From c4494b49ff54efb3bc1fa8e2b3bb48fae693b8e4 Mon Sep 17 00:00:00 2001
From: holden093 <kevin@nixit.it>
Date: Mon, 29 Jun 2026 14:24:41 +0200
Subject: [PATCH] fix(chat): use DEFAULT_MAX_TOKENS consistently across all
 call sites
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Replace hardcoded 0 and 4096 in skills_routes.py with the canonical
constant, and bump LLMConfig.DEFAULT_MAX_TOKENS from 0 to 4096 so
function defaults match. The rewrite path (chat_routes.py:1572) keeps
its intentional 0 — local reasoning models share their token budget
with <think> blocks and need to be uncapped there.

Root cause: max_tokens=0 caused the OpenAI-compat payload builder to
omit the param entirely, so upstreams like DeepSeek defaulted to a
very short completion.
---
 routes/skills_routes.py | 9 +++------
 src/llm_core.py         | 2 +-
 2 files changed, 4 insertions(+), 7 deletions(-)

diff --git a/routes/skills_routes.py b/routes/skills_routes.py
index 711baa2e5..146f2c053 100644
--- a/routes/skills_routes.py
+++ b/routes/skills_routes.py
@@ -18,6 +18,7 @@ from pydantic import BaseModel, Field
 
 from services.memory.skills import SkillsManager
 from src.auth_helpers import get_current_user
+from src.constants import DEFAULT_MAX_TOKENS
 from core.middleware import require_admin
 
 logger = logging.getLogger(__name__)
@@ -440,7 +441,7 @@ async def _run_skill_test_job(key, name, md, task, url, model, headers, owner, s
     try:
         async for chunk in stream_agent_loop(
             url, model, messages, headers=headers,
-            temperature=0.3, max_tokens=0, max_rounds=8, owner=owner,
+            temperature=0.3, max_tokens=DEFAULT_MAX_TOKENS, max_rounds=8, owner=owner,
         ):
             if not chunk.startswith("data: ") or chunk.strip() == "data: [DONE]":
                 continue
@@ -701,12 +702,8 @@ async def _run_skill_test_once(md: str, task: str, url, model, headers, owner) -
         {"role": "user", "content": task},
     ]
     try:
-        # max_tokens explicitly set: passing 0 lets some upstreams (Ollama,
-        # OpenAI-compat) generate an empty completion, which manifested as
-        # the skill test returning nothing while chat (which carries its
-        # preset's max_tokens) worked. 4096 matches the chat default.
         async for chunk in stream_agent_loop(url, model, messages, headers=headers,
-                                             temperature=0.3, max_tokens=4096, max_rounds=8, owner=owner):
+                                             temperature=0.3, max_tokens=DEFAULT_MAX_TOKENS, max_rounds=8, owner=owner):
             if not chunk.startswith("data: ") or chunk.strip() == "data: [DONE]":
                 continue
             try:
diff --git a/src/llm_core.py b/src/llm_core.py
index 4dec32376..f9e0525f5 100644
--- a/src/llm_core.py
+++ b/src/llm_core.py
@@ -94,7 +94,7 @@ class LLMConfig:
     """Configuration constants for LLM operations."""
     DEFAULT_TIMEOUT = 30
     DEFAULT_TEMPERATURE = 1.0
-    DEFAULT_MAX_TOKENS = 0
+    DEFAULT_MAX_TOKENS = 4096
     MAX_RETRIES = 3
     RETRY_DELAY = 0.5
     STREAM_TIMEOUT = 300