Repository navigation
Releases
Botte Secrète 1.11.2
Botte Secrète 1.11.2 réconcilie les sources de la release 1.11.1 avec main et corrige la répartition MCP : coûts pondérés, validation des entrées, limites de taille, recherche bornée et repli LPT explicite. La documentation précise le traitement des éléments trop grands pour bin_pack.
Validation : CI de la PR et du commit de fusion réussie ; suite locale complète de 129 suites et 2 476 tests avant les derniers ajustements ciblés ; 44 tests ciblés et 40 sous-tests après ajustement. Le wheel joint a été installé isolément et vérifié pour les coûts, le repli et les limites. Ses 798 fichiers skills ont été comparés aux sources retenues en normalisant les fins de ligne.
Les archives sont liées au commit 986a579. La qualification native d'une installation fraîche dans Codex Desktop reste distincte ; aucun gain de coût, d'énergie ou de qualité n'est extrapolé. Le changelog du dépôt indique encore « candidate » pour cette version ; les preuves précises figurent dans VALIDATION-1.11.2.json.
Botte Secrète 1.11.1
Botte Secrète 1.11.1 — version corrective officielle.
Le plugin annonce désormais le catalogue MCP complet avec
BOTTE_MCP_LAZY_TOOLS=0. Les outils spécialisés sont enregistrés directement
dans Codex ; find_tool n'est plus présenté comme un mécanisme d'enregistrement
de méthodes. Le mode compact reste disponible dans le runtime pour les autres
clients. Les guides liés par le README sont inclus dans le plugin autonome.
Validation de ces octets :
- Windows : 129 suites, 2 478 contrôles annoncés, zéro échec.
- Ubuntu/WSL : 129 suites, 2 481 contrôles annoncés, zéro échec.
- Contrats Memory Hub supplémentaires : 127 réussis sur chaque plateforme.
- Même wheel installé isolément sur Windows et Linux : fichiers vérifiés,
dépendances et CLI compatibles, trois scénarios MCP réussis par plateforme. - La nouvelle régression vérifie les outils et leurs schémas dans le catalogue
réel avant les appels ; elle a échoué sur l'ancienne déclaration du plugin.
Commit : 2704094f5917b0f5d384b571f195db5968c4811b. Les 1 261 fichiers de l'inventaire validé correspondent
exactement aux blobs de ce commit. Les anciennes archives 1.11.0 sont conservées.
Pièces jointes : wheel Python, sources, catalogue Codex, manifeste des fichiers,
preuves résumées et sommes SHA-256. Installer le wheel dans un environnement
dédié et enregistrer le catalogue Git au tag v1.11.1 ou à ce commit exact.
Le runtime botte-mcp doit être disponible sur le PATH de Codex.
Les manifestes de construction décrivent l'état avant publication ;
PUBLICATION.json les relie au commit et au tag. Les annotations de sauts propres
aux plateformes et les limites de couverture sont conservées dans VALIDATION.json.
Linux est validé sous WSL ; aucune qualification ARM/Pi ou GPU ni aucun gain
facturé ou de qualité générale n'est déduit de ces tests. La vérification native
de l'installation officielle Odin est consignée séparément après publication.
Botte Secrète 1.11.0
Botte Secrète 1.11.0 — version stable.
Le cache persistant conserve les écritures concurrentes et respecte les
effacements et expirations. Les lectures de preuves et verrous Windows sont
corrigés, et les tests sensibles aux fins de ligne et aux chemins sont portables.
Les fonctions de supervision sémantique et les empreintes de contexte du cache
sont conservées.
Validation de cette source :
- Windows : 129 suites, 2 476 contrôles annoncés par le runner, aucun échec.
- Ubuntu sous WSL : 129 suites, 2 479 contrôles, aucun échec.
- Memory Hub : 127 tests supplémentaires réussis sur chaque plateforme.
- Même wheel installé isolément sur Windows et Linux : 794 fichiers contrôlés,
CLI et dépendances vérifiées, trois scénarios MCP stdio réussis par plateforme. - Cache : concurrence entre processus et threads, effacement et expiration testés.
