[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7B2Z5B5Z70VB1RW98RP2HPW":3,"skill-same:01M7B2Z5B5Z70VB1RW98RP2HPW":48},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":25,"commitSha":26,"skillDirectory":27,"forksCount":28,"repositoryHtmlUrl":29,"repositoryHomepageUrl":6,"ownerHtmlUrl":30,"ownerBlogUrl":31,"ownerDescription":32,"updatedAt":33,"install":34,"resources":47},"01M7B2Z5B5Z70VB1RW98RP2HPW","hamster-e2b-live-validation",null,"Prove a claim about sandbox-side behavior by running it in a real E2B sandbox instead of reading comments. Use when a delivery or environment boot fails inside the sandbox, when a probe\u002Fhealth-check times out or hangs, when you need to know what is actually installed in a snapshot and at what version, when a comment asserts how a third-party CLI behaves, or when someone proposes raising a timeout. Covers the throw-shapes E2B uses for exit-vs-timeout, cold snapshot-resume timing, deliberate break-state testing, hang bisection via stdin, and offline npm lineage checks.","gethamster\u002Fskills","gethamster","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F215697643?v=4","01M7B2R6PGB3SDZZ71E0WE2S74","Hamster","01M7B2R6Q21HWJRXZZWB78BNDG",4,"MIT","https:\u002F\u002Fgithub.com\u002Fgethamster\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002F.experimental\u002Fhamster-e2b-live-validation\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: hamster-e2b-live-validation\ndescription: \"Prove a claim about sandbox-side behavior by running it in a real E2B sandbox instead of reading comments. Use when a delivery or environment boot fails inside the sandbox, when a probe\u002Fhealth-check times out or hangs, when you need to know what is actually installed in a snapshot and at what version, when a comment asserts how a third-party CLI behaves, or when someone proposes raising a timeout. Covers the throw-shapes E2B uses for exit-vs-timeout, cold snapshot-resume timing, deliberate break-state testing, hang bisection via stdin, and offline npm lineage checks.\"\nversion: \"1.0.0\"\n---\n\n# Live E2B Validation\n\n## When to Apply\n\nReach for this the moment a question is about **what actually happens inside the sandbox**:\n\n- A boot\u002Fdelivery step fails with `deadline_exceeded`, `CommandExitError`, or a hang.\n- A comment claims a third-party CLI behaves some way (`--version` exits, `--help` hangs, the binary lives at path X).\n- Someone proposes raising a timeout. Measure first — the timeout is usually not the bug.\n- You need the installed version \u002F package identity \u002F file layout of a baked toolchain.\n- You changed a sandbox-side command string and want the shipped string proven, not the one you retyped.\n\nThe governing rule: **sandbox-side installs are unpinned** (`sandbox-agent install-agent \u003Cagent>` takes npm-latest at bake time), and environment snapshots are baked from user sandboxes months earlier. So comments about third-party behavior rot silently and invisibly. Anything load-bearing gets measured.\n\nRelated: `hamster-e2b-studio` for the template\u002Fsnapshot architecture itself, [`sandbox-agent`](..\u002Fsandbox-agent\u002FSKILL.md) for the agent runtime.\n\n## Credentials and Entry Point\n\n`apps\u002Fhorde-v2\u002F.env` already carries a working key and the template deliveries launch from:\n\n```\nE2B_API_KEY=...\nE2B_TEMPLATE_ID=hamster-pre-production\u002Fhamster-base-snapshot\n```\n\nDeployed envs share one E2B team, so this snapshot name is what most deliveries actually boot. Run probes with an **absolute** env-file path (the shell's cwd moves between calls):\n\n```bash\nnpx tsx --env-file=\u002Fabs\u002Fpath\u002Fto\u002Fapps\u002Fhorde-v2\u002F.env \u002Fabs\u002Fpath\u002Fto\u002Fscratchpad\u002Fprobe.ts\n```\n\nWrite probe scripts into the **scratchpad, never the repo**. They are throwaway evidence, not deliverables.\n\n## The Harness\n\nCopy this. The `run()` wrapper is not boilerplate — it is required, for the reason in the next section.