# The work ahead.

> Where looprig is going: what is in progress, what is designed, and what must come together next. Statuses reflect actual module state.

- Human: https://looprig.com/roadmap
- Machine index: https://looprig.com/llms.txt

doc lr-rm-001 / roadmap

This page tracks the rig's forward work, grouped by how real it is today. Statuses come from the module specs and design docs, not from wishes. No dates: items ship when they meet the spec.

## in progress

plate r1 / in progress implementation underway

- 01 token usage and context occupancy usage reported on every model call, measured against the context window, feeding durable compaction
- 02 hustle mechanism harness-owned auxiliary model calls; compaction summarization first, classifiers next
- 03 rig lifecycle and workspace snapshots immutable rig composition, per-loop restore, content-addressed workspace snapshots
- 04 carbon modularization the Carbon reference coding product slims to loop definitions and rig assembly; reusable machinery moves into modules
- 05 tui package boundaries unified rail, runtime context, loop footer

## designed, not yet built

plate r2 / designed, not yet built specs approved

- 01 web client one Go binary: BFF plus embedded SPA plus a framework-neutral TypeScript SDK over the serve /v1 contract
- 02 acp bridge bidirectional Agent Client Protocol: expose looprig agents, drive foreign agents, one foreignloop module with claude, codex, and acp drivers
- 03 sandbox v1 OS-enforced command confinement: Seatbelt on macOS; namespaces, Landlock, seccomp, nftables, cgroups on Linux
- 04 structured output schema-bound loop and delegate results with strict validation

## on the horizon

plate r3 / on the horizon design first, then build

- 01 mcp with durable events authentication boundaries, permissions, and elicitation on the same session model
- 02 agent tasks in flow agents as first-class task kinds inside durable workflows
- 03 complete opentelemetry tracing across models, loops, tools, gates, delegates, workspaces, and workflows
- 04 guardrails and evals prompt-injection classifiers, provenance-aware policy, evaluation harnesses
- 05 secret references tools resolve secrets without values ever entering model context
- 06 document parsing multimodal document parsing through a dedicated parser service
- 07 clients everywhere web, desktop, and mobile clients that work with looprig and acp agents out of the box
- 08 memory and instructions persistent memory and scoped instruction files for long-lived agents

The architecture is credible. The remaining work is to make it visible, coherent, and immediately useful.

[read the mission](/mission) [read the docs](/docs)