the result. Coordination runs over a shared blackboard: the lead curates a per-feature .team/<feature>.md (spec, contract, security requirements, findings) that specialists read, instead of re-pasting context into every prompt. Security is traceable end to end — security-research emits numbered requirements (SR-n) that be/fe build to and security-audit verifies against.

flowchart TD
    Human(["Human"])
    Lead{{"Orchestrator / Lead<br/>· also PM scoping"}}
    BB[("Shared blackboard<br/>.team/&lt;feature&gt;.md")]
    Arch["architect"]
    Eng["be · fe"]
    Sec["security-research · security-audit"]

    Human <-->|"goal · synthesis"| Lead

    Lead <--> Arch
    Lead <--> Eng
    Lead <--> Sec

    Lead ==>|"sole writer"| BB
    BB -.->|"spec"| Arch
    BB -.->|"contract · SR-n"| Eng
    BB -.->|"contract · diff"| Sec
Loading

The lead is the hub: it delegates to each specialist and synthesises their reports (double-headed arrows). Specialists never talk to each other — the only shared state is the blackboard, which the lead alone writes and specialists read (dashed arrows: e.g. be/fe build against the ## Contract and the SR-n requirements). Simplified for clarity — see the full roster (incl. pm, uiux-research, uiux-audit) in The team above.

The full operating manual ships as each harness's instruction file (CLAUDE.md, AGENTS.md, GEMINI.md); the canonical source is src/manual.md.

Contributing

The repo is the source + build system, not a pile of hand-written configs:

  • Edit roles in src/roles/*.md and the shippable manual in src/manual.md.
  • Per-harness mapping lives in src/harnesses.json.
  • generate.py (Python 3, standard library only) renders src/ → dist/<harness>/. The committed dist/ is what install.sh copies.
  • Never hand-edit dist/. Run python3 generate.py; CI fails on drift (python3 generate.py --check).

See CLAUDE.md for the full contributor guide.

License

MIT

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages

0