Model Routing

The G-MODEL-ROUTING-AT-EXEC gate. When a task fires, this tree picks the model. Wrong tier = governance violation OR wasted budget.

What each tier handles

Opus 4.7 stays in the main conversation. It does:

• Plan authoring + amendments
• Audit verdict synthesis (mapping raw evidence → decisions)
• Multi-phase coordination + dependency sequencing
• Novel architectural decisions
• Henry interaction + decision gating
• Rule compliance gating

Sonnet 4.6 runs as a subagent dispatched via the Agent tool. It handles:

• Tool-driven mechanical execution (e.g. run a shipped tool against a target file)
• Per-file code transforms against pre-specified rubric
• Batch operations with deterministic success criteria
• Test authoring against behavior spec Opus wrote
• Grep-heavy evidence gathering
• Documentation generation from structured data

Haiku 4.5 runs as a subagent for trivial reads only:

• File-existence checks
• Glob searches
• Single-value fetches

The 8-element dispatch prompt

When Opus dispatches Sonnet, the prompt MUST include all 8 of:

1. Task goal (one sentence)
2. Evidence already gathered (file paths, line numbers, grep counts)
3. Tool to use (shipped lib/tool path)
4. Test requirements (how many, exact behavior)
5. Acceptance criteria (deterministic checks)
6. Rollback command (exact revert steps)
7. Rule constraints the subagent must self-check
8. Output format (what to report back)

After Sonnet returns, Opus runs an independent verification (grep, test count, file existence). Trusted-but-verified.

Never dispatch Sonnet for anything novel

If the pattern isn’t already proven (prior site migrated successfully, prior test passing, prior edit verified), Opus does the first instance itself to establish the pattern.