\n\n```ts\nimport { Sandbox } from 'e2b';\n\nasync function main() {\n  const sbx = await Sandbox.create(process.env.E2B_TEMPLATE_ID!, {\n    apiKey: process.env.E2B_API_KEY!,\n    timeoutMs: 300_000,\n  });\n  console.log('Sandbox:', sbx.sandboxId);\n\n  const run = async (label: string, cmd: string, timeoutMs = 45_000) => {\n    const t = Date.now();\n    try {\n      const r = await sbx.commands.run(cmd, { timeoutMs });\n      console.log(\n        `### ${label}: exit=${r.exitCode} ms=${Date.now() - t} ` +\n          `stdout=${JSON.stringify(r.stdout.slice(0, 300))} ` +\n          `stderr=${JSON.stringify(r.stderr.slice(0, 300))}`,\n      );\n    } catch (e: any) {\n      \u002F\u002F Both a nonzero exit AND a timeout arrive here, as different shapes.\n      console.log(\n        `### ${label}: THREW ms=${Date.now() - t} ${e?.name} ` +\n          `exitCode=${e?.exitCode} ` +\n          `stdout=${JSON.stringify(String(e?.stdout ?? ''))} ` +\n          `stderr=${JSON.stringify(String(e?.stderr ?? '').slice(0, 300))}`,\n      );\n    }\n  };\n\n  try {\n    await run('presence', 'test -x \u002Fpath\u002Fto\u002Fthing');\n    \u002F\u002F ...\n  } finally {\n    await sbx.kill(); \u002F\u002F bills by wall clock — always kill\n  }\n}\n\nmain().catch((e) => {\n  console.error(e);\n  process.exit(1);\n});\n```\n\nAlways time every command. `ms=` is the whole point: it separates \"slow\" from \"hung\" and gives you the headroom number to argue about a timeout with.\n\n## Gotcha: `commands.run` throws on both failure modes, differently\n\nYou cannot read `result.exitCode` — there is often no result. Two distinct throws:\n\n| Failure | `e.name` | `e.exitCode` | Where the diagnostic lives |\n|---|---|---|---|\n| Nonzero exit | `CommandExitError` | the exit code | `e.stdout` \u002F `e.stderr` |\n| Timeout (`timeoutMs` expired) | `TimeoutError` | **`undefined`** | `e.message`, containing `[deadline_exceeded]` |\n\nConsequences that bite:\n\n- A probe that lands in an \"exit unknown\" branch was a **timeout**, not an exit — the process never returned.\n- A `TimeoutError` carries no stdout\u002Fstderr at all. If you need output from a hung command, you cannot get it this way; restructure the probe.\n- Production code wrapping these needs `toCommandErrorContext()` (`e2b\u002Fcommand-error.ts`); `toErrorContext()` only reads `.message`\u002F`.code` and silently drops the command output.\n\n## Import the real constants — do not retype the command\n\nA probe that validates a string you typed by hand proves nothing about what ships. Import from the source module:\n\n```ts\nimport {\n  CODEX_ACP_PRESENCE_PROBE,\n  CODEX_ACP_VERSION_PROBE,\n} from '\u002Fabs\u002Fpath\u002Fto\u002Fapps\u002Fhorde-v2\u002Fsrc\u002Fdelivery\u002Fsandbox\u002Fe2b\u002Fagent-versions';\n```\n\n`tsx` resolves the workspace deps fine from an absolute path. If the module pulls Nest\u002Fwebpack asset imports it may not load — that constraint is why `agent-versions.ts` documents that it must stay free of `*.sh` asset imports.\n\n## Cold Snapshot-Resume Timing\n\nThe claim \"the timeout is too short on a cold boot\" is testable, and usually false. Memory snapshots resume warm.\n\n```ts\nconst sbx = await Sandbox.create(template, { apiKey, timeoutMs: 300_000 });\n\u002F\u002F ... warm baseline ...\nawait sbx.betaPause();\n\n\u002F\u002F Resume, and make the command under test the FIRST command after resume.\nconst resumed = await Sandbox.connect(sbx.sandboxId, {\n  apiKey,\n  timeoutMs: 300_000,\n  autoResume: true,\n} as any);\nawait run('COLD first exec', commandUnderTest);\nawait resumed.betaPause();\n```\n\nRules:\n\n- **One candidate per resume.** The first command pays the page-in cost; everything after it is warm and tells you nothing.\n- `Sandbox.resume` **does not exist** in e2b 2.38.3. The statics are `list \u002F create \u002F connect \u002F fork`. Resuming is `connect(id, { autoResume: true })`.\n- Reference numbers measured on `hamster-base-snapshot`: cold first exec of a Node one-liner ≈ **515 ms**, of a bundled Node CLI ≈ **795 ms**. If your probe's budget is 45 s, cold-start is not your problem.\n\n## Break-State Testing\n\nThe happy path is the easy half. A health check earns trust by proving its failure paths, so corrupt the install on purpose and re-probe:\n\n```ts\n\u002F\u002F Dangling symlink: bin survives, target package removed.\nawait run('break', `rm -rf $(dirname $(dirname $(readlink -f ${BIN})))`);\nawait run('presence after break', PRESENCE_PROBE);\nawait run('version after break', VERSION_PROBE);\n\n\u002F\u002F Fully absent.\nawait run('break', `rm -f ${BIN}`);\nawait run('presence when absent', PRESENCE_PROBE);\n```\n\nRecord the exit code and stderr of each state and put the table in the PR. That is what distinguishes \"fails loud\" from \"we hope it fails loud\".\n\n## Bisecting a Hang\n\nWhen a command hangs, find out **why** before changing anything:\n\n1. **Bound it tightly** (`timeoutMs: 15_000`) so iteration is fast.\n2. **Re-run with stdin closed**: `cmd \u003C \u002Fdev\u002Fnull`. This is the sharp discriminator —\n   a process that **exits instantly with `\u002Fdev\u002Fnull` but hangs without it** is a stdio\n   server waiting on input, not a slow process. That is a fall-through into a JSON-RPC \u002F\n   ACP bridge, and no timeout will ever fix it.\n3. **Compare sibling flags.** `--version` vs `--help` vs no args. If one exits and another\n   hangs, the tool has an unbounded default branch and *every* unrecognised flag reaches it.\n4. Conclude: a probe whose safety depends on a third party's argv table is a hang waiting\n   for a release. Prefer reading static metadata (a `package.json` on disk) over launching\n   the binary at all.\n\n## Offline Checks (no sandbox needed)\n\nCheaper than a sandbox, and the right tool for \"did upstream change behavior\":\n\n```bash\nnpm view \u003Cpkg> version bin                 # entry point + current version\nnpm view \u003Cpkg> versions --json             # full release list\nnpm view \u003Cpkg> time --json                 # release dates — map a snapshot to its era\nnpm pack \u003Cpkg>@\u003Cversion>                   # then untar and read dist\u002F or bin\u002F\n```\n\nFor a bundled JS CLI, grep the shipped bundle for its arg handling rather than trusting the README:\n\n```bash\ngrep -o 'process.argv[^;]\\{0,200\\}' package\u002Fdist\u002Findex.js\n```\n\nFor platform-specific native binaries, pack the variant matching your machine (e.g. `-darwin-arm64`) and run it locally. **macOS has no `timeout(1)`** — bound it with:\n\n```bash\nperl -e 'alarm 15; exec @ARGV' .\u002Fbinary --version \u003C \u002Fdev\u002Fnull; echo \"exit=$?\"\n```\n\nRun the same flags across several versions in a loop to find exactly where behavior changed. This is how you catch a package **lineage switch** (upstream moving a tool to a different scope with different arg parsing), which is invisible from the install command alone.\n\n## Turn the Finding Into a Test\n\nLive validation proves it once. A unit test keeps it proven.\n\nParameterise the shipped command by its variable input so a spec can execute the **real** string:\n\n```ts\nexport function codexAcpVersionProbe(binPath: string): string { \u002F* ... *\u002F }\nexport const CODEX_ACP_VERSION_PROBE = codexAcpVersionProbe(CODEX_ACP_BIN);\n```\n\n```ts\n\u002F\u002F Build a fixture install in a tmpdir, then run it through a shell, as E2B does —\n\u002F\u002F this covers the quoting too.\nconst result = spawnSync(codexAcpVersionProbe(fixtureBin), {\n  shell: true,\n  encoding: 'utf8',\n});\n```\n\nThen confirm the new tests **fail against the old implementation** before accepting them. String-matching a command (`expect(cmd).not.toContain('--version')`) checks syntax; executing it checks behavior. Keep a structural assertion only where the *shape* is the invariant (e.g. \"must never exec the binary\").\n\n## Hygiene\n\n- `sbx.kill()` in a `finally`. Sandboxes bill by wall clock.\n- Do not rely on the wall-clock timeout to clean up. The production helper `createE2BSandbox` deliberately sets `lifecycle.onTimeout: 'pause'` so a timed-out delivery stays recoverable — an abandoned sandbox there **pauses rather than dies** and lingers.