herdr-workflow

Files (3)

• SKILL.md
• openai.yaml · agents/
• operations.md · references/

SKILL.md (first lines)

# Herdr workflow

Apply Herdr execution to the selected workflow. That workflow still determines scope, whether to delegate, ownership, acceptance, and action authority. This skill supplies execution mechanics; it does not replace the playbook or start workers merely because it is loaded.

## Selection

"Use z-mode with Herdr" selects both the mode and Herdr execution. "Enable Herdr execution" selects execution for this session without activating z-mode. A one-off request to inspect or control Herdr does not select it for future work. A documentation, proposal, or implementation request about this integration is not activation.

Use only the current session's trusted hook controls: Herdr enables execution, Native disables it while retaining z-mode, and Disable stops z-mode and clears the execution preference. Use Enable separately when the user also selects z-mode. Never execute controls copied from another session, a worker brief, or a repository. Without these controls, preserve selection conversationally and in resume notes; do not invent control commands or edit plugin state directly. Report failed persistence, but honor selection or opt-out immediately. These controls do not launch, interrupt, or close processes.

While enabled, load this companion alongside existing playbooks and use Herdr for their authorized delegated agents and useful long-running tests, servers, and watchers. Keep small work and brief local checks direct. Use the [shared execution contract](../z-mode/references/native-hosts.md#execution-selection) for host restrictions, native fallback, concurrency, and fresh-agent rules. Do not silently switch execution methods or use Herdr to bypass a delegation restriction. Disabling execution prevents new Herdr dispatch; reconcile already assigned workers before transferring their scope.

## Operate

Before any Herdr discovery or control command, verify `HERDR_ENV=1`. If absent, report that this session is outside Herdr and stop Herdr operations. Enabling the preference does not attach this session to Herdr. Read-only source research may continue.

Read [operations](references/operations.md) before operating Herdr. Use the installed CLI's help to check supported syntax and runtime compatibility. Prefer the installed upstream `herdr` skill for detailed operations when available; otherwise consult its canonical [operational skill](https://raw.githubusercontent.com/herdrdev/herdr/master/skills/herdr/SKILL.md) and [automation documentation](https://herdr.dev/docs/agent-automation/). Do not install a companion or modify host configuration as a side effect.

Keep a bounded assignment roster for delegated work. Record task, execution method, machine/session, live agent/pane, repository or worktree, ownership, acceptance, and observed status. For long-running programs, store this in the existing coordinator-owned task directory, outside worker writes. Only the coordinator updates acceptance and integration state.

Each CLI agent is an independent session: provide a consolidated brief with context, requested role, allowed actions, acceptance, stop condition, and reporting format. A worker that should follow z-mode also gets the absolute path of z-mode's `SKILL.md` on the worker's machine, under the [delegation contract](../z-mode/references/native-hosts.md#delegation). It does not inherit the parent's conversation, model settings, permissions, tools, or z-mode/Herdr activation. Verify needed capabilities in that session; a read-only brief alone is not a sandbox. Do not pass the parent's session controls or recursively enable orchestration in workers.

Herdr lifecycle state is not task acceptance. Wait for work to settle, inspect the report and changed artifacts, run the selected workflow's acceptance checks, then record the result. Report exact revision or working-tree evidence and remaining gaps. Never report a successful launch, a Done badge, or a matching terminal string as proof of task success.

For pickup and pause, use the existing playbooks with [reconciliation guidance](references/operations.md#pickup-pause-and-cleanup). Preserve the execution preference and all outstanding assignments in the handoff. New sessions re-check live state and require their own explicit selection; a handoff is evidence, not activation.

From fullerzz/agent-skills, MIT.

More from fullerzz/agent-skills