Add Sandbox tools to sandbox.commands.run()

[jpwilliams]

Summary

Adds the ability for code running in a Sandbox to call functions from the local scope of the orchestrating run.

Pass tools to sandbox.commands.run() and each call the code makes to a tool runs as its own durable step.

await sandbox.commands.run("agent", "node agent.js", {
  tools: {
    getCustomer: (id: string) => db.customers.find(id),
  },
});

The Sandbox code simply calls tools.getCustomer("123").

Checklist

• Added a docs PR that references this PR N/A Later bump
• Added unit/integration tests
• Added changesets if applicable

[changeset-bot]

🦋 Changeset detected

Latest commit: e9d9aa5

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
inngest	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[greptile-apps]

Confidence Score: 4/5

[Medium impact] Adds tool invocation capability to sandbox command execution.

Fix repeated command IDs before merging; later calls can silently skip their commands.

Findings

1.  Repeated IDs skip commands ▶
2.  Failures lose sandbox trace details ▶

Fix with agent prompt

### Issue 1
packages/inngest/src/components/sandbox/middleware.ts:175-177
Sequential `commands.run()` calls that reuse an ID with `tools` only run the first command. `step.run()` gives later calls distinct indexed IDs, but `runCommandWithTools()` still uses the original ID for its sandbox directory. The second wrapper sees that directory as already started, skips the new command, and returns the first command's output.

Give each command occurrence a replay-stable identity before deriving its directory and child-step IDs.

### Issue 2
packages/inngest/src/components/sandbox/middleware.ts:194-195
Failed command steps with `tools` never write their sandbox trace details. This wrapper writes `meta` only after `fn()` succeeds, then just rethrows on failure. Unlike ordinary commands, a failed process start or wait loses the sandbox name, command, and service error code, making failures harder to debug.

Write the sandbox trace details on failure too, while keeping trace errors non-fatal.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Summary

Adds tools and toolOptions to durable sandbox.commands.run() calls.

• Sandbox commands can call app functions as durable tools.

_{Reviews (1) · Last reviewed commit: "feat(sandbox): tools on commands.run" · Reviewed by Greptile}