Symptoms
Gateway (Feishu, etc.) returns HTTP 400 on API calls with errors like:
- DeepSeek:
Error from provider (DeepSeek): Tool names must be unique.
- Xiaomi MiMo:
tools contains duplicate names: lcm_expand
- Moonshot/Kimi:
function name lcm_grep is duplicated
TUI sessions are unaffected.
Environment:
Root Cause
Commit #17098 (perf(tools): memoize get_tool_definitions + TTL-cache check_fn results) introduced a quiet_mode cache in model_tools.py. The caching logic on cache-hit (line 278 return list(cached)) is safe; but the first uncached call (lines 280-283) returns the same object that gets stored into the cache. Any caller that mutates the returned list also mutates the cache.
Later, run_agent.py line 1986-1993 appends LCM context engine tool schemas to self.tools without checking for duplicates:
for _schema in self.context_compressor.get_tool_schemas():
_wrapped = {"type": "function", "function": _schema}
self.tools.append(_wrapped) # mutates shared cached listNote: memory tools injection (lines 1728-1748) already has dedup logic — context engine tools injection does NOT.
Why TUI is fine: TUI either uses quiet_mode=False (no caching) or short-lived processes.
Reproduction
Verified with a local script. Clear cache, call get_tool_definitions(quiet_mode=True), mutate the returned list, call again — the mutation is visible in the cache. In a long-lived Gateway process, each agent init appends LCM tools again, causing accumulation.
Suggested Fix
Two places to harden:
-
model_tools.py: On the first non-cached call, return a shallow copy and cache the original — same defensive pattern as the cache-hit path (line 278).
-
run_agent.py: Add dedup logic to context engine tool injection (lines 1986-1993), mirroring the memory tools dedup (lines 1728-1748).
Either fix alone stops the symptom; both together provide defense-in-depth.
Workaround
Restarting the Gateway temporarily clears the in-process cache, but the bug will recur as agents accumulate LCM tool appends.
Symptoms
Gateway (Feishu, etc.) returns HTTP 400 on API calls with errors like:
Error from provider (DeepSeek): Tool names must be unique.tools contains duplicate names: lcm_expandfunction name lcm_grep is duplicatedTUI sessions are unaffected.
Environment:
context.engine: lcm,hermes-lcmplugin enabledRoot Cause
Commit #17098 (perf(tools): memoize get_tool_definitions + TTL-cache check_fn results) introduced a
quiet_modecache inmodel_tools.py. The caching logic on cache-hit (line 278return list(cached)) is safe; but the first uncached call (lines 280-283) returns the same object that gets stored into the cache. Any caller that mutates the returned list also mutates the cache.Later,
run_agent.pyline 1986-1993 appends LCM context engine tool schemas toself.toolswithout checking for duplicates:Note: memory tools injection (lines 1728-1748) already has dedup logic — context engine tools injection does NOT.
Why TUI is fine: TUI either uses
quiet_mode=False(no caching) or short-lived processes.Reproduction
Verified with a local script. Clear cache, call
get_tool_definitions(quiet_mode=True), mutate the returned list, call again — the mutation is visible in the cache. In a long-lived Gateway process, each agent init appends LCM tools again, causing accumulation.Suggested Fix
Two places to harden:
model_tools.py: On the first non-cached call, return a shallow copy and cache the original — same defensive pattern as the cache-hit path (line 278).run_agent.py: Add dedup logic to context engine tool injection (lines 1986-1993), mirroring the memory tools dedup (lines 1728-1748).Either fix alone stops the symptom; both together provide defense-in-depth.
Workaround
Restarting the Gateway temporarily clears the in-process cache, but the bug will recur as agents accumulate LCM tool appends.