th-1cc9fa: model-output ceiling clamp + raise starvation defaults (TypeScript parity)

[brentrager]

Why

A budget/policy max_tokens can exceed what a model can physically emit. When it does, reasoning models burn the whole budget on reasoning_content and return empty content, or the upstream 400s (e.g. groq-compound caps output at 8192). The fix clamps every request's max_tokens to min(configured, model.max_output_tokens), where the per-model ceiling comes from the LiteLLM gateway's /model/info endpoint (model_info.max_output_tokens).

What (TypeScript core)

Mirrors the already-merged Rust reference (rust/smooth-operator-core/src/llm.rs LlmClient::with_model_ceiling / effective_max_tokens; src/agent.rs AgentConfig.model_max_output, PR #69) in TS idioms:

• AgentOptions gains modelMaxOutput?: number — the active model's hard output ceiling (agent-config threading, the TS analog of AgentConfig.model_max_output).
• New exported effectiveMaxTokens(configured, ceiling?) = min(configured, ceiling), floored at 1; undefined/0/non-positive ⇒ graceful passthrough (zero behaviour change). This is the TS analog of LlmClient::effective_max_tokens.
• Both request-build sites — the non-streaming run loop and the streaming runStream loop — now send effectiveMaxTokens(...) for max_tokens.

The ceiling itself is sourced by the consumer (the server) from /model/info and passed in via modelMaxOutput, kept out of the published engine so it takes no LiteLLM-specific HTTP — same split as Rust.

Mirroring Rust

None/unset ⇒ no clamp (passthrough), identical to the Rust reference. The clamp lives at the request build (the TS engine builds the OpenAI request directly rather than through a separate LlmClient wrapper class, so the natural home for the clamp is the agent's request build + AgentOptions, rather than a withModelCeiling builder on a client object).

Tests

• effectiveMaxTokens: clamp-down (ceiling < budget), passthrough (ceiling ≥ budget, undefined, 0, negative), never returns 0.
• SmoothAgent: asserts the recorded request body's max_tokens is clamped on both the non-streaming and streaming paths, and unclamped when no ceiling is set.

pnpm typecheck, pnpm build, pnpm test all green (141 tests).

🤖 Generated with Claude Code

[changeset-bot]

🦋 Changeset detected

Latest commit: e7ddcb2

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
@smooai/smooth-operator-core	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR