bee

Overview ¶

`bee agents` subcommand — opens the parallel-agents overview. Each chat submission spawns a detached headless agent in its own git worktree; the overview lists status + tokens + last thought across all of them.

`bee bench`: run a task suite through the headless /goal loop against the configured (local) model and emit a scored scoreboard. Measurement only — tuning bee from the scores is the caller's job.

Background-task spawner. `bee bg <message>` re-execs the bee binary in headless mode with a pinned session id, redirects stdout+stderr to a log file under ~/.bee/sessions/bg/, detaches via Setsid, prints the id to the user, and exits.

Introspection:

bee bg --list            list background bees
bee bg --tail <id>       follow a background log
bee bg --kill <id>       stop a background bee

`bee doctor` — read-only preflight check.

Reports environment health: bee dirs, sandbox helper, configured providers' creds, active config/profile/caveman level. Prints a human-readable table by default; --json emits one machine record.

Pure-Go subcommand-dispatch model — no shim sprays to check.

Local environment probes for `bee doctor`: bee dirs, sandbox helper, loaded config + per-provider credentials. Pure read-only.

Ollama-specific probes for `bee doctor`. Daemon liveness, model availability, num_ctx cache warm-up. Only runs when ollama is the active provider.

`bee fan` — spawn N parallel bees over a workload. Each bee owns its own Engine + session; results stream back via internal/hive's Pool.

CLI hive lister. `bee hive` prints all known bees: background tasks and recent sessions, with status and a preview of the first user message.

`bee hyperplan` — adversarial planning swarm.

5 Critic engines attack a plan from distinct angles (security, perf, complexity, scope, edge cases), then a Synthesizer (Queen) consolidates the critiques into a refined plan.

First-run init: ensure ~/.bee/skills exists and seed bundled skills when the user has none. Silent on success; never fatal.

Command bee is a pure-Go self-learning minimalist coding agent.

Subcommand dispatch is intentionally tiny — stdlib only. Unknown arg[1] falls through to skill lookup so `bee <skill> ...` runs the named skill non-interactively.

remote-control: a local web relay. bee binds an HTTP server on the LAN; open the printed URL (or scan the QR) on another device to drive this session. Everything executes locally on this machine — no cloud proxy.

Headless run path: parse flags, build engine, drive one turn, exit.

Skill shims and CI smoke all exec through here. Pipes user message in from args or stdin; streams assistant text to stdout; prints session id to stderr on exit so callers can resume or inspect.

Background-loop runner for `bee run --bg-loop`. Drives turns across inbox polls, writes status sidecars, exits on ctx cancel.

Knowledge-store adapter wiring for `bee run`. Implements loop.KnowledgeStore with a two-phase query: deterministic scoring first, optional small-model tag extraction second when phase 1 is thin.

Provider construction and input plumbing for `bee run`. Wire-api routing, test-stub short-circuits, prewarm probes, stdin/flag resolution.

Tool registry construction for `bee run`. Wires the built-in tool set, optional approval gate, write-path filter, user-defined tools, and disabled-tool exclusion.

`bee swarm` subcommand: queen + N workers for one complex task.

Wires loop.Engine instances (one planner, N workers) into hive.Queen, runs the decompose → dispatch → synthesize pipeline, prints progress.

Interactive TUI entry point. Builds the same Engine as `bee run` but hands it to the bubbletea program in internal/tui.

`bee waggle` inspects and curates the procedure-memory library: routes the agent crystallized from repeated read-only tool use during sessions.

Subcommand `bee zzz`: overnight autonomous-commit loop. gnhf port — see internal/zzz for the inner loop. Builds an engine the same way runHeadlessReal does, then hands control to zzz.Drive.