Profile System¶
The profile system enables the same multi-agent framework to review different types of artifacts (code vs. strategy documents) with appropriate specialists, templates, and reference modules.
Profile structure¶
Each profile lives in profiles/<profile>/ and contains:
profiles/
code/
config.yml # Profile configuration
agents/ # Specialist prompts
security-auditor.md
performance-analyst.md
code-quality-reviewer.md
correctness-verifier.md
architecture-reviewer.md
devils-advocate.md
templates/ # Output format definitions
finding-template.md
challenge-response-template.md
report-template.md
...
references/ # Knowledge base modules
security/
owasp-top10-2025.md
agentic-ai-security.md
...
strat/
config.yml
agents/
feasibility-analyst.md
architecture-reviewer.md
security-analyst.md
user-impact-analyst.md
scope-completeness-analyst.md
testability-analyst.md
devils-advocate.md
templates/
finding-template.md
challenge-response-template.md
report-template.md
requirements-template.md
references/
all/
rhoai-platform-constraints.md
rhoai-auth-patterns.md
productization-requirements.md
architecture/
security/
operations/
config.yml¶
Each profile's config.yml defines:
name: code
description: Source code review
agents:
- tag: SEC
name: Security Auditor
file: security-auditor.md
flag: --security
- tag: PERF
name: Performance Analyst
file: performance-analyst.md
flag: --performance
# ...
quick_preset:
agents: [SEC, CORR]
iterations: 2
budget: 150000
thorough_preset:
agents: all
iterations: 3
budget: 800000
default:
iterations: 2
budget: 350000
The profile configuration is read by profile-config.sh and used by SKILL.md to determine which agents to spawn, which templates to use, and which presets apply.
How profiles are selected¶
- Default:
codeprofile - With
--profile strat:stratprofile - Profile determines: agent list, templates, references, presets, validation rules
Key differences between profiles¶
| Aspect | Code | Strategy |
|---|---|---|
| Evidence type | file:line references |
Text citations (section, paragraph) |
| Findings have | Source Trust field | Verdict field (Approve/Revise/Reject) |
| Report includes | Remediation roadmap | Per-document verdicts |
| Specialists | 5 | 6 |
| Reference scope | security/ only |
all/, architecture/, security/, operations/ |
| Supplementary outputs | None | Threat surface, NFR scan, requirements |
| Quick preset agents | SEC + CORR | SEC + FEAS |
Adding a new profile¶
To create a new profile (e.g., docs for documentation review):
- Create
profiles/docs/config.ymlwith agent definitions - Create agent prompts in
profiles/docs/agents/ - Create templates in
profiles/docs/templates/ - Add reference modules in
profiles/docs/references/(optional) - Update
profile-config.shif needed - Update SKILL.md to recognize the new profile flag
- Add validation rules to
validate-output.sh
The phase procedures (self-refinement, challenge, resolution, report) are profile-agnostic. They read the active profile's configuration and adapt automatically.
Shared components¶
These components work identically across profiles:
- Phases: Self-refinement, challenge, resolution, report, remediation
- Protocols: Isolation, injection resistance, mediated communication, convergence
- Scripts: Validation, deduplication, budget tracking, cache management
- Devil's advocate: Both profiles include one, with profile-specific context handling