Handoff That to Agent Pane

Treat the handoff as a brief: give the specialist the relevant context, a bounded assignment, and a checkable deliverable.

Take the snapshot first: load and follow the tmp-snapshot skill before any other step in this workflow, preflight included. Load and follow the herdr skill before issuing any Herdr command.

Resolve the brief

Determine these inputs from the request:

• Agent: claude, codex, agent (Cursor Agent), or pi. Default to claude when the user omits it. Take the executable's Herdr --kind from the launch matrix: read ~/.agents/skills/herdr-orchestration/references/agent-launch.md. When that file is not installed, take the kind from herdr agent's installed kind list instead — the kinds are named for agents, and the executable name agent is not among them.
• Model: normalize natural-language model and effort names before choosing the launch command; otherwise use the table below.
• Context slice: include only the part of the conversation and current work the specialist needs.
• Assignment: state the work to perform, its boundaries, whether edits are allowed, the deliverable, and how to verify it.
• Delegation: carry any requested subagent strategy into the assignment, including requests to use as many subagents as useful.
• Return path: decide whether the user only wants the pane started or wants this session to collect and integrate the result.

Ask once before mutating anything only when the assignment cannot be inferred safely.

Model defaults

Each command above is complete and verbatim: it already embeds its agent's autonomous posture, so dispatch needs no extra permission or sandbox flags. The launch matrix (~/.agents/skills/herdr-orchestration/references/agent-launch.md) is the canonical explanation of those posture flags — including why the Codex command needs no separate --sandbox or --ask-for-approval flag and why pi's --approve is a trust flag, not a permission flag. Consult it instead of re-deriving flags; when it is not installed, the commands above still stand as written.

Treat GPT 5.6, GPT 5.6 Sol, and Codex 5.6 as the model ID gpt-5.6-sol. Treat ultra, maximum, and max as max reasoning effort, never as part of a model ID. Thus “GPT 5.6 ultra” means:

codex -m gpt-5.6-sol --approve-for-me -c 'model_reasoning_effort="max"'

Treat xhigh and extra high as xhigh reasoning. Use another literal Codex model ID only when the user supplies the exact ID explicitly. Never manufacture an ID from prose.

For Cursor Agent, use this exact command when the user asks for Grok 4.6 Extra High Fast or does not override the agent default:

agent --model cursor-grok-4.6-xhigh-fast --yolo --sandbox disabled

Grok 4.6 exposes low, medium, high, and xhigh, each with a -fast twin; match on the ID rather than the label, since cursor-grok-4.6-high is labelled “Cursor Grok 4.6” and cursor-grok-4.6-high-fast is labelled “Cursor Grok 4.6 Fast”. Grok 4.5 exposes low, medium, and high, each with a -fast twin, and has no xhigh. Treat xhigh and “extra high” as the literal model ID cursor-grok-4.6-xhigh-fast when “fast” is requested, and cursor-grok-4.6-xhigh when it is not. Do not infer any other Cursor model ID from prose.

When the chosen agent is agent (Cursor Agent) or pi, confirm before splitting that the selected literal model ID is listed. Run only the check for the agent you are dispatching, and skip this block entirely for claude and codex, which expose no local model list:

agent --list-models
pi --version
pi --list-models k3

pi --list-models k3 currently surfaces kimi-coding/k3 (1.0M context), kimi-coding/k3-256k (262.1K context), and openai-codex/gpt-5.3-codex-spark. Our pi default is kimi-coding/k3-256k.

For the Cursor default, require the exact cursor-grok-4.6-xhigh-fast entry. If it is absent, report the failed precondition; never silently fall back to the non-fast variant, a lower Grok 4.6 tier, or Grok 4.5. For pi's default, require version 0.80.10 or later, configured Kimi Code access, and the exact kimi-coding/k3-256k entry. If k3-256k is absent, report the failed precondition and never silently use k3, K2.7, or another listed model. If any requested default is unavailable, stop before creating the pane and substitute only with the user's approval.

Keep the launch interactive and preserve the current working directory. Run each agent in the autonomous posture its table command already carries, per the launch matrix. Honor an explicit permission override from the user.

Preflight with local, non-inference commands only: command -v, --version, --help, or the model-list commands above. Never run codex exec, a print-mode agent turn, or any paid probe to test a model. Do not inspect personal config to replace defaults pinned by this skill.

Dispatch

1. Invoke tmp-snapshot to save the relevant context slice as a standalone handoff in /tmp. Include current state, relevant decisions, concrete paths or artifacts, constraints, and unresolved questions. Treat its returned path and outline as a sub-step result and continue this workflow.
2. Confirm the snapshot path is readable and contains enough context to perform the assignment without the chat.
3. Follow herdr to verify HERDR_ENV=1, learn the installed CLI, and preflight the chosen executable and model. Complete this step only with a valid interactive launch command.
4. Split the calling pane to the right without moving focus:

``bash herdr pane split --current --direction right --cwd "$PWD" --no-focus ``

1. Read the new pane ID from the JSON response. Rename the pane for the assignment, then start the agent by name and kind, passing the launch command's arguments after --:

``bash herdr pane rename <returned-pane-id> "<short-label>" herdr agent start <name> --kind claude --pane <returned-pane-id> -- --model fable --effort xhigh --dangerously-skip-permissions ``

--kind selects the executable, so everything after -- is the remainder of the chosen command from the table above. agent start returns once Herdr detects the agent ready for input. If it returns agentnotready, the agent is blocked at a startup dialog such as a trust or permissions prompt: read it with herdr agent read <name> --source visible --format ansi, surface it to the user, and prompt only once the agent is idle. Use a few words for the pane label; the full brief goes in the prompt template below, not in the label. Agent names must be unique among all live agents, so derive a short kebab-case name from the assignment and check herdr agent list first; if agent start returns agentnametaken, choose another name.

1. Send one task prompt with herdr agent prompt <name> "<text>". Use this shape:

```text Read <absolute-snapshot-path>. It is the handoff context for this assignment.

Assignment: <specific work> Scope: <boundaries and whether edits are allowed> Delegation: <requested subagent strategy, or "decide as needed"> Deliverable: <review, second opinion, implementation, investigation, or side-task result> Verification: <evidence or checks required>

Treat this prompt as authoritative if it conflicts with the snapshot. Surface any blocker instead of expanding scope. ```

1. Confirm the agent moved to working. If it did not, inspect its status and transcript before reporting the failure.

Return

When the user only asked to open or hand off, finish after dispatch starts. Report the pane ID, agent and model, snapshot path, and assignment in one compact status.

When the user asked for the result, wait with herdr agent wait <name>, whose defaults settle on idle, done, or blocked, handle blockers through the user, read the transcript, and return or integrate the requested deliverable.

Finish only when the correct right-hand pane exists, the requested model is running, the snapshot-backed brief has been submitted, and the return path requested by the user is complete.