martin-loop 0.1.3 → 0.1.5

package/README.md

@@ -2,25 +2,27 @@
 
 <img src="./docs/assets/martinloop-logo.png" alt="MartinLoop" width="260">
 
-### A governed runtime for autonomous AI coding agents. ⭐⭐⭐
+### The cross agent governance layer for autonomous AI coding agents.⭐
 
 [![License: MIT](https://img.shields.io/badge/license-MIT-7c3aed?style=flat-square)](./LICENSE)
 [![TypeScript](https://img.shields.io/badge/TypeScript-strict-3178c6?style=flat-square&logo=typescript&logoColor=white)](./tsconfig.base.json)
 [![Node](https://img.shields.io/badge/node-%3E%3D20-3c873a?style=flat-square&logo=nodedotjs&logoColor=white)](#quick-start)
 [![npm](https://img.shields.io/badge/npm-martin--loop-cc3534?style=flat-square&logo=npm&logoColor=white)](https://www.npmjs.com/package/martin-loop)
 
+MartinLoop has been accepted into the NVIDIA Inception program.
+
 <br>
 
 **Your overnight AI pipeline estimated $2.40.**  
 **You woke up to a $65 bill.** 
  <br> 47 retries. No hard stop. No rollback. No audit trail. Nothing merged.  
  MartinLoop exists so that never happens again.✅ <br> <br>
- If you think autonomous AI coding agents need budgets, brakes, and receipts, ⭐ the repo so more builders can find it.
+ If you think autonomous AI coding agents need budgets, brakes, and receipts, Plase star ⭐ the repo so more builders can find it.
 <br>
 
 > AI coding agents are useful. Unbounded retry loops are not.
 >
-> MartinLoop wraps agent runs with budgets, policy checks, verifier gates, rollback evidence, and inspectable run records.
+> MartinLoop wraps agent runs with budgets, policy checks, verifier gates, rollback evidence, and inspectable run records. Built for Enterprise Coding Agents, Agentic Teams, and Autonomous Companies. 
 <br>
 <img src="./docs/assets/cli-animated.svg" alt="MartinLoop CLI — governed agent run" width="720">
 
@@ -58,6 +60,7 @@ It does not try to replace the agent pattern. It makes that pattern safe to run.
 | Verifier gate | A run only reaches `completed` when the adapter result and verifier state pass. Unsafe verifier commands are blocked before agent execution. |
 | Failure taxonomy | Classifies failures across 11 current classes, including hallucination, test regression, scope creep, repo grounding failure, environment mismatch, and budget pressure, that distinguishes real success from unsafe, invalid, or terminal behavior.|
 | Safety leash | Evaluates verifier commands, file scope, dependency or migration changes that require approval, and secret-like values in task text. **Policy-as-code**. |
+| Context integrity | Scans user prompts and tool output for injection patterns (authority inversion, instruction override, identity redefinition) before any attempt is admitted. Aborts with human escalation on detection. |
 | Rollback evidence | Captures rollback boundaries and restore outcomes for repo-backed attempts when a persistence store is configured. |
 | Context distillation | Carries a distilled summary of recent attempts and remaining constraints into subsequent attempts. |
 | Run records | The CLI appends JSONL loop records under `~/.martin/runs/<workspaceId>.jsonl`; lower-level stores can also persist contracts, ledgers, and attempt artifacts.
@@ -66,7 +69,7 @@ It does not try to replace the agent pattern. It makes that pattern safe to run.
 ⭐The result is a runtime that can complete good work, refuse unsafe work, stop uneconomical work, and leave evidence behind.✅
 ---
 
-## The Ralph Loop, explained
+## Ralph-Style Loops Need a Control Layer
 
 **"Everybody has gotten infatuated with what we call these Ralph Wiggum loops, just like send the thing off and it'll just go figure something out..A, It never figures anything out. And B, you just get this ginormous bill...**" - Chamath Palihapitiya, All-In Podcast #263, March 2026
 
@@ -82,7 +85,7 @@ The pattern is simple: attempt the task, run checks, retry on failure, repeat. T
 - it rolls back failed runs instead of leaving broken state behind
 - it reduces runaway token growth with context distillation
 
-If Ralph ever burned $165.70 on your dime, you're in the right place. Martin stopped him at $4.97 with a full audit trail. LFG! 🚀 Finally a Martin Prince leash for Ralph Wiggums! :)  
+If a Ralph-style loop has ever burned budget without producing a verified result, MartinLoop is designed to stop that failure mode before the next unsafe attempt runs. $165.70 on your dime, you're in the right place. Martin stopped him at $40.97 with a full audit trail.
 
