Agents

Agents in security-assessment are all judgment agents — they are never invoked directly by the user, but are dispatched by the plugin's orchestrating commands during assessment phases.

For how these agents sequence into the pipeline, see Workflows and the authoritative phase graph in security-assessment-pipeline/SKILL.md.

Judgment Agents

Agent	File	Purpose	Invocation
authorization-logic-review	authorization-logic-review.md	Top-down authorization review. Maps the access-control model (RBAC/ABAC/ACL/tenancy), then verifies enforcement at every layer.	Dispatched by /security-assessment Phase 1b
business-logic-domain-review	business-logic-domain-review.md	Business-logic review for ML/fraud services. Detects fail-open, score manipulation, and other domain-specific anti-patterns.	Dispatched by /security-assessment Phase 1b
compliance-edge-annotator	compliance-edge-annotator.md	Edge annotator for compliance findings whose pattern row has llm_review_trigger=true. Refines pattern-table citations; never invents new findings.	Dispatched by compliance-mapping skill
cross-repo-synthesizer	cross-repo-synthesizer.md	Synthesizes attack-chain narratives from multi-repo RECON + shared-cred matches + service-comm diagram. Produces named attack chains.	Dispatched by /cross-repo-analysis
deep-code-reasoning	deep-code-reasoning.md	Context-aware vulnerability detection beyond static patterns. RECON-scoped freeform reasoning about IDOR, confused deputy, TOCTOU, privilege escalation, and similar classes.	Dispatched by /security-assessment Phase 1b
exec-report-generator	exec-report-generator.md	Synthesizes the publication-ready executive report from upstream artifacts. Emits a 7-section per-repo report plus cross-repo summary for multi-target runs.	Dispatched by /security-assessment Phase 5 and /cross-repo-analysis
fp-reduction	fp-reduction.md	Applies the six-stage FP-reduction rubric to a unified-finding stream, producing a disposition register with confidence field.	Dispatched by false-positive-reduction skill (Phase 2)
recon-driven-scan	recon-driven-scan.md	Bridges RECON narrative risk claims to file:line evidence. Emits findings only when the source actually exhibits the described pattern.	Dispatched by /security-assessment Phase 1b
redteam-evasion-analyzer	redteam-evasion-analyzer.md	Interprets probe 05 (evasion) alongside probe 03 (sensitivity) and probe 04 (boundaries). Rates adversarial realism and explains evasion mechanisms.	Dispatched by /redteam-model Phase 4 (parallel with other analyzers)
redteam-extraction-analyzer	redteam-extraction-analyzer.md	Interprets probe 07 (model extraction) alongside probe 03 (sensitivity). Translates R² into extraction fidelity and extracts decision-rule signatures.	Dispatched by /redteam-model Phase 4 (parallel with other analyzers)
redteam-recon-analyzer	redteam-recon-analyzer.md	Interprets probe 01 (API recon). Severity-rates info leaks, identifies the framework, and recommends a feature-discovery strategy for subsequent probes.	Dispatched by /redteam-model Phase 4 (parallel with other analyzers)
redteam-report-generator	redteam-report-generator.md	Refines the red-team adversarial-report.md into an executive document. Assigns RED/AMBER/GREEN rating and produces remediation with effort estimates.	Dispatched by /redteam-model Phase 5
tool-finding-narrative-annotator	tool-finding-narrative-annotator.md	Weaves findings into four narrative domains (PII flow, ML edge cases, messaging auth, crypto). Produces prose for the executive report.	Dispatched by security-assessment-pipeline Phase 3