\n- A sandbox you paused yourself (for cold-resume timing) is still alive, and a script that throws between `betaPause()` and its cleanup leaves it that way. Kill it explicitly:\n  `await (await Sandbox.connect(id, { apiKey, autoResume: true } as any)).kill()`.\n- Sweep for strays afterwards. `Sandbox.list()` returns a **paginator, not an array** — it has no `.data` and is not iterable:\n\n  ```ts\n  const p: any = await Sandbox.list({ apiKey } as any);\n  const all: any[] = [];\n  while (p.hasNext) all.push(...(await p.nextItems()));\n  ```\n\n  The account holds hundreds of legitimately paused sandboxes (user environments), so filter to the ids **your** run printed rather than judging by age, and never kill one you did not create.\n- Probe scripts live in the scratchpad. Never commit them.\n- Put the measured numbers in the commit\u002FPR body. \"Refuted by measurement: cold resume was 795 ms against a 45 s bound\" is reviewable; \"should be fine\" is not.\n","# Live E2B Validation\n\n## When to Apply\n\nReach for this the moment a question is about **what actually happens inside the sandbox**:\n\n- A boot\u002Fdelivery step fails with `deadline_exceeded`, `CommandExitError`, or a hang.\n- A comment claims a third-party CLI behaves some way (`--version` exits, `--help` hangs, the binary lives at path X).\n- Someone proposes raising a timeout. Measure first — the timeout is usually not the bug.\n- You need the installed version \u002F package identity \u002F file layout of a baked toolchain.\n- You changed a sandbox-side command string and want the shipped string proven, not the one you retyped.\n\nThe governing rule: **sandbox-side installs are unpinned** (`sandbox-agent install-agent \u003Cagent>` takes npm-latest at bake time), and environment snapshots are baked from user sandboxes months earlier. So comments about third-party behavior rot silently and invisibly. Anything load-bearing gets measured.\n\nRelated: `hamster-e2b-studio` for the template\u002Fsnapshot architecture itself, [`sandbox-agent`](..\u002Fsandbox-agent\u002FSKILL.md) for the agent runtime.\n\n## Credentials and Entry Point\n\n`apps\u002Fhorde-v2\u002F.env` already carries a working key and the template deliveries launch from:\n\n```\nE2B_API_KEY=...\nE2B_TEMPLATE_ID=hamster-pre-production\u002Fhamster-base-snapshot\n```\n\nDeployed envs share one E2B team, so this snapshot name is what most deliveries actually boot. Run probes with an **absolute** env-file path (the shell's cwd moves between calls):\n\n```bash\nnpx tsx --env-file=\u002Fabs\u002Fpath\u002Fto\u002Fapps\u002Fhorde-v2\u002F.env \u002Fabs\u002Fpath\u002Fto\u002Fscratchpad\u002Fprobe.ts\n```\n\nWrite probe scripts into the **scratchpad, never the repo**. They are throwaway evidence, not deliverables.\n\n## The Harness\n\nCopy this. The `run()` wrapper is not boilerplate — it is required, for the reason in the next section.\n\n```ts\nimport { Sandbox } from 'e2b';\n\nasync function main() {\n  const sbx = await Sandbox.create(process.env.E2B_TEMPLATE_ID!, {\n    apiKey: process.env.E2B_API_KEY!,\n    timeoutMs: 300_000,\n  });\n  console.log('Sandbox:', sbx.sandboxId);\n\n  const run = async (label: string, cmd: string, timeoutMs = 45_000) => {\n    const t = Date.now();\n    try {\n      const r = await sbx.commands.run(cmd, { timeoutMs });\n      console.log(\n        `### ${label}: exit=${r.exitCode} ms=${Date.now() - t} ` +\n          `stdout=${JSON.stringify(r.stdout.slice(0, 300))} ` +\n          `stderr=${JSON.stringify(r.stderr.slice(0, 300))}`,\n      );\n    } catch (e: any) {\n      \u002F\u002F Both a nonzero exit AND a timeout arrive here, as different shapes.\n      console.log(\n        `### ${label}: THREW ms=${Date.now() - t} ${e?.name} ` +\n          `exitCode=${e?.exitCode} ` +\n          `stdout=${JSON.stringify(String(e?.stdout ?? ''))} ` +\n          `stderr=${JSON.stringify(String(e?.stderr ?? '').slice(0, 300))}`,\n      );\n    }\n  };\n\n  try {\n    await run('presence', 'test -x \u002Fpath\u002Fto\u002Fthing');\n    \u002F\u002F ...