Commit publié : d5139c08090d49d4eefa9eb43f303704275af710. Les 1 258 fichiers de l'inventaire validé correspondent
exactement aux blobs de ce commit. La branche main a reçu entre-temps des
modifications de documentation et de CI ; la release conserve la source testée.
Les pièces jointes comprennent le wheel Python, les sources, le catalogue Codex,
les empreintes SHA-256 et les preuves. Installer le wheel dans un environnement
virtuel dédié ; ses dépendances Python sont déclarées dans le paquet. Le plugin
Codex nécessite séparément botte-mcp sur le PATH.
Les manifestes de construction joints décrivent l'état local avant publication ;
PUBLICATION.json relie cette construction au commit et au tag publiés.
Le ZIP des sources exclut l'historique Git et l'ancien artefact sous home/.
Les tests du benchmark qui exigent Git ont été exécutés avec un index temporaire
au commit de base et ses objets en lecture seule, sans retrait de tests.
Périmètre : Linux sous WSL ; Pi/ARM et hôte Codex natif non requalifiés ici.
Le contrôle des liens Markdown a une couverture partielle documentée. Cette
validation ne démontre pas de gain facturé ni de qualité sur un pilote humain
indépendant et n'active aucune politique apprise. Aucune installation active
n'a été remplacée.
2.0.0rc2 — External Pilot
Botte Secrète 2.0.0rc2 — External Pilot
Pre-release for bounded external pilot testing. Latest stable release: 1.9.0.
Changes from RC1
- Failed test subprocesses override passing-looking summaries (#126).
- Exact cache identity preserves whitespace and rejects ambiguous legacy keys (#127).
- Portable sandbox test fixtures and Windows/Ubuntu wheel validation.
Factory Assurance and its pilot limits are inherited from the frozen RC1. This
release does not merge the outstanding development stack into main.
Install
python -m pip install "git+https://github.com/zedarvates/botte-secrete.git@69c71b83c75ed8bde31ffd2927dcd5a8cdbadd88"Package version: 2.0.0rc2.
Source commit: 69c71b83c75ed8bde31ffd2927dcd5a8cdbadd88.
Source tree: b45914380878f1c37905847b6ecace4c11b9d885.
Validation
Release Candidate run
passed on Windows and Ubuntu with Python 3.12:
- Full suite: 971 passed, 0 failed on each OS, across 66 suites.
- Built wheel installation, metadata and import-location checks outside source imports.
- CLI, MCP stdio, dependency consistency, 35 Factory Assurance checks and six exact-cache regressions.
- README commands: 11 passed, 0 failed, 11 skipped on each OS.
- Documentation: 65 Markdown files, 0 broken local links on each OS.
Synthetic PR commit 52b2247110ca9d12451d7728e547ccdde3a2f61d has the same
tree as the durable source commit. Both wheel hashes were verified against the
downloaded CI artifacts. The attached JSON records their identities.
The local Windows/Python 3.14 full run separately reported 974 passed, 0 failed;
that environment-dependent total is not substituted for the CI counts above.
Limits and reproduction
Start in observe, consultative or shadow mode. Independent human pilot review
and a genuinely private/external holdout remain unproven. These checks do not
certify general production readiness. RC1's public slugify pilot has not been
rerun for RC2 and remains historical RC1 evidence.
Reproduction instructions and candidate notes
include the build, installed smoke and regression commands.
Candidate PR #128.
2.0.0rc1 — External Pilot
Botte Secrète 2.0.0rc1 — External Pilot
Pre-release for bounded external pilot testing. Latest stable release: 1.9.0.
What is included
- Deterministic Factory Assurance with mission boundaries and evidence gates.
- Builder/independent-judge separation and named agent governance.
- Positive and negative/mutation controls, public-safe holdout attestations,
and conservative consequence handoff without automatic merge or deployment. - External pilot protocol and reproducible source identity.
