gnhf (Good Night, Have Fun)

What does this agent do, and when should you use it?

gnhf is a command-line orchestrator that turns an existing coding agent CLI into a continuous overnight loop. On start it verifies a clean Git working tree, creates a gnhf/ branch (or stays put with --current-branch), writes the objective to prompt.md, then repeatedly builds an iteration prompt, invokes the agent non-interactively, and decides what to do with the result. Each successful iteration becomes its own unsigned git commit with a summary appended to notes.md; failed iterations are rolled back with git reset --hard, and three consecutive failures abort the run. Run metadata — prompt, notes, stop condition, end-state.json and a JSONL debug log — lives under the local, git-ignored .gnhf/runs/ directory so the branch only holds intentional work. It is built for unattended use, with iteration and token caps, sleep prevention, worktree parallelism, usage-limit waiting and a permanent exit summary.

Running gnhf "<prompt>" inside a Git repository checks that the working tree is clean, writes prompt.md and creates or reuses a gnhf/ branch. In every loop it assembles an iteration prompt from the objective, the accumulated notes.md context and a final output JSON schema, then runs the chosen agent non-interactively — claude, codex exec, Copilot in JSONL mode, pi in JSON mode, cursor-agent in --print stream-json mode, a locally started opencode serve or acli rovodev serve process, or an ACP target through the bundled acpx. On a structured success it makes one unsigned commit and appends a note; on failure it resets the tree, counts the failure, backs off exponentially for retryable errors, and aborts immediately on permanent errors such as exhausted credit while printing the run log path. When a Claude usage window is exhausted it reads the reset time from the rate-limit event, waits, and retries the same iteration, bounded by --max-rate-limit-wait. Runtime caps are --max-iterations, --max-tokens and --stop-when; the interactive TUI keeps the terminal title updated with elapsed time, token totals and commit count, and Ctrl+C requests a graceful stop before a second press force-stops. At exit it prints a permanent summary with branch, elapsed time, iteration and token totals, branch diff stats, notes and log paths and review commands, while a JSONL debug log is written to .gnhf/runs/<runId>/gnhf.log.

1. Developers who want an agent to push a long objective — reducing complexity, adding tests — overnight and review a ready branch in the morning.
2. Teams running several independent tasks in one repository at once using --worktree for isolated branches and directories.
3. Loose projects where --current-branch --push keeps a deployed or locally watched branch updated during the run.
4. Users with an existing Claude Code, Codex or Copilot subscription who want to put idle quota to work.
5. Teams reaching non-native agents (Gemini, Grok, custom ACP servers) while keeping one unified log and rollback semantics.
6. Cost-conscious users who want hard budgets (iterations, tokens, rate-limit wait) in place before letting go.

How do you install or deploy this agent?

Requires Node.js, Git and a signed-in coding agent CLI; run it inside a Git repository.

npm install -g gnhf

From source:

git clone https://github.com/kunchenguid/gnhf.git
cd gnhf
corepack enable
pnpm install
pnpm run build
pnpm link --global

The first run creates ~/.gnhf/config.yml with defaults (agent: claude); edit as needed:

agent: claude

maxConsecutiveFailures: 3

preventSleep: true

How do you use this agent?

From a clean Git repository, hand it a single-sentence objective:

$ gnhf "reduce complexity of the codebase without changing functionality"

Add an iteration and token budget for a shorter nap:

$ gnhf "reduce complexity of the codebase without changing functionality" \
    --max-iterations 10 \
    --max-tokens 5000000

Run several tasks in parallel:

$ gnhf --worktree "implement feature X" &
$ gnhf --worktree "add tests for module Y" &

Commit and push on the current branch after each successful iteration:

$ gnhf --current-branch --push "keep improving this app"

Switch agent or model:

$ gnhf --agent codex --model gpt-5.4 "fix failing tests"
$ gnhf --agent 'acp:gemini' "polish the docs"

Disable anonymous telemetry:

export GNHF_TELEMETRY=0

What are this agent's strengths and limitations?

How does this agent compare with similar options?

Key facts side by side with the most closely related agents.

How does FollowAgents rate this agent?

FAQ