From 2ed904fd0031ba2ab1a2228944687078a4458649 Mon Sep 17 00:00:00 2001
From: autodev-bot <autodev@memtensor.local>
Date: Sun, 30 Aug 2026 05:29:29 +0800
Subject: [PATCH] fix(llm): propagate opts.op through facade to providers
 (#2308)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

`LlmClient.buildCallInput()` dropped `opts.op` when constructing the
`ProviderCallInput` object handed to `provider.complete()` / `stream()`.
Callers set `op` for every logical call site (e.g. `capture.summarize`,
`retrieval.filter`, `skill.evolve`) but only the audit / metrics paths
inside the facade saw the tag — providers never did.

As a result any per-op provider behavior (request-body tweaks, routing
overrides, reasoning kill-switches keyed on
`opts.op === "capture.summarize"`) silently could never fire. The
issue-reported symptom: OpenRouter DeepSeek reasoning models burn full
token budget and double latency on `capture.summarize` because the
recommended `thinking: { type: "disabled" }` switch is unreachable.

Fix:
- extend `ProviderCallInput` with optional `op?: string`
- copy `opts?.op` in `buildCallInput()`

Field stays optional so providers must not assume it is set. No public
LlmClient surface change; consumers keep calling `complete` /
`completeJson` / `stream` unchanged.

Added 4 unit tests in `tests/unit/llm/client.test.ts` under an
"op propagation (issue #2308)" describe block covering complete /
completeJson / stream forwarding and the "no op supplied" case.
---
 apps/memos-local-plugin/core/llm/client.ts    |  1 +
 apps/memos-local-plugin/core/llm/types.ts     |  8 ++++
 .../tests/unit/llm/client.test.ts             | 45 +++++++++++++++++++
 3 files changed, 54 insertions(+)

diff --git a/apps/memos-local-plugin/core/llm/client.ts b/apps/memos-local-plugin/core/llm/client.ts
index ee456ac12..a94d05556 100644
--- a/apps/memos-local-plugin/core/llm/client.ts
+++ b/apps/memos-local-plugin/core/llm/client.ts
@@ -282,6 +282,7 @@ export function createLlmClientWithProvider(
       maxTokens: opts?.maxTokens ?? config.maxTokens ?? DEFAULT_MAX_TOKENS,
       jsonMode,
       stop: opts?.stop,
+      op: opts?.op,
     };
   }
 
diff --git a/apps/memos-local-plugin/core/llm/types.ts b/apps/memos-local-plugin/core/llm/types.ts
index 2dd761f32..ef61c439f 100644
--- a/apps/memos-local-plugin/core/llm/types.ts
+++ b/apps/memos-local-plugin/core/llm/types.ts
@@ -257,6 +257,14 @@ export interface ProviderCallInput {
   maxTokens: number;
   jsonMode: boolean;
   stop?: string[];
+  /**
+   * Logical call site (e.g. `capture.summarize`, `retrieval.filter`,
+   * `skill.evolve`). Forwarded from `LlmCallOptions.op` so providers
+   * can apply per-op behavior (request-body tweaks, routing overrides,
+   * reasoning kill-switches, per-op budget caps). Optional — providers
+   * must not assume it is set.
+   */
+  op?: string;
 }
 
 /** What providers return — pre-facade post-processing. */
diff --git a/apps/memos-local-plugin/tests/unit/llm/client.test.ts b/apps/memos-local-plugin/tests/unit/llm/client.test.ts
index cd125ecd8..084ec5d6f 100644
--- a/apps/memos-local-plugin/tests/unit/llm/client.test.ts
+++ b/apps/memos-local-plugin/tests/unit/llm/client.test.ts
@@ -322,6 +322,51 @@ describe("llm/client", () => {
     );
   });
 
+  // ─── op propagation to providers (issue #2308) ────────────────────────
+  //
+  // The facade must forward `opts.op` onto the `ProviderCallInput` handed to
+  // `provider.complete()` / `provider.stream()` so per-op provider behavior
+  // (e.g. request-body tweaks, routing overrides, reasoning kill-switches
+  // keyed on `capture.summarize`) can fire. Dropping it silently makes
+  // those switches unreachable.
+  describe("op propagation (issue #2308)", () => {
+    it("complete forwards opts.op onto the provider input", async () => {
+      const fake = new FakeProvider("openai_compatible", () => ({ text: "ok", durationMs: 1 }));
+      const client = createLlmClientWithProvider(cfg(), fake);
+      await client.complete("hi", { op: "capture.summarize" });
+      expect(fake.lastInput?.op).toBe("capture.summarize");
+    });
+
+    it("completeJson forwards opts.op onto the provider input", async () => {
+      const fake = new FakeProvider("openai_compatible", () => ({
+        text: '{"a":1}',
+        durationMs: 1,
+      }));
+      const client = createLlmClientWithProvider(cfg(), fake);
+      await client.completeJson<{ a: number }>("score", { op: "retrieval.filter" });
+      expect(fake.lastInput?.op).toBe("retrieval.filter");
+    });
+
+    it("stream forwards opts.op onto the provider input (non-streaming provider)", async () => {
+      // FakeProvider has no stream(); the facade wraps complete() in a
+      // single-chunk iterable, which still exercises buildCallInput.
+      const fake = new FakeProvider("openai_compatible", () => ({ text: "one", durationMs: 1 }));
+      const client = createLlmClientWithProvider(cfg(), fake);
+      const parts: string[] = [];
+      for await (const c of client.stream("x", { op: "skill.evolve" })) {
+        if (!c.done) parts.push(c.delta);
+      }
+      expect(fake.lastInput?.op).toBe("skill.evolve");
+    });
+
+    it("omits op field when caller supplies no op (contract stays optional)", async () => {
+      const fake = new FakeProvider("openai_compatible", () => ({ text: "ok", durationMs: 1 }));
+      const client = createLlmClientWithProvider(cfg(), fake);
+      await client.complete("hi");
+      expect(fake.lastInput?.op).toBeUndefined();
+    });
+  });
+
   // ─── Circuit breaker (issue #1897) ──────────────────────────────────────
   describe("circuit breaker", () => {
     function statusSink(): { rows: LlmStatusDetail[]; push: (d: LlmStatusDetail) => void } {