The Lagra workbench.

CLI harness · lagra-harness

A standalone agentic harness for Lagra scene work. Designed to run from a shell, the IDE, or behind an MCP / WebSocket gateway. Model-agnostic by design — the Lagra tool contract, session ledger, verifier manifest, and publish-link evidence all exist before any LLM policy is trusted.

The harness treats correctness as an artifact, not a chat claim. A scene is not done until a verifier manifest exists and the checks pass against the actual Lagra binary. Publishability is a separate claim: the scene must exist, carry wasm true, the landing repo must expose the /loader/<scene>.lagra entrypoint, the wasm bundle must exist and not be older than the source, and the verifier result must pass.

defaults · engine: ~/dev/qdesics/lagra-core · landing: ~/dev/lagra-landing · binary: $LAGRA_BIN → target/release/lagra

Commands

# run + verify an existing scene
lagra-harness run --scene NAME [--project PROJ] --task "…"

# construct a NEW scene from a base, with a policy file or command
lagra-harness run --mode construct \
  --scene BASE [--target-project PROJ] \
  [--policy-file policy.json | --policy-command CMD] \
  [--publish-build] --task "…"

# verify only (skip the construct/repair loop)
lagra-harness verify --scene NAME [--project PROJ] [--steps N]

# publish to the landing /loader/<name>.lagra entrypoint
lagra-harness publish --scene NAME [--build] [--handoff]

# plan a task without executing
lagra-harness plan  --task "…" [--scene NAME]

# print the full schema (for agent grounding)
lagra-harness schema --json

What each run writes

# .lagra-harness/sessions/<id>/
events.jsonl                # full event ledger
inspection.json             # scene shape + node counts
verifier.json               # Noether-derived checks + manifest
verifier-result.json        # PASS / FAIL per gate

# construct runs add:
base-inspection.json
policy-decision.json
working-document.json
repair-attempts.json        # if a repair loop ran

# publish runs add:
publish-build.json
publish-manifest.json
publish-handoff.{json,md}

Desktop IDE · crates/lagra-ide

Live scene work in an Electron + React shell that drives the same engine as the wasm demos. Validator overlays, per-particle drill-down, scene-source editing, and an embedded MCP agent terminal — all backed by the native Lagra runtime via IPC.

Scenes spawn sub-agents that hold a typed handle to the running World and register actions as typed InterventionKind events. The LagrangianGraph inspector renders every .lagra term as a Chrome-DevTools-style typed object tree; the Tree / node-graph view renders the same scene as Blueprint-style topology (preset → entities → forces → solver) wired with bezier connections.

MCP gateway

Claude Code, Codex, Albert — every agent speaks Lagra through the same MCP server the harness exposes. Tools are typed; arguments are validated against the schema; every call writes to the same session ledger as a CLI run.

Agents can run, verify, construct, and publish scenes; inspect a running World; register typed InterventionKind events; or drop directly into a scene's loader URL. Same contract, regardless of whether the caller is a human, a shell script, or an LLM.

Wire it up

# ~/.claude/config.json (Claude Code)
{
  "mcpServers": {
    "lagra": {
      "command": "lagra-harness",
      "args": ["mcp"],
      "env": {
        "LAGRA_ENGINE_ROOT": "~/dev/qdesics/lagra-core"
      }
    }
  }
}

Once connected, agents call

lagra.run        # step + verify a scene
lagra.verify     # run conservation / contract gates
lagra.construct  # build a new scene from a base + policy
lagra.publish    # push to /loader/<name>.lagra
lagra.inspect    # return scene shape + node counts
lagra.schema     # return the full tool schema