Herdr Peer Agents

Herdr Peer Agents

Use this skill when you need another coding agent to help with implementation, review changes, verify assumptions, or run in parallel inside Herdr.

Herdr panes are real terminals. Herdr has four levels:

• Workspace: project/task context
• Tab: a view inside a workspace
• Pane: a terminal area inside a tab
• Agent: a detected coding agent running in a pane

Agent communication is terminal-based:

1. start or identify a peer agent
2. send text to its terminal
3. submit the prompt with Enter
4. wait for the agent to become idle
5. read the peer agent's terminal output

Important: herdr agent send writes text into the target terminal. It may not submit the prompt by itself. For reliable submission, follow it with herdr pane send-keys Enter. Codex may require pressing Enter twice to submit its composer.

Use Herdr to organize parallel work instead of piling commands into the current pane. Do not steal visible focus unless the user asked for it; prefer --no-focus.

Check Herdr is available

herdr status
herdr workspace list
herdr tab list
herdr pane list
herdr agent list

Use the pane_id, workspace_id, tab_id, and terminal_id returned by Herdr for later commands. If Herdr is not running, tell the user to start Herdr first.

Optional helper script

This skill includes a small wrapper at scripts/herdr-peer.sh for common operations. The wrapper tries herdr agent start first, caches the returned pane/terminal IDs, waits for Herdr to detect the spawned process as an agent, and then renames it so later calls can use the friendly name reliably. If the agent start pane disappears or never detects, it falls back to herdr pane split + herdr pane run .

# from the skill directory, or with an absolute path
./scripts/herdr-peer.sh list
./scripts/herdr-peer.sh start reviewer --cwd "$PWD" --split right -- pi
./scripts/herdr-peer.sh ask reviewer "Review the git diff. Do not edit files."
./scripts/herdr-peer.sh close reviewer

Use raw Herdr commands below when you need finer control.

Helper panes for commands, servers, and tests

For ordinary commands, servers, and tests, use pane commands instead of agent commands. Prefer pane run over raw key injection when submitting a complete shell command:

herdr pane split  --direction right --cwd "$PWD" --no-focus
herdr pane rename  tests
herdr pane run  "just test"
herdr pane read  --source recent-unwrapped --lines 120

Use herdr wait output for normal processes and servers:

herdr wait output  --match "ready" --source recent-unwrapped --lines 120 --timeout 120000

Use agent waits only for recognized coding agents.

Spawn a peer agent

Start a named peer agent in the same project:

herdr agent start reviewer \
  --cwd "$PWD" \
  --split right \
  --no-focus \
  -- pi

Start an implementation helper below the current pane:

herdr agent start implementer \
  --cwd "$PWD" \
  --split down \
  --no-focus \
  -- pi

Use --cwd PATH to set the peer agent's working directory.

Use --workspace WORKSPACE_ID and --tab TAB_ID when you need exact placement in an existing Herdr workspace/tab:

herdr agent start reviewer \
  --cwd "$PWD" \
  --workspace w123 \
  --tab w123:1 \
  --split right \
  --no-focus \
  -- pi

Herdr pane placement supports split directions (right or down) plus workspace/tab selection. It does not require stealing user focus; prefer --no-focus.

Use clear names like reviewer, implementer, verifier, tests, server, or docs. Names make later operations auditable:

herdr agent rename  reviewer
herdr pane rename  tests

Send a prompt and wait for response

Get the target pane id:

herdr agent get reviewer

Send a prompt:

herdr agent send reviewer "Review the current git diff for correctness, safety, and missing tests. Reply with concise findings and a final APPROVE or CHANGES_REQUESTED."

Submit it with Enter using the peer's pane_id from agent get:

herdr pane send-keys  Enter

For Codex peers, send Enter twice if the prompt appears in the composer but does not run:

herdr pane send-keys  Enter
herdr pane send-keys  Enter

Wait for completion. Use idle, not done, for herdr agent wait CLI completion waits:

herdr agent wait reviewer --status idle --timeout 300000

For low-level pane waits, Herdr also supports semantic agent status waits:

herdr wait agent-status  --status idle --timeout 300000
herdr wait agent-status  --status blocked --timeout 300000

Read the response:

herdr agent read reviewer --source recent-unwrapped --lines 160

One-shot review workflow

herdr agent start reviewer --cwd "$PWD" --split right --no-focus -- pi
sleep 3
herdr agent get reviewer
herdr agent send reviewer "You are a code reviewer. Inspect the current repository changes. Focus on bugs, regressions, security issues, and missing tests. Do not edit files. Return findings as bullets plus final verdict."
herdr pane send-keys  Enter
herdr agent wait reviewer --status idle --timeout 300000
herdr agent read reviewer --source recent-unwrapped --lines 200

Implementation helper workflow

Use this when the main agent wants a peer to implement a bounded slice.

herdr agent start implementer --cwd "$PWD" --split right --no-focus -- pi
sleep 3
herdr agent get implementer
herdr agent send implementer "Help implement the smallest safe change for: . Before editing, summarize your plan. Keep changes minimal. When done, summarize files changed and tests run."
herdr pane send-keys  Enter
herdr agent wait implementer --status idle --timeout 600000
herdr agent read implementer --source recent-unwrapped --lines 240

After reading the response, inspect any file changes yourself before reporting success.

Bidirectional communication pattern

A peer can message back by using the same Herdr commands if it knows the other agent's name or pane id:

herdr agent list
herdr agent send  "I found an issue: ..."
herdr pane send-keys  Enter

For clean collaboration, tell each spawned agent:

• its role (reviewer, implementer, verifier)
• whether it may edit files
• how to report completion
• whether it should message another Herdr agent
• an explicit end state to avoid loops

Safety rules

• Prefer --no-focus so the user's active pane is not stolen.
• Use short, bounded prompts with an explicit stop condition.
• For review agents, say Do not edit files.
• For implementation agents, say exactly what files/scope are allowed when possible.
• Always read the peer output before acting on it.
• Close temporary peer panes when finished:

herdr pane close 

Troubleshooting

If herdr agent start -- ... returns agent_status: unknown and herdr agent get fails immediately, wait for detection and/or use the returned pane_id/terminal_id. The helper script does this automatically: it caches both IDs, polls herdr pane get , and renames the detected terminal back to the requested friendly name. If that pane disappears or never detects, the helper falls back to the manual start pattern: split a pane, run the agent command, then rename after detection.

Manual recovery:

# If the pane exists and the agent is detected
herdr pane get 
herdr agent rename  
herdr agent get 

# If agent start failed/disappeared
herdr pane split  --direction right --cwd "$PWD" --no-focus
herdr pane rename  
herdr pane run  "pi"
herdr agent rename  

If a prompt is visible in the target composer but does not submit, send Enter again:

herdr pane send-keys  Enter
herdr pane send-keys  Enter

If herdr agent wait --status done fails, use:

herdr agent wait  --status idle

If sending text appears in the peer pane but nothing happens, submit Enter explicitly:

herdr pane send-keys  Enter

If the agent target name is ambiguous, use its terminal_id or pane_id from:

herdr agent list

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: msadig
• Source: msadig/herdr-peer-agents-skill
• License: MIT

Install and usage instructions live in the source repository linked above.