cave-teams

Drive the whole cave-teams library from one function, as DATA — build/run any team from a spec in any sequence, save a team to your proven library, reuse a proven team anywhere, and discover teams across projects. Use when you want to serialize a team, run a team from JSON/dict, save+reuse a proven team across repos ("golden"), or have a higher system / global app drive cave-teams. NOT how you normally program (use the cave-teams DSL for that). Triggers: "cave()", "metacontrol", "run from a spec", "team as data", "golden team", "reuse a proven team", "scan for .cave", "global cave".

Build and run programmable agent teams with cave-teams — the provider-agnostic version of Claude Code Teams. Use when the user wants to wire multiple agents together, run agents in sequence or parallel, build a multi-agent pipeline/workflow, make a "team" of coding agents (Claude Code, Codex, MiniMax, or any), loop an agent until a check passes, run a contest between agents, or compose agents with the >> / | DSL. Triggers: "agent team", "multi-agent", "orchestrate agents", "wire these agents", "run agents in parallel", "cave-teams", "agent pipeline", "agent workflow".

Program WHEN each agent may be messaged — conditions on messages, the GUARDRAILS of a leader-driven team run. Use to gate agents on runtime state: after these two respond, only when a flag holds, any predicate over the message log. THIS is what makes cave-teams programmable where Claude Code Teams' flow is fixed.

Run agents that craft, compete, and evolve over generations — the user buys the best, winners reproduce. A compiler (a domain → a trained agent) and a research lab (selection discovers what works) in one engine. `crafter_sim()`.

N agents read and write a shared state over rounds, with a mutator that folds each contribution in and an optional adjudicator (stigmergy arena). Use for agents collaborating on a shared board, a market, a workspace they all see. `blackboard()`.

Run different agents depending on a condition (guarded choice / router). Use for 'if X do this else that', routing, triage, picking a path by the state. `choice()` / `router()`.

Run agents on a dependency graph — each fires when its inputs are ready (the general scheduler; sequential and parallel are its limits). Use for 'merge after frontend AND backend', diamond/fan-in dependencies, partial orders. `dag()`.

Hand off specific named data from one agent to the next, typed (load files, validate expected outputs, map outputs→inputs). Use when step B needs particular named inputs produced by step A, not just the raw text. `dovetail()`.

Reproduce a generation: copy each winner's whole agent directory into the next generation and wipe its session memory (children inherit structure, not history). Use for evolving agents/teams across generations. `evolve()`.

Loop an agent until a checker approves (a bounded fixpoint that always terminates). Use for 'draft then review until approved', refinement loops, write→critique→revise, retry-until-pass. `gate()` / `loop_refine()`.

A self-improving agent stack that gets better as it works: an executor does the work, an observer extracts patterns, a static meta-observer anchors, and a skill-editor persists the learnings. Use for teams that should compound their own capability over cycles. `metacog_shell()`.

Run agents at the same time on the same input (fan-out / scatter). Use for 'in parallel', 'at once', 'concurrently', 'fan out', multiple independent reviewers or checks. The `|` operator / `par()`.

Run an arena over a fixed number of epochs (seasons) that carry, reset, or ratchet a standard between them. Use for tournaments/worlds that run several rounds where the bar climbs each season. `season()`.

Run agents one after another (a pipeline); each reads the previous one's output. Use when one step feeds the next, or the user says 'in order', 'then', 'pipeline', 'A then B then C', 'sequential'. The `>>` operator / `seq()`.

N agents attempt the same task in parallel and a judge selects the winner (also scatter-gather / map-reduce). Use for 'best of N', contests, judge panels, generate-and-select, picking the strongest output. `tournament()` / `synthesis_gate()`.

Compose an arena + gate + season into one GameWorld — a whole simulation/economy that is itself a Link, so worlds nest inside worlds. Use for 'a world of agents', simulations, economies, agents that craft/compete/evolve. `GameWorld`.