 <div align="center">
   <img src="./docs/assets/martin-raplph.png.jpg" alt="Martin vs Ralph — governed vs ungoverned agent loop" width="240">
@@ -119,6 +122,8 @@ pnpm --filter @martin/benchmarks eval
 pnpm --filter @martin/benchmarks eval:phase12
 ```
 
+Challenge page: [Can your AI coding agent finish this task under $3?](./docs/distribution/UNDER-3-CHALLENGE.md)
+
 ---
 
 ## Quick Start
@@ -127,7 +132,9 @@ pnpm --filter @martin/benchmarks eval:phase12
 npm install -g martin-loop
 ```
 
-This installs both the `martin-loop` package and the `martin` command alias. The package is currently published on npm as version `0.1.2`.
+This installs both the `martin-loop` package and the `martin` command alias. The package is currently published on npm as version `0.1.4`.
+
+Want a safe sandbox first? Run `npx martin-loop demo` and MartinLoop will copy a disposable local workspace into `./martin-loop-demo`.
 
 ### Public Package Surface
 
@@ -136,8 +143,23 @@ The frozen public package surface for this release candidate is:
 - Install target: `npm install martin-loop`
 - CLI target: `npx martin-loop`
 - SDK target: `import { MartinLoop } from "martin-loop"`
+- MCP target (registry-ready package): `npx -y @martinloop/mcp`
 
 The `martin` command alias is installed for local operator convenience, but the public CLI surface is `npx martin-loop`.
+The standalone MCP server package is smoke-validated locally with `pnpm --filter @martinloop/mcp smoke:pack` and is ready for registry publication as a separate release step.
+
+### Claude Code MCP install
+
+Use the published MCP package directly:
+
+- macOS/Linux: `claude mcp add --scope user martin-loop -- npx -y @martinloop/mcp`
+- Windows PowerShell/cmd: `claude mcp add --scope user martin-loop -- cmd /c "npx -y @martinloop/mcp"`
+
+If you just want to launch the server manually, the one-line command is:
+
+```sh
+npx @martinloop/mcp
+```
 
 ### Run a governed task
 
@@ -196,7 +218,7 @@ martin run <objective> [options]
   --metadata <key=value>  Attach metadata to the run record; repeatable
 ```
 
-The public CLI also includes `inspect`, `resume`, and a `bench` redirect that points reviewers to the workspace benchmark harness.
+The public CLI also includes `demo`, `inspect`, `resume`, and a `bench` redirect that points reviewers to the workspace benchmark harness.
 
 <div align="center">
   <img src="./docs/assets/cli-static.svg" alt="MartinLoop CLI terminal output" width="720">
@@ -287,20 +309,22 @@ The lower-level `runMartin` function is also exported for callers that want to a
 | `@martin/core` | Runtime controller, policy engine, safety leash, grounding, persistence, and rollback logic. |
 | `@martin/adapters` | Claude CLI, Codex CLI, direct-provider, and stub adapter surfaces. |
 | `@martin/cli` | Local CLI implementation for `run`, `inspect`, `resume`, and the benchmark redirect. |
-| `@martin/mcp` | MCP server tools: `martin_run`, `martin_inspect`, and `martin_status`. |
+| `@martinloop/mcp` | MCP server tools: `martin_run`, `martin_inspect`, and `martin_status`. |
 | `benchmarks/` | Workspace-only deterministic benchmark and RC validation harness. |
 | `apps/control-plane/` | Hosted control-plane workstream, outside the initial npm package surface. |
 | `apps/local-dashboard/` | Local dashboard/read-model viewer, not currently packaged as public npm API. |
 
-The `@martin/core`, `@martin/adapters`, and `@martin/contracts` package manifests are still private workspace packages; the public install target is the root `martin-loop` facade.
+The `@martin/core`, `@martin/adapters`, and `@martin/contracts` package manifests are still private workspace packages. The public runtime install target is the root `martin-loop` facade, while `@martinloop/mcp` is packaged as a standalone MCP server with vendored internal runtime dependencies for registry publication.
 
 ---
-
 ## Development
 
-Requirements: Node 20+ and pnpm 10.x.
+Requirements:
 
-```sh
+- Node 20+
+- pnpm 10.x
+
+```bash
 git clone https://github.com/Keesan12/martin-loop.git
 cd martin-loop
 pnpm install
@@ -308,9 +332,9 @@ pnpm install
 pnpm test
 pnpm lint
 pnpm build
+
 ```
 
