Hermes + Ollama HTTP 500 — "no user query found in messages"¶

Status: Diagnosed 2026-09-12 on ai-nuc
Symptom: Open Cowork / plain Ollama chat works. Hermes queries fail with HTTP 500.

What it is (and is not)¶

This is not Tailscale, UFW, or a missing API key. The NUC received the Hermes request and rejected the message shape.

Confirmed in journalctl -u ollama at 2026-09-12 04:17:

level=ERROR source=routes.go:2706 msg="chat prompt error" error="no user query found in messages"

Hermes (agent harness) sends a tool-loop continuation after the first tool call. Typical shape:

[system, assistant(tool_calls), tool(result)]

or a user turn whose content is [] (empty list).

qwen3.8:27b uses RENDERER qwen3.8. That renderer’s validateMessages() requires a plain user message that is not a <tool_response> wrapper. Same weights with RENDERER qwen3.5 accept the request.

Upstream:

ollama#17778 — qwen3.8 renderer rejects tool-loop continuations
ollama#18107 — context truncation can also drop the user turn and produce the same 500

A second, related failure: Hermes reports the model’s max context (262144) from /api/show, but Ollama’s default runtime window is often 4096. Truncation then drops the user message. Hermes docs say set context length to 64000 and match num_ctx on the server.

Fix on ai-nuc (done)¶

Created a same-weights alias:

Name	Renderer	`num_ctx`	Purpose
`qwen3.8:27b`	qwen3.8	default	Chat / Open Cowork
`qwen3.8:27b-hermes`	qwen3.5	65536	Hermes agent tool loops

Modelfile: scripts/Modelfile.qwen3.8-27b-hermes

ollama create qwen3.8:27b-hermes -f ~/office-plan/scripts/Modelfile.qwen3.8-27b-hermes

Does not re-download the 17GB weights. Disk cost is a small extra manifest.

What to change in Hermes (MacBook / wherever Hermes runs)¶

Model name: qwen3.8:27b-hermes (not qwen3.8:27b)
Endpoint can stay http://100.73.71.29:11434/v1
Context length: 64000 (Hermes minimum; matches the alias num_ctx)
API key: any non-empty string (ollama is fine)

CLI equivalent:

hermes model
# Custom endpoint: http://100.73.71.29:11434/v1
# Model: qwen3.8:27b-hermes
# Context length: 64000

Or in ~/.hermes/config.yaml:

model: qwen3.8:27b-hermes
provider: custom
base_url: http://100.73.71.29:11434/v1
context_length: 64000

Start a new Hermes session after the switch. An old session still holds the failing transcript.

Verify¶

On ai-nuc:

# Stock model fails this Hermes-like tool continuation
curl -s -w "\nHTTP %{http_code}\n" http://127.0.0.1:11434/api/chat -d '{
  "model":"qwen3.8:27b","stream":false,
  "messages":[
    {"role":"assistant","content":"","tool_calls":[{"function":{"name":"web_search","arguments":{"q":"ollama"}}}]},
    {"role":"tool","tool_name":"web_search","content":"{\"results\":[]}"}
  ]
}'

# Alias should return HTTP 200
curl -s -w "\nHTTP %{http_code}\n" http://127.0.0.1:11434/api/chat -d '{
  "model":"qwen3.8:27b-hermes","stream":false,
  "messages":[
    {"role":"assistant","content":"","tool_calls":[{"function":{"name":"web_search","arguments":{"q":"ollama"}}}]},
    {"role":"tool","tool_name":"web_search","content":"{\"results\":[]}"}
  ]
}'

From Hermes: send a short query that would use a tool (check unread email, list files). Expect a reply, not the 500.

If it still 500s¶

Confirm Hermes is actually using qwen3.8:27b-hermes (hermes config show).
Confirm context is 64000, not 262144 or auto-detected max.
Check journalctl -u ollama -n 30 on ai-nuc while sending the query.
Update Hermes — later builds sanitize empty content: [] user turns (hermes-agent#99234).

Which models can hit this¶

The same HTTP 500 text comes from two different failures. Check both when you switch models.

Cause A — `RENDERER qwen3.8` (model-family bug)¶

Ollama’s qwen3.8 renderer requires a plain user message that is not a <tool_response> wrapper. Hermes tool loops often send [assistant(tool_calls), tool(result)] with no user text. That 500s even when context is huge.

Family / tag	Risk	Notes
`qwen3.8:*` (27b, 4b, mlx, cloud, any size)	Yes	Confirmed. Includes `qwen3.8:27b` on this box
`qwen3.8:*-hermes` aliases with `RENDERER qwen3.5`	No	Same weights, different renderer
`qwen3.6:*`	No	Same request returns 200 (#17778)
`qwen3.5:*`	No	This is the renderer we switched to
Qwen3 MoE GGUFs that do not set `RENDERER qwen3.8` (your 35B-A3B)	Unlikely	Check with `ollama show MODEL --modelfile \\| grep RENDERER`
Llama, Mistral, Gemma, DeepSeek, Phi, Command-R	No for Cause A	Different templates; no qwen3.8 `validateMessages`

Future official qwen3.8 pulls will come back with Cause A until Ollama relaxes that check (a later commit accepts user or tool turns; you are on 0.33.3, which still 500s).

