AGENT SKILL

sim-zero-mode

Use when building/testing an app or agent that depends on paid APIs or LLM calls and you must exercise plumbing/UI/harness without spend or real intelligence, for CI, demos, local dev. Activates on test without burning API budget, demo without live calls, or fakes/offline mode.

License 许可
MIT
最近更新
2026-11-15
Stars
★ 0
Forks
3

01 前言 Front Matter

name
sim-zero-mode
description
Use when building/testing an app or agent that depends on paid APIs or LLM calls and you must exercise plumbing/UI/harness without spend or real intelligence, for CI, demos, local dev. Activates on test without burning API budget, demo without live calls, or fakes/offline mode.

03 SKILL.md 正文

Sim-Zero Mode

A whole-application mode where everything runs on fake data with NO real LLM/paid/external calls. The app still works end-to-end — it just knows nothing. Every screen renders, every flow completes, every agent step fires; only the intelligence and the spend are absent.

This is fakes-over-mocks elevated to a first-class runtime mode, not a test-only scaffold.

What sim-zero proves (and doesn't)

  • Proves: plumbing, wiring, UI rendering, navigation, the agent harness/orchestration, state flow, error handling, contracts between components.
  • Does NOT prove: answer quality, model behavior, real third-party correctness. (That's what live evals are for — keep them separate.)

So: a green sim-zero run means "the machine moves correctly," not "the machine is smart." Both matter; sim-zero isolates the first cheaply.

Design rules

  1. One switch, whole app. A single config/flag (SIM_ZERO=1) flips every external boundary to a fake — not per-call opt-in. The boundary, not the caller, decides.
  2. Fakes return plausible, deterministic, shaped data that satisfies the real contract (same schema/types), so downstream code and UI behave exactly as in production. Deterministic → CI-stable.
  3. No network/secret required. Sim-zero must run with zero credentials and zero egress — that's what makes it safe for CI and untrusted demos.
  4. Honest labeling. The UI/logs should make it discoverable that data is simulated (so a demo isn't mistaken for real intelligence).
  5. Cover the error paths too. Fakes should be able to simulate failures/timeouts/empties, not just happy-path data.

Where it pays off

  • CI: run the full app/agent harness on every PR with no spend and no flakiness from external services.
  • Local dev: work offline, fast, free.
  • Demos: show the product working safely without live keys or unpredictable model output.
  • Plumbing regression: catch "the wiring broke" independently of "the model changed."

The analog to generalize from

A trading/agent system's paper-mode (executes the full pipeline against simulated fills, no real orders/money) is sim-zero for that domain. Generalize the pattern: every agent and app should ship a sim-zero mode. If you can't run your app with all external intelligence/spend switched off, your boundaries aren't clean enough — fixing that is itself valuable.

Anti-patterns

  • Per-call mocks scattered through tests instead of one app-wide boundary switch → drift, partial coverage, "works in tests, not in the app".
  • Fakes that don't match the real contract → green sim-zero, broken production.
  • Non-deterministic fakes → flaky CI.
  • Sim-zero that secretly still calls something paid → defeats the purpose; assert zero egress.

04 同仓库更多 Skills