\n  } finally {\n    await sbx.kill(); \u002F\u002F bills by wall clock — always kill\n  }\n}\n\nmain().catch((e) => {\n  console.error(e);\n  process.exit(1);\n});\n```\n\nAlways time every command. `ms=` is the whole point: it separates \"slow\" from \"hung\" and gives you the headroom number to argue about a timeout with.\n\n## Gotcha: `commands.run` throws on both failure modes, differently\n\nYou cannot read `result.exitCode` — there is often no result. Two distinct throws:\n\n| Failure | `e.name` | `e.exitCode` | Where the diagnostic lives |\n|---|---|---|---|\n| Nonzero exit | `CommandExitError` | the exit code | `e.stdout` \u002F `e.stderr` |\n| Timeout (`timeoutMs` expired) | `TimeoutError` | **`undefined`** | `e.message`, containing `[deadline_exceeded]` |\n\nConsequences that bite:\n\n- A probe that lands in an \"exit unknown\" branch was a **timeout**, not an exit — the process never returned.\n- A `TimeoutError` carries no stdout\u002Fstderr at all. If you need output from a hung command, you cannot get it this way; restructure the probe.\n- Production code wrapping these needs `toCommandErrorContext()` (`e2b\u002Fcommand-error.ts`); `toErrorContext()` only reads `.message`\u002F`.code` and silently drops the command output.\n\n## Import the real constants — do not retype the command\n\nA probe that validates a string you typed by hand proves nothing about what ships. Import from the source module:\n\n```ts\nimport {\n  CODEX_ACP_PRESENCE_PROBE,\n  CODEX_ACP_VERSION_PROBE,\n} from '\u002Fabs\u002Fpath\u002Fto\u002Fapps\u002Fhorde-v2\u002Fsrc\u002Fdelivery\u002Fsandbox\u002Fe2b\u002Fagent-versions';\n```\n\n`tsx` resolves the workspace deps fine from an absolute path. If the module pulls Nest\u002Fwebpack asset imports it may not load — that constraint is why `agent-versions.ts` documents that it must stay free of `*.sh` asset imports.\n\n## Cold Snapshot-Resume Timing\n\nThe claim \"the timeout is too short on a cold boot\" is testable, and usually false. Memory snapshots resume warm.\n\n```ts\nconst sbx = await Sandbox.create(template, { apiKey, timeoutMs: 300_000 });\n\u002F\u002F ... warm baseline ...\nawait sbx.betaPause();\n\n\u002F\u002F Resume, and make the command under test the FIRST command after resume.\nconst resumed = await Sandbox.connect(sbx.sandboxId, {\n  apiKey,\n  timeoutMs: 300_000,\n  autoResume: true,\n} as any);\nawait run('COLD first exec', commandUnderTest);\nawait resumed.betaPause();\n```\n\nRules:\n\n- **One candidate per resume.** The first command pays the page-in cost; everything after it is warm and tells you nothing.\n- `Sandbox.resume` **does not exist** in e2b 2.38.3. The statics are `list \u002F create \u002F connect \u002F fork`. Resuming is `connect(id, { autoResume: true })`.\n- Reference numbers measured on `hamster-base-snapshot`: cold first exec of a Node one-liner ≈ **515 ms**, of a bundled Node CLI ≈ **795 ms**. If your probe's budget is 45 s, cold-start is not your problem.\n\n## Break-State Testing\n\nThe happy path is the easy half. A health check earns trust by proving its failure paths, so corrupt the install on purpose and re-probe:\n\n```ts\n\u002F\u002F Dangling symlink: bin survives, target package removed.\nawait run('break', `rm -rf $(dirname $(dirname $(readlink -f ${BIN})))`);\nawait run('presence after break', PRESENCE_PROBE);\nawait run('version after break', VERSION_PROBE);\n\n\u002F\u002F Fully absent.\nawait run('break', `rm -f ${BIN}`);\nawait run('presence when absent', PRESENCE_PROBE);\n```\n\nRecord the exit code and stderr of each state and put the table in the PR. That is what distinguishes \"fails loud\" from \"we hope it fails loud\".\n\n## Bisecting a Hang\n\nWhen a command hangs, find out **why** before changing anything:\n\n1. **Bound it tightly** (`timeoutMs: 15_000`) so iteration is fast.\n2. **Re-run with stdin closed**: `cmd \u003C \u002Fdev\u002Fnull`. This is the sharp discriminator —\n   a process that **exits instantly with `\u002Fdev\u002Fnull` but hangs without it** is a stdio\n   server waiting on input, not a slow process. That is a fall-through into a JSON-RPC \u002F\n   ACP bridge, and no timeout will ever fix it.