Cause B — context truncation (almost any model)¶

Ollama default live window is 4096 unless the Modelfile or OLLAMA_CONTEXT_LENGTH sets num_ctx. Hermes sends system + tool schemas + tool output (often 6k–36k tokens). Truncation can drop the user turn and keep the tool turn. Same 500.

Reported on models that are not qwen3.8:

Model	Source
`qwen2.5-coder:14b`	ollama#18107
Any pull with no `PARAMETER num_ctx`	Same issue, their repro step 1
`llama3.2:3b` (your box)	Cause A: no. Cause B: yes if you leave default 4096
Qwythos 9B / Qwen3.6 35B on this box	Cause A: no. Cause B: yes without a 64k alias

Hermes advertising 131k / 262k / 1M does not raise Ollama’s live num_ctx.

Cause C — client message shape (any model)¶

Hermes (older builds) can send user.content: []. Ollama treats an empty list as “no user query.” A Hermes update that sanitizes that is hermes-agent#99234. Start a new session after /model.

HuggingFace GGUFs (`hf.co/...` or `ollama pull hf.co/...`)¶

Ollama does not treat HF imports as a separate API. After import it still picks a template/renderer from GGUF metadata (general.architecture, chat template, family). The 500 is about that assigned renderer + live num_ctx, not about “official library vs HuggingFace.”

What you pulled	Typical assignment	Hermes risk
Official `qwen3.8:27b`	`RENDERER qwen3.8`	Cause A + B
HF Qwen3.8 GGUF (any quant, any uploader)	Often same qwen3.8 renderer if architecture is detected as 3.8	Cause A likely
HF Qwen3.6 / Qwen3.5 GGUF (your 35B-A3B, Qwythos 9B)	`qwen35` / `qwen35moe`, no `RENDERER qwen3.8` on this box	Cause A unlikely; Cause B still yes
HF Llama / Mistral / Gemma / DeepSeek GGUF	Their own template	Cause A no; Cause B yes
Fine-tunes that keep Qwen3.8 chat template in the GGUF	May still get qwen3.8 renderer	Treat as Cause A until the curl test says otherwise
Merges / “uncensored” / “claude-mythos” quants	Renderer follows base architecture, not the marketing name	Always inspect

Your current HF models (checked 2026-09-12):

hf.co/empero-ai/Qwythos-9B-... — architecture qwen35, no RENDERER qwen3.8
hf.co/HauhauCS/Qwen3.6-35B-A3B-... — architecture qwen35moe, no RENDERER qwen3.8

Both can still 500 in Hermes if you leave Ollama’s default 4096 context.

After every HF pull:

ollama show hf.co/OWNER/REPO:QUANT --modelfile | grep -E '^(FROM |RENDERER |PARSER |PARAMETER num_ctx|TEMPLATE )'

If you see RENDERER qwen3.8, make a -hermes alias (FROM that HF tag, RENDERER qwen3.5, PARAMETER num_ctx 65536). If you do not see qwen3.8, still add num_ctx 65536 before using it as a Hermes agent.

Quick check before testing a new model¶

# 1. Renderer (Cause A)
ollama show MODEL --modelfile | grep -E '^(RENDERER|PARAMETER num_ctx)'

# 2. Hermes-shaped tool continuation (Cause A)
curl -s -w "\nHTTP %{http_code}\n" http://127.0.0.1:11434/api/chat -d "{
  \"model\":\"MODEL\",\"stream\":false,
  \"messages\":[
    {\"role\":\"assistant\",\"content\":\"\",\"tool_calls\":[{\"function\":{\"name\":\"web_search\",\"arguments\":{\"q\":\"test\"}}}]},
    {\"role\":\"tool\",\"tool_name\":\"web_search\",\"content\":\"{\\\"results\\\":[]}\"}
  ]
}"

HTTP 500 + no user query found on that tiny payload → Cause A. Make a -hermes alias with RENDERER qwen3.5.
HTTP 200 on that payload, but Hermes still 500s on a real query → Cause B. Add PARAMETER num_ctx 65536 and set Hermes context to 64000.

Rule of thumb for Hermes testing¶

Never point Hermes at a stock qwen3.8:* tag. Use a -hermes alias.
Every model you use as an agent gets num_ctx 65536 (or higher if VRAM allows) and Hermes context 64000.
After /model, start a new session.
If a new Qwen 3.8-class model lands, run the curl above before trusting it in Hermes.

Last updated: 2026-09-12