Thinking level

Avenor exposes one optional thinking setting for backend-native reasoning controls. Use it when a run needs a specific reasoning effort and the selected backend can apply that setting without changing the prompt.

avenor --backend pi --thinking high --prompt "Review this change"
avenor --backend codex-app-server --thinking xhigh --prompt "Implement the plan"
avenor --backend claude --thinking medium --prompt "Summarize the design"

The control is also available in stable spawn JSON, avenor control spawn, the Go and TypeScript MCP tools, and the core, Pi, and OpenCode TypeScript spawn tools.

Values

thinking accepts these exact lowercase values:

• off
• minimal
• low
• medium
• high
• xhigh
• max

Omit the setting to use the backend default. Avenor passes the same explicit spawn-level value to retries, loop and team phases, queued prompts, and follow-up runs. Avenor does not normalize case, clamp values, or add thinking instructions to prompts. An invalid value returns an error that lists the accepted values.

Backend support

The native mechanisms below were verified with Codex CLI 0.144.6, Pi 0.83.0, and Claude Code 2.1.220. A required feature check that does not match returns the same unsupported-parameter error instead of guessing.

Backend	New session	Explicit resume	Native control
codex-app-server	all seven values	all seven values	turn/start.effort
pi	all seven values	all seven values	fresh client: --thinking; reused client: set_thinking_level
claude	low, medium, high, xhigh, max	unsupported	--effort at startup
claude-channel	low, medium, high, xhigh, max	unsupported	--effort at startup
opencode-acp	unsupported	unsupported	none verified
opencode-http	unsupported	unsupported	none verified
gemini-acp	unsupported	unsupported	none verified
cursor-acp	unsupported	unsupported	none verified
agy	unsupported	unsupported	none guaranteed across its transports
pony	unsupported	unsupported	none

Every explicit value on an unsupported path fails with an error such as:

backend "agy" does not support parameter "thinking"

The rejection occurs before Avenor starts a provider, reserves a stable runtime, creates session artifacts, or sends a prompt. The MCP and control clients validate the enum for early feedback. The stable supervisor and direct CLI apply the authoritative backend policy.

Native behavior

Codex app-server

Avenor stores an explicit effort as thread-scoped outbound state and sends the exact value on the first and each later turn/start request. It does not add reasoning_effort to thread/start, query model/list as a gate, require a model, or clamp the value.

The app-server validates effort against the active thread model on the actual turn. If Codex rejects it, Avenor adds the backend and requested value while retaining the app-server error text.

An explicit resume replaces Avenor's outbound override. A resume with omitted thinking clears only that override and omits effort; it does not reset the effort retained by Codex.

Pi

For each fresh explicit start or resume, Avenor first checks that pi --help advertises --thinking <level>. It then starts Pi with --thinking <level>. It does not send a setter to a fresh client.

For a reused client, Avenor sends set_thinking_level on that client and requires a successful response before state lookup, session registration, or the next prompt. Avenor does not start an extra Pi process to probe the setter. Omitted thinking sends neither the flag nor the setter.

Pi owns model-specific mapping and clamping. The seven Avenor controls do not promise identical effective behavior for every Pi model.

Claude and Claude Channel

Both Claude backends check that claude --help advertises --effort <level> before an explicit start. They pass --effort for low, medium, high, xhigh, and max. Claude has no native startup value for off or minimal, so Avenor rejects those values.

The effort flag is startup-only. Both backends reject every explicit resume value, even if the live in-process session appears to use the same value. Omitted thinking keeps the existing in-memory resume behavior.

Unsupported backends

OpenCode ACP 1.18.3 has no verified thinking config option. OpenCode HTTP SDK 1.15.3 exposes model variants, but a variant name is not a canonical thinking level. Avenor does not infer a mapping or add an unverified payload field.

Gemini ACP and Cursor ACP have no verified session control for this setting. Pony's model request has no thinking field. Agy 1.1.8 advertises --effort for its headless transport, but the Avenor backend can select headless, RPC, or auto transport. This release therefore rejects thinking for Agy as a whole; transport-specific support can be added only after a separate contract guarantees it.

None of these backends receives a prompt prefix, system instruction, warning-only fallback, or silent no-op.