\n3. **Compare sibling flags.** `--version` vs `--help` vs no args. If one exits and another\n   hangs, the tool has an unbounded default branch and *every* unrecognised flag reaches it.\n4. Conclude: a probe whose safety depends on a third party's argv table is a hang waiting\n   for a release. Prefer reading static metadata (a `package.json` on disk) over launching\n   the binary at all.\n\n## Offline Checks (no sandbox needed)\n\nCheaper than a sandbox, and the right tool for \"did upstream change behavior\":\n\n```bash\nnpm view \u003Cpkg> version bin                 # entry point + current version\nnpm view \u003Cpkg> versions --json             # full release list\nnpm view \u003Cpkg> time --json                 # release dates — map a snapshot to its era\nnpm pack \u003Cpkg>@\u003Cversion>                   # then untar and read dist\u002F or bin\u002F\n```\n\nFor a bundled JS CLI, grep the shipped bundle for its arg handling rather than trusting the README:\n\n```bash\ngrep -o 'process.argv[^;]\\{0,200\\}' package\u002Fdist\u002Findex.js\n```\n\nFor platform-specific native binaries, pack the variant matching your machine (e.g. `-darwin-arm64`) and run it locally. **macOS has no `timeout(1)`** — bound it with:\n\n```bash\nperl -e 'alarm 15; exec @ARGV' .\u002Fbinary --version \u003C \u002Fdev\u002Fnull; echo \"exit=$?\"\n```\n\nRun the same flags across several versions in a loop to find exactly where behavior changed. This is how you catch a package **lineage switch** (upstream moving a tool to a different scope with different arg parsing), which is invisible from the install command alone.\n\n## Turn the Finding Into a Test\n\nLive validation proves it once. A unit test keeps it proven.\n\nParameterise the shipped command by its variable input so a spec can execute the **real** string:\n\n```ts\nexport function codexAcpVersionProbe(binPath: string): string { \u002F* ... *\u002F }\nexport const CODEX_ACP_VERSION_PROBE = codexAcpVersionProbe(CODEX_ACP_BIN);\n```\n\n```ts\n\u002F\u002F Build a fixture install in a tmpdir, then run it through a shell, as E2B does —\n\u002F\u002F this covers the quoting too.\nconst result = spawnSync(codexAcpVersionProbe(fixtureBin), {\n  shell: true,\n  encoding: 'utf8',\n});\n```\n\nThen confirm the new tests **fail against the old implementation** before accepting them. String-matching a command (`expect(cmd).not.toContain('--version')`) checks syntax; executing it checks behavior. Keep a structural assertion only where the *shape* is the invariant (e.g. \"must never exec the binary\").\n\n## Hygiene\n\n- `sbx.kill()` in a `finally`. Sandboxes bill by wall clock.\n- Do not rely on the wall-clock timeout to clean up. The production helper `createE2BSandbox` deliberately sets `lifecycle.onTimeout: 'pause'` so a timed-out delivery stays recoverable — an abandoned sandbox there **pauses rather than dies** and lingers.\n- A sandbox you paused yourself (for cold-resume timing) is still alive, and a script that throws between `betaPause()` and its cleanup leaves it that way. Kill it explicitly:\n  `await (await Sandbox.connect(id, { apiKey, autoResume: true } as any)).kill()`.\n- Sweep for strays afterwards. `Sandbox.list()` returns a **paginator, not an array** — it has no `.data` and is not iterable:\n\n  ```ts\n  const p: any = await Sandbox.list({ apiKey } as any);\n  const all: any[] = [];\n  while (p.hasNext) all.push(...(await p.nextItems()));\n  ```\n\n  The account holds hundreds of legitimately paused sandboxes (user environments), so filter to the ids **your** run printed rather than judging by age, and never kill one you did not create.\n- Probe scripts live in the scratchpad. Never commit them.\n- Put the measured numbers in the commit\u002FPR body. \"Refuted by measurement: cold resume was 795 ms against a 45 s bound\" is reviewable; \"should be fine\" is not.