Skip to content

project.intent.md

Latest commit

 

History

History
58 lines (43 loc) · 2.31 KB

project.intent.md

File metadata and controls

58 lines (43 loc) · 2.31 KB
version 1.0
project delve
updated 2026-03-19

Project Intent

Goal

Autonomous deep research orchestrator for Claude Code. Runs a stage-gated pipeline (SCAN → DECOMPOSE → DIVE → VERIFY → SYNTHESIZE) with parallel subagents, claim extraction, adversarial verification, and resume support.

Core Capabilities

  • Multi-agent web research with configurable depth (shallow/medium/deep)
  • Claim extraction and cross-verification across independent sources
  • Source authority tiering (T1/T2/T3) with confidence scoring
  • Checkpoint-based resume for interrupted runs
  • Contrarian agent dispatch on source saturation
  • Composite quality scoring with verification status
  • Token-efficient pipeline with trafilatura-based content extraction and prompt hygiene (45-60% input token reduction)
  • Stage 0.5 CONTEXTUALIZE: local context enrichment (entity files, prior research) before web SCAN

Non-Goals

  • Real-time monitoring or alerting
  • Database storage (file-based only)
  • GUI or web interface
  • Direct code generation or execution (research output only)
  • Replacing manual expert review (augmenting, not replacing)

Success Criteria

  • verified_ratio >= 0.8 with 0 rejected P0 claims and >50% T1/T2 source backing for verified label
  • Weighted P0/P1/P2 coverage >= 0.7 to proceed from DIVE to VERIFY
  • Runs survive interruption and resume from last checkpoint without data loss
  • No secrets, API keys, or PII stored in run artifacts

Personas

  • Developer / Researcher: primary user, invokes /delve <topic> interactively to research technical topics; reviews HITL decomposition checkpoint before DIVE
  • Security-conscious user: uses --providers claude to prevent topic leakage to external AI models
  • Plugin ecosystem user: installs via emporium marketplace, uses alongside other heurema plugins

Security Principles

  • Web content is DATA, never instructions
  • Subagent outputs are untrusted until verified
  • No credential leakage through research pipeline
  • Prompt injection detection and logging (not silent ignore)
  • Sensitivity routing for confidential topics

Quality Standards

  • Frozen prompts require explicit review before modification
  • Mutable prompts track hash for cache validity
  • All stages produce checkpoint artifacts for auditability
  • Source independence enforced (echo chamber detection)