Packages

Harness wrapping local AI coding CLIs (Claude Code, Codex, Antigravity) as atomr-agents Callables. Headless mode parses structured streams; interactive mode bridges a tmux session to clients.

Axum + embedded SPA companion for the coding-cli harness. Start runs, browse results, follow SSE events, and attach an xterm.js terminal to tmux-wrapped interactive sessions over WebSocket.

Isolator trait + Local and Docker implementations for the coding-cli harness. Spawns CLI processes on the host or in a container with optional PTY.

Google Antigravity CLI (agy) adapter for the coding-cli harness: stream-json parser + system-instruction projection. Successor to the Gemini CLI adapter.

Claude Code adapter for the coding-cli harness: stream-json parser + concept projection (CLAUDE.md, skills, MCP, settings).

OpenAI Codex CLI adapter for the coding-cli harness: `codex exec` headless invocation + AGENTS.md projection.

Google Gemini CLI adapter for the coding-cli harness: stream-json parser + system-instruction projection.

AgentContext + ContextAssembler for atomr-agents.

Core types for the atomr-agents framework: ids, budgets, context, events, errors.

Uniform input/output contract for deep-research harnesses: request, result, citation, plan, telemetry types.

Pluggable deep-research harness: clarify→plan→search→write→verify, multi-agent parallel, and iterative-deepening strategies behind a uniform contract.

Axum + embedded SPA companion for the deep-research harness. Kick off runs, browse results, follow live SSE event streams.

Two-tier outer shell for atomr-agents deep research: intent classifier routes between a fast shallow path and the full deep harness.

EmbeddingActor and pluggable AnnIndex (CPU default via atomr-accel-agents).

Eval suites + regression detection + deterministic event-stream replay.

Harness: tested, packaged, persistent execution loop.

atomr-agents host runtime — long-lived agent identity, memory, skills, hooks, schedules, channels, atop atomr actors.

Axum + embedded React SPA companion for the atomr-agents host runtime. Manage agents, SOUL/MEMORY/RULES/USER docs, skills, curator proposals, crons, hooks, channels/routing, branches, registry, evals, MCP servers, config, and a live event stream.

Document loaders, text splitters, and CachedEmbedder for atomr-agents.

InstructionStrategy + ComposedInstructionStrategy<P, T, B> assembler.

Meetings harness: takes a diarized STT conversation and produces structured attendees, notes, and actions with owners. Supports batch and live-streaming extraction with a dynamic, tiered summarization ledger.

Optional axum backend + embedded React SPA for reviewing and editing meeting analyses (attendees, notes, actions, summaries).

Memory stores and strategies (recency, semantic, summarizing, chained).

Structured event taxonomy + tracing/metric exporters for atomr-agents.

Organizational hierarchy (Org/Department/Team) over cluster sharding.

Output parsers + auto-repair + streaming partial-JSON for atomr-agents.

Persona strategies: structural (Static, Big Five, MBTI, Jungian, Composite) + emphasis.

PyO3 bindings: host (run harnesses) and guest (Python-defined tools/strategies).

Versioned artifact registry: tool sets, skills, personas, agents, workflows, harnesses.

Retriever zoo: Vector, BM25, MultiQuery, ContextualCompression, ParentDocument, Ensemble-RRF, SelfQuery, EmbeddingsFilter, TimeWeighted.

Docker (bollard) backend for the atomr-agents microVM sandbox — the cross-platform "Insecure Dev Mode" fallback. One container per sandbox; exec, file I/O, commit-based snapshot/fork.

Backend-agnostic microVM sandbox traits, profile/budget/request types, lifecycle events, and a deterministic in-memory mock backend for the atomr-agents sandbox subsystem.

In-VM guest agent for the atomr-agents microVM sandbox: a static Rust daemon that runs as PID 1 inside a Firecracker guest, serving the sandbox-proto command set over AF_VSOCK.

Orchestrator for the atomr-agents microVM sandbox: owns a pluggable SandboxBackend, a sandbox registry, a bin-packing scheduler and snapshot pool, broadcasts SandboxEvents, and is itself a Callable.

Axum REST + SSE companion for the atomr-agents sandbox harness: create/exec/fork/snapshot/destroy endpoints and a SandboxEvent /events stream.

Wire protocol shared between the sandbox host and the in-VM guest agent: length-prefixed postcard frames over AF_VSOCK, with the GuestRequest/GuestResponse command set.

The `execute_in_sandbox` Tool: exposes the atomr-agents microVM sandbox harness to agents, with profile resolution, Rust-floor budget scaling, and EventBus telemetry.

Typed security context: clearance/compartment walls, mandate checks, a capability/secrets broker, and reusable WalledTool middleware.

Skills: bundled instruction+tool+sub-agent capabilities.

Channelled state, reducers, and per-super-step checkpointing for atomr-agents.

Strategy traits and combinators for atomr-agents.

Audio decoding (symphonia) and microphone capture (cpal) for atomr-agents speech-to-text.

Core types and traits for atomr-agents speech-to-text: SpeechToText / StreamingSession traits, Capabilities, AudioInput, Transcript, MockSpeechToText.

Local speaker diarization via sherpa-onnx for atomr-agents speech-to-text. Plugs into any backend whose Capabilities::diarization is None.

STT harness: an agentic streaming pipeline that diarizes, transcribes, and accumulates a conversation record aligned to agentic structures.

Optional axum backend + embedded React SPA for reviewing and editing diarized STT conversations.

Shared HTTP / WebSocket plumbing for atomr-agents speech-to-text remote backends: client builder, secret refs, retry/backoff, rate limits, error classification.

AssemblyAI STT backend (REST + Universal-Streaming WebSocket) for atomr-agents.

Deepgram STT backend (REST + WebSocket) for atomr-agents.

OpenAI Whisper / gpt-4o-transcribe REST backend for atomr-agents speech-to-text.