Tools
PoggioAI_MSc Openclaw
By Plugin for OpenClaw of PoggioAI/MSc
Install
npm install
npmConfiguration Example
{
"plugins": {
"pai-msc-openclaw": {
"enabled": true,
"config": {
"consortiumPath": "",
"condaEnvName": "poggioai-msc",
"defaultPreset": "max-quality",
"defaultMode": "local",
"defaultModel": "claude-opus-4-6",
"defaultBudgetUsd": 300,
"progressPollIntervalMs": 15000,
"steeringBasePort": 5001,
"uploadTimeoutMs": 60000
}
}
}
}README
# pAI/MSc-openclaw
Native OpenClaw plugin for the [pAI/MSc](https://github.com/PoggioAI/PoggioAI_MSc) autonomous research pipeline. Transforms a research hypothesis into a conference-grade manuscript with a single command — zero config, zero human steers required.
```
================================================================
pAI/MSc-openclaw — Autonomous Research Pipeline
================================================================
Thanks from the PoggioAI Team for using this tool!
Contact us:
Discord: https://discord.gg/Pz7spPPY
Email: pierb@mit.edu
Please acknowledge PoggioAI in your papers and cite our
technical report if you use this tool:
https://poggioai.github.io/papers/poggioai-msc-v0.pdf
================================================================
```
---
## Quick Start
```
/pai-msc "Investigate whether batch normalization implicitly regularizes the spectral norm of weight matrices in shallow ReLU networks"
```
That's it. The plugin:
1. Auto-installs the pAI/MSc Python backend on first use
2. Passes your existing OpenClaw API keys (no separate `.env` needed)
3. Creates an isolated run workspace with all inputs in `initial_context/`
4. Prompts you for reference files (papers, datasets) via Telegram/interface
5. Injects 25 backtested quality prompts + a 647-line author style guide
6. Runs the full 22-agent pipeline with quality-maximizing defaults
7. Streams progress updates to your chat as stages complete
8. Delivers the finished paper (PDF or markdown) back to you
---
## What's In This Repository
```
├── openclaw.plugin.json # Plugin manifest — name, version, config schema
├── package.json # TypeScript package (build with npm run build)
├── tsconfig.json # TypeScript compiler config
├── README.md # This file
│
├── src/ # TypeScript source (3,000+ lines)
│ ├── index.ts # Entry point: definePluginEntry() — registers
│ │ # 4 commands, 4 tools, 1 background service,
│ │ # and a shutdown hook
│ │
│ ├── commands/ # User-facing slash commands
│ │ ├── research.ts # /pai-msc "hypothesis" [flags] — the main entry
│ │ ├── pai-msc-status.ts # /pai-msc-status — show current stage + budget
│ │ ├── pai-msc-stop.ts # /pai-msc-stop — kill a running pipeline
│ │ └── pai-msc-list.ts # /pai-msc-list — list all runs in this session
│ │
│ ├── tools/ # Agent-callable tools (programmatic access)
│ │ ├── run-pipeline.ts # pai-msc.runPipeline — start a run
│ │ ├── steer-pipeline.ts # pai-msc.steerPipeline — inject instructions
│ │ ├── get-results.ts # pai-msc.getResults — retrieve status + paper
│ │ └── approve-milestone.ts # pai-msc.approveMilestone — gate responses
│ │
│ ├── services/ # Core services
│ │ ├── workspace-manager.ts # Creates per-run workspace with initial_context/,
│ │ │ # logs/, uploads/ directories. All run data is
│ │ │ # isolated under ~/.openclaw/poggioai-msc/runs/
│ │ ├── upload-handler.ts # Prompts user for reference files via Telegram/
│ │ │ # interface with 3-strategy fallback. Saves to
│ │ │ # initial_context/uploads/
│ │ ├── installer.ts # Auto-install: clone repo → conda env → pip →
│ │ │ # patch prompts → preflight check → sentinel
│ │ ├── process-manager.ts # Spawn Python subprocess with workspace-aware
│ │ │ # env. Logs to logs/stdout.log + stderr.log
│ │ ├── progress-poller.ts # Background service: polls every 15s for stage
│ │ │ # changes, budget thresholds, completion/failure.
│ │ │ # Also handles the narrative voice hook and
│ │ │ # review score escalation.