-```md
 Current RC gate commands:
 
 ```sh
@@ -320,9 +344,7 @@ pnpm repo:smoke
 pnpm rc:validate
 pnpm pilot:prep:validate
 pnpm release:matrix:local
-Caution: Registry Publication
-
-This package is published through the public martin-loop package surface. Treat registry publication as a guarded release step: verify the RC gate commands, confirm the version follows semantic versioning, and document breaking changes before publishing.
+```
 
 > **Caution:** This package is live on npm. Treat registry publication as a guarded release step — verify the RC gate commands, confirm semantic versioning, and document breaking changes before publishing.
 
@@ -332,6 +354,11 @@ Helpful docs:
 
 - [OSS quickstart](./docs/oss/QUICKSTART.md)
 - [OSS examples](./docs/oss/EXAMPLES.md)
+- [Under-$3 benchmark challenge](./docs/distribution/UNDER-3-CHALLENGE.md)
+- [Directory submission pack](./docs/distribution/DIRECTORY-SUBMISSIONS.md)
+- [Integration outreach pack](./docs/distribution/INTEGRATION-OUTREACH.md)
+- [Claude Code walkthrough](./docs/oss/CLAUDE-CODE-WALKTHROUGH.md)
+- [Ralph-style loop safety guide](./docs/oss/RALPH-LOOP-SAFETY.md)
 - [OSS boundary report](./docs/oss/OSS-BOUNDARY-REPORT.md)
 - [Release surface report](./docs/oss/RELEASE-SURFACE-REPORT.md)
 
@@ -360,3 +387,12 @@ Conventional commit prefixes: `feat:`, `fix:`, `chore:`, `docs:`, `refactor:`, a
 *"AI coding accountability: completes good work, refuses unsafe work, stops uneconomical work."*
 
 </div>
+
+<div align="center">
+  <br>
+  <picture>
+    <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/Keesan12/martin-loop/main/docs/assets/nvidia-inception-program.png">
+    <img src="https://raw.githubusercontent.com/Keesan12/martin-loop/main/docs/assets/nvidia-inception-program-light.png" alt="NVIDIA Inception Program logo" width="280">
+  </picture>
+  <br>
+</div>

package/demo/seeded-workspace/README.md

@@ -0,0 +1,35 @@
+# MartinLoop Demo Sandbox
+
+This workspace is the safe public demo copied by `martin-loop demo`.
+
+It is intentionally small:
+
+- `npm test` is green out of the box
+- `martin.config.yaml` keeps the budget tiny
+- the first suggested MartinLoop run can stay in stub mode with `MARTIN_LIVE=false`
+
+## Files
+
+- `src/invoice-summary.js`: tiny module used by the demo task
+- `test/invoice-summary.test.js`: Node test suite
+- `TASKS.md`: suggested objectives for a stub-safe run or a live adapter run
+- `martin.config.yaml`: low-risk governance defaults
+
+## Suggested flow
+
+```sh
+npm install
+npm test
+```
+
+Safe first run:
+
+```sh
+MARTIN_LIVE=false npx martin-loop run "Summarize the demo workspace and confirm the verifier is green" --verify "npm test"
+```
+
+Optional live run:
+
+```sh
+npx martin-loop run "Add support for a discount percentage to summarizeInvoice and update the tests" --verify "npm test" --engine codex
+```

package/demo/seeded-workspace/TASKS.md

@@ -0,0 +1,29 @@
+# Suggested Demo Tasks
+
+## Stub-safe first run
+
+Use this when you want to see MartinLoop create a governed run record without spending provider budget:
+
+```text
+Summarize the demo workspace, confirm the verifier command is green, and explain the safest next change to make.
+```
+
+Verifier:
+
+```sh
+npm test
+```
+
+## Optional live run
+
+Use this when you want a real coding task in the sandbox:
+
+```text
+Add support for a discount percentage to summarizeInvoice and update the tests while keeping the existing tax behavior intact.
+```
+
+Verifier:
+
+```sh
+npm test
+```

package/demo/seeded-workspace/martin.config.yaml

@@ -0,0 +1,11 @@
+policyProfile: strict_local
+budget:
+  maxUsd: 2
+  softLimitUsd: 1
+  maxIterations: 2
+  maxTokens: 12000
+governance:
+  destructiveActionPolicy: approval
+  telemetryDestination: local-only
+  verifierRules:
+    - npm test

package/demo/seeded-workspace/package.json

@@ -0,0 +1,8 @@
+{
+  "name": "martin-loop-demo-sandbox",
+  "private": true,
+  "type": "module",
+  "scripts": {
+    "test": "node --test"
+  }
+}

