fix(providers): send maxOutputTokens under the name the SDK reads

[will-lamerton]

The bug

The AI SDK renamed this setting to maxOutputTokens in v5. We're on ai@6.0.193 and still sending maxTokens (chat-handler.ts). The v6 typings contain 8 references to maxOutputTokens and zero maxTokens: declarations, so the value was being dropped on the floor.

It typechecked because the call site spreads it:

...(modeOverrides?.modelParameters && {
  maxTokens: modeOverrides.modelParameters.maxTokens,  // silently ignored
}),

Object spreads bypass excess-property checking. No compile error, no runtime error, no log - the ceiling just quietly became whatever the provider inferred from the model id.

/tune's max-tokens control has therefore been inert for every provider and every user since the v6 upgrade.

Why it's worse than it sounds

@ai-sdk/anthropic derives the ceiling from the model id, with this fallback (dist/index.mjs:5222):

} else {
  return { maxOutputTokens: 4096, ..., isKnownModel: false };
}

Point sdkProvider: "anthropic" at an Anthropic-compatible endpoint serving a non-Claude model and every reply is capped at 4096 tokens, truncated mid-sentence with no error.

That is what has been breaking nc-review on Nano-Collective/prompt-scrubber: reviews died mid-analysis without ever writing their verdict. nanocoder#1332 made the truncation visible and recoverable; this removes the cause.

Changes

• Send the setting as maxOutputTokens, restoring /tune.
• Add maxOutputTokens to provider entries in agents.config.json. Headless runs build modeOverrides without modelParameters, so /tune is not a route CI can use - the provider entry is. /tune still wins where set.
• The user-facing name stays maxTokens in ModelParameters; it's persisted in tune preferences, so renaming would silently discard existing user config. The SDK name is applied at the boundary.

Behaviour change worth flagging

Fixing this means tune values start applying. The Nano (low-end hardware) preset sets maxTokens: 2048, which until now did nothing and will now actually cap output. That is the documented intent of the preset, but anyone on it will notice shorter replies.

Tests

Three in chat-handler.spec.ts, asserting the request body over an intercepted fetch rather than trusting types - typechecking demonstrably cannot catch this class of bug:

• an unrecognised model id yields max_tokens: 4096 (pins the default that makes this necessary)
• maxOutputTokens: 32000 reaches the wire as max_tokens: 32000
• maxTokens: 32000 does not - a guard that fails loudly if the SDK ever renames back

Plus maxOutputTokens plumbing asserted in client-factory.spec.ts.

chat-handler.spec.ts (29), client-factory.spec.ts (43), config-schema.spec.ts (47) pass; tsc --noEmit and biome ci clean. I did not run the full ava suite locally.

🤖 Generated with Claude Code

[github-actions]

nc-review: nothing to raise

@will-lamerton — nothing to raise from the automated review.

Renames the SDK-bound ceiling setting from maxTokens (v4, silently dropped by ai@6) to maxOutputTokens, threads maxOutputTokens through loadProviderConfigs, the schema and the type, and resolves /tune's maxTokens to the SDK name at the boundary so persisted preferences are not invalidated. Three new wire-level tests pin the SDK contract (default 4096 for unrecognised Anthropic model ids, maxOutputTokens reaching the wire as max_tokens, and the v4 spelling being ignored) and one client-factory test asserts plumbing. No findings.

---

_{🔴 blocking · 🟠 a reviewer would ask for a change · ⚪ optional}

_{Automated code review — correctness, security, design, tests, plus duplicates and scope. A human still decides; this is not a substitute for review and is not exhaustive. The required status checks separately cover lint, formatting, types, unused dependencies, the test suite and the build. This bot never merges. Maintainers can rerun with /re-review.}