vitya
962b0beddd
- README: remove autogen/crewai references, add v2 architecture diagram, code-review scenario, cross-LLM config examples, roadmap - wiki/overview: event-driven architecture, Pydantic v2, roadmap - wiki/index: add entities/roles, concepts/code-review, packages/pydantic - wiki/log: Phase 1 completion, code-review decision - New: entities/roles.md, concepts/code-review.md, packages/pydantic.md Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
45 lines
1.5 KiB
Markdown
45 lines
1.5 KiB
Markdown
---
|
|
title: Discussion Roles
|
|
type: entity
|
|
updated: 2026-05-04
|
|
---
|
|
|
|
# Discussion Roles
|
|
|
|
Roles define participants in a meeting-room discussion. Each role binds to a provider, model, temperature, tool set, and system prompt.
|
|
|
|
## General discussion roles
|
|
|
|
| Role | Purpose | Default tools |
|
|
|------|---------|-------------|
|
|
| moderator | Guide conversation, summarize, keep on-topic | none |
|
|
| skeptic | Find weaknesses, ask hard questions, counter-argue | readonly |
|
|
| idea_generator | Propose unconventional solutions, search web for inspiration | web |
|
|
| analyst | Structure arguments, assess risks, propose action plans | all |
|
|
|
|
## Code review roles
|
|
|
|
| Role | Purpose | Default tools |
|
|
|------|---------|-------------|
|
|
| defender | Defend code decisions, explain rationale | readonly |
|
|
| attacker | Ruthlessly find bugs, style violations, dead code | readonly |
|
|
| security | Check vulnerabilities, injection, CVEs | web |
|
|
| moderator | Structure review findings into CRITICAL/IMPORTANT/MINOR | none |
|
|
|
|
## Role configuration
|
|
|
|
```yaml
|
|
roles:
|
|
my_role:
|
|
name: "Display Name"
|
|
provider: openai # provider key from providers section
|
|
model: "gpt-4o"
|
|
temperature: 0.7 # 0.0 = deterministic, 1.0 = creative
|
|
tools: "readonly" # tool set name or list ["read_file", "web_search"]
|
|
system_prompt: |
|
|
Your instructions here...
|
|
```
|
|
|
|
## Cross-LLM review
|
|
|
|
Assign different `provider` values to roles so separate LLMs cross-check each other. The defender uses one model, the attacker uses another — each catches what the other misses. |