Bug: Context compressor ignores reasoning field — empty summaries with thinking models (Ollama 0.22+)

Summary

context_compressor.py reads only response.choices[0].message.content (line 871) when extracting the summary from the auxiliary LLM. When using thinking/reasoning models (DeepSeek v4, GLM-5.1, Qwen 3, etc.) via Ollama 0.22+, these models return their output in the reasoning field with content as an empty string — especially when max_tokens is constrained (which compression does). The compressor gets an empty summary and falls back to a static context marker, losing the entire compaction.

Root Cause

Ollama 0.22.x changed how thinking models return responses. The reasoning field is now always populated for models with a thinking renderer (DeepSeek v4, GLM-5.1, Qwen 3.5, etc.), even without think: true. With limited max_tokens, the model's reasoning tokens consume the entire budget, and content comes back empty.

Hermes already has extract_content_or_reasoning() in auxiliary_client.py (line 3561) that handles this exact case — it checks content first, then falls back to reasoning, reasoning_content, and reasoning_details. But context_compressor.py doesn't use it.

Affected Code

agent/context_compressor.py line 871:

content = response.choices[0].message.content  # BUG: ignores reasoning field

Should be:

from agent.auxiliary_client import call_llm, extract_content_or_reasoning
# ...
content = extract_content_or_reasoning(response)

Reproduction

1. Configure Hermes with Ollama 0.22.1+ and a thinking model (e.g., deepseek-v4-flash:cloud or glm-5.1:cloud) as the compression auxiliary
2. Have a conversation long enough to trigger context compression
3. Observe that the compression summary is empty — the model's output went entirely to reasoning while content is ""
4. Compressor falls back to a static context marker: "Summary generation failed — inserting static fallback context marker"

Evidence

In our deployment, 3 of 25 compression events produced fallback markers (12% failure rate). After upgrading to Ollama 0.22.1, testing with deepseek-v4-flash:cloud and max_tokens: 100 returns:

• content: "" (empty)
• reasoning: "..." (928 chars of actual summary content)

With max_tokens: 300, content returns 75 chars while reasoning takes 929 chars — the reasoning field holds the actual compressed content the compressor needs.

Impact

• All thinking models used for compression produce empty summaries when max_tokens is insufficient to cover both reasoning and content
• The Caveman Compressor plugin also inherits this bug via super()._generate_summary()
• No recovery path — the fallback marker is static text, not a summary

Fix

One-line change + one import:

# Line 27:
from agent.auxiliary_client import call_llm, extract_content_or_reasoning

# Line 871:
content = extract_content_or_reasoning(response)

The extract_content_or_reasoning() function already exists and handles content, reasoning, reasoning_content, and reasoning_details with appropriate fallback logic and inline think-tag stripping.

Environment

Machine 1 (Hazel):

• Hermes Agent: 0b76d23 (2026-04-30), 39 commits behind upstream main
• OS: Ubuntu 25.10 (Questing Quokka), kernel 6.17.0-23-generic, x86_64
• CPU: 12th Gen Intel i7-12700H (14 cores: 6P+8E)
• GPU: NVIDIA GeForce RTX 3060 Laptop 6GB, driver 595.58.03
• RAM: 32 GB (2x16 GB DDR5, 30.5 GiB usable after kernel reservation)
• Python: 3.13.7
• Model: glm-5.1:cloud via custom provider (Ollama, http://127.0.0.1:11434/v1)
• Ollama: 0.22.1
• Models affected in testing: deepseek-v4-flash:cloud, glm-5.1:cloud, qwen3.5-397b-cn-think:latest

Machine 2 (Ember/Nova):

• Hermes Agent: 0b76d23 (2026-04-30), 39 commits behind upstream main
• OS: Ubuntu 26.04 LTS, kernel 7.0.0-15-generic, x86_64
• CPU: AMD Ryzen 9 5900XT 16-Core
• GPU: NVIDIA RTX PRO 4000 Blackwell 24GB, driver 595.58.03
• RAM: 64 GB (60.7 GiB usable after kernel reservation)
• Python: 3.14.4
• Model: glm-5.1:cloud via custom provider (Ollama, http://127.0.0.1:11434/v1)
• Ollama: 0.22.1
• Same models affected

Both machines have the local fix applied and verified. The bug is reproducible on stock Hermes Agent without the patch.

Related Issues

• [Bug] Thinking model (glm-5-turbo) reasoning tokens exhaust output budget, producing empty responses with no recovery path #9344 — Broader "thinking model reasoning tokens exhaust output budget" bug in the main agent loop. This issue ([Bug] Context compressor ignores reasoning field — empty summaries with thinking models (Ollama 0.22+) #19003) is a specific instance of that pattern in the compressor subsystem.