package/demo/seeded-workspace/src/invoice-summary.js

@@ -0,0 +1,11 @@
+export function summarizeInvoice(items, taxRate = 0) {
+  const subtotal = items.reduce((sum, item) => sum + item.quantity * item.unitPrice, 0);
+  const tax = Number((subtotal * taxRate).toFixed(2));
+  const total = Number((subtotal + tax).toFixed(2));
+
+  return {
+    subtotal: Number(subtotal.toFixed(2)),
+    tax,
+    total
+  };
+}

package/demo/seeded-workspace/test/invoice-summary.test.js

@@ -0,0 +1,20 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+
+import { summarizeInvoice } from "../src/invoice-summary.js";
+
+test("summarizeInvoice returns subtotal, tax, and total", () => {
+  const result = summarizeInvoice(
+    [
+      { quantity: 2, unitPrice: 19.99 },
+      { quantity: 1, unitPrice: 5.5 }
+    ],
+    0.13
+  );
+
+  assert.deepEqual(result, {
+    subtotal: 45.48,
+    tax: 5.91,
+    total: 51.39
+  });
+});

package/dist/vendor/adapters/claude-cli.d.ts

@@ -15,15 +15,18 @@ import type { MartinAdapter } from "../core/index.js";
 import { type SpawnLike } from "./cli-bridge.js";
 /**
  * Given a prompt string, returns the full argv array to pass to spawn().
- * Example for Claude:  (p) => ["--print", p, "--dangerously-skip-permissions"]
- * Example for Codex:   (p) => ["--full-auto", p]
+ * Example for Claude:  () => ["--output-format", "json", "--print"]
+ * Example for Codex:   () => ["exec", "--sandbox", "workspace-write", "-"]
  */
 export type CliArgsBuilder = (prompt: string) => string[];
+export type CliStdinBuilder = (prompt: string) => string | undefined;
 export interface AgentCliAdapterOptions {
     /** The executable to spawn (e.g. "claude", "codex"). */
     command: string;
     /** Converts a prompt string into the argv array passed to spawn(). */
     argsBuilder: CliArgsBuilder;
+    /** Optional stdin payload for CLIs that accept prompt input via stdin or `-`. */
+    stdinBuilder?: CliStdinBuilder;
     /** Adapter ID suffix. Defaults to command. */
     adapterIdSuffix?: string;
     /** Working directory for all subprocesses. Defaults to process.cwd(). */
@@ -63,8 +66,16 @@ export interface CodexCliAdapterOptions {
     label?: string;
     /** Override the model passed via --model flag. */
     model?: string;
-    /** Run in full-auto mode (--full-auto). Defaults to true. */
+    /**
+     * Deprecated no-op retained for compatibility.
+     *
+     * Codex CLI's supported non-interactive entrypoint is `codex exec`.
+     * MartinLoop now uses explicit sandboxing instead of the legacy
+     * `--full-auto` compatibility path, which can exit before verifier execution.
+     */
     fullAuto?: boolean;
+    /** Codex sandbox mode for model-generated commands. Defaults to workspace-write. */
+    sandbox?: "read-only" | "workspace-write" | "danger-full-access";
     /** Extra args appended after core args (before prompt). */
     extraArgs?: string[];
     spawnImpl?: SpawnLike;
@@ -81,7 +92,11 @@ export declare function createAgentCliAdapter(options: AgentCliAdapterOptions):
  */
 export declare function createClaudeCliAdapter(options?: ClaudeCliAdapterOptions): MartinAdapter;
 /**
- * Spawns `codex [--full-auto] [--model <model>] "<prompt>" [extraArgs]`.
+ * Spawns `codex exec --cd <workspace> --sandbox <mode> [--model <model>] [extraArgs] -`.
+ *
+ * The prompt is delivered via stdin so Windows shell quoting cannot truncate or
+ * reinterpret long MartinLoop prompts that contain paths, deny rules, or budget
+ * context.
  *
  * Requires the Codex CLI to be installed and authenticated:
  *   npm install -g @openai/codex

package/dist/vendor/adapters/claude-cli.js

@@ -129,15 +129,12 @@ export function createAgentCliAdapter(options) {
                 }
             }
             const args = options.argsBuilder(prompt);
-            // stdinPrompt: if argsBuilder signals stdin delivery by returning args ending with "--stdin-prompt",
-            // remove that sentinel and pass the prompt via stdin instead (avoids Windows shell-escaping issues).
