Cluide — Claude Entry Point

@AGENTS.md

The imported AGENTS.md is the canonical shared repository policy. If this interface does not resolve imports, explicitly read the root AGENTS.md before working. Do not duplicate shared rules here.

Claude-only

Dispatch Overrides

Claude routing exceptions for work delegated from sessions in this repository — subagents, workflow stages, scheduled tasks. The dispatch skill (skills/dispatch/SKILL.md) applies wherever this section is silent.

• Before executing any task with bulk, parallel or mechanical parts, load the dispatch skill — even when another skill (a maintenance playbook, a review protocol) has already supplied the procedure. Playbooks say what to do; dispatch says which tier does it. Under the inline floor, doing it inline is correct; log it as inline.
• Default worker tier: sonnet, replacing the table's haiku rows — only the known-safe list below runs on haiku. Pass model on every spawn of an unpinned agent type; an omitted one inherits the session's tier.
• Never below opus for: writing or reviewing guide prose that ships, and review verdicts on guides, skills or templates — publication quality is this repository's whole product.
• Known-safe on haiku: link checks, byte-identity mirror sweeps (diff -q), guide-number citation greps, file inventories — anything the merge gate defines mechanically.
• The default tier and the known-safe list are cost rules: raise the tier when a task is harder than its archetype. The never-below line is a tier floor and is not traded against cost.
• Log dispatches to development/ROUTING_LOG.md (gitignored working state: this section is the published example, the log is not).

For conversational Claude setup, see PLATFORM_SETUP.md.