Oktopi Research Team

oktopi-research-team plugin

PDP (Product Development Plan) gap-analysis for Claude Code, designed as an agentic, multi-agent review system (inspired by Anthropic's multi-agent research system) and grounded in the shared Oktopi Taxonomy-config.

Architecture

           ┌──────────────────────────┐
           │      pdp-reviewer        │   ← orchestrator (Lead Reviewer)
           │  scope → dispatch → sync │     model: opus
           └─────────────┬────────────┘
                         │  Task tool, parallel fan-out
     ┌──────────┬────────┼────────┬──────────────────┐
     ▼          ▼        ▼        ▼                  ▼
  cmc-rev   pharm-tox  …      commercial-rev     pm-rev    (12 function reviewers — sonnet)
     │          │        │        │                  │
     └──────────┴────────┴────────┴──────────────────┘
                         │
                         ▼
              structured JSON verdicts
                         │
                         ▼
              gate-readiness report

Each reviewer:

• Embodies the role's goal (seasoned pharma function lead persona)
• Anchors on the Oktopi rubric via data/questions/<modality>/<FN>.json
• Asks adaptive follow-ups when the rubric doesn't cover a novel risk
• Returns a structured JSON the orchestrator can reconcile

What's inside

plugins/oktopi-research-team/
├── agents/                                 # 12 function reviewers + 1 orchestrator
│   ├── pdp-reviewer.md                     # Lead Reviewer (orchestrator)
│   ├── cmc-reviewer.md
│   ├── commercial-reviewer.md
│   └── ... (10 more functions)
├── skills/
│   ├── stage-gate-sg1/ … stage-gate-sg9/   # 9 gate-goal skills (concise)
│   └── function-<slug>/                    # 12 function-mandate skills
├── commands/
│   ├── review-stage-gate.md                # /review-stage-gate
│   └── review-function.md                  # /review-function
├── data/
│   ├── functions.json                      # role + mission + mandate per function
│   ├── stage-gates.json                    # SG1..SG9 with goal + focus
│   ├── stage-gate-index.json               # counts per (SG, mode, function) + domains
│   ├── modes.json                          # SR / OE / DD / RS
│   ├── heatmap/<modality>.json             # question → {mode → {sg → priority}}
│   └── questions/<modality>/<FN>.json      # 1,492 questions with priorities + rubric
└── scripts/
    └── build_taxonomy_data.py              # regenerate everything from Taxonomy-config

Why this design

• Auto-routing by natural language. Every agent and skill description lists the user phrases that should trigger it ("Use PROACTIVELY when the user asks about: GMP manufacturing, tech transfer, or process validation..."). Claude matches the user's question against these triggers and auto-invokes the right specialist — no explicit /command needed. A top-level oktopi-research-team router skill catches any pharma dev question and delegates.
• Agents are goal-embodied, not question-parroting. Each reviewer knows why they exist (their mission) and what they own (their mandate). The 1,492-question rubric is their floor, not their ceiling — they're explicitly instructed to add adaptive questions when a novel modality or fresh regulatory signal demands it.
• Orchestrator owns parallelism and reconciliation. Like a Lead Researcher, pdp-reviewer scopes the work, dispatches subagents concurrently, and synthesizes one gate-readiness report with cross-functional risk clustering.
• Skills describe intent, not data. Stage-gate and function skills are concise goal statements (< 10 KB each) that trigger naturally when the user mentions a gate or function. Question-level data lives in JSON that agents load on demand.
• Every finding is citable. Question IDs (COM5, BBSTAT18, etc.) link back to the Oktopi Expert Toolkit rubrics; adaptive questions are tagged [adaptive] with a rationale.

Expanding an agent's tooling and knowledge

Each reviewer is designed to grow — add reference material and tools without touching the build script:

• Per-function knowledge lives under data/knowledge/<CODE>/ (one directory per function, scaffolded on build). Drop SOPs, playbooks, guideline summaries, template questionnaires in markdown or JSON. The matching reviewer is instructed to scan this folder alongside the rubric.
• External tools (MCP): add servers to the plugin's .mcp.json (e.g. ClinicalTrials.gov, PubMed, an internal CMC database). Then extend the tools: frontmatter in the specific <slug>-reviewer.md agent to grant access (e.g. tools: Read, Grep, Glob, mcp__pubmed__search).
• Sub-specialists: spawn a narrower agent under agents/<slug>-<subspeciality>.md (e.g. commercial-hta-specialist). Reference it from the parent reviewer's workflow.
• Trigger tuning: if a function should catch more phrases, edit that function's triggers: list in build_taxonomy_data.py and rerun the script — descriptions and the router skill regenerate automatically.

Usage

/review-stage-gate SG5 OE small-molecule ~/Desktop/acme-pdp.pdf

The command hands off to pdp-reviewer, which:

1. Loads stage-gate-index.json to see which functions carry Critical question load at SG5 × OE
2. Dispatches those function reviewers in parallel
3. Each reviewer loads its question bank, filters on Critical at SG5/OE, evaluates evidence, and returns structured JSON
4. Orchestrator reconciles into one readiness report with cross-functional risk clusters

For a single-function pass:

/review-function commercial SG6 DD biologics ~/Desktop/acme-dataroom/

Regenerating data

git clone https://github.com/oktopi-org/Taxonomy-config.git
python3 plugins/oktopi-research-team/scripts/build_taxonomy_data.py \
    --taxonomy ./Taxonomy-config

Requires Python 3.10+ and openpyxl. The script regenerates:

• All 13 agent markdown files (12 reviewers + orchestrator)
• All 22 skill files (9 stage-gate + 12 function + 1 top-level router)
• All JSON data under data/, including scaffolded data/knowledge/<CODE>/ directories

Extending

• Change a role's mission or mandate → edit FUNCTIONS in build_taxonomy_data.py and rebuild.
• Tune the orchestrator → edit render_orchestrator_agent().
• Add a new function → add a FUNCTIONS entry + FUNCTION_SLUG + rubric file mapping, and the script will generate the agent, skill, and question JSON.