Install the validated candidate
python -m pip install "git+https://github.com/zedarvates/botte-secrete.git@d98505fd6586ff3d0d32a2cb4ebd5f0408d4142a"Package version: 2.0.0rc1.
Source commit: d98505fd6586ff3d0d32a2cb4ebd5f0408d4142a.
Git tree: 9edca7e514e1f12c13302b630545973167e55a3d.
Validation
Release Candidate run #5
passed on the same Git tree as the published source commit:
- Build and installation; installed metadata
2.0.0rc1; assurance import. - Full regression: 960 passed, 0 failed across 64 suites.
- README commands: 11 passed, 0 failed, 11 skipped.
- Documentation: 64 Markdown files, 0 broken local links.
- Explicit Factory Assurance core, integration, and named-role tests.
GitHub tested synthetic commit 9386f2f37486d1e299cb58382961a1918707cb51;
its tree matches the durable source commit above.
Pilot limits
Start in observe, consultative, or shadow mode. This is not production
certification or authorization for unbounded autonomy. The StoryCore dark lap
still lacks a proven genuinely private/external holdout. No authenticated
holdout signer infrastructure or external pilot feedback is claimed.
The candidate comes from the frozen release branch in PR #120.
Publishing this pre-release does not merge the Factory Assurance development
branch into main. The README badge in the frozen source identifies the stable
line; package metadata identifies this RC.
Read the full candidate notes
and external pilot protocol.
Reproducible validation reports
Versioned installation and public-pilot reports include pinned reproduction scripts and explicit evidence limits.
GitHub Windows and Ubuntu reproduction passed: installation, CLI/MCP, 35 Factory Assurance checks, and the public python-slugify pilot (82 baseline tests passing, all three controlled mutations detected). This does not establish independent human review or a private holdout.
Rerun via Actions: RC1 report reproduction > Run workflow, or the documented local command.
1.9.0
What's Changed
- feat(llm_mcp): lazy tool loading — the biggest amont cut by @zedarvates in #35
- Feat/lazy tools by @zedarvates in #36
- feat: package for external users — pip install + botte/botte-mcp cons… by @zedarvates in #37
- fix: checkup CLI hangs indefinitely on repos with build caches / huge assets by @zedarvates in #38
- feat: add safe loop optimizer and release candidate by @zedarvates in #39
- ci: install declared dependencies before test suites by @zedarvates in #40
- ci: install declared dependencies before test suites by @zedarvates in #41
- feat: add deterministic loop policy features by @zedarvates in #42
Full Changelog: v1.7.0...1.9.0
v1.7.0
What's Changed
- feat(nn_audit): audit micro-NNs — grounded in real data or synthetic rule-copies? by @zedarvates in #31
- feat(botte_nn): distill error_classifier from real labelled errors by @zedarvates in #30
- fix(security_scanner): correct fail_on filter + Python 3.12+ AST compat by @zedarvates in #29
- feat(nn_audit): wired vs orphan — make the synthetic-NN verdict actionable by @zedarvates in #32
- feat(nn): remove 2 orphan synthetic micro-NNs + wire nn_audit into /checkup by @zedarvates in #33
- feat(context_profiler): measure the always-on prefix vs small local windows by @zedarvates in #34
Full Changelog: v1.5.0...v1.7.0
v1.5.0 — Belt over the harness
🧦 Botte Secrète v1.5.0 — Belt over the harness
Tiny local neural nets + an anti-hallucination harness now sit above the LLM stack: trivial routing/triage is decided by trained micro-models in ~0 ms, local models are constrained to structured output and verified before they're trusted, and the belt sharpens itself from real outcomes. The security track also closes. 0 cloud tokens throughout.
🧠 The belt (micro-NN)
- Micro-NN belt (
botte_nn+auto_router/nn_belt) — trained nets (binary_router, effort/token/priority/error classifiers, anomaly_detector; Rust core + numpy fallback) decide local-vs-cloud routing/triage at ~0 ms, 0 tokens; abstain below a confidence floor. - Self-improving loop —
active_learningretrains from logged outcomes (warm-start + honest eval);auto_routerimplicit feedback gives the loop ground truth.
