Kimi K2.5 via aggregators (synthetic.new, OpenRouter) gets no max_tokens/reasoning_effort → empty response

[liyoungc]

Kimi K2.5 via aggregators (synthetic.new, OpenRouter, Together…) silently truncates to empty response

Symptom

When model: hf:moonshotai/Kimi-K2.5 (or any Kimi slug) is configured to go through an aggregator base URL (e.g. https://api.synthetic.new/v1), long agentic prompts — especially cron jobs with multiple tool calls — consistently fail with:

⚠️  Response truncated (finish_reason='length') - model hit max output tokens
⚠️  Truncated tool call detected — retrying API call...
⚠️  Response truncated (finish_reason='length') - model hit max output tokens
⚠️  Truncated tool call response detected again — refusing to execute incomplete tool arguments.

Eventually the cron scheduler records last_status: error, last_error: "Agent completed but produced empty response (model error, timeout, or misconfiguration)". Direct LINE / Discord chat with the same model usually works because short prompts have enough budget to leak past the truncation, but anything longer than ~30 tool-bearing turns dies.

Root cause

_is_kimi (run_agent.py:8309 in current main) is base-URL-only:

_is_kimi = (
    base_url_host_matches(self.base_url, "api.kimi.com")
    or base_url_host_matches(self.base_url, "moonshot.ai")
    or base_url_host_matches(self.base_url, "moonshot.cn")
)

It misses every aggregator that routes to Moonshot inference. The downstream effect is in agent/transports/chat_completions.py:240–280:

• The Kimi-specific max_tokens=32000 default (line ~259) isn't applied → falls through to "send no max_tokens" → aggregator picks a small server default (synthetic.new behaves as if it's ~4K).
• The Kimi-specific reasoning_effort=medium hint (line ~272) isn't sent → Kimi K2.5 runs with its built-in default thinking effort.
• Kimi K2.5 has a thinking/reasoning mode that pre-spends output tokens on hidden reasoning before producing visible text. With a small max_tokens cap and no effort hint, it spends the entire budget on reasoning and emits zero visible tokens.
• finish_reason: length with empty content → hermes retries the API call with continuation → same outcome → "Truncated tool call response detected again — refusing to execute" → empty response.

The interesting thing is is_moonshot_model(self.model) already exists in agent/moonshot_schema.py:171 and correctly recognises model-name-based slugs (hf:moonshotai/Kimi-K2.5, nous/moonshotai/kimi-k2.5, …). It's even used immediately above the is_kimi check in chat_completions.py to sanitize tools. But the runtime detection path doesn't reuse it.

Repro

1. Set model.default: hf:moonshotai/Kimi-K2.5 and a synthetic.new (or OpenRouter, Together) credential in ~/.hermes/auth.json.
2. Create any cron job with kind: cron, deliver: discord, prompt requiring 3–5 tool calls (e.g. multiple curl + a python3 script).
3. Wait for fire — every fire ends in last_status: error, last_error: "Agent completed but produced empty response". Container logs show repeated Response truncated (finish_reason='length') warnings with no visible content between them.

The same model works fine when configured against https://api.moonshot.ai/v1 directly because base URL matches and the Kimi defaults kick in.

Proposed fix

--- a/run_agent.py
+++ b/run_agent.py
@@ -8307,11 +8307,21 @@ class Agent:
         _is_nous = "nousresearch" in self._base_url_lower
         _is_nvidia = "integrate.api.nvidia.com" in self._base_url_lower
+        # Detect Kimi models routed through aggregators (synthetic.new,
+        # OpenRouter, Together, ...).  Without this branch, those routes
+        # miss the Kimi-specific max_tokens=32000 default and the
+        # reasoning_effort=medium hint, leaving the model with whatever
+        # tiny output budget the aggregator defaults to and free-running
+        # thinking mode that swallows the entire budget — visible response
+        # ends up empty.
+        try:
+            from agent.moonshot_schema import is_moonshot_model as _is_moonshot
+        except Exception:  # pragma: no cover — optional helper
+            _is_moonshot = lambda _m: False  # noqa: E731
         _is_kimi = (
             base_url_host_matches(self.base_url, "api.kimi.com")
             or base_url_host_matches(self.base_url, "moonshot.ai")
             or base_url_host_matches(self.base_url, "moonshot.cn")
+            or _is_moonshot(self.model)
         )

This makes _is_kimi consistent with the model-name-based detection already used elsewhere (is_moonshot_model in agent/moonshot_schema.py:171). Both max_tokens=32000 and reasoning_effort=medium then route correctly regardless of which aggregator the user goes through.

Related

This bug compounded with two other cron robustness issues filed separately (null next_run_at skip + non-dict origin AttributeError) — taken together they explain a class of "cron quietly does nothing" reports on aggregated Kimi deployments.

Environment

• hermes-agent: upstream/main as of 2026-05-02
• Model: hf:moonshotai/Kimi-K2.5
• Provider: synthetic.new (https://api.synthetic.new/v1)
• Python 3.14, croniter installed
• Encountered on a chococlaw VPS Docker deployment (Linux Debian)