botte-secrete

Memory as a learnable skill — store, recall, compress, and consolidate agent memories. Inspired by Stanford AutoMem.

Auto-decide whether a task runs on a LOCAL model or a CLOUD model (DeepSeek, GLM, Nemotron, Grok, Gemma, …) from an automatic effort estimate, and run multi-model fusion (cascade, draft→refine, vote). Use when the user wants automatic local-vs-cloud routing, to add cloud LLM providers, to make local and cloud models collaborate, or mentions effort-based routing, model fusion/ensemble, OpenRouter, DeepSeek, GLM, Nemotron, or Grok.

Fenêtres de contexte pour boucles rétroactives — charge seulement les deltas.

Optimisations DAG/RAG — waves, pruning, memoization, routing.

Generate one self-contained, timestamped HTML dashboard of the system's cost picture — routing savings (control loop), metric trends, current metrics, and the cost of outstanding fixes. Also renders as a live ANSI terminal view (--tui, --watch) and serves a live HTTP API (api.py). Use when the user wants a single visual view of cost/savings/health over time, or a live terminal view they don't have to open a browser for.

Append-only JSONL decision log (.botte/events.jsonl) that every filter in the belt writes to — routing, cache hits, escalations, micro-NN outputs. The single source of truth demo mode, the live dashboard, and session replay all read from. Use when you want to see or emit a live feed of routing/cache/escalation decisions, or when building a tool that needs to watch the belt work in real time.

Vérification différentielle — ne vérifie que les sections modifiées.

MCP server that lets Claude Code (or any MCP client) discover and call local LLM servers (LM Studio, Ollama, …) as tools, to offload cheap tasks off the cloud. Use when the user wants to wire local models into their agent, register an MCP server, or have the agent automatically route simple tasks to local hardware.

Compression token-level — hashing sémantique + byte-pair pruning.

Audit the local cluster's hardware/software/MCP setup and recommend changes that cut token cost — GPU upgrades, Hailo NPU for vision, moving the inference node to Linux, running Qdrant locally, wiring MCP — with an ASCII cluster diagram. Also provides an auto one-pass audit (directives + infra + code duplication + skills) on the project. Use when the user asks how to reduce token/usage cost via hardware/infra, wants cluster setup tips, an ASCII diagram of their setup, or a quick all-in-one audit.

Cost-focused project metrics, broken down per component — LOC by language and component, duplicate-function groups, directive health, always-on context cost (CLAUDE.md tokens × turns), local-routing posture, skill-search tokens avoided, and the audit's own (near-zero) cost. Use when the user wants to quantify a project's token/cost profile, see LOC/health per component, or measure what the toolkit saves.

Scan Python skills and MCP servers for malicious code — dangerous imports, network exfiltration, filesystem abuse, subprocess injection, obfuscation, crypto weakness, environment leaks, and supply-chain attacks. Use when auditing code for security, running pre-commit hooks, or preparing CI/CD security gates.

Clean development caches (.botte-cache, .pytest_cache, __pycache__, .mypy_cache)

Estimate what a task or a fix will cost — tokens, model/tier, money ($), and wall-time — using the tiered cost model. Use whenever the user wants to know the cost of a correction, audit, or task before running it, or to compare local vs cloud cost.

Headroom-inspired multi-type compression — text, JSON, logs, tool output, code. Reversible. MCP server compatible.

Transparent LLM compression proxy for botte-secrete — sit between any AI agent and its LLM API to compress requests by 40-95%. Use when you want token savings without changing agent code.

Run the canonical, already-optimal project checkup in one command — policy presence, directives health, per-component metrics, infra tips, duplication, and drift detection — so you never have to hand-write a good checkup prompt. Use when the user says "do a complete checkup", "botte doctor", after a component update, when onboarding to a project, or when multiple agents/devs may have caused drift.

Measure a project's always-on prefix (agent directives + core rules + MCP tool schemas + skill catalogue) in tokens and as a % of small local-model windows (64k/128k/256k), with a concrete reduction plan (lazy tool loading, on-demand skill search). Use to see how much of a modest machine's usable context is spent before any real work, and how to shrink it so weaker machines can run local LLMs usably.

Live ANSI dashboard of the belt's decisions — routing, token savings, micro-NN outputs, escalations, cache hits — either a built-in scripted scenario (no LLM, no network, works on a bare machine) or tailing a real project's event log. Use when the user wants to see/demo what the routing belt is doing, record a README GIF, or watch live decisions while an agent works.

