@mousedev/harness-core

The Mouse completion loop and everything it needs, with zero runtime dependencies. After each model turn the loop probes the workspace deterministically (did anything change, do the repository's checks pass, was a test or workflow file deleted), then either stops with evidence or sends a continue prompt to the same session with the failing output and a request for a structured MOUSE_AUDIT self-audit. It is engine-agnostic: an Engine is anything that can take a prompt and report steps and text, and a Workspace is a directory plus a way to run a shell command in it.

This package is what a host embeds. The CLI in @mousedev/harness wires it to a local checkout and to opencode run through @mousedev/harness-opencode; a hosted product wires the same functions to its own sandbox. Nothing here touches the network, and the only filesystem writes are the trace file under ~/.mouse/runs/ and the temp marker the CLI creates. Prompt bytes in prompt.ts and loop.ts are cache-sensitive and pinned by a golden test; see AGENTS.md before editing them.

Main API

import {
  localWorkspace, loadPolicy, headSha, makeProbe, runCompletionLoop,
  openTrace, traceFile, parseAuditBlock, buildContinuePrompt, buildAgentPrompt,
  DEFAULT_POLICY, parsePolicy, permissionRulesForTool,
  detectEcosystem, checksFromEcosystem, classifyInferenceFailure,
} from "@mousedev/harness-core";

| Module | Exports | |---|---| | loop.ts | runCompletionLoop(input), buildContinuePrompt(kind, ctx), parseAuditBlock(text), MIN_ROUND_MS, types CompletionProbe, CompletionBudget, CompletionOutcome, ContinueKind, CompletionRound, CompletionResult | | engine.ts | Engine, PromptResult, PromptOptions, TokenTotals | | workspace.ts, exec.ts | Workspace, localWorkspace({ root }), shellPath, LOCAL_EXEC_TIMEOUT_CODE | | probe.ts | makeProbe({ workspace, runStartSha, marker, signal, checks?, checkTimeoutMs? }) | | policy.ts | DEFAULT_POLICY, parsePolicy, loadPolicy(ws), parseChecks, parsePermissions, permissionRulesForTool | | detect.ts | detectEcosystem(ws), checksFromEcosystem, checksFromPackageJson, pytestCommand, detectEcosystemCommand | | git.ts | headSha, fingerprint, deletedVerificationFiles, VERIFICATION_PATH | | mea.ts | initTaskState, applyAudit, auditFromSandboxProbe, types TaskState, AuditReport | | prompt.ts | buildAgentPrompt(mode, { profile }), AGENT_PROMPT_MAX_CHARS, types Mode, Profile | | trace.ts, paths.ts | openTrace(file), parseTrace(text), TRACE_VERSION, mouseHome, runsDir, traceFile, workspaceSlug | | events.ts | LoopEvent | | version.ts | MOUSE_VERSION | | failures.ts | classifyInferenceFailure(message, raw?) |

A complete example that drives the loop from your own code is examples/sdk-run. The loop's behaviour, step by step, is in docs/loop.md; the policy file in docs/config.md.

Requires Node 22 or newer. MIT.