brainwires-autonomy

Autonomous agent operations for the Brainwires Framework — self-improvement, Git workflows, environment interaction, and human-out-of-loop execution.

Features

Agent Operations — attention management, health monitoring, hibernation, parallel execution, supervision
Self-Improvement — feedback-driven strategy selection, code quality scanning, test coverage analysis, crash recovery with AI-powered diagnostics
Git Workflow Automation — branch lifecycle, PR management, merge policies, webhook handling
CI/CD Orchestrator — community-driven automation: GitHub Issues → investigate → fix → PR → merge
Cron Scheduler — recurring autonomous tasks with failure policies and rate limiting
File System Reactor — watch directories for changes, debounce events, trigger autonomous actions
System Service Management — controlled systemd, Docker, and process management with safety guardrails
GPIO Hardware Access — re-exported from brainwires-hardware — safe GPIO pin management with allow-lists and auto-release for embedded/IoT

Feature Flags

Feature	Description
`self-improve`	Self-improvement controller, strategies, and crash recovery
`eval-driven`	Eval-driven feedback loop + empirical scoring eval cases for entity importance and tiered memory
`supervisor`	Agent supervisor with health monitoring and restart
`attention`	Attention mechanism with RAG integration
`parallel`	Parallel coordinator with optional MDAP
`training`	Autonomous training loop
`git-workflow`	Automated Git workflow pipeline (issue → PR → merge)
`webhook`	Webhook server + CI/CD orchestrator for Git forge events
`scheduler`	Cron-based scheduled autonomous tasks
`reactor`	File system event reactor with debouncing
`services`	System service management (systemd, Docker, processes)
`gpio`	GPIO hardware access via `brainwires-hardware` (Linux)
`full`	All features enabled

Empirical Eval Cases (`eval-driven`)

The eval-driven feature includes eval cases that validate scoring heuristics produce correct relative orderings, measured via NDCG. These plug directly into AutonomousFeedbackLoop:

use brainwires_autonomy::eval::{entity_importance_suite, multi_factor_suite};

let cases = [entity_importance_suite(), multi_factor_suite()].concat();
let loop_ = AutonomousFeedbackLoop::new(config, cases, provider);

Case	Category	What it validates
`EntityImportanceRankingCase`	`entity_resolution`	Hub entities rank above peripheral (NDCG ≥ 0.8)
`EntitySingleMentionCase`	`entity_resolution`	Single-mention entities have non-zero importance despite ln(1)=0
`EntityTypeBonusCase`	`entity_resolution`	File > Type > Function > Error > Concept > Command > Variable (NDCG ≥ 0.95)
`MultiFactorRankingCase`	`memory`	4 scenarios: similarity, recency, fast-decay, and importance ordering (NDCG ≥ 0.99 each)
`TierDemotionCase`	`memory`	Lowest-retention entries are ranked first for demotion (NDCG ≥ 0.99)

All cases are deterministic (no LLM calls) and complete in under 1 ms.

Examples

# Core (no feature flags required)
cargo run -p brainwires-autonomy --example safety_guard
cargo run -p brainwires-autonomy --example health_monitor
cargo run -p brainwires-autonomy --example session_metrics

# Self-improvement
cargo run -p brainwires-autonomy --example self_improve_strategies --features self-improve
cargo run -p brainwires-autonomy --example crash_recovery --features self-improve

# Git workflow & CI/CD
cargo run -p brainwires-autonomy --example git_workflow_pipeline --features git-workflow
cargo run -p brainwires-autonomy --example cicd_orchestrator --features webhook

# Environment interaction
cargo run -p brainwires-autonomy --example cron_scheduler --features scheduler
cargo run -p brainwires-autonomy --example fs_reactor --features reactor
cargo run -p brainwires-autonomy --example service_manager --features services

Safety

All environment-interaction features are designed with strict safety defaults:

Services: read-only by default, hardcoded deny-list for critical system services (sshd, dbus, systemd-*, etc.)
GPIO: empty allow-list by default (no pins accessible), auto-release on agent timeout — see brainwires-hardware for GPIO examples
Scheduler: budget tracking, circuit breakers, per-task failure policies
Reactor: rate limiting, debouncing, path allow/deny lists
Crash Recovery: meta-crash detection (aborts if the crash handler itself keeps crashing), max fix attempts

License

Licensed under either of Apache License, Version 2.0 or MIT License at your option.

brainwires-autonomy 0.8.0