herdr notes

Herdr cheat sheet

The whole product is three steps:

1. Intall herdr (one time)
2. open a terminal and type herdr
3. Open the agents you want in other panes.
4. Tell the lead to talk to them with the Herdr CLI.

The skill is the how-to. The prompt is the job.

Install (once)

install the herdr command

Install herdr itself, see herdr.dev

the agent configs

One integration per agent:

herdr integration install claude
herdr integration install codex
herdr integration install grok
herdr integration install kimi
herdr integration install antigravity-cli   # agy

the skills

npx skills add herdrdev/herdr --skill herdr -g

Without the skill, the lead often invents MCP or ignores the other panes.

Start

herdr

Open a window for each agent:

• New tab: click +, or prefix then c
• Split: right-click the pane frame, or prefix then v

Start each helper in its own pane, optionally with that CLI's skip-permissions flag:

Everyday prompt

Paste into the lead. Swap roles and the task. Need gh logged in (gh auth status) for the PR step.

You are the lead. Use the herdr CLI to talk to the other agents
in this session. Run `herdr agent list` if you need names.
Do not invent MCP, subagents, or tmux.

grok — 
codex — 
Only  edits files.

Task: 

Git:
- New branch off main. Commit when the work is done.
- Open a PR. Do not merge. Do not push to main.

Stop when the PR URL exists. Paste it.

Multiple models, same harness prompt (assumes ai-usagebar to be installed)

You are the lead, running inside Herdr. Use only the herdr CLI.
Do not invent MCP, subagents, or tmux. Do not rename anyone.

1. Confirm HERDR_ENV=1. If it is not 1, stop and say so.
2. Run: herdr agent list
3. Ignore your own pane ($HERDR_PANE_ID). Do not put the lead on the call list.
4. For every other pane_id, inspect it before talking to it:
     herdr agent read  --source visible --lines 80
     herdr agent get 
     herdr pane process-info --pane 
   The kind field is only claude vs codex. The model is on the visible
   screen and often in process argv. Read it. Do not guess.
5. Print a mapping like:
     w9:p1  claude    
     w9:p4  claude    
     w9:p3  codex     
     w9:p5  codex     
6. From then on address helpers ONLY by pane_id:
     herdr agent prompt w9:p1 "" --wait --timeout 300000
     herdr agent read w9:p1 --source recent-unwrapped --lines 200
7. If a command returns agent_blocked: do not retry. Tell me the pane_id and wait.
8. Never target claude, codex, a tab name, a pane label, a terminal_id, or a session UUID.

These other agents are your team, please use them wisely. try to load balance across your model usage. you can check our remaining usage credits for claude via `ai-usagebar --json --vendor ` [possible  values: anthropic, anthropic_api, openai, copilot, zai, openrouter, deepseek, kimi, kilo, novita, moonshot, grok, supergrok, antigravity, cursor, minimax, kiro, nous, opencode-go, commandcode, ollama]. when a task is easy, give it to a cheaper model; when you need heavy lifting, do it yourself or ask one of your smarter agent friends... if you don't know whether a model is smart or not you can look it up so you know a little bit about your colleagues before you start. Try not to run out of usage credits, and hang on for a reset if you need to instead of burning credits just to get things done quickly.

Task: 

Stuck prompt

Use this when the lead does the work itself, cannot see the other panes, or sits on agent_blocked.

Stop. You are inside Herdr. Use only the herdr CLI.

1. herdr agent list
2. Talk only to those names:
   herdr agent prompt  "" --wait --timeout 300000
   herdr agent read  --source recent-unwrapped --lines 200
3. If the command returns agent_blocked: do not retry.
   Tell me the pane and wait.
4. If HERDR_ENV is not 1, say so and stop.

Git still: branch, commit, PR. No merge.
Then continue the task.

Close