RedAmon on Proxmox.
From deployment to agent debugging.

Objective and outcome

The objective was to evaluate whether RedAmon could operate in an existing, memory-constrained Proxmox home lab without purchasing additional server hardware or hosting a language model inside the RedAmon VM.

Two-machine architecture

Application services run on Proxmox; model inference remains on a separate Windows GPU machine. The LLM connection uses Ollama's OpenAI-compatible endpoint over the local network.

Deployment sequence

# Initial deployment
cd ~/redamon
./redamon.sh install

# Post-install checks
./redamon.sh status
free -h
docker stats --no-stream

# Verify Ollama reachability from the guest
curl http://192.168.1.54:11434/api/tags

Secrets and administrator passwords are intentionally excluded from this portfolio artifact.

Problems encountered and resolution

# Persistent build session and retained log
tmux new -s redamon-install
cd ~/redamon
set -o pipefail
./redamon.sh install 2>&1 | tee ~/redamon-install.log

# After a disconnect
tmux attach -t redamon-install

# Investigate a failed build
grep -iE 'error|failed|ECONNRESET|ETIMEDOUT' \
  ~/redamon-install.log | tail -40

Container and resource validation

RedAmon's status command reported nine healthy containers, no OOM kills, no restarts, and no memory-cap drift at the time of validation.

First project: markcardiff.me

A project named First Test was created for the operator-owned static website markcardiff.me, hosted on Azure Storage static website infrastructure behind Azure Front Door. The initial reconnaissance pipeline ran and populated Neo4j. This establishes that the non-agent reconnaissance path can collect and persist basic asset relationships.

Performance during reconnaissance

With the reconnaissance workload active, the guest showed approximately 3.3 GiB used RAM and 4.5 GiB available. A vmstat 2 5 sample showed zero swap-in and swap-out activity. Proxmox's high displayed memory percentage largely reflected Linux caching rather than evidence of acute memory pressure.

# Useful point-in-time checks during a controlled run
free -h
vmstat 2 5
docker stats --no-stream
./redamon.sh status

When the model can describe a tool but cannot invoke it

The AI Agent repeatedly reported that no reconnaissance data had been gathered, despite the populated project graph. Prompts requesting graph inspection and later Subfinder enumeration produced plausible reasoning in the UI, but no corresponding successful tool execution was observed.

Reproducible log evidence

Parse attempt 1/3 failed: Validation error for LLMDecision
Input should be 'use_tool', 'plan_tools', 'transition_phase',
'complete', 'ask_user', 'deploy_fireteam' or 'switch_skill'
input_value='execute_subfinder'

Decision: action=use_tool, tool=None
THOUGHT: Starting with subdomain enumeration using subfinder
ACTION: use_tool
No valid action in decision: use_tool, tool: None

The same failure pattern was reproduced in two attempts on 30 October 2026. The AI UI showed a Subfinder plan followed by “Preparing response...” and a message that the scan had not begun.

Source-level trace

Diagnostic commands used

# Find decision schema, parser and routing code
grep -Rnl --include='*.py' 'class LLMDecision' agentic
grep -Rnn --include='*.py' 'def try_parse_llm_decision' agentic
grep -Rnn --include='*.py' -E 'No valid action in decision|Parse attempt .*failed' agentic

# Correlate user-visible agent activity with actual tool decisions
docker logs redamon-agent --since 15m 2>&1 |   grep -E 'Decision:|ACTION:|TOOL:|No valid action|Parse attempt|Executing tool|subfinder' |   tail -60

Settings discovered during follow-up

In a new project, the Enable AI in Pipeline master switch was observed OFF. Its tooltip indicates it controls per-tool AI hooks within the reconnaissance pipeline. This setting is distinct from the AI Agent, which was demonstrably making Ollama calls already. Domain Verification was also OFF, and the Engagement Limits panel indicated no active limits. For future controlled runs, domain verification and rate ceilings warrant deliberate configuration.

Resume with controlled validation

Deployment and initial reconnaissance are complete. The next milestone is a single, observable AI Agent tool call, not a broad autonomous assessment.

Engineering lessons from deployment and investigation