research-pipeline

[RESEARCH PIPELINE - ITERATION {{ITERATION}}/{{MAX}}]

Your previous attempt did not output the completion promise. Continue working on the pipeline.

Research Pipeline is a multi-agent pipeline that systematically analyzes academic papers, identifies research gaps, and proposes novel methods with experiment plans. It processes a directory of reference papers through 8 stages: 3 sequential analysis stages, 3 parallel idea-proposal stages, a synthesis stage, and an experiment planning stage. Modeled after the ralph execution pattern with persistent state tracking.

<Use_When>

• User wants systematic research paper analysis with gap identification
• User says "research pipeline", "analyze papers", "paper review", "research assistance"
• User has a directory of reference papers and a research goal
• User wants to go from literature review to experiment design automatically </Use_When>

<Do_Not_Use_When>

• User wants to review a single paper only -- just read and analyze it directly
• User wants a general literature survey without method proposal -- use sciomc instead
• User wants to write a paper -- this generates research direction, not paper text </Do_Not_Use_When>

<Why_This_Exists> Research ideation requires structured analysis that most researchers do manually over weeks: reading papers, finding weaknesses, searching for complementary work, blending ideas, formalizing methods, designing architectures, and planning experiments. This pipeline automates the full sequence with specialized agents at each stage. Stages 4-6 run in parallel because idea blending, mathematical formulation, and architecture design are independent creative proposals that benefit from diverse perspectives — the synthesizer then unifies them. </Why_This_Exists>

<Pipeline_Overview>

[Sequential Phase: Analysis]
Stage 1: Paper Review ──────→ weaknesses per paper
Stage 2: Gap Search ────────→ papers compensating weaknesses
Stage 3: Summarization ─────→ structured summaries (all papers)
                │
                ▼
[Parallel Phase: Idea Proposal]  ← Stages 4, 5, 6 run SIMULTANEOUSLY
┌───────────────┼───────────────┐
│               │               │
▼               ▼               ▼
Stage 4:      Stage 5:      Stage 6:
Idea          Math          Architecture
Blender       Expert        Expert
│               │               │
└───────────────┼───────────────┘
                │
                ▼
[Sequential Phase: Convergence]
Stage 7: Proposal Synthesis ─→ unified method from 4+5+6
Stage 8: Experiment Plan ────→ baselines & evaluation design

Stages 1-3 are sequential (each depends on previous). Stages 4-6 are parallel (all read from stages 1-3, independent of each other). Stages 7-8 are sequential (7 synthesizes 4-6, 8 plans experiments from 7). State is persisted to .omc/research-pipeline/{session-id}/ for resume capability. </Pipeline_Overview>

<Paper_Directory_Convention>

• All reference PDFs must be in the specified directory
• Baseline paper: prefix filename with baseline_ (e.g., baseline_attention_is_all_you_need.pdf)
• Alternative: place the baseline in a baseline/ subdirectory
• If no baseline is detected, STOP and ask the user which paper is the baseline before proceeding
• The baseline paper is the primary method the user wants to improve upon or compare against </Paper_Directory_Convention>

<Execution_Policy>

• Stages 1-3 execute strictly sequentially -- each must complete before the next begins
• Stages 4, 5, 6 execute in parallel -- fire all three agents simultaneously after Stage 3 completes
• Stage 7 waits for ALL of stages 4, 5, 6 to complete before starting
• Stage 8 executes after Stage 7 completes
• Use Opus-tier agents for stages requiring deep analysis (1, 4, 5, 6, 7)
• Use Sonnet-tier agents for standard work (2, 3, 8)
• Read PDF files using the Read tool (Claude Code supports PDF reading)
• For the parallel phase, read agent definition files from agents/ directory for detailed prompts
• Persist all outputs to the session directory
• If a stage fails, retry once with more context before reporting failure </Execution_Policy>

0. **Initialization**: a. Parse arguments: extract `` and `--goal ""` from `{{PROMPT}}` b. Verify the paper directory exists and contains PDF files c. List all PDFs and identify the baseline paper (look for `baseline_` prefix or `baseline/` subdirectory) d. If no baseline detected, STOP and ask the user which paper is the baseline e. Generate a session ID: `research-{YYYYMMDD}-{short-hash}` f. Create session directory: `.omc/research-pipeline/{session-id}/` g. Create `pipeline.json` state file with all 8 stages marked `pending` h. Save the list of papers and the identified baseline to `pipeline.json`

---

Sequential Phase: Analysis (Stages 1-3)

1. Stage 1 - Paper Review (Identify weaknesses): a. Read EACH paper in the directory using the Read tool (PDF support) b. Read the agent prompt template from agents/paper-reviewer.md c. For each paper, analyze and document:

