Skip to content

Never stop coding. Free MIT AI gateway: one endpoint, 375 providers (150+ free), 1,400+ models Kimi, Claude, GPT, Gemini, GLM, DeepSeek, MiniMax. Works with Claude Code, Codex, Cursor, OpenCode, Cline & Copilot. Quota-aware auto-fallback, RTK+Caveman compression saves 15-95% tokens, MCP/A2A, Desktop/PWA. Built by 900+ contributors

5.3 KB
Raw Blame
/**
 * #6274 — the reasoning-token buffer must not inflate probe-sized max_tokens.
 *
 * Claude Code's `/model` capability check sends `max_tokens: 1`; for a thinking-
 * capable model with a large output cap (e.g. glm-5.2) the #3587 headroom heuristic
 * (`max(current + 1000, ceil(current * 1.5))`) rewrote it to 1001 and forwarded that
 * upstream. A tiny explicit budget below REASONING_BUFFER_MIN_TRIGGER (256) is a
 * probe and must pass through verbatim; genuine budgets keep the #3587 headroom
 * only when the full headroom fits inside an explicit output cap.
 *
 * Kept standalone against the pure `resolveReasoningBufferedMaxTokens` rather than
 * extending the frozen `combo-routing-engine.test.ts` god-file.
 *
 * #9507 update: the #3587 headroom heuristic was removed — the buffer never
 * enlarges an explicit client max_tokens. The assertions at/above the trigger
 * threshold now expect pass-through (256 -> 256, 32000 -> 32000).
 */
import test from "node:test";
import assert from "node:assert/strict";
import fs from "node:fs";
import os from "node:os";
import path from "node:path";

const TEST_DATA_DIR = fs.mkdtempSync(path.join(os.tmpdir(), "omniroute-reasoning-buffer-"));
process.env.DATA_DIR = TEST_DATA_DIR;

const core = await import("../../src/lib/db/core.ts");
const { saveModelsDevCapabilities, clearModelsDevCapabilities } =
  await import("../../src/lib/modelsDevSync.ts");
const { resolveReasoningBufferedMaxTokens, REASONING_BUFFER_MIN_TRIGGER } =
  await import("../../open-sse/services/reasoningTokenBuffer.ts");

function capabilityEntry(limitContext: unknown, overrides: Record<string, unknown> = {}) {
  return {
    tool_call: true,
    reasoning: false,
    attachment: false,
    structured_output: true,
    temperature: true,
    modalities_input: JSON.stringify(["text"]),
    modalities_output: JSON.stringify(["text"]),
    knowledge_cutoff: null,
    release_date: null,
    last_updated: null,
    status: null,
    family: null,
    open_weights: false,
    limit_context: limitContext,
    limit_input: limitContext,
    limit_output: 4096,
    interleaved_field: null,
    ...overrides,
  };
}

test.before(() => {
  // A thinking-capable model with a large output cap: the #3587 guards all pass.
  saveModelsDevCapabilities({
    zhipu: {
      "glm-5.2": capabilityEntry(200000, { reasoning: true, limit_output: 65536 }),
      // Deliberately NOT prefixed with a real MODEL_SPECS key (e.g. "glm-5.2") —
      // getStaticSpec()/getCanonicalModelSpecId() does prefix matching (#6714),
      // so a fixture id like "glm-5.2-no-output-cap" would silently fall through
      // to the real glm-5.2 static spec's 131072 cap and defeat this fixture's
      // "no cap anywhere" premise.
      "totally-fictitious-model-6714-no-output-cap": capabilityEntry(200000, {
        reasoning: true,
        limit_output: null,
      }),
      "glm-5.2-output-cap-40000": capabilityEntry(200000, {
        reasoning: true,
        limit_output: 40000,
      }),
    },
  });
});

test.after(() => {
  clearModelsDevCapabilities();
  core.resetDbInstance();
  fs.rmSync(TEST_DATA_DIR, { recursive: true, force: true, maxRetries: 5, retryDelay: 100 });
});

test("#6274 reasoning buffer does not inflate probe-sized max_tokens", () => {
  // The Claude-Code `/model` probe (max_tokens: 1) must pass through (was 1001).
  assert.equal(
    resolveReasoningBufferedMaxTokens("zhipu/glm-5.2", 1),
    1,
    "probe-sized max_tokens=1 must not be inflated"
  );
  // Just below the trigger threshold is still treated as a probe.
  assert.equal(
    resolveReasoningBufferedMaxTokens("zhipu/glm-5.2", REASONING_BUFFER_MIN_TRIGGER - 1),
    REASONING_BUFFER_MIN_TRIGGER - 1,
    "budgets below REASONING_BUFFER_MIN_TRIGGER are respected verbatim"
  );
  // Issue #9507: the buffer must NEVER enlarge a client's explicit max_tokens.
  // Previously the #3587 headroom heuristic rewrote these upward
  // (256 -> 1256, 32000 -> 48000); that violated the #1761 contract that
  // upward adjustment must be opt-in. The over-cap clamp still narrows.
  assert.equal(
    resolveReasoningBufferedMaxTokens("zhipu/glm-5.2", REASONING_BUFFER_MIN_TRIGGER),
    REASONING_BUFFER_MIN_TRIGGER,
    "budgets at the threshold are forwarded verbatim (#9507)"
  );
  // A realistic reasoning budget is forwarded verbatim, not enlarged.
  assert.equal(
    resolveReasoningBufferedMaxTokens("zhipu/glm-5.2", 32000),
    32000,
    "genuine reasoning budgets are forwarded verbatim (#9507)"
  );
});

test("reasoning buffer requires an explicit cap and preserves near-cap budgets", () => {
  assert.equal(
    resolveReasoningBufferedMaxTokens("zhipu/totally-fictitious-model-6714-no-output-cap", 32000),
    null,
    "missing model output cap should disable heuristic token inflation"
  );

  // Known cap below the heuristic result: preserve the caller's in-range budget
  // rather than inflating to a value that may reduce response room unexpectedly.
  assert.equal(
    resolveReasoningBufferedMaxTokens("zhipu/glm-5.2-output-cap-40000", 32000),
    32000,
    "known model output cap should preserve in-range near-cap budgets"
  );

  // Known cap below the caller value still clamps the requested value itself.
  assert.equal(
    resolveReasoningBufferedMaxTokens("zhipu/glm-5.2-output-cap-40000", 41000),
    40000,
    "requested max_tokens above the model output cap should be capped"
  );
});