Releases

Cross-platform Claude Code smart session manager

Find duplicate audio files using acoustic fingerprinting (chromaprint/AcoustID)

Web-search + URL-fetch tools for harness-rs agents: DuckDuckGo → Bing fallback, HTML → text, no API keys required.

Allowlisted read-only shell tool for harness-rs. Per-program safe-args predicate.

CortexDB-backed Memory for harness-rs: implements the harness-core Memory trait over CortexDB's MCP server (memory_save / memory_search), giving any harness agent semantic recall + a shared knowledge brain. Optional.

MCP client for harness-rs: expose a remote MCP server's tools as harness tools

Deterministic + agent hybrid state-machine executor for harness-rs. Stripe-Minions-style with branch-on-failure retries.

Permission modes + per-tool allow/deny rules for harness-rs. Layers over the existing Hook system (PreToolUse gating) — gives ai-ledger-style multi-user backends a first-class way to map user tier → tool access without writing a hook per tool.

MCP (Model Context Protocol) stdio server exposing the harness-rs tool registry to Claude Code, Cursor, and Codex.

Experience memory for harness-rs: record each run as an episode (situation → tools used → outcome) and semantically recall similar past experiences to guide the next run. Backend-agnostic over the Memory trait. Optional.

Single-machine async Run orchestrator for harness-rs: a concurrent DAG of sub-agent Jobs with dependency gating, per-job retry/backoff/dead-letter, dynamic replanning, a run-level token budget, and crash-resumable Run/Job state. Optional.

Loop engineering for harness-rs: maturity levels (L1/L2/L3), human gates, token budgets, and schedulable production-loop patterns (daily triage, PR babysitter, CI sweeper, …) composed over the scheduler, sandbox, sub-agent, and memory primitives. Optional.

Optional standalone scheduler for the harness-rs agent framework. Runs declarative TOML-configured jobs (daily / weekly / interval / cron) that invoke any shell command — typically a harness-rs-built agent binary.

Sandbox tiers for harness-rs — git Worktree, Docker Container, and a trait for deployment-owned isolation backends.

ReAct agent loop, subagent isolation, and session record/replay (JSONL) for the harness-rs framework.

LLM-facing skill-authoring tool for harness-rs agents: skill_manage (create/patch/edit/delete SKILL.md). Used by the learning loop.

Jail-safe filesystem tools (read_file, write_file, edit_file, list_dir, grep, glob) for harness-rs. Symlink-escape protection.

LLM model adapters for harness-rs — OpenAiCompat (DeepSeek/Groq/Together/Ollama), AnthropicNative (Messages API), MockModel.

HookBus for harness-rs. Dispatches 27 lifecycle events with Allow/Deny/Inject/Mutate semantics. Optional OpenTelemetry feature.

Five-stage progressive context compactor for harness-rs — DefaultCompactor (structural) and ModelBackedCompactor (semantic).

Integration with bevy_quinnet for bevy_replicon

Default World runtime (Tokio process runner, system clock, in-memory KV) and Context helpers for harness-rs.

agentskills.io-compliant SKILL.md loader, validator, linter, exporter, and in-process registry for harness-rs.

Rust agent framework based on harness-engineering principles (Lopopolo, Böckeler). Facade re-exporting the public surface of the harness-rs-* crates.

Procedural macros for harness-rs: #[skill], #[tool], #[guide], #[sensor], #[hook]. agentskills.io-compliant registration via inventory.

Core traits and types for the harness-rs agent framework: Model, Tool, Guide, Sensor, Hook, Compactor, Skill, plus Context, World, Block, Event, FixPatch and 27 lifecycle events.

Async Rust client for OpenAI-compatible LLM provider APIs

Protobuf UDP runtime and time-series state kernel for aerospace simulation

Graph neural network primitives: GCN and hyperbolic GCN on Burn tensors

semantic versioning data query tool