🛡️ The harness (anti-hallucination)
local_harness— a 5-layerHarnessSpecexecutor + deterministic verifier, structured-output constraints on local models, and a with/without-harness hallucination bench. Escalate, don't hallucinate.
🔍 Security track (RepoAudit/DeepAudit-inspired)
- CWE knowledge base (
cwe_kb) — local RAG that enriches/de-noises findings with name/description/mitigation (cwe_explain+ wired intosecurity_scan). - Malicious-pattern scan in CI —
security_scanner(regex + AST: obfuscation/exfiltration/remote-exec) wired into/checkup+ the PR comment, high-signal subset only.
🧩 Also
pipeline_dsl(custom local-first YAML workflows),audit_dag(one canonical audit DAG), trajectory learning for solver optimization.
528 tests · 0 failed · stdlib/numpy · local-first · MIT
v1.4.0 — Security + determinism
🧦 Botte Secrète v1.4.0 — Security + determinism
A neuro-symbolic security analyzer wired into CI, scoped docs management for multi-component projects, and the OR-Tools-inspired determinism program — exact computation replacing LLM reasoning for structured decisions. 0 cloud tokens throughout.
🛡️ Security
- Taint / data-flow analyzer (neuro-symbolic, RepoAudit/DeepAudit-inspired): traces attacker-controlled sources → dangerous sinks, CWE-tagged (78/89/94/502/918); symbolic pass 0 tokens, optional local-model
--judge.security_scanMCP tool. - Security in
/checkup(and CI): every PR gets a 🛡️ security verdict via the existing checkup workflow — no extra wiring.
📚 Docs management
- docs_steward — scoped documentation map for monorepos (server/client/tools): a bounded coder loads only its scope + linked globals, not every component's docs. Plus a docs lifecycle (prune finished tasks, archive reports).
docs_map/docs_lifecycleMCP tools.
🧮 Determinism program (OR-Tools-inspired)
Every structured decision once delegable to an LLM is now exact, repeatable, 0-token computation:
- context_budget — optimal skills/docs to load under a token budget (exact 0/1 knapsack). ~85% less always-on context per task on this repo.
- nlp_deterministic — classify/extract without an LLM (rules + gazetteers + local embedding).
- solvers — assignment (LPT), bin-packing (FFD), precedence scheduling (DAG + parallel waves).
- Deterministic LLM tests — the flaky live-model test branches now run against a stubbed local client, so the suite is green regardless of which model is loaded.
380 tests · 0 failed · stdlib-first · local-first · MIT
v1.3.0 — Autonomy iteration
🧦 Botte Secrète v1.3.0 — Autonomy iteration
The conductor now runs safe plans, ingest gains real local semantic recall, and every PR gets a local-first checkup. 0 cloud tokens throughout.
Added
- Conductor executor — the conductor can now run a plan, not just produce it. Read-only analysis steps run unattended; mutating/cloud steps are confirm-gated; steps with an unfilled
<placeholder>are skipped. New--execute/--confirm/--dry-runCLI flags and anexecute_planMCP tool. - Real local embeddings for ingest —
ingest/searchauto-resolve a local/v1/embeddingsendpoint from the backend registry (real semantic vectors), falling back to the deterministic hash vector when none is available. Newresolve_embed(),--embed-url/--embed-modelflags, overrides on theingest_sourceMCP tool; result reportsembed: endpoint|hash. - GitHub Action — checkup on PR — the
🧦 Botte Checkup (PR)workflow runs the canonical/checkupon every pull request and posts (or updates in place) a single verdict comment viagh. Newcheckup --pr-commentflag + pureformat_pr_comment(). Reusable by any project that deployed botte-secrète.
Removed
- 4 genuinely-unused symbols flagged by our own dead-code analyzer (
check_tool,CodeFingerprint.is_same,SkillInfo.token_cost/qualified_name).
277 tests · 27 MCP tools · stdlib-first · local-first · MIT