Delve

Autonomous deep research with claim-level verification

claude plugin marketplace add heurema/emporium
claude plugin install delve@emporium

What it does

Research tasks need more than a single web search. Delve runs a stage-gated pipeline that scans for sources, decomposes the topic into sub-questions, dispatches parallel research subagents, then verifies every claim through adversarial fact-checking before producing a synthesis report. Each stage checkpoints to disk so interrupted runs resume from where they left off.

Install

claude plugin marketplace add heurema/emporium
claude plugin install delve@emporium

Manual install from source

git clone https://github.com/heurema/delve
cd delve
claude plugin install .

Quick start

# Full research pipeline (default: medium depth, 3-4 agents)
/delve "WebSocket vs SSE for real-time updates"

# Fast scan + synthesize (skip dive & verify)
/delve "OAuth 2.1 changes" --quick

# Maximum depth with 6 parallel agents
/delve "Rust async runtimes" --depth deep

Commands

Command	Description
/delve <topic>	Full research pipeline
/delve <topic> --quick	Scan + synthesize only
/delve <topic> --depth <level>	shallow (2) / medium (4) / deep (6) agents
/delve <topic> --providers claude	Single-model mode, no external AI
/delve <topic> --output <path>	Custom output location
/delve resume [run_id]	Resume interrupted run
/delve status	List recent runs

Pipeline

SCAN -> DECOMPOSE -> DIVE -> VERIFY -> SYNTHESIZE
 |          |          |        |          |
 |          |          |        |          +- Merge + annotate -> synthesis.md
 |          |          |        +- Adversarial claim verification
 |          |          +- Parallel research subagents
 |          +- Topic decomposition + HITL approval
 +- Web search + existing research check

Depth levels

Depth	Agents	Verify	Time
--quick	0	skip	~60-90s
shallow	2	top 30%	~3-5 min
medium	3-4	all factual	~6-10 min
deep	5-6	all + cross	~10-20 min

Quality model

Two independent axes:

• Verification: verified / partially-verified / unverified
• Completion: complete / incomplete / synthesis_only / draft

Reports include a methodology section with exact quality labels, claim counts, and source attribution.

Resume

Runs checkpoint after each stage to ~/.cache/delve/runs/. Resume interrupted runs:

/delve resume              # pick from recent runs
/delve resume <run_id>     # resume specific run

Requirements

• Claude Code with skill support
• WebSearch and WebFetch tools available (for SCAN and DIVE stages)
• Git repository context (for saving reports to docs/research/)

Privacy

Delve uses Claude Code subagents and built-in WebSearch/WebFetch tools. Your research topic and web content are processed through Claude's standard API under your existing authentication. DIVE agents use an optional Python script (scripts/fetch_clean.py) to extract clean article text via the trafilatura library — this makes direct HTTP requests from your machine to fetch web pages, separate from Claude Code's WebFetch. If the script is unavailable (no uv installed) or fails, agents fall back to WebFetch automatically. When --providers claude is active, no external AI models are invoked and the topic is redacted in local event logs.

Feedback

If the verifier returns incorrect verdicts, the pipeline stalls, or resume doesn't work — file it from Claude Code with Reporter:

claude plugin install reporter@emporium

Then: /report bug or /report feature or /report question

See also

• skill7.dev/plugins/delve — plugin page and changelog
• github.com/heurema/emporium — plugin registry
• docs/how-it-works.md — architecture, data flow, verification model
• docs/reference.md — all commands, options, output format, troubleshooting

License

MIT