-            const useStdin = args.at(-1) === "--stdin-prompt";
-            const spawnArgs = useStdin ? args.slice(0, -1) : args;
-            const agentResult = await runSubprocess(options.command, spawnArgs, {
+            const stdinData = options.stdinBuilder?.(prompt);
+            const agentResult = await runSubprocess(options.command, args, {
                 cwd: workingDirectory,
                 timeoutMs,
                 spawnImpl: options.spawnImpl,
-                ...(useStdin ? { stdinData: prompt } : {})
+                ...(stdinData === undefined ? {} : { stdinData })
             });
             if (agentResult.timedOut) {
                 return {
@@ -157,18 +154,19 @@ export function createAgentCliAdapter(options) {
                 };
             }
             if (agentResult.exitCode !== 0 && agentResult.stdout.trim().length === 0) {
+                const failureMessage = formatPreVerifierSubprocessFailure(options.command, agentResult.stderr, agentResult.exitCode);
                 return {
                     status: "failed",
-                    summary: `${options.command} subprocess exited with an error.`,
+                    summary: `${options.command} subprocess exited before verifier execution.`,
                     usage: normalizeUsage({
                         actualUsd: 0,
                         tokensIn: 0,
                         tokensOut: 0,
                         provenance: "unavailable"
                     }),
-                    verification: { passed: false, summary: "Subprocess error." },
+                    verification: { passed: false, summary: `Verifier not run: ${failureMessage}` },
                     failure: {
-                        message: `${agentResult.stderr.trim() || `Exit code ${String(agentResult.exitCode)}`}. environment_mismatch`
+                        message: failureMessage
                     }
                 };
             }
@@ -355,40 +353,52 @@ export function createClaudeCliAdapter(options = {}) {
             "--print",
             "--dangerously-skip-permissions",
             ...modelArgs,
-            ...extraArgs,
-            "--stdin-prompt" // sentinel: tells execute() to deliver prompt via stdin
-        ]
+            ...extraArgs
+        ],
+        stdinBuilder: (prompt) => prompt
     });
 }
 // ---------------------------------------------------------------------------
 // Pre-configured: OpenAI Codex CLI
 // ---------------------------------------------------------------------------
 /**
- * Spawns `codex [--full-auto] [--model <model>] "<prompt>" [extraArgs]`.
+ * Spawns `codex exec --cd <workspace> --sandbox <mode> [--model <model>] [extraArgs] -`.
+ *
+ * The prompt is delivered via stdin so Windows shell quoting cannot truncate or
+ * reinterpret long MartinLoop prompts that contain paths, deny rules, or budget
+ * context.
  *
  * Requires the Codex CLI to be installed and authenticated:
  *   npm install -g @openai/codex
  */
 export function createCodexCliAdapter(options = {}) {
-    const fullAuto = options.fullAuto !== false;
     const modelArgs = options.model ? ["--model", options.model] : [];
     const extraArgs = options.extraArgs ?? [];
+    const sandbox = options.sandbox ?? "workspace-write";
+    const workingDirectory = options.workingDirectory ?? process.cwd();
     return createAgentCliAdapter({
         command: "codex",
         adapterIdSuffix: "codex",
         model: options.model ?? "codex",
         label: options.label ?? "Codex CLI adapter",
-        workingDirectory: options.workingDirectory,
+        workingDirectory,
         timeoutMs: options.timeoutMs,
         verifyTimeoutMs: options.verifyTimeoutMs,
         supportsJsonOutput: false,
         spawnImpl: options.spawnImpl,
-        argsBuilder: (prompt) => [
-            ...(fullAuto ? ["--full-auto"] : []),
+        argsBuilder: () => [
+            "exec",
+            "--cd",
+            workingDirectory,
+            "--sandbox",
+            sandbox,
+            "--color",
+            "never",
             ...modelArgs,
-            prompt,
-            ...extraArgs
-        ]
+            ...extraArgs,
+            "-"
+        ],
+        stdinBuilder: (prompt) => prompt
     });
 }
 // ---------------------------------------------------------------------------
@@ -402,14 +412,23 @@ export function createCodexCliAdapter(options = {}) {
 // ---------------------------------------------------------------------------
 function buildPrompt(request) {
     const lines = [];
+    const mutationMode = request.context.mutationMode ?? "edit";
     lines.push("You are running in autonomous agentic mode.");
-    lines.push("MAKE ALL REQUIRED FILE EDITS NOW. Do not ask for confirmation. Do not ask clarifying questions.");
-    lines.push("Do not explain what you found without also making the changes. Edit the files and complete the task.");
+    if (mutationMode === "verify_only") {
+        lines.push("DO NOT EDIT FILES. Run the verifier only and report whether it passes.");
+        lines.push("Do not ask for confirmation. Do not ask clarifying questions.");
+    }
+    else {
+        lines.push("MAKE ALL REQUIRED FILE EDITS NOW. Do not ask for confirmation. Do not ask clarifying questions.");
+        lines.push("Do not explain what you found without also making the changes. Edit the files and complete the task.");
+    }
     lines.push("");
     lines.push("If PROGRESS.md exists in your working directory, read it first for context from prior attempts.");
     lines.push("If it does not exist, proceed with the objective below.");
     lines.push("");
-    lines.push("Complete the following coding task. Make all necessary file changes.");
+    lines.push(mutationMode === "verify_only"
+        ? "Complete the following verification-only task without making file changes."
+        : "Complete the following coding task. Make all necessary file changes.");
     lines.push("When you are done, the verification commands listed below must pass.");
     lines.push("");
     lines.push("OBJECTIVE:");
@@ -447,7 +466,9 @@ function buildPrompt(request) {
     lines.push(`  Attempt ${String(attemptNumber)}`);
     lines.push(`  Remaining budget: $${String(request.context.remainingBudgetUsd)} USD`);
     lines.push(`  Remaining iterations: ${String(request.context.remainingIterations)}`);
-    lines.push("  Do not expand scope beyond what is needed to pass verification.");
+    lines.push(mutationMode === "verify_only"
+        ? "  Do not modify files; only run verification."
+        : "  Do not expand scope beyond what is needed to pass verification.");
     lines.push("");
     if (request.previousAttempts.length > 0) {
         lines.push("PRIOR FAILED ATTEMPTS (learn from these — do not repeat the same mistakes):");
@@ -494,6 +515,16 @@ function truncate(text, maxLength) {
     }
     return `...${text.slice(-(maxLength - 3))}`;
 }
+function formatPreVerifierSubprocessFailure(command, stderr, exitCode) {
+    const detail = stderr.trim() || `Exit code ${String(exitCode)}`;
+    const lowerDetail = detail.toLowerCase();
+    const codexLaunchBlocked = command === "codex" &&
+        /\b(full-auto|sandbox|approval|permission|trusted|safety|unexpected argument)\b/u.test(lowerDetail);
+    if (codexLaunchBlocked) {
+        return `Codex CLI failed before patch completion, likely due to its launch/sandbox configuration. MartinLoop invokes Codex through "codex exec --sandbox workspace-write"; verify Codex CLI auth and configuration if this persists. ${detail}. environment_mismatch`;
+    }
+    return `${detail}. environment_mismatch`;
+}
 const INJECTION_PATTERNS = [
     /\[INST\]/gi,
     /<\/?system>/gi,

package/dist/vendor/adapters/cli-bridge.d.ts

@@ -26,3 +26,4 @@ export declare function readGitExecutionArtifacts(repoRoot: string, timeoutMs: n
     changedFiles?: string[];
     diffStats?: ReturnType<typeof diffStatsFromNumstat>;
 }>;
+export declare function splitCommand(command: string): string[];