Skip to content

feat: Integrate Anthropic plugin domain knowledge as 5 document-analysis subagents #20

Description

@Number531

Summary

Integrate domain knowledge from Anthropic's open-source enterprise plugins (anthropics/knowledge-work-plugins, anthropics/financial-services-plugins) into 5 new on-demand document-analysis subagents + 1 financial-analyst enhancement.

Spec: docs/pending-updates/plugin-updates.md

New Agents (5)

Agent Purpose MCP Domains
nda-triage-analyst 10-point NDA screening → GREEN/YELLOW/RED classification None (document analysis only)
contract-review-analyst Clause-by-clause playbook review + redline generation None (document analysis only)
legal-risk-assessment-analyst 5×5 severity-likelihood risk matrix → risk register CourtListener
compliance-analyst GDPR/CCPA/DPA audit checklists FTC, Federal Register
cim-builder 8-section Confidential Information Memorandum generation SEC, financial-disclosure, BLS, code-execution

Enhancement: financial-analyst — append merger model (accretion/dilution) + deal tracker methodology

Feature Flag

DOCUMENT_ANALYSIS_PLUGINS=false (default off). Agent count: 40 → 45 when enabled. Zero impact to existing memo pipeline.

Files (14 total)

  • 7 new: 5 agent definitions + 1 test file + 3 structured output schemas (in existing file)
  • 7 modified: featureFlags.js, _promptConstants.js, financial-analyst.js, index.js, domainMcpServers.js, hookSSEBridge.js, sdkHooks.js

Integration Points (all covered in spec)

  • Feature flag design (MODULAR_SUBAGENTS dependency noted)
  • Agent definitions with trigger keywords + output path overrides
  • Structured output schemas (NDA_TRIAGE, CONTRACT_REVIEW, RISK_ASSESSMENT)
  • GATE_CHECK_CONFIGS with stateFile/stateFileDir fields
  • findSessionDirectory() update for document-analysis/ path detection
  • classifyAgent() entries before broad analyst catch-all
  • classifyDocument() handler for document-analysis/ directory
  • SUBAGENT_DOMAIN_MAP entries with rationale
  • REPORT_SAVING_INSTRUCTIONS path override (specialist-reports/ → document-analysis/)
  • Per-agent prompt constant table (MCP_FALLBACK, DATABASE_URL_TEMPLATES)
  • Test plan (~25 assertions across 7 groups)

Source

Apache 2.0 licensed SKILL.md files from Anthropic enterprise plugins (Feb 24, 2026 launch). Domain knowledge extracted and restructured into subagent system prompts — not used as plugins (incompatible execution model for server-side orchestration).

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions