Upgrading to 0.11.0

0.11.0 adds a verbatim vendor-parameter passthrough and removes the model-name suffix convention it replaces. This page is the whole migration.

Breaking: the -low / -medium / -high model-name suffix is gone

If any agent's model ends in -low, -medium or -high, its behaviour changes on upgrade. Everyone else is unaffected.

What changed, and why

Before 0.11.0, reasoning effort for Anthropic and Gemini was selected by decorating the model name. The provider parsed it back off with rsplit("-", 1):

AgentConfig(name="advisor", model="claude-sonnet-5-high")
# provider sent: model="claude-sonnet-5", output_config={"effort": "high"}

That convention had three defects, and the first is the one that forced its removal:

1. It silently mangled legitimate model names. rsplit("-", 1) cannot tell an effort suffix from the end of a real model name. Any model genuinely called something ending in -high was truncated, and the tail was misread as an effort level — with nothing in the request or the response to say so. As vendors keep shipping names, that becomes more likely, not less.
2. It was invisible. Nothing in the API surface said the model string was being parsed, so a caller had no way to know it had happened.
3. It did not exist on openai_compat. The same suffixed model meant reasoning effort on one provider and a mangled model name on another, so identical configuration produced different requests depending on who served them.

Migrating

Restore the model name to its real value and pass the effort through params.

AnthropicGeminiFront matter

# before
AgentConfig(name="advisor", model="claude-sonnet-5-high")

# after
AgentConfig(
    name="advisor",
    model="claude-sonnet-5",
    params={"thinking": {"type": "adaptive"}, "output_config": {"effort": "high"}},
)

# before
AgentConfig(name="advisor", model="gemini-3-pro-high")

# after
AgentConfig(
    name="advisor",
    model="gemini-3-pro",
    params={"thinking_config": {"thinking_level": "high"}},
)

---
name: advisor
model: claude-sonnet-5
params:
  thinking: {type: adaptive}
  output_config: {effort: high}
---

Anthropic's rule that temperature must be absent whenever thinking is active is preserved. It now keys off params containing thinking rather than off a parsed model name, so enabling adaptive thinking still suppresses temperature automatically.

How to tell whether you are affected

Search your agent configuration for a model ending in one of the three suffixes:

grep -rn "model:.*-\(low\|medium\|high\)$" path/to/your/agents/

• No matches — nothing to do. params is purely additive for you.
• Matches — apply the migration above. If you leave them, the model name is now sent to the vendor verbatim, which is almost certainly not a real model, and the call will fail rather than silently do the wrong thing. That noisy failure is deliberate.

What you gain

params is forwarded verbatim to the vendor SDK call, so anything a vendor supports is reachable without waiting for AgentFlow to name it — including reasoning_effort on openai_compat, which had no effort mechanism at all before 0.11.0.

await provider.chat(messages, params={"reasoning_effort": "high"})

Two rules apply, both covered in LLM Providers:

• A parameter a model ignores is your business, not an AgentFlow error.
• A parameter that would override AgentFlow's own arguments — model, messages, system, tools — is refused with a ValueError naming every offending key. params changes how a call is made, never what is asked.