   • Key contributions: What novel ideas does this paper introduce?
   • Methodology: What approach/algorithm/model is used?
   • Results: What are the main experimental results?
   • Weaknesses: What are the limitations, assumptions, failure cases?
   • Open questions: What does the paper leave unanswered? d. For the baseline paper, provide extra-detailed analysis:
   • Specific quantitative limitations (e.g., "O(n²) complexity for sequence length")
   • Assumptions that could be relaxed
   • Scenarios where the method underperforms
   • Missing comparisons or evaluations e. Spawn an Opus-tier agent:

   Agent(prompt="[Read agents/paper-reviewer.md for full instructions]
   You are an expert research paper reviewer. Read and analyze each of these papers thoroughly.
   Papers: [list all PDF paths]
   Baseline paper: [baseline filename]
   Research goal: [goal]
   Output a structured review with weaknesses for each paper.
   Use weakness IDs (W-B1, W-B2 for baseline; W-{N}1 for others) for cross-referencing.",
   subagent_type="general-purpose", model="opus")
   

   f. Save output to stage-1-paper-review.md g. Update pipeline.json: stage 1 status = complete h. Verification: Confirm output contains weakness analysis for EVERY paper

2. Stage 2 - Gap Search (Find compensatory papers + download PDFs): a. Read Stage 1 output (stage-1-paper-review.md) b. Read the agent prompt template from agents/gap-scout.md c. Extract the top weaknesses across all papers (prioritize baseline weaknesses) d. Create the related_papers/ directory inside the paper directory:

   mkdir -p {paper-directory}/related_papers
   

   e. For each major weakness, search for papers that address it:

   • Use WebSearch to find relevant papers (Semantic Scholar, arXiv, Google Scholar)
   • Cross-disciplinary search is encouraged (the area doesn't matter)
   • For each found paper, note: title, authors, year, venue, URL, and how it addresses the weakness
   • Download the paper PDF into {paper-directory}/related_papers/ using curl f. Spawn a Sonnet-tier agent:

   Agent(prompt="[Read agents/gap-scout.md for full instructions]
   You are a research paper scout. Given these weaknesses from our literature review:
   [Include weaknesses from Stage 1]
   Research goal: [goal]
   Search the web for papers that address each weakness. Use WebSearch and WebFetch.
   Cross-disciplinary papers are welcome -- the area doesn't matter as long as the technique is relevant.
   
   IMPORTANT - Paper Downloads:
   For every relevant paper you find, download its PDF into this directory:
     {paper-directory}/related_papers/
   Use Bash with curl to download:
     curl -L -o '{paper-directory}/related_papers/{FirstAuthor}_{Year}_{ShortTitle}.pdf' '{pdf_url}'
   Track download status for each paper: [DOWNLOADED], [DOWNLOAD_FAILED], or [NO_PDF_AVAILABLE].
   Verify each download with ls -la.",
   subagent_type="general-purpose", model="sonnet")
   

   g. Save output to stage-2-gap-search.md h. Update pipeline.json: stage 2 status = complete i. Verification: Confirm at least 1 compensatory paper found per major weakness j. Verification: Confirm downloaded PDFs exist in {paper-directory}/related_papers/ with non-zero size

3. Stage 3 - Paper Summarization (Structured summaries): a. Read Stage 1 output (original paper analyses) and Stage 2 output (found papers) b. Read the agent prompt template from agents/paper-summarizer.md c. Read downloaded related papers from {paper-directory}/related_papers/ using the Read tool (PDF support) d. Create structured summaries for ALL papers (both original references and downloaded related papers) e. For each paper, produce:

   • One-paragraph summary (problem, approach, key result)
   • Core technique (the main algorithmic/methodological contribution)
   • Key equations/formulations (if applicable)
   • Strengths and limitations (from Stage 1 or new analysis)
   • Relevance to goal (how this paper relates to the research goal) f. Create a comparative table showing all papers side-by-side g. Spawn a Sonnet-tier agent:

   Agent(prompt="[Read agents/paper-summarizer.md for full instructions]
   You are an expert research summarizer. Create structured summaries of all these papers.
   Paper analyses: [Include Stage 1 output]
   Found papers: [Include Stage 2 output]
   Research goal: [goal]
   
   IMPORTANT - Downloaded Related Papers:
   Stage 2 downloaded related paper PDFs into: {paper-directory}/related_papers/
   Read these PDFs using the Read tool to produce deeper summaries of the found papers
   (not just abstract-level summaries from Stage 2 URLs).
   Also create a comparative table and complementarity map.",
   subagent_type="general-purpose", model="sonnet")
   

   h. Save output to stage-3-summaries.md i. Update pipeline.json: stage 3 status = complete

---

Parallel Phase: Idea Proposal (Stages 4, 5, 6)

CRITICAL: Fire all three agents SIMULTANEOUSLY after Stage 3 completes. Each agent receives the same inputs (stages 1-3 outputs) but works independently on a different aspect of the proposal.

