| name | model-dispatch |
| description | Choose which AI model to use for a given Sounio task: type-checker changes, render examples, PGO passes, docs/skills updates, bootstrap debugging, codebase sweeps, and paper sections. |
Sounio Model Dispatch Guide
Workflow
- Identify the task type from the Decision Flowchart or Quick Reference table below.
- Assign the task to the indicated model.
- If the result is uncertain or the first attempt fails, escalate via the Fallback Chain.
- For multi-file compiler changes that touch formal semantics, enable Multi-Model Review.
Decision Flowchart
- Touches 3+ files across compiler stages? -> Opus 4.6
- Requires formal reasoning / soundness? -> Opus 4.6
- Long autonomous run or prose writing? -> GPT-5.3-Codex
- Whole-codebase sweep or impact analysis? -> Gemini 3.1 Pro
- System-level debugging or cross-check? -> GLM-5
- Render example completion (self-contained, no imports)? -> Sonnet 4.6
- Touches self-hosted frontend (check/resolve) + correctness proof? -> Opus 4.6
- PGO pass (IR module, new opcodes, self-test)? -> Opus 4.6
- Website/docs/skills update? -> Sonnet 4.6
- Native backend single-file pass (peephole/frame/encode)? -> Sonnet 4.6
- Multi-file native pipeline change (codegen+lower_ir+encode)? -> Opus 4.6
- Everything else -> Sonnet 4.6
Quick Reference
| Question | Model |
|---|
| Change the type checker | Opus 4.6 |
| Write a test | Sonnet 4.6 |
| Write a paper section | GPT-5.3-Codex |
| What will break if I change this? | Gemini 3.1 Pro |
| Why is the bootstrap failing? | GLM-5 |
| Refactor one stdlib file | Sonnet 4.6 |
| Add keyword to lexer+parser+checker+HLIR | Opus 4.6 |
| Second opinion on checker logic | GLM-5 |
| Name a new language feature | GPT-5.3-Codex |
| Find every use of X in 30 files | Gemini 3.1 Pro |
| Complete a render example (heatmap, DAG, 3D) | Sonnet 4.6 |
| Add rasterize_* primitive to stdlib/render/ | Sonnet 4.6 |
| Fix self-hosted --check failure | Opus 4.6 |
| Debug --native-compile bootstrap crash | Opus 4.6 + GLM-5 cross-check |
| Add a new PGO pass (new ir/*.sio module) | Opus 4.6 |
| Run sprint gate and triage failures | Sonnet 4.6 |
| Generate website SVG previews | Sonnet 4.6 |
| Add a new peephole pattern to peephole.sio | Sonnet 4.6 |
| Wire a new native pass in compile_ir_function | Opus 4.6 |
| Debug native binary crash / ELF reloc error | Opus 4.6 + GLM-5 cross-check |
| Add emit_* helper to encode.sio | Sonnet 4.6 |
Multi-Model Review Triggers
Enable manually on: check/epistemic.sio, check/causal.sio, hlir/ir.sio type enums, bootstrap seed updates.
Fallback Chain
Opus 4.6 -> GPT-5.3-Codex -> Gemini 3.1 Pro -> Sonnet 4.6. GLM-5 for cross-checks.
Targets
- Opus first-compile success rate: >= 90%
- Sonnet first-compile success rate: >= 80%
- Sonnet-to-Opus escalation rate: < 15%
- Paper draft revision cycles: <= 2