fix(providers): pass max reasoning effort through on the Responses API for probe-verified models

Author: jddunn

src/core/llm/providers/__tests__/model-effort.test.ts

@@ -4,8 +4,10 @@ import {
   isEffortLevel,
   EFFORT_LEVELS,
   mapEffortToOpenAiReasoningEffort,
+  mapEffortToOpenAiReasoningEffortForModel,
   mapEffortToOpenAiResponsesEffort,
   modelAcceptsXhighResponsesEffort,
+  modelAcceptsMaxResponsesEffort,
 } from '../model-effort.js';
 
 describe('modelSupportsEffort', () => {
@@ -85,11 +87,23 @@ describe('mapEffortToOpenAiResponsesEffort (model-aware /v1/responses effort)',
     expect(mapEffortToOpenAiResponsesEffort('gpt-5.5', 'xhigh')).toBe('xhigh');
   });
 
-  it('passes xhigh through for gpt-5.6 / gpt-5.6-sol (max -> xhigh, allow-listed)', () => {
-    expect(mapEffortToOpenAiResponsesEffort('gpt-5.6', 'max')).toBe('xhigh');
+  it('passes the real max tier through for the gpt-5.6 family (live-probed 2026-08-06)', () => {
+    expect(modelAcceptsMaxResponsesEffort('gpt-5.6')).toBe(true);
+    expect(modelAcceptsMaxResponsesEffort('gpt-5.6-sol')).toBe(true);
+    expect(mapEffortToOpenAiResponsesEffort('gpt-5.6', 'max')).toBe('max');
+    expect(mapEffortToOpenAiResponsesEffort('gpt-5.6-sol', 'max')).toBe('max');
+    // xhigh stays xhigh — the unlock only touches the max tier.
     expect(mapEffortToOpenAiResponsesEffort('gpt-5.6-sol', 'xhigh')).toBe('xhigh');
   });
 
+  it('keeps max clamped to xhigh off the allow-list (gpt-5.5) and chat-side everywhere', () => {
+    expect(modelAcceptsMaxResponsesEffort('gpt-5.5')).toBe(false);
+    expect(modelAcceptsMaxResponsesEffort('gpt-5.4')).toBe(false);
+    expect(mapEffortToOpenAiResponsesEffort('gpt-5.5', 'max')).toBe('xhigh');
+    // Chat Completions rejects max for the 5.6 family (probed 2026-07-20 + 2026-08-06).
+    expect(mapEffortToOpenAiReasoningEffortForModel('max', 'gpt-5.6')).toBe('xhigh');
+  });
+
   it('caps xhigh -> high for a non-allow-listed gpt-5 model', () => {
     expect(mapEffortToOpenAiResponsesEffort('gpt-5.4', 'max')).toBe('high');
     expect(mapEffortToOpenAiResponsesEffort('gpt-5-mini', 'xhigh')).toBe('high');

src/core/llm/providers/model-effort.ts

@@ -68,18 +68,44 @@ export function modelAcceptsXhighResponsesEffort(modelId: string): boolean {
   return /^gpt-5\.[56]/i.test(modelId);
 }
 
+/**
+ * Responses-API models whose `reasoning.effort` accepts `'max'` — the tier
+ * ABOVE `xhigh`, and a Responses-API-only surface (Chat Completions rejects
+ * `max` for the same family; see {@link CHAT_MAX_EFFORT_MODELS} below).
+ * Live-probed per entry — record request shape, status, response summary,
+ * date, and the exact model alias in this comment when adding one.
+ *
+ * 2026-08-06 probe (POST /v1/responses, `reasoning: {effort: 'max'}`,
+ * max_output_tokens 2000): HTTP 200 `status: completed` on `gpt-5.6` (bare
+ * alias), on `gpt-5.6-sol`, and on `gpt-5.6-sol` WITH function tools
+ * attached (the agentic shape). Boundary probes the same day: `'ultra'` is
+ * rejected on both aliases (`Invalid value: 'ultra'. Supported values are:
+ * 'none', 'minimal', 'low', 'medium', 'high', 'xhigh', and 'max'.`), and
+ * chat.completions still 400s on `max` for the family. Widen only after a
+ * fresh Responses probe returns HTTP 200 on the new id.
+ */
+const RESPONSES_MAX_EFFORT_MODELS: readonly string[] = ['gpt-5.6'];
+
+/** Whether `modelId` is on the probe-verified Responses `max` allow-list. */
+export function modelAcceptsMaxResponsesEffort(modelId: string): boolean {
+  return RESPONSES_MAX_EFFORT_MODELS.some((f) => modelId.toLowerCase().startsWith(f));
+}
+
 /**
  * Map the agentos effort scale onto OpenAI's `/v1/responses` `reasoning.effort`
- * for `modelId`. Identical to {@link mapEffortToOpenAiReasoningEffort} EXCEPT it
- * caps `xhigh` → `high` for any model NOT on
- * {@link modelAcceptsXhighResponsesEffort} — a model-aware guard so a request
+ * for `modelId`. Identical to {@link mapEffortToOpenAiReasoningEffort} EXCEPT:
+ * `max` passes through as the real `'max'` tier on models probe-verified via
+ * {@link modelAcceptsMaxResponsesEffort} (elsewhere it clamps to `'xhigh'` as
+ * before), and `xhigh` caps to `high` for any model NOT on
+ * {@link modelAcceptsXhighResponsesEffort} — model-aware guards so a request
  * for maximum depth degrades one step instead of 400ing the whole Responses
- * call on a model that rejects `xhigh`.
+ * call on a model that rejects the tier.
  */
 export function mapEffortToOpenAiResponsesEffort(
   modelId: string,
   effort: unknown,
 ): string | undefined {
+  if (effort === 'max' && modelAcceptsMaxResponsesEffort(modelId)) return 'max';
   const base = mapEffortToOpenAiReasoningEffort(effort);
   if (base === undefined) return undefined;
   return base === 'xhigh' && !modelAcceptsXhighResponsesEffort(modelId) ? 'high' : base;