Configuration Files

protoPen uses several configuration files located in the config/ directory. This directory is bind-mounted into the container at /opt/protopen/config/, so changes on the host take effect after a container restart.

engagement-config.json

Primary configuration for pen testing operations: hardware devices, engagement defaults, and tool risk levels.

Path: config/engagement-config.json

Structure

{
  "instance": {
    "name": "protoPen",
    "description": "Mobile pen testing agent — Steam Deck + peripherals"
  },
  "devices": {
    "portapack": {
      "type": "serial",
      "serial_number": "",
      "fallback_port": "/dev/ttyACM0",
      "prompt": "ch>",
      "timeout": 3
    },
    "flipper": {
      "type": "serial",
      "serial_number": "",
      "fallback_port": "/dev/ttyACM1",
      "prompt": ">: ",
      "timeout": 3
    },
    "marauder": {
      "type": "serial",
      "serial_number": "",
      "fallback_port": "/dev/ttyUSB0",
      "baud_rate": 115200,
      "timeout": 3
    },
    "wifi_adapter": {
      "type": "network",
      "interface": "wlan1",
      "monitor_interface": "wlan1mon"
    }
  },
  "engagement": {
    "default_mode": "passive",
    "workspace_dir": "/home/deck/engagements",
    "alert_webhook": "",
    "alert_channel_id": ""
  },
  "tool_risk": {
    "rf_scan": 0,
    "wifi_deauth": 2,
    "...": "..."
  }
}

Key Sections

Section	Description
instance	Identity metadata for this protoPen node
devices	Serial connection settings for each hardware device
engagement	Default engagement mode, workspace directory, Discord alert webhook
risk_levels	Maps mode names to integer levels (passive=0, active=1, redteam=2)
tool_risk	Maps individual tool actions to their required risk level

---

langgraph-config.yaml

Configuration for the LangGraph agent backend (used when AGENT_BACKEND=langgraph).

Path: config/langgraph-config.yaml

Structure

model:
  provider: gateway
  name: claude-sonnet-4-6
  api_base: http://ava:4000/v1
  api_key: ""
  temperature: 0.3
  max_tokens: 4096
  max_iterations: 75

subagents:
  threat_scanner:
    enabled: true
    tools: [cve_search, security_feeds, github_trending, browser, security_memory]
    max_turns: 30
  vuln_analyst:
    enabled: true
    tools: [cve_search, security_feeds, security_memory, browser, target_intel]
    max_turns: 40
  # ... (intel_reporter, recon, exploit, reporter)

middleware:
  knowledge: true
  audit: true
  memory: true

knowledge:
  db_path: /sandbox/knowledge/security.db
  embed_model: qwen3-embedding
  top_k: 10
  search_mode: hybrid
  enrich_chunks: true
  inject_min_trust: 1   # trust floor for auto-injected memory (ADR 0069 D8): 1 = re-rank
                        # curated-first only; 2 = refuse tier-1 EXTERNAL (scraped/OSINT) memory

# Skills (SKILL.md methodology). Progressive disclosure (default on): a lightweight
# <available_skills> catalog (name + description) is injected each turn and the
# agent pulls a full body on demand via the load_skill tool — instead of injecting
# matched skill bodies. user_only skills are catalog-hidden, reachable only via
# /skill. Set progressive_disclosure: false for the legacy top-k full-body inject.
skills:
  enabled: true
  dir: /sandbox/skills
  db_path: /sandbox/skills.db
  top_k: 5
  progressive_disclosure: true

# Progressive tool disclosure (ADR 0005). OFF by default. protoPen binds ~80
# tools; with deferral on, most tool schemas are withheld from the model each
# turn (every tool stays callable) and the agent loads them on demand via the
# `search_tools` meta-tool. `keep` overrides the always-exposed base set
# (orchestration + task/schedule management + search_tools).
tools:
  deferred:
    enabled: false
    keep: []          # e.g. [task, run_workflow, create_task, search_tools]

# Goal mode (autonomy) — re-invoke the agent toward a verifier-backed condition
# (set with /goal in chat, or by the agent via the set_goal tool). Verifiers are
# read-only, no shell: findings / targets / task / llm. See explanation/autonomy.
goals:
  enabled: true
  max_iterations: 10
  no_progress_limit: 4
  dream_cadence_cron: ""      # 5-field cron to seed a recurring /dream pass (blank = off)

# Watches (ADR 0067) + working-state injection (ADR 0079) — the self-driving loop.
# N concurrent condition-watches, each on its own cadence, that run a follow-up turn
# on trip; the manager polls this often and evaluates only watches that are due.
watch:
  enabled: true
  poll_interval_s: 5
  working_state: true         # inject the per-turn <working_state> snapshot (OODA)

# Background sub-agents (ADR 0070). auto_resume: on completion a job self-wakes its
# origin session to brief the operator (coalesced) and indexes its result to the KB.
background:
  auto_resume: true

Key Sections

Section	Description
model	LLM provider, model name, API base, generation parameters
subagents	Per-subagent tool allowlists, enable/disable, max turns
middleware	Toggle knowledge injection, audit logging, memory consolidation
knowledge	Embedding model, vector search config, contextual enrichment, inject_min_trust (recalled-memory trust floor, ADR 0069 D8)
skills	SKILL.md retrieval — enable, dir, db path, top-k, progressive_disclosure (catalog + load_skill)
workflows	Declarative subagent recipes — enable, writable dir
tools.deferred	Progressive tool disclosure (ADR 0005) — enabled, keep
goals	Goal mode (autonomy) — enabled, max_iterations, no_progress_limit, dream_cadence_cron (scheduled memory hygiene)
watch	Condition-watches (ADR 0067) + working_state OODA injection (ADR 0079) — enabled, poll_interval_s, working_state
background	Background sub-agents (ADR 0070) — auto_resume (self-wake briefing + KB durability)

---

security-config.json

Default security topics and feed configuration seeded on startup.

Path: config/security-config.json

Contains a topics array (CVE tracking, wireless security, exploit techniques, etc.), each with name, keywords, and priority. Also includes feeds configuration for scan intervals and tracked_repos for security tool monitoring. Topics are seeded into the knowledge store on first launch.

---

SOUL.md

The agent's identity and personality prompt.

Path: config/SOUL.md

This file defines who protoPen is, its values, and behavioral guidelines. It is injected as the first section of the system prompt. Edit this file to customize the agent's personality.