Skip to content

FAQ.md

Latest commit

 

History

History
111 lines (72 loc) · 3.85 KB

FAQ.md

File metadata and controls

111 lines (72 loc) · 3.85 KB
Raw

FAQ

Is NightPilot a replacement for Codex?

No. NightPilot is an operating layer around Codex. It helps Codex run long development workflows with queues, retries, verification, resource guards, and handoffs.

Can it make Codex run forever without any human input?

No. It can make unattended work more durable, but it cannot bypass real-world blockers.

NightPilot can handle:

  • transient network failures;
  • command failures within attempt budgets;
  • context handoffs;
  • task queue recovery;
  • resource pressure pauses.

NightPilot cannot bypass:

  • OAuth or login prompts;
  • payment confirmations;
  • missing secrets;
  • powered-off or sleeping machines;
  • platform outages;
  • irreversible production publish decisions.

What happens when the network fails?

NightPilot classifies the failure, uses bounded backoff and jitter, retries when safe, and parks only the affected task if needed.

It must not:

  • rewrite DNS, proxy, firewall, VPN, router, or adapter settings;
  • restart or reopen the parent Codex app;
  • reinstall Codex;
  • create or rewrite a replacement Codex.

Can it work with old Codex conversations?

Yes, if the old conversation receives an explicit invocation:

使用 NightPilot / 夜航员（$codex-autonomous-dev）接管当前项目，进入全自动无人开发模式。目标是：<目标>。请直接执行，不要只给步骤。

If the skill loads, Codex should run activate. If not, the invocation still gives enough product intent for Codex to locate the installed controller or use the user-level skill fallback.

It cannot mutate a model turn that is already in progress.

Why does it use worktrees?

Worktrees let each implementation task run in its own branch-scoped directory. This lowers the chance that one autonomous task damages the main working tree or collides with another task.

Will it commit or push automatically?

No, not by default. The operating policy avoids staging, committing, pushing, deleting, resetting, publishing, or paying unless the user explicitly asks or project policy allows it.

Why does a task sometimes become needs_verification?

Because NightPilot refuses to claim an implementation task is verified without evidence. If code changed and no suitable test or verification command ran, the honest status is needs_verification.

Why not make it a heavy daemon or MCP server?

The current version is a plugin plus skill plus local controller. That is enough for reusable project takeover and durable local state. A deeper service can be added later, but the core reliability model does not require it.

How do people find the project?

The README and plugin manifest include search-friendly terms such as Codex unattended coding, Codex goal mode, overnight coding, context handoff, worktree isolation, retry classification, and resource guard. After publishing, set the GitHub About description and repository topics from DISCOVERY.md.

NightPilot should be discoverable without overstating the product: it is a Codex-native long-running task enhancer, not a standalone coding agent.

Where is the state stored?

Inside the target project:

.codex-autonomous/

This makes the workflow resumable from files instead of chat memory.

Is it safe to run in any project?

It is designed to be cautious around destructive actions. Still, autonomous coding can modify files. For important repositories, keep backups and use Git. NightPilot defaults to worktree isolation when the project is a Git repository.

What is the difference between harden-project and harden-global?

harden-project changes the current NightPilot controller configuration for this project.

harden-global changes Codex defaults for future sessions.

Neither can change a Codex execution that is already running.

activate does not run harden-global by default. Use activate --harden-global when that is intended.