otrebu Fable Herdr Orchestration
Read ~/.agents/skills/herdr-orchestration/SKILL.md and follow all of it, including worker agent, worker permissions, topology, the dispatch loop, and completion and cleanup; it sets disable-model-invocation, so reach it by file rather than through the skill tool. If that path does not exist, try ../herdr-orchestration/SKILL.md relative to this file. If neither resolves, say that the orchestration layer is not installed and stop; do not proceed on this file alone.
This skill binds that one to Fable and Codex, exercising its "a binding skill pins one" exception: the worker CLI is pinned per role in the table below and does not track the host agent. The sections below override that skill only where they conflict; everything it says otherwise still applies, including its requirement to state the resolved agent in the first status report; state each role's model and effort there too, since this skill pins them.
Control tower
Treat the current Fable session as the control tower. Run it from a trusted directory with:
claude --model fable --effort xhigh --dangerously-skip-permissions
herdr pane current --current reports the agent kind but not the model, so read this session's own launch line with ps -o command= -p $PPID and confirm it carries --model fable. If it does not, give the command above to the user and stop. If the launch line cannot be read at all, say so and ask the user to confirm before dispatching anything.
Role binding
Bind the capability roles from herdr-orchestration to these agents:
| Role | Agent |
|---|---|
| Control tower | Fable, xhigh — the current session. |
| Default worker | Codex GPT-5.6, xhigh. |
| Specialist worker | Fable, xhigh, for architecture or synthesis that benefits from a second Fable perspective. |
| Maximum worker | Codex GPT-5.6, max for the hardest isolated implementation, debugging, or adversarial review; Fable, max for the hardest cross-cutting reasoning, recovery from stalled work, or an independent final review. |
The -m values in the launch commands below are literal model ids, not display names: "Codex GPT-5.6" is gpt-5.6-sol.
Start uncertain work at xhigh. Reserve max for work whose observed difficulty or stakes justify its cost. When escalating, choose Codex for isolated execution and Fable for cross-cutting reasoning.
Launch commands
Start workers with herdr agent start <name> --kind <kind> --pane <pane-id> -- <agent-args>. --kind selects the executable, so pass each command below without its leading executable name as the arguments after --:
# Default worker — Codex, xhigh
codex -m gpt-5.6-sol --approve-for-me -c 'model_reasoning_effort="xhigh"'
# Maximum worker — Codex, max
codex -m gpt-5.6-sol --approve-for-me -c 'model_reasoning_effort="max"'
# Specialist worker — Fable, xhigh
claude --model fable --effort xhigh --dangerously-skip-permissions
# Maximum worker — Fable, max
claude --model fable --effort max --dangerously-skip-permissions
A default Codex worker therefore starts as:
herdr agent start impl-1 --kind codex --pane <pane-id> -- -m gpt-5.6-sol --approve-for-me -c 'model_reasoning_effort="xhigh"'
For Codex, -c key=value overrides one config value for that worker without changing global config. Values are parsed as TOML, so string values retain their inner quotes.
The permission and sandbox flags in these commands come verbatim from the launch matrix at ~/.agents/skills/herdr-orchestration/references/agent-launch.md; that file, not this one, defines what they do. The model and effort flags are this binding's own.