Expose auto_route/local_chat/fusion/find_skills/infra_tips to Hermes-Agent (or any framework that expects OpenAI-function-calling tool specs instead of MCP) — plus a one-call MCP config generator for the zero-code path if the framework already speaks MCP. Use when connecting botte-secrète's routing belt to another agent framework.

One-line summary of the belt's session activity (tokens saved, cache hits, local/cloud split, escalations) for a terminal statusline — Claude Code's statusLine hook, tmux, or any shell prompt. Reads .botte/events.jsonl. Use when the user wants a persistent, passive view of savings while they work, or asks to set up a statusline.

MCP Gateway — expose toutes les skills Botte comme outils MCP. Découverte automatique, schémas d'entrée, transport stdio. Compatible Claude Code, Codex, Cursor, et tout client MCP.

Meta-Harness orchestre les skills Botte comme des agents interchangeables dans un pipeline gouverné. Planifie → exécute en sandbox → review croisé → applique avec garde-fous. Inspiré d'Omnigent mais 100% Botte-native.

Distillation automatique cloud → micro-NN.

Segmentation multi-window du contexte — chargement sélectif.

Arbre des préfixes de prompts — diffing entre agents.

Agents autobudgétaires — gèrent leur propre budget token.

Shaping dynamique per-turn — adapte la compression à l'effort.

Cache les réponses des agents pour skipper l'exécution quand l'output est prédictible. Use when you want to cut 10-15% by avoiding redundant agent runs.

Prefix prune le contexte — arbre de préfixes, diffing, élague les sections inutilisées. Use when you want to cut 5-10% more tokens by removing dead context.

Ponytail-inspired YAGNI enforcement — climb the decision ladder before writing any code. stdlib → regex → existing module → new code.

Local CWE knowledge base (RAG) to enrich and de-noise security findings — match a finding (or any text) to the relevant weakness by exact id or local-embedding similarity, and attach the weakness name, description, and concrete mitigation. Deterministic, offline, 0 cloud tokens. Use to explain a CWE, find the likely weakness for a code snippet, or enrich taint/security findings with "why + how to fix".

Lightweight SBOM scanner for Python/Rust/Node dependencies

Reproducible token/cost benchmark — runs a fixed task corpus through the real auto_router decision logic and compares it against a "no routing, everything to cloud STANDARD" baseline. Turns the README's savings claim into a checkable number instead of an assertion. Use when the user wants proof of token savings, a benchmark for a PR/README, or numbers to back an integration pitch (e.g. Hermes).

Discover, audit and use local LLM servers (LM Studio, Ollama, LocalAI, vLLM, llama.cpp) on this machine or the network to offload work from the cloud and save tokens. Use when the user mentions local models, LM Studio, Ollama, "run it locally", token savings via local hardware, or wants to know what models their machine can run.

Audit the micro-NNs — is each model grounded in REAL data, or a synthetic copy of a hand-coded rule? Scans skills/botte_nn and reports, per model, the training data source (real/synthetic/unknown), whether the model file records provenance (trained_on/eval_accuracy), whether a test guards a real-world output, and a grounded/synthetic verdict. Deterministic, 0 cloud tokens. Use to tell which learned components are real vs placeholder, and which should be grounded or replaced by the rule they imitate.

FastContext Agent — exploration repo déterministe. Parse une requête d'exploration → READ/GLOB/GREP ciblés → rapport compact (fichier:ligne:score). Remplace 56% des appels LLM de type "read/search" par des opérations stdio à ~5ms, 0 token. Use when the agent needs to understand a codebase, find patterns, locate imports, or gather context without an LLM call.

Trajectory Learning for Botte Secrète — stores solver trajectories and searches similar past optimizations to inform future decisions

Deterministic combinatorial solvers in stdlib — balance work across workers/backends (assignment, LPT), pack items under a capacity (bin-packing, FFD), and order plan steps under dependencies into a sequence + parallel waves (DAG topological scheduling, cycle-detecting). 0 cloud tokens, repeatable. Use to spread cluster work, pack tasks under a budget/capacity, or order a plan's steps — instead of asking an LLM to figure out the assignment/order.

Classify and extract from text WITHOUT an LLM — intent classification (lexical overlap + local embedding), entity extraction (regex/gazetteers for urls/emails/ips/paths/env vars/flags/numbers), and stopword-filtered keyword frequency. Deterministic, instant, 0 cloud tokens. Use instead of asking a model to classify text or extract entities, and as the routing/intent layer that keeps cheap language decisions off the LLM.