{
  "markdown": "# PoC -- John Stawinski\n\n# deep-research\n\nMulti-agent research pipelines for Claude Code — stop leaving your session to go research elsewhere.\n\n## Who This Is For\n\nYou're mid-session, you need real research, and you don't want to break flow to Perplexity or ChatGPT and paste results back. Or you want to deeply understand an open-source repo before building on it. Or you need structured data on 20 competitors in a consistent schema.\n\nThese pipelines delegate research to agent teams so your top-level Claude stays free. Results come back as committed markdown in `docs/research/` — artifacts, not chat messages.\n\n## Install\n\nThis is a Claude Code plugin. The first-class install path is to hand the job to your agent — paste the prompt below into a Claude Code session and let it follow the playbook in [`docs/install.md`](docs/install.md).\n\n> Please install the deep-research-claude plugin from https://github.com/dbc-oduffy/deep-research-claude. The repo's `docs/install.md` is the install playbook — read it and follow the steps. Verify by running `/deep-research setup` at the end and report the result.\n\nAfter the agent finishes, restart Claude Code (so the Agent Teams env var takes effect) and try:\n\n```\n/deep-research web \"agent orchestration patterns in LLM frameworks\"\n```\n\n## Pipelines\n\n### Pipeline A — Internet Research\n\nHaiku scout builds a source corpus via web search. 3-5 Sonnet specialists deep-read sources, verify claims, and challenge each other adversarially. Opus sweep agent checks coverage, fills gaps, and writes the final document. An optional iterative deepening pass targets high-severity gaps identified in the first sweep.\n\n```\n/deep-research web \"topic\"\n```\n\n### Pipeline B — Repository Research\n\n2 Haiku scouts inventory every file in assigned chunks. 4 Sonnet specialists deep-read, analyze architecture, and optionally compare against a second project. Opus synthesizer writes the final assessment.\n\n```\n/deep-research repo /path/to/repo [--compare /path/to/mine] [--deeper] [--deepest]\n```\n\n- `--compare` — gap-analysis artifact comparing target repo to your project\n- `--deeper` — dependency-weighted repomap during scoping; specialists prioritize structurally central files\n- `--deepest` — adds a Sonnet atlas agent producing architecture artifacts (file index, system map, connectivity matrix)\n\n### Pipeline C — Structured Research\n\nSchema-conforming batch research across N entities. Haiku scout maps findings to schema fields. 1-5 Sonnet verifiers challenge each other's values (CONFIRMED / UPDATED / REFUTED / CONTESTED). Opus synthesizer resolves contested fields and outputs validated YAML/JSON.\n\n```\n/deep-research structured tasks/research/spec.yaml subject-key\n```\n\n### Pipeline D — NotebookLM Research\n\nResearch YouTube videos, podcasts, and media Claude can't access directly, via NotebookLM. Haiku scout ingests sources. Sonnet workers query on focused sub-questions. Opus sweep writes the synthesis.\n\nRequires the [notebooklm-mcp-cli](https://github.com/jacob-bd/notebooklm-mcp-cli) MCP server and a Google account with NotebookLM access.\n\n```\n/notebooklm-research \"topic\"\n```\n\n## Commands\n\n| Command | Pipeline |\n|---------|----------|\n| `/deep-research web <topic>` | Internet research with iterative deepening |\n| `/deep-research repo <path>` | Repository analysis (with optional `--compare`, `--deeper`, `--deepest`) |\n| `/deep-research structured <spec> <key>` | Schema-conforming batch research |\n| `/notebooklm-research <topic>` | NotebookLM media research |\n\nAll pipelines are fire-and-forget — the EM spawns the team and is freed. Results are committed to `docs/research/` automatically.\n\n## Agents\n\n| Agent | Model | Role |\n|-------|-------|------|\n| **research-scout** | Haiku | Web search, source vetting, shared corpus |\n| **research-specialist** | Sonnet | Source verification, adversarial peer challenges, structured claims |\n| **research-synthesizer** | Opus | Coverage check, gap-filling, final document |\n| **repo-scout** | Haiku | File inventory with signatures, constants, data flow |\n| **repo-specialist** | Sonnet | Architecture analysis, optional project comparison |\n| **structured-synthesizer** | Opus | Schema validation, contested field resolution, final YAML/JSON |\n\n## Integration with coordinator\n\nThis plugin works standalone. When used alongside the [coordinator plugin](https://github.com/dbc-oduffy/coordinator-claude), the EM automatically suggests research pipelines via a `PreToolUse` hook when Claude reaches for ad-hoc web search — nudging toward these structured pipelines instead of one-off `WebFetch` calls that consume the coordinator's context window.\n\n### Optional dependency: `coordinator-safe-commit`\n\nPipeline commands (`/web`, `/repo`, `/structured`) invoke `~/.claude/plugins/coordinator-claude/coordinator/bin/coordinator-safe-commit` for phase-end commits. The helper provides scoped staging (per-session audit-trail integrity) for users running multiple concurrent agent sessions on the same branch.\n\n**If you have the coordinator plugin installed,** no action needed — the helper is on the expected path.\n\n**If running deep-research standalone,** substitute either form when you encounter the command:\n- `git add <explicit-paths> && git commit -m \"<subject>\"` — manual scoped staging\n- `git add -A && git commit -m \"<subject>\"` — blanket staging (acceptable for solo single-session use)\n\n## Research Backing\n\nPipeline design derives from published guidance (OpenAI, Perplexity, Google, Anthropic, Stanford STORM) and is validated through [controlled experiments](docs/research/2026-03-31-deep-research-pipeline-evidence.md). Anthropic independently built a [production multi-agent research system](https://www.anthropic.com/engineering/multi-agent-research-system) using the same core pattern — their eval showed 90.2% improvement over single-agent. We converged on the same architecture independently; this system extends it with Haiku scouts for cost efficiency, adversarial peer dynamics between specialists, and asynchronous orchestrator dispatch.\n\n## Source of Truth\n\nThis is the canonical home of the deep-research plugin. Originally developed as part of [coordinator-claude](https://github.com/dbc-oduffy/coordinator-claude) and extracted for independent distribution.\n\n## Acknowledgements\n\nPipeline D is built on [notebooklm-mcp-cli](https://github.com/jacob-bd/notebooklm-mcp-cli) by [jacob-bd](https://github.com/jacob-bd) — an MCP server that provides programmatic access to Google NotebookLM.\n\n---\n\n[Dónal O'Duffy](https://github.com/dbc-oduffy) & Claude\n",
  "bytes": 6575,
  "sha": "cf7c1c28f1b0537ed6d8ad422a949612a6fe282b5b581371ac0849edefd5b7c7",
  "repo_slug": "oduffy-delphi/deep-research-claude",
  "fonte": "repo",
  "truncated": false,
  "api": "https://agentalog.com/api/listings/plg_oduffy_delphi_deep_research_claude_deep__e0c3cb9f/readme"
}