Roadmap

Larb ships in vertical slices. Phases 0–7 are built and tested (the engine, the security model, the ecosystem, and the eval tooling); the path ahead turns the alpha into a tool developers install and trust.

Status legend: ✅ shipped · 🟡 partial / unverified live · 🔜 planned

Timeline

What's shipped (phases 0–7)

Area	Status	Notes
Trust-before-anything boot	✅	No config-as-code, no network before consent
Capability tools + permission engine	✅	Layered allow/deny, project policy, every grant logged
Hard spend governor	✅	Halts the agent before overspend (run / session / day)
Model-agnostic providers (11)	✅	One config line; conformance suite
Orchestration + verification loop	✅	Mandatory lint/build/test before "done"
Multi-agent delegation	✅	Strong orchestrator → cheap worker
Real container sandbox	🟡	Built + unit-tested; needs live verification on a host with docker/podman
Governed network egress	✅	http_fetch + container egress allow-list proxy
Durable run state	✅	larb runs / larb resume
True streaming (SSE / NDJSON)	✅	OpenAI + Ollama incremental; Anthropic caching
Signed, manifested skills	✅	Install from dir / tarball / git; install ≠ trust
MCP (Model Context Protocol)	✅	stdio servers as permission-gated tools; larb mcp / larb mcp probe
Project instructions (AGENTS.md)	✅	AGENTS.md / .larb/AGENTS.md loaded as advisory context
Secret broker	✅	Single redacting env boundary
Benchmark harness	✅	Resolution rate + cost/task, worktree isolation
SWE-bench harness	🟡	Loader + grading primitives; full graded runs need dataset repos
Live model runs (local + cloud)	✅	Verified end-to-end against Ollama (local) and an OpenAI-compatible provider
Surgical edits + ranged reads	✅	edit_file (exact, occurrence-checked) — cost scales with the change, not the file
Project memory	✅	remember tool; inspectable markdown, loaded into future sessions
Post-edit diagnostics	✅	check commands run after every edit; failures feed straight back
Headless mode	✅	larb run --yes [--json] for CI/scripts (requires prior full trust)
Interactive session	✅	Bare larb: conversation across tasks, Esc-interrupt + steer, /plan
Windows support	✅	Required CI job (ubuntu·macos·windows); child-process env baselines
CLI · TUI · editor bridge	✅	Streaming, diff review, approvals, live cost meter
VS Code extension	🟡	MVP in-repo over the bridge protocol; marketplace listing pending

What's next (toward 1.0)

Release readiness

• ✅ Confirmed end-to-end live runs — validated against Ollama (local) and an OpenAI-compatible cloud provider.
• ✅ Published to npm — npm i -g @larb/cli@alpha (currently 0.1.0-alpha.4); promote to latest after alpha feedback.
• 🟡 Live container-sandbox verification on a host with docker/podman.

Trust & scale 🔜

• MicroVM sandbox backend behind the existing seam — airtight raw-socket egress blocking, not just proxy-respecting clients.
• Community skill registry with provenance and a public manifest schema.
• Full SWE-bench grading wired to the dataset + per-repo test commands, and a published resolution-rate + cost/task number.
• Parallel multi-agent using isolated git worktrees, with deliberate merge.

Reach

• 🟡 VS Code extension — MVP ships in-repo over the larb bridge protocol; marketplace listing (and JetBrains) planned.
• 🔜 Single-binary distribution (Bun / pkg) for zero-Node installs.
• 🔜 Deno runtime spike — its --allow-* model maps onto our capability sandbox.
• 🔜 Editor-native plan view and richer diff review.

Quality

Success is measured, not asserted:

• Resolution rate — SWE-bench Verified, via the larb bench harness.
• Safety — zero "config-triggered execution" or "secret-before-consent" findings in red-team review; % of skills with manifest enforcement.
• Cost — dollars per resolved task vs. a Claude-Code / Codex baseline.
• Portability — number of providers passing the conformance suite.
• Ecosystem — community skills published; % signed / verified.

Non-goals (for now): a hosted SaaS backend, training our own foundation model, a full IDE, and multi-channel chat gateways.

See the architecture and security model.