From cbc170818f39630671a1e4405850bbc87edd70ad Mon Sep 17 00:00:00 2001
From: chouti <chouti@users.noreply.github.com>
Date: Mon, 8 Jun 2026 19:06:03 +0800
Subject: [PATCH] fix(config): add maxTokens to LlmSchema & SkillEvolverSchema

Reasoning models (deepseek-reasoner, o1*, gpt-5-thinking) consume
hundreds of tokens on chain-of-thought. The hard-coded 1024 cap on
LLM JSON reflection caused 73+ 'llm.json malformed' errors per
episode, blocking episode closure and disconnecting the bridge.

- Add maxTokens: NumberInRange(4000, 1024, 32768) to LlmSchema
- Add maxTokens: NumberInRange(4000, 1024, 32768) to SkillEvolverSchema
- Add corresponding defaults in defaults.ts (backward compat)

TypeBox's Value.Default preserves user-supplied fields, so old
configs without maxTokens get the 4000 default; new configs can
override.
---
 apps/memos-local-plugin/core/config/defaults.ts |  2 ++
 apps/memos-local-plugin/core/config/schema.ts   | 15 +++++++++++++++
 2 files changed, 17 insertions(+)

diff --git a/apps/memos-local-plugin/core/config/defaults.ts b/apps/memos-local-plugin/core/config/defaults.ts
index 1cf2d2cf6..fe43147a2 100644
--- a/apps/memos-local-plugin/core/config/defaults.ts
+++ b/apps/memos-local-plugin/core/config/defaults.ts
@@ -41,6 +41,7 @@ export const DEFAULT_CONFIG: ResolvedConfig = {
     apiKey: "",
     timeoutMs: 45_000,
     maxRetries: 3,
+    maxTokens: 4_000,
   },
   skillEvolver: {
     // Empty by default — falls back to the shared `llm` settings.
@@ -52,6 +53,7 @@ export const DEFAULT_CONFIG: ResolvedConfig = {
     apiKey: "",
     temperature: 0,
     timeoutMs: 60_000,
+    maxTokens: 4_000,
   },
   algorithm: {
     lightweightMemory: {
diff --git a/apps/memos-local-plugin/core/config/schema.ts b/apps/memos-local-plugin/core/config/schema.ts
index 7c9ff193b..daffa0f75 100644
--- a/apps/memos-local-plugin/core/config/schema.ts
+++ b/apps/memos-local-plugin/core/config/schema.ts
@@ -65,6 +65,14 @@ const LlmSchema = Type.Object({
   timeoutMs: NumberInRange(45_000, 1_000),
   /** Max retries on transient errors. */
   maxRetries: NumberInRange(3, 0, 10),
+  /**
+   * Max output tokens per LLM call. Default 4_000 — sufficient budget for
+   * reasoning models (deepseek-reasoner, o1*, gpt-5-thinking) that burn
+   * hundreds of tokens on chain-of-thought before content. Below ~2_000
+   * the JSON reflection / synth prompts get truncated mid-string and the
+   * bridge logs `llm.json malformed`.
+   */
+  maxTokens: NumberInRange(4_000, 1_024, 32_768),
 }, { default: {} });
 
 /**
@@ -89,6 +97,13 @@ const SkillEvolverSchema = Type.Object({
   apiKey: StringWithDefault(""),
   temperature: NumberInRange(0, 0, 2),
   timeoutMs: NumberInRange(60_000, 1_000),
+  /**
+   * Max output tokens per skill-evolver LLM call. Same reasoning budget
+   * as `llm.maxTokens` — 4_000 covers typical crystallisation JSON
+   * (multi-policy evidence rollup + decision rationale) without hitting
+   * the 32k OpenAI ceiling.
+   */
+  maxTokens: NumberInRange(4_000, 1_024, 32_768),
 }, { default: {} });
 
 const AlgorithmSchema = Type.Object({