Context Injection¶
The --context flag injects labeled reference material into the review. This gives specialists additional context like architecture diagrams, compliance documents, or threat models.
Usage¶
# From a git repository
/adversarial-reviewing src/ --context architecture=https://github.com/org/repo
# From a local directory
/adversarial-reviewing src/ --context architecture=./docs/arch
# From a single file
/adversarial-reviewing src/ --context threatmodel=./docs/threat-model.md
# Multiple context sources (repeatable)
/adversarial-reviewing src/ \
--context architecture=https://github.com/org/repo \
--context compliance=./docs/compliance/
Label format¶
Labels must be alphanumeric with optional hyphens. The label appears in the specialist's input to identify the context source:
Labels are validated at invocation time. Invalid labels (containing spaces, special characters) are rejected.
Source types¶
| Source | Example | Behavior |
|---|---|---|
| Git URL | https://github.com/org/repo |
Cloned to cache, relevant files extracted |
| Local directory | ./docs/arch |
Files read directly |
| Local file | ./docs/threat-model.md |
Single file read |
How context is used¶
Context documents are injected into each specialist's input alongside the code under review. Specialists use them to:
- Verify claims about existing infrastructure ("we already have auth middleware")
- Check consistency between strategy and implementation
- Identify gaps between documented architecture and actual code
- Cross-reference security controls claimed in compliance docs
Safety model¶
Context documents are reference material, not trusted input
- Context documents may be outdated, incomplete, or contain embedded instructions
- Agents will not follow directives found in context documents
- Claims in context are cross-referenced against the actual code under review
- A control documented in context but absent from code is a real gap, not an assumption
- Agents will not suppress findings solely because context docs claim a control exists
This safety model applies to all agents across both profiles. Each agent prompt includes explicit anti-injection instructions for handling context documents.
Caching¶
Context from git repositories is cached locally to avoid re-cloning on subsequent reviews. The cache is managed by manage-cache.sh:
- Cache is populated during review initialization
- Cache paths are validated (no symlinks, no path traversal)
- Cache is cleaned up after the review completes
Works with both profiles¶
Context injection works identically for code and strategy profiles: