🥾 b00t — Agentic Hive OS

"Tell me what I'm running on, what tools are available, what I'm allowed to do, what goals I should optimize for, and where the boundaries are."

b00t is a loop harness for agents. It bestows a fresh model with typed context, a compendium of skills, formal verification surfaces, and a playbook of low-cognitive-cost deterministic commands — so the expensive tokens go to judgment, not rediscovery.

Design stance, MBSE-style: every capability is a typed artifact (a datum) with a schema, a lifecycle, a verifier, and an evidence trail. The model proposes; the type system, grammars, and solvers dispose. If the harness can check it, the model never gets to lie about it.

---

⚡ Install

curl -fsSL https://raw.githubusercontent.com/elasticdotventures/_b00t_/main/install.sh | bash

Downloads the binary from GitHub Releases and verifies its SHA256 checksum. Linux x86_64/aarch64/armv7, macOS Intel/Apple Silicon.

cargo install b00t-cli          # from crates.io
# or from source:
git clone https://github.com/elasticdotventures/_b00t_.git && cd _b00t_ && cargo install --path b00t-cli

---

🧠 The agent boot sequence

A fresh agent runs four deterministic commands and is operational:

b00t whoami                        # identity, role, session context, boundaries
b00t blessing --manifest --role=X  # prerequisite graph → tool authorization manifest
b00t learn <skill>                 # load exactly the blessings the task needs (context is finite)
b00t task next                     # what to do; b00t task done <id> when verified

Learning a skill datum unlocks the tools in its unlocks field. No learning = no auth. Skills discover by concept, not name — b00t learn "constrained decoding" finds grammar-verify via DWIW fanout (datum search + ontology graph adjacency, weighted).

---

🧬 The type system is the operating system

Every tool, model, skill, role, gate, and lesson is a datum: TOML dialects (.toml < .tomllm < .tomllmd, rank-shadowed by key) with schema stanzas and a machine-readable tail-map. 22+ DatumType variants (cli, mcp, skill, ai, k8s, verifier…).

b00t ontology sparql --subject <X> --predicate all   # walk the triple graph
b00t learn chalk-interner                            # DatumStore ⇄ Chalk Interner mapping
just validate-datums                                 # CI gate over the whole datum tree

Datums are a language. b00t-lsp (tower-lsp over b00t-datum-core) gives editors and agents diagnostics (parse spans, tail-map contract, rank shadowing, unknown types), hover, and cross-datum references (depends_on, composes_with). Tier-1 taplo support ships JSON Schemas generated from the same constants — schema and diagnostics cannot drift. Registered in serena's solidlsp, so symbolic code tools work on the datum graph itself.

---

🔬 Verification: LLM proposes, grammar constrains, Z3 disposes

Structural hallucination is not discouraged — it is made unrepresentable.

b00t learn grammar-verify                      # the full pattern, with recipes
echo '(assert (and (> x 0) (< x 0)))(check-sat)' | z3 -smt2 -in   # → unsat

• Decode-time constraints: GBNF grammars (_b00t_/b00t-verify.gbnf) force claim-shaped output through a verify call; JSON-Schema constraints derive from the consuming Rust type (schemars) and stamp per-backend dialects (vLLM guided_json, llama-server json_schema, OpenAI response_format) via one abstraction.
• Runtime verification: the verify MCP tool routes SMT2 through Z3 (~50ms round-trip); b00t-mcp's LLM proxy executes model-emitted verify tool-calls and audits grammar-shaped claims — a hallucinated sat gets rewritten to the real verdict before the client sees it.
• Gates + evidence: _b00t_/gates/*.gate.toml validate on commit; every PASS appends to the evidence log; PASS evidence converts to verified training examples (just ai-finetune::evidence-train).

---

🧭 Cognitive tiers & the loop

Route work by complexity; compress ruthlessly at every boundary:

Tier	Models	Work	Output contract
sm0l	qwen3.6-A3B, haiku	tests, lint, classify, route	PASS or FAIL: <5-line excerpt>
ch0nky	qwen3.6 local (vLLM/llamacpp)	implement, refactor, debug	diff + test result
frontier	claude opus/sonnet	architecture, novel design	structured decision

b00t-loop -n 10                    # ralph-style iteration loop
b00t ooda run                      # typed Observe→Orient→Decide→Act pipeline nodes
b00t hive activate=<profile>       # resource-gated system state (GPU exclusion groups)

Knowledge flows both directions: b00t lfmf <tool> "<lesson>" memoizes tribal knowledge (salvage-first: malformed input degrades to a tagged lesson, never bail-and-discard; b00t lfmf stats all reports hit/salvage/miss telemetry). Lessons are endurants — temporal bugs go to b00t task add "bug: ..." instead.

---

🔎 Semantic code ops (serena, c0re)

Symbol-scoped reading and patching via LSP — measured 83–96% context savings vs whole-file reads on this repo. Packaging ladder: k8s > podman > host binary > uvx (encapsulation bounds the reasoning surface).

podman build -t serena:latest vendor/serena/     # Containerfile — auditable surface
kubectl apply -f _b00t_/k8s/serena.yaml          # b00t-serena namespace, live pod
b00t mcp install serena claudecode               # datum → registry → client config
scripts/serena-smoke.sh <launch-cmd...>          # same handshake drives every rung

find_symbol, find_referencing_symbols, replace_symbol_body, insert_after_symbol — edits address the symbol graph, not line numbers, so they survive file drift.

---

🤖 MCP integration

# b00t-mcp: compile-time generated tools + exec/discover proxy
claude mcp add b00t -- b00t-mcp --stdio
b00t_discover("<keyword>")         # find the command   → b00t_exec("task list") runs it
b00t_mcp_stack_load("serena")      # dynamic capability extension at runtime

# OpenAI-compatible LLM gateway (b00t-server): soul-registry upstream discovery,
# 🎂 cake budget hard gate, spotlight usage telemetry, agentic verify loop
b00t-mcp --llm -p 3000

Assimilate the outside world: b00t grok assimilate -t <topic> --source-url <url> --b00tyverse distills content to git-blob datums, registers vendor forks, and feeds the ontology.

---

🐝 Hive coordination

b00t agent capability              # announce role + skills
b00t agent discover --role=qa      # find peers
b00t agent message / vote          # A2A messaging + consensus
just compile-agent <role> 3 /tmp/agent.md && claude --agent /tmp/agent.md

Sub-agent output contract: compressed summaries only. Raw output never enters executive context.

---

🎯 Mission

The convergence target: a ch0nky model that speaks native b00t — every workstream feeds one of three legs: training signal (evidence→train, transcript harvest), decode-time constraint (grammars/schemas), or tool surface (verify, serena, b00t-lsp). Don't train the model or evolve the harness — do both, with one shape.

---

🌐 Platform

Linux x86_64/aarch64/armv7 · macOS Intel/AS · single-node k8s (k0s/k3s) friendly · rootless podman (--userns=keep-id) · Python via uv only.

🛠 Development

cargo build --workspace
cargo test -p b00t-cli --lib       # 927 tests
just -l                            # recipe survey
just validate-datums               # datum tree gate

📖 Docs

AGENTS.md (agent protocol) · CLAUDE.md (harness boilerplate) · _b00t_/ (the datum compendium — the real documentation) · b00t learn <anything> (ask the system itself).