d-research-ultra

name

d-research-ultra

description

Prebuilt multi-agent deep research skill for AI agents. Use for source-backed web research, public-data collection, literature reviews, due diligence, policy or standards analysis, market or technical research, evidence ledgers, execution gates, blocker reports, and audit-grade synthesis. Ships the D Research core methodology plus a runtime-neutral orchestrator and six bundled worker-agent role definitions. Uses ephemeral parallel or sequential workers when the host supports them, falls back to manual role checklists when it does not, and creates persistent workers only when the user explicitly requests installation.

D Research Ultra

D Research Ultra is the prebuilt multi-agent distribution of D Research. It keeps the D Research core methodology self-contained in this package and adds a runtime-neutral orchestrator plus six bundled worker roles.

Use this skill when the user needs source-backed research with a clear audit trail, especially when the task is broad, high-recall, contested, long-horizon, multilingual, or evidence-sensitive.

Do not use it to bypass login walls, paywalls, captchas, anti-bot systems, robots restrictions, rate limits, private groups, private profiles, or access controls. Blocked sources become blocker reports, not escalation attempts.

Package Model

This repository is a standalone skill package:

• Core D Research methodology: AGENTS.md, references/, templates/, scripts/, adapters/, examples/, and docs/.
• Ultra multi-agent layer: agents/orchestrator.md, agents/spawn-contract.md, agents/manifest.json, and the six role definitions in agents/subagent-roles/.
• Optional helper scripts remain small, local, and auditable. They support the workflow but do not replace the agent.

D Research Ultra is not tied to any specific CLI, vendor, or runtime. The main agent delegates work through whichever subagent, task, worker, background-agent, or agent-roster mechanism the host provides. If the host has no real subagent mechanism, the main agent runs the same checklists in single-agent mode and discloses that limitation.

When To Use

Use this skill for:

• deep web research and source discovery
• public web data collection and structured extraction
• academic, literature, scoping, rapid, or PRISMA-style reviews
• market, competitor, product, and technical research
• due diligence, red-flag, risk, or public-claim review
• policy, standards, RFC, governance, and compliance analysis
• creative, cultural, media, trend, reception, and archive research
• public URL analysis and lawful public snapshot capture
• fact verification and contradiction checks
• evidence-ledger, blocker-report, and reproducibility workflows
• long-horizon research that benefits from isolated worker contexts

Refuse or reframe requests for private-person profiling, doxxing, stalking, harassment, deanonymization, private data, or access-control bypass.

Execution Modes

Classify the request first with references/research-intake.md, then choose the lightest mode that can meet the evidence standard.

Fast Mode

Use for one atomic fact, one URL, one clearly scoped source, or a quick answer where completeness is not requested.

• Main agent runs the D Research core workflow directly.
• Optionally delegate only to Evidence Verifier if the host has a cheap verifier subagent and the claim is important.
• Skip the full six-agent pipeline unless the user requests audit-grade depth.

Standard Mode

Use for normal multi-source research.

• Main agent performs intake and source-map planning.
• Delegate Source Mapper, Data Extractor, Evidence Verifier, and Report Synthesizer when real worker agents are available.
• Apply references/execution-gates.md before final synthesis.

Completeness-First Mode

Use when the user asks for audit-grade work, "full pipeline", "be thorough", due diligence, contested claims, low-recall topics, or a long-horizon task.

Run the six-worker pipeline in waves:

1. Wave 1, in parallel when the host supports it: Source Mapper and Public Web & Community Hunter.
2. Wave 2, in parallel after the first source batch: Recall Auditor and Data Extractor.
3. Wave 3: Evidence Verifier.
4. Wave 4: Report Synthesizer.

If the host only supports sequential subagents, run the same waves in order. If a worker fails or is unavailable, run that role's checklist manually and disclose the fallback.

Worker Roster

The portable worker definitions live in agents/subagent-roles/. Use the role IDs and names from agents/manifest.json as the canonical roster.

Role ID	Canonical worker name	Mission
source-mapper	D Research Source Mapper	Build source maps, search matrices, candidate URLs, blockers, identity risks, and next-query plans.
recall-auditor	D Research Recall Auditor	Run adversarial second-pass recall for missed sources, alternate language/register paths, archives, mirrors, and contradictions.
public-web-community-hunter	D Research Public Web & Community Hunter	Find lawful public web, community, forum, official social, video, and archive leads with strict privacy and identity labels.
data-extractor	D Research Data Extractor	Extract auditable structured data from public files, APIs, tables, embedded JSON, metadata, PDFs, OCR, and visible text.
evidence-verifier	D Research Evidence Verifier	Verify atomic claims against exact URLs, primary sources, contradiction checks, staleness checks, confidence, and ledger rows.
report-synthesizer	D Research Report Synthesizer	Turn verified findings into source-backed reports with caveats, blockers, gaps, confidence, and no unsupported claims.

Read agents/orchestrator.md for the full dispatch procedure and agents/spawn-contract.md for the host-runtime contract.

Runtime-Neutral Spawn Contract

When a real worker mechanism exists, the main agent should:

1. Prefer ephemeral, task-scoped workers initialized from the bundled role definitions.
2. Run independent workers in parallel when the host supports it.
3. Otherwise run the same workers sequentially while preserving wave dependencies.
4. Request compact, structured final outputs from each worker.
5. Merge only returned evidence and explicit caveats, not hidden worker assumptions.
6. Fall back to manual role checklists when no real worker can run.

Do not create persistent worker files or registrations by default. Persistent installation is an optional optimization that requires an explicit user request and an explicit host scope. Dynamic workers inherit the host's default model unless the user or runtime configuration selects another model.

Do not hard-code one vendor's command names in the core skill. Runtime adapters may live outside this core layer, but the skill body should remain portable.

Core Workflow

For single-agent mode and all fallback paths, follow AGENTS.md. Important references:

• Intake and routing: references/research-intake.md
• Source discovery: references/source-discovery.md
• Query fanout: references/query-patterns.md
• Browser-first crawl: references/browser-first-crawl.md
• API workflow: references/api-access-workflow.md
• Extraction: references/data-extraction-toolbox.md
• Multi-format extraction: references/multi-format-extraction.md
• Evidence ledger: references/evidence-ledger.md
• Execution gates: references/execution-gates.md
• Report generation: references/report-generation.md
• Reproducibility: references/reproducibility-checklist.md
• Tool adapter policy: references/tool-adapter-policy.md
• Long-horizon planning: references/research-plan-protocol.md

For specialized branches, load only the relevant reference file:

• atomic facts: references/fact-verification.md
• public posts: references/social-media-archival.md
• public-role named-person research: references/person-aggregation.md
• systematic reviews: references/systematic-review-protocol.md
• research bibliography: references/research-bibliography.md
• multilingual/local research: references/multilingual-research.md
• Vietnamese source discovery: references/vietnamese-source-discovery.md
• register/jargon recall: references/register-and-jargon-expansion.md
• frontier search: references/frontier-search.md

Output Standard

Final answers should separate:

• verified facts
• source-backed inference
• unknowns and gaps
• blockers and manual retrieval paths
• contradictions or stale-source risks
• confidence by claim or section
• exact source URLs or local artifact paths

Do not present a result as complete until source mapping, recall, coverage, evidence verification, blockers, and confidence have been handled or explicitly marked out of scope.