│ │ └── quality-injector.ts # Copies backtested prompts + style guide into
│ │ # initial_context/ and paper_workspace/
│ │
│ ├── bridge/ # Integration layer between plugin and Python backend
│ │ ├── env-passthrough.ts # Reads API keys from OpenClaw env → writes .env
│ │ ├── config-writer.ts # Generates .llm_config.yaml from preset + flags
│ │ ├── steering-client.ts # HTTP client for live steering API
│ │ │ # (POST /interrupt, /instruction, GET /status)
│ │ └── result-reader.ts # Reads run_summary.json, budget_state.json,
│ │ # review_verdict.json, finds paper file
│ │
│ ├── defaults/ # Configuration defaults
│ │ ├── quality-presets.ts # QUALITY_MAX and QUALITY_FAST presets with all
│ │ │ # CLI flag values pre-configured
│ │ └── stage-names.ts # 24 pipeline stage constants + human-readable
│ │ # display names for progress messages
│ │
│ └── types/ # TypeScript type definitions
│ ├── openclaw-api.ts # OpenClawApi interface + runtime guards
│ ├── pipeline.ts # RunHandle, PipelineOptions, StageEvent, RunSummary
│ ├── config.ts # PluginConfig interface + DEFAULT_CONFIG
│ ├── budget.ts # BudgetState, BudgetEntry, BUDGET_THRESHOLDS
│ └── steering.ts # SteeringInstruction, SteeringStatus, ReviewVerdict
│
├── assets/ # Quality artifacts (ported from backtested Claude skill)
│ │
│ ├── author_style_guide_default.md # 647-line ML theory writing standard
│ │ # Contains: non-negotiable principles, anti-patterns
│ │ # (paper/section/sentence/epistemic), abstract rules
│ │ # (120-180 words, no theorem refs), related-work rules
│ │ # (organize by ideas not authors), concrete lints,
│ │ # epistemic lints, deletion pass, self-audit checklist,
│ │ # case studies with diagnosis + fix
│ │
│ ├── state_template.json # Pipeline state machine template (32 fields)
│ │
│ └── prompts/ # 25 backtested agent prompts
│ ├── 01-persona-practical.md # Practical Compass persona
│ ├── 02-persona-rigor.md # Rigor & Novelty persona
│ ├── 03-persona-narrative.md # Narrative Architect persona
│ ├── 04-persona-synthesis.md # Synthesis coordinator (min 3 debate rounds)
│ ├── 05-literature-review.md # Adversarial novelty falsification
│ ├── 06-brainstorm.md # 3-phase: divergent → convergent → dependency
│ ├── 07-formalize-goals.md # Goal formalization + track decomposition
│ ├── 08-math-literature.md # Theory-specific literature search
│ ├── 09-math-proposer.md # Claim graph construction
│ ├── 10-math-prover.md # Proof construction with technique library
│ ├── 11-math-verifier.md # Adversarial proof auditor + numerical checks
│ ├── 12-experiment-design.md # Experiment design with anti-hallucination
│ ├── 13-experimentation.md # Experiment execution
│ ├── 14-experiment-verify.md # Cross-seed stability, verdict annotation
│ ├── 15-formalize-results.md # Conservative results synthesis
│ ├── 16-duality-check.md # Dual-lens: actionability + soundness (>= 6/10)
│ ├── 17-resource-prep.md # Figures, tables, bibliography
│ ├── 18-writeup.md # 260-line writeup: 12 passes, 2 full edit cycles
│ ├── 19-proofreading.md # AI-voice detection checklist (9 categories)
│ ├── 20-reviewer.md # Hard blockers B1-B5, AI voice risk assessment
│ ├── 21-research-plan-writeup.md
│ ├── 22-track-merge.md # Theory-experiment unified summary
│ ├── 23-verify-completion.md # 3-way routing: COMPLETE/INCOMPLETE/RETHINK
│ ├── 24-followup-lit-review.md # Gap-specific targeted follow-up
│ └── 25-narrative-voice.md # Pre-writeup tone/voice guidance
│
├── scripts/
│ ├── install-consortium.sh # Manual installer (normally auto-runs)
│ └── check-prereqs.sh # Verify conda, python, pdflatex, API keys
│
└── examples/
├── quickstart-task.txt # Example research hypothesis
└── custom-style-guide-example.md # How to write a custom style guide
```
---
## How It Works
### The Pipeline
When you run `/pai-msc "hypothesis"`, the plugin orchestrates a 22-agent pipeline:
```
/pai-msc "hypothesis"
│
│ ── Plugin Layer ──────────────────────────────────────────────
│
├─ 1. Auto-install pAI/MSc Python backend (first time only)
├─ 2. Create isolated run workspace under ~/.openclaw/poggioai-msc/runs/
├─ 3. Copy prompts + style guide → initial_context/
├─ 4. Write task.txt + pipeline_options.json → initial_context/
├─ 5. Prompt user for reference files → initial_context/uploads/
├─ 6. Write .env + .llm_config.yaml (per-run, not shared)
├─ 7. Spawn: python launch_multiagent.py --resume {workspace} [flags]
│
│ ── pAI/MSc Pipeline (Python/LangGraph) ──────────────────────
│
├─ Phase 1: Persona Debate (3-5 rounds)
│ 3 personas (Practical, Rigor, Narrative) debate in escalating
│ rounds. Each round must be HARDER than the last. Minimum 3
│ rounds; extends to 5 if any persona still rejects. Synthesis
│ produce
... (truncated)
tools
Comments
Sign in to leave a comment