4. Stage 4 - Idea Blending (Creative synthesis — runs in PARALLEL): a. Read ALL previous stage outputs (Stages 1-3) b. Read the agent prompt template from agents/idea-blender.md c. Identify complementary techniques across papers and propose 3-5 concrete blending strategies d. Each strategy includes: blend name, source papers, core idea, expected benefit, challenges, novelty e. Rank by expected impact and feasibility f. Spawn simultaneously with Stages 5 and 6:

   Agent(prompt="[Read agents/idea-blender.md for full instructions]
   You are a creative research ideation expert.
   Paper reviews: [Stage 1 output]
   Gap search results: [Stage 2 output]
   Paper summaries: [Stage 3 output]
   Research goal: [goal]
   Propose 3-5 ways to blend ideas from these papers into novel approaches.
   Think creatively -- cross-disciplinary combinations are encouraged.
   Rank strategies by impact and feasibility.",
   subagent_type="general-purpose", model="opus")
   

   g. Save output to stage-4-idea-blending.md h. Update pipeline.json: stage 4 status = complete

5. Stage 5 - Mathematical Formulation (Formal method — runs in PARALLEL): a. Read ALL previous stage outputs (Stages 1-3) b. Read the agent prompt template from agents/math-expert.md c. Based on the paper analyses and identified gaps, independently propose a mathematical formulation:

   • Problem formulation: Define the problem formally (input space, output space, objective)
   • Proposed method: Define the method with equations, using LaTeX notation
   • Loss function / Objective: Define the training objective or optimization goal
   • Theoretical properties: Analyze complexity, convergence, bounds if applicable
   • Key assumptions: State assumptions clearly d. NOTE: This agent works INDEPENDENTLY from Stage 4 — it proposes its own mathematical approach based on the paper analyses, not on blending strategies e. Spawn simultaneously with Stages 4 and 6:

   Agent(prompt="[Read agents/math-expert.md for full instructions]
   You are a mathematical expert in machine learning and optimization.
   Paper reviews with weaknesses: [Stage 1 output]
   Compensatory papers: [Stage 2 output]
   Paper summaries with key equations: [Stage 3 output]
   Research goal: [goal]
   Based on the identified weaknesses and available techniques, propose a rigorous
   mathematical formulation for a novel method. Use LaTeX notation.
   Define the problem, propose the method formally, specify the loss function,
   and analyze theoretical properties.
   You are working INDEPENDENTLY -- propose your own approach from the paper analyses.",
   subagent_type="general-purpose", model="opus")
   

   f. Save output to stage-5-math-formulation.md g. Update pipeline.json: stage 5 status = complete

6. Stage 6 - Architecture Design (System architecture — runs in PARALLEL): a. Read ALL previous stage outputs (Stages 1-3) b. Read the agent prompt template from agents/arch-expert.md c. Based on the paper analyses and identified gaps, independently propose a model/system architecture:

   • High-level architecture diagram (ASCII or structured description)
   • Component breakdown: Each module/component with its role
   • Data flow: How data moves through the system
   • Training and inference procedures
   • Comparison with baseline architecture d. NOTE: This agent works INDEPENDENTLY from Stages 4 and 5 — it proposes its own architectural approach e. Spawn simultaneously with Stages 4 and 5:

   Agent(prompt="[Read agents/arch-expert.md for full instructions]
   You are a senior ML systems architect.
   Paper reviews with weaknesses: [Stage 1 output]
   Compensatory papers: [Stage 2 output]
   Paper summaries with architectures: [Stage 3 output]
   Research goal: [goal]
   Based on the identified weaknesses and available techniques, propose a concrete
   model/system architecture for a novel method that addresses the key gaps.
   Include component breakdown, data flow, training procedure, and comparison with baseline.
   You are working INDEPENDENTLY -- propose your own architectural approach from the paper analyses.",
   subagent_type="general-purpose", model="opus")
   

   f. Save output to stage-6-architecture.md g. Update pipeline.json: stage 6 status = complete

Parallel execution code pattern:

// Fire ALL THREE simultaneously — do NOT wait between them
Agent(prompt="[Stage 4: Idea Blending]...", model="opus")   // Fire
Agent(prompt="[Stage 5: Math Expert]...", model="opus")      // Fire
Agent(prompt="[Stage 6: Arch Expert]...", model="opus")      // Fire
// Wait for ALL three to complete before proceeding to Stage 7

