Gnhf

GNHF

Overview

GNHF is an agent orchestrator: it repeatedly calls another coding agent until a natural-language stop condition is met. This skill teaches the host agent to prepare one durable run and, in Companion mode, steer or review it.

Core rule: the host agent orchestrates; GNHF executes. Do not manually implement inside the same scope while a GNHF worker is responsible for it unless the user explicitly changes the delegation.

In Companion mode, GNHF completion is not user acceptance. "Stop condition met" only means the worker stopped; the host still compares the result to the user's latest requirements and fresh verification.

Modes

Choose exactly one mode for the run.

Hands-Off

Use when the task is bounded, verification is clear, and the user wants one configured run to proceed without steering.

• Prepare a precise prompt with constraints, non-goals, verification, and stop condition.
• Launch GNHF and wait for completion.
• Intervene early only for hard failure, runaway scope, destructive behavior, or impossible prerequisites.
• Report the final GNHF status after exit.

Examples:

• English: "I'm going to bed. Use GNHF with Copilot to keep working on this branch and stop when the test suite passes."
• Chinese: "我要睡了。用 GNHF 接着跑这个分支，测试都过了就停。"

Companion

Use when the task is uncertain, exploratory, design-heavy, research-heavy, or likely to need course correction.

Default to Companion when the user asks to iterate until satisfied, requests multi-round work, provides review findings, asks for design/skill/documentation improvement, or asks for supervision.

• Keep a note of original intent, branch, session id, and last known result.
• Poll the active GNHF process until exit or until an intervention point appears.
• Intervene when the worker optimizes the wrong thing, repeats failed fixes, skips requested research, drifts scope, or claims success without evidence.
• Treat review findings as the next acceptance criteria.
• Prefer a new bounded GNHF prompt over manually taking over implementation.

Examples:

• English: "Run GNHF for a few rounds on this onboarding flow. Check the diff between rounds and tighten the next prompt if it starts polishing the wrong thing."
• Chinese: "用 GNHF 多跑几轮这个 onboarding 流程。每轮看一下 diff，如果它开始改偏了，就收窄下一轮 prompt。"

Launch

Check the installed CLI before relying on flags:

gnhf --help

Known shape:

gnhf \
  --agent <agent> \
  --max-iterations <n> \
  --stop-when "<observable completion condition>" \
  --prevent-sleep on \
  "<worker prompt>"

Use --model <model> to select a model for a run. For config defaults and agent-specific restrictions, follow the README Configuration.

Before launch:

git status --short
git branch --show-current
git log --oneline --max-count=5

Prompt skeleton:

Objective: <one concrete outcome>.

Use <agent/model requirement>. Work in this repo. Treat this as a long-running GNHF task.

Before coding, inspect the current repo, relevant docs, and recent commits. Preserve user changes. Do not make unrelated refactors.

After each meaningful slice, run relevant verification. If blocked, commit no fake success; leave notes with the blocker and evidence.

Stop only when: <observable completion condition>.

Steer

In Companion mode, evaluate after each iteration or meaningful output chunk:

Steering prompt:

Continue from the current repo state. The previous run partially succeeded: <evidence>.

Do not redo completed work. Focus only on <bounded correction>.

The issue to fix now is <specific observed issue>. Verify with <commands/checks>.

Stop only when <observable condition>.

Companion Review

Use only in Companion mode, when the host is supervising quality or deciding whether to continue with another bounded run.

1. Inspect branch, status, commits, changed files, and diff.
2. Read GNHF notes/logs as claims, not evidence.
3. Run independent verification: tests, lint, build, typecheck, manual QA, or domain-specific checks.
4. Compare the result to the stop condition and the user's latest feedback.
5. Decide: Mergeable, Needs follow-up GNHF run, or Do not merge.

If the result needs follow-up, continue in Companion mode instead of presenting the run as complete. Do not merge unless explicitly authorized.