research

Research Pipeline Skill

You are running the Grocery-Bot research pipeline. Your job is to systematically find, score, and catalog academic papers and implementations relevant to our multi-agent grocery delivery optimization problem.

Workflow

1. Load State

Read these files to understand current progress:

• research/QUEUE.md — what to research next, what's done, search queries already run
• research/PAPERS.md — all papers found so far with scores and metadata
• research/NOTES.md — deep reading notes for high-relevance papers

2. Pick Next Item

Find the first unchecked [ ] item in the Priority Queue section of QUEUE.md. If the queue is empty, check the Sub-problem Backlog for new research directions and generate queue items from them.

3. Execute the Item

Each queue item is one of these types:

Deep-read a paper (e.g., "Deep-read RMCA paper [ASSIGN-001]"):

1. Find the paper's entry in PAPERS.md
2. WebFetch the arxiv abstract page (use https://arxiv.org/abs/XXXX.XXXXX, NOT the PDF)
3. Extract: problem formulation, algorithm, complexity, experimental results
4. Search for code repositories: "paper title" github, "paper title" code
5. Update the paper entry in PAPERS.md with findings
6. If relevance >= 4: add deep notes to NOTES.md (see format below)
7. Change status from FOUND → SKIMMED or READ

Forward/backward citation search (e.g., "Forward-cite RMCA via Semantic Scholar"):

1. Use Semantic Scholar API via WebFetch:
   • Forward citations: https://api.semanticscholar.org/graph/v1/paper/search?query=PAPER+TITLE&fields=title,authors,year,venue,url,citationCount,externalIds
   • Paper details: https://api.semanticscholar.org/graph/v1/paper/ARXIV:XXXX.XXXXX?fields=title,authors,year,venue,citations.title,citations.year,citations.venue,citations.externalIds,references.title,references.year,references.venue,references.externalIds
2. Filter citations by relevance (warehouse, MAPF, delivery, pickup, multi-agent)
3. Add relevant papers to PAPERS.md as FOUND entries with initial scores
4. Add promising papers as new queue items

Search query (e.g., 'Search "order batching warehouse robots"'):

1. Run the search via WebSearch
2. Scan results for academic papers, code repos, competition reports
3. For each relevant find, add to PAPERS.md
4. For very promising finds, add deep-read items to the queue
5. Log the query in "Search Queries Run" section of QUEUE.md (with date)

Competition/benchmark check (e.g., "Check League of Robot Runners 2024"):

1. WebSearch for the competition results
2. WebFetch the results page
3. Document winning approaches, code repos, key techniques
4. Add relevant papers/repos to PAPERS.md
5. Add follow-up items to the queue

4. Update Tracking Files

After completing each item:

1. QUEUE.md: Mark the item as [x] done. Add any new items discovered. Move completed items to the "Completed" section with date.
2. PAPERS.md: Update or add paper entries.
3. NOTES.md: Add deep notes for high-relevance papers.

5. Parallelize with Research Agents

When the queue has multiple independent items, use the Agent tool to process them in parallel for maximum throughput:

When to parallelize: Launch background agents when you have 3+ independent queue items (e.g., multiple deep-reads, multiple searches, or a mix).

How to parallelize:

1. Process the first 2-3 items yourself (to warm up context)
2. Launch remaining independent items as background agents using:

   Agent(subagent_type="general-purpose", run_in_background=true, model="sonnet")
   

3. Each agent gets a clear prompt with:
   • The specific queue item to process
   • The paper entry format and PDIA rubric (copy from this skill)
   • Instructions to read files first, then edit (not overwrite)
   • Which section of PAPERS.md to add entries to
4. Wait for agents to complete, then do a consistency pass

Agent prompt template:

You are continuing a research pipeline for the Grocery-Bot project.
Your task is to [specific queue item].

Steps:
1. [specific research steps]
2. Update /Users/.../research/PAPERS.md — [add/update entries]
3. Update /Users/.../research/NOTES.md — [add deep notes if relevant]

[Include paper format, PDIA rubric, and file paths]

IMPORTANT: Read the existing files first before editing to preserve
existing content. Append to files, don't overwrite.

What NOT to parallelize: Items that depend on each other (e.g., a forward-cite search that depends on a deep-read finding the arxiv ID first).

Consistency pass: After all agents complete, read PAPERS.md and QUEUE.md to verify no duplicate entries, no conflicting edits, and all cross-references are consistent.

6. Continue or Stop

• If there are more unchecked items in the queue, continue to the next one (or launch more agents for parallel processing).
• Stop when: queue is empty.
• Before stopping, ensure all files are saved and consistent.
• Print a brief summary of what was done and what's next.

PDIA Relevance Scoring Rubric

Score each paper on four dimensions (1-5):

Overall Relevance = round(mean(P, D, I, A))

Priority read threshold: Problem >= 4 AND Applicability >= 3

Paper Entry Format (for PAPERS.md)

### [TAG-NNN] Paper Title
- **Authors**: X et al. | **Year**: 2021 | **Venue**: AAAI
- **URL**: https://arxiv.org/abs/...
- **Code**: https://github.com/... (or "None found")
- **Sub-problem**: task-assignment | pathfinding | congestion | batching | joint | learning
- **Relevance**: 4 | **Quality**: 3 | **Has code**: Y/N
- **PDIA**: P=4 D=4 I=3 A=5
- **Key idea**: One sentence summary
- **Applicable to us**: How it connects to Grocery-Bot specifically
- **Cites**: [ID, ID] | **Cited by**: [ID, ID]
- **Status**: FOUND | SKIMMED | READ | REJECTED

Deep Notes Format (for NOTES.md)

## [TAG-NNN] Paper Title

**Read date**: YYYY-MM-DD

### Problem & Formulation
What exactly does this paper solve? Mathematical formulation if relevant.

### Algorithm
Step-by-step description of the core algorithm.

### Key Results
Performance numbers, comparisons, scalability.

### Relevance to Grocery-Bot
How does this apply to our specific problem? What component would it improve?

### Implementation Sketch
Concrete steps to integrate this into our codebase.
Which files would change? What's the expected improvement?

### Open Questions
What's unclear? What would we need to test?

Tag Prefixes

• MAPF-NNN — general MAPF papers
• ASSIGN-NNN — task assignment papers
• PATH-NNN — pathfinding/routing papers
• CONG-NNN — congestion/traffic management
• BATCH-NNN — order batching/trip composition
• JOINT-NNN — joint assignment + pathfinding
• LEARN-NNN — learning-based approaches
• COMP-NNN — competition reports/benchmarks
• SURVEY-NNN — survey/overview papers
• SCHED-NNN — scheduling/dispatching papers

Related Skills

When research is complete and you're ready to implement findings, use /implement to start the implementation pipeline. It reads from the same research files (IMPLEMENTATION_PLAN.md, NOTES.md) and tracks progress in research/PROGRESS.md.

Important Guidelines

• Never re-run a search query already listed in QUEUE.md "Search Queries Run"
• Always log new search queries with the date
• Don't fetch PDFs — use arxiv abstract pages, Semantic Scholar, or author pages
• Follow citation chains — when a paper cites or is cited by something promising, add it
• Tag by sub-problem — every paper must map to one of our sub-problems
• Be honest about scores — a 5 means it directly solves our exact problem
• Dead ends are valuable — log them in QUEUE.md so we don't revisit
• Update, don't duplicate — if a paper already exists in PAPERS.md, update its entry