Multi-Adapter Model Routing

Route different LLM models to different pipeline steps for cost optimization and provider resilience.

Model Tier System

Wave classifies pipeline steps into three complexity tiers based on persona and step characteristics:

Tier	Intent	Use Case
cheapest	Cost-optimized	Navigation, summarization, scanning
balanced	Latency-optimized	Balanced speed/cost for standard tasks
strongest	Capability-optimized	Complex reasoning, code generation

Adapter Tier Models

Each adapter can define tier-specific model mappings:

adapters:
  claude:
    binary: claude
    default_model: sonnet
    tier_models:
      cheapest: haiku
      balanced: ""
      strongest: opus
  opencode:
    binary: opencode
    default_model: opencode/big-pickle
    tier_models:
      cheapest: opencode/big-pickle
      balanced: opencode/big-pickle
      strongest: opencode/big-pickle

When auto_route: true is enabled, Wave uses these tier mappings to select models automatically.

Auto-Routing

Enable automatic model selection based on step complexity:

routing:
  auto_route: true
  complexity_map:
    cheapest: haiku
    balanced: ""
    strongest: opus

Override specific tiers in wave.yaml routing section, or use adapter-level tier_models for per-adapter mappings.

Per-Step Model Assignment

steps:
  - id: analyze
    persona: navigator
    model: haiku          # explicit override

  - id: implement
    persona: craftsman
    # no model — uses adapter tier model based on complexity

  - id: review
    persona: reviewer
    model: haiku          # cheap model for review

Adapter Resolution (strongest to weakest)

Priority	Source	Scope	Example
1	CLI --adapter flag	Entire run	wave run impl-issue --adapter opencode
2	Step-level adapter:	Single step	adapter: codex in pipeline YAML
3	Persona-level adapter:	Steps using that persona	adapter: claude in persona definition
4	Adapter default	Steps with no override	Falls back to first configured adapter

adapters:
  claude:
    binary: claude
    mode: headless
  opencode:
    binary: opencode
    mode: headless
  codex:
    binary: codex
    mode: headless
  gemini:
    binary: gemini
    mode: headless

CLI Adapter Override

Use --adapter <name> to override the adapter for all steps in a run:

wave run ops-hello-world --adapter opencode --model "zai-coding-plan/glm-5-turbo"

Override per step in pipeline YAML:

- id: implement
  adapter: codex              # use OpenAI Codex for this step
  model: gpt-4o

Model Format Differences

Each adapter uses a different model identifier format:

Adapter	Format	Examples
claude	Short names	sonnet, haiku, opus
opencode	provider/model	zai-coding-plan/glm-5-turbo, anthropic/claude-sonnet-4-20250514
gemini	Plain names	gemini-2.0-pro, gemini-2.5-flash
codex	OpenAI identifiers	gpt-4o, o3

Fallback Chains

When a provider fails, try alternatives:

runtime:
  fallbacks:
    anthropic: [openai, gemini]
    openai: [anthropic]

Fallback only triggers on transient failures (rate limits, timeouts). Permanent failures (auth errors, missing binary) do not fallback.

Persona-Level Defaults

Personas can declare a preferred model, overridable at step level:

# .wave/personas/analyst.yaml
name: analyst
model: haiku
adapter: claude