\n",{"name":5,"version":24,"description":7},"1.0.0","main","fe3bb1476c3ec23d8a366eac0fa8cd9f0852360a","skills\u002F.experimental\u002Fhamster-e2b-live-validation",0,"https:\u002F\u002Fgithub.com\u002Fgethamster\u002Fskills","https:\u002F\u002Fgithub.com\u002Fgethamster","https:\u002F\u002Ftryhamster.com","We're building the AI-native future of work, unlocking the promised acceleration for your whole team, from idea to impact.","2026-11-05 11:33:33",{"source":8,"skillName":5,"commands":35,"prompt":45,"zipAvailable":46},[36,39,42],{"tool":37,"command":38},"npx","npx skills add gethamster\u002Fskills --skill hamster-e2b-live-validation",{"tool":40,"command":41},"bunx","bunx skills add gethamster\u002Fskills --skill hamster-e2b-live-validation",{"tool":43,"command":44},"pnpm","pnpm dlx skills add gethamster\u002Fskills --skill hamster-e2b-live-validation","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fgethamster\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002F.experimental\u002Fhamster-e2b-live-validation\u002FSKILL.md",true,[],[49,56,63,70,77,84],{"skillNo":50,"name":51,"nameZh":6,"description":52,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":53,"categories":54,"keywords":55,"publishedAt":19},"01M7B5MWYW0KE251276S56KQDZ","writing-waterfall-requirements-documents","Write a waterfall requirements document: elicit, structure and review testable, traceable requirements, then baseline the specification before design.","https:\u002F\u002Fgithub.com\u002Fgethamster\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fwriting-waterfall-requirements-documents\u002FSKILL.md",[],[],{"skillNo":57,"name":58,"nameZh":6,"description":59,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":60,"categories":61,"keywords":62,"publishedAt":19},"01M7B5MRCXWWPKREPB1CHQ7H90","writing-vision-and-values-statements","Write a V2MOM vision a reader can test and a ranked list of values that settles real tradeoffs, so the rest of the document has a clear anchor.","https:\u002F\u002Fgithub.com\u002Fgethamster\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fwriting-vision-and-values-statements\u002FSKILL.md",[],[],{"skillNo":64,"name":65,"nameZh":6,"description":66,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":67,"categories":68,"keywords":69,"publishedAt":19},"01M7B5MKH2CH22YZAHQW3080VQ","writing-system-prompts-for-claude","Write Claude prompts and CLAUDE.md files that supply exact inputs, scope the task, and define done so the first attempt can be checked.","https:\u002F\u002Fgithub.com\u002Fgethamster\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fwriting-system-prompts-for-claude\u002FSKILL.md",[],[],{"skillNo":71,"name":72,"nameZh":6,"description":73,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":74,"categories":75,"keywords":76,"publishedAt":19},"01M7B5ME1Y3X9RNRKP5HKHKACZ","writing-page-specific-website-copy","Website copywriting by page type: write homepage, landing page, pricing page and feature page copy that fits each page's visitor and single job.","https:\u002F\u002Fgithub.com\u002Fgethamster\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fwriting-page-specific-website-copy\u002FSKILL.md",[],[],{"skillNo":78,"name":79,"nameZh":6,"description":80,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":81,"categories":82,"keywords":83,"publishedAt":19},"01M7B5M8MQNM8MN8PYH6BVWJ2B","writing-internal-press-releases","How to write an internal press release: the one-page working backwards press release that opens a PR\u002FFAQ document before anything is built.","https:\u002F\u002Fgithub.com\u002Fgethamster\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fwriting-internal-press-releases\u002FSKILL.md",[],[],{"skillNo":85,"name":86,"nameZh":6,"description":87,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":88,"categories":89,"keywords":90,"publishedAt":19},"01M7B5M3P3YH1XF37R8PWXXSBK","writing-email-sequences-that-sell","Email sequence copywriting for nurture and sales: give each email one job, write clear subject lines, and close with one specific call to action.","https:\u002F\u002Fgithub.com\u002Fgethamster\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fwriting-email-sequences-that-sell\u002FSKILL.md",[],[]]