Copilot Chat Participant

Overview

The extension registers a Chat Participant (@weapon) in GitHub Copilot Chat. Users type @weapon in the Copilot Chat panel to get context-aware pentest assistance. The participant reads engagement state (hosts, credentials) from Context and builds prompts that include this context alongside the user's query.

Gated by weaponized.ai.enabled (default true). Requires GitHub Copilot.

---

Registration

registerAIFeatures() in src/features/ai/index.ts creates the participant:

• Participant ID: weapon.chat
• Handler: weaponChatHandler — receives the request, builds context-enriched prompts, streams responses
• Followup provider: suggests relevant next actions based on the command just used
• Icon: images/icon.png

Registered in the activation sequence inside a try/catch — failure doesn't block other subsystems.

---

Commands

Users invoke commands via @weapon /command:

Command	Purpose	Behavior
/analyze	Analyze tool output	Asks LLM to extract findings, next steps, and commands from pasted output
/suggest	Suggest next steps	Asks LLM for 3-5 prioritized actions based on current engagement state
/generate	Generate commands	Asks LLM to output exact commands using $TARGET, $RHOST, etc.
/explain	Explain a concept	Asks LLM to explain a tool, technique, or concept in pentest context
/report	Engagement summary	Renders a markdown table of hosts and credentials directly (no LLM call)
(default)	Free chat	Passes the user's query with engagement context to the LLM

---

Prompt Construction

Each request builds a message array:

1. System prompt — role definition, available tools, guidelines (from prompts/systemPrompt.ts)
2. Engagement context — current hosts and credentials, formatted by prompts/hostContext.ts and prompts/userContext.ts
3. Referenced content — any files or locations the user attached via #file or #selection
4. Task prompt — command-specific instructions + user's query

The LLM is selected via vscode.lm.selectChatModels({ vendor: "copilot", family: "gpt-4o" }). Responses are streamed chunk-by-chunk to the chat panel.

---

Security

• Credential redaction: AIService.redactCredentials() replaces passwords and NT hashes with [REDACTED] in any text before sending to the LLM
• Context metadata: host/user context includes structure (hostname, IP, login) but buildUserContext shows auth type ("password" / "NT hash") without the actual values
• No tool execution: the participant only generates text — it does not run commands. Users must copy and execute commands themselves.

---

Followup Flow

After each command, the followup provider suggests logical next steps:

• After /analyze → "Suggest next steps" or "Generate commands"
• After /suggest → "Generate the first suggested command"
• After /generate → "Explain what this command does"

---

File Structure

src/features/ai/
  index.ts              — registerAIFeatures() entry point
  participant.ts        — Chat handler, command routing, LLM interaction
  service.ts            — AIService: state access + credential redaction
  prompts/
    systemPrompt.ts     — Base system prompt text
    hostContext.ts       — Formats hosts into context string
    userContext.ts       — Formats credentials into context string (redacted)