---

Sequential Phase: Convergence (Stages 7-8)

7. Stage 7 - Proposal Synthesis (Unify parallel proposals): a. Read ALL outputs from the parallel phase: Stage 4 (idea blending), Stage 5 (math formulation), Stage 6 (architecture design) b. Also read Stages 1-3 outputs for reference c. Synthesize the three independent proposals into a unified, coherent method:

   • Identify convergence: Where do the three proposals agree? These are high-confidence directions.
   • Resolve conflicts: Where do they disagree? Choose the strongest approach with justification.
   • Integrate strengths: Take the best ideas from each:
     • From Stage 4 (Idea Blender): The most creative and novel combination strategies
     • From Stage 5 (Math Expert): The most rigorous mathematical formulation and theoretical properties
     • From Stage 6 (Arch Expert): The most practical and implementable architecture design
   • Produce a unified proposal that includes:
     • Method name and one-paragraph description
     • Core idea (from idea blending, refined)
     • Mathematical formulation (from math expert, refined to match the chosen architecture)
     • Architecture design (from arch expert, refined to implement the chosen math formulation)
     • Novelty statement: What is genuinely new about this unified approach
     • Comparison with baseline: How this improves on the baseline paper d. Spawn an Opus-tier agent:

   Agent(prompt="You are a senior research director who synthesizes diverse technical proposals
   into unified, coherent research directions.
   
   Three independent expert agents have analyzed the same papers and proposed different aspects
   of a novel method. Your job is to synthesize their outputs into ONE unified proposal.
   
   Idea Blending proposals (creative combinations): [Stage 4 output]
   Mathematical formulation (formal method): [Stage 5 output]
   Architecture design (system design): [Stage 6 output]
   
   Background:
   Paper reviews: [Stage 1 output summary]
   Research goal: [goal]
   
   Instructions:
   1. Identify where the three proposals CONVERGE (high-confidence directions)
   2. Resolve CONFLICTS between proposals (choose strongest with justification)
   3. Integrate the best from each:
      - Best creative ideas from the Idea Blender
      - Most rigorous math from the Math Expert
      - Most practical architecture from the Arch Expert
   4. Produce a UNIFIED proposal with: method name, core idea, math formulation,
      architecture, novelty statement, and baseline comparison.
   
   The unified proposal must be internally consistent -- the math must match the architecture,
   and both must implement the core idea.",
   subagent_type="general-purpose", model="opus")
   

   e. Save output to stage-7-synthesis.md f. Update pipeline.json: stage 7 status = complete g. Verification: Confirm the synthesis references all three input proposals and produces a coherent unified method

8. Stage 8 - Experiment Planning (Baseline experiments): a. Read Stage 7 output (unified proposal) and Stage 1 output (baseline paper analysis) b. Also read Stage 3 (all paper summaries) for baseline and comparison methods c. Read the agent prompt template from agents/experiment-planner.md d. Design a comprehensive experiment plan:

   • Datasets: Which datasets to use (from baseline paper + additional relevant ones)
   • Baselines: List of methods to compare against, including:
     • The baseline paper's method (primary comparison)
     • Other methods from reference papers
     • Relevant SOTA methods from Stage 2
     • Ablation variants of the proposed method
   • Metrics: Evaluation metrics (from baseline paper + additional ones)
   • Main experiments: Core comparison table specification
   • Ablation studies: What components to ablate and why
   • Analysis experiments: Visualization, qualitative analysis, failure case analysis
   • Compute budget estimate: Approximate GPU hours and resources needed
   • Expected results: Hypothesis for what the results should show e. Spawn an Opus-tier agent:

   Agent(prompt="[Read agents/experiment-planner.md for full instructions]
   You are an expert experiment designer for ML research.
   Unified method proposal: [Stage 7 output]
   Baseline paper analysis: [Stage 1 baseline section]
   All paper summaries: [Stage 3 output]
   Research goal: [goal]
   Design a comprehensive experiment plan with baselines, datasets, metrics, ablations,
   and expected results. The baseline paper's method is the PRIMARY comparison target.",
   subagent_type="general-purpose", model="opus")
   

   f. Save output to stage-8-experiments.md g. Update pipeline.json: stage 8 status = complete

---

9. Final Report Generation: a. Read all 8 stage outputs b. Generate a synthesized final-report.md containing:
   • Executive summary of the research direction
   • Literature analysis (from Stages 1-3)
   • Independent proposals overview (from Stages 4-6)
   • Unified method proposal (from Stage 7 — the core output)
   • Experiment plan (from Stage 8)
   • Full references list c. Update pipeline.json: overall status = complete d. Print a summary of what was produced to the user

<State_Management>

pipeline.json Format

{
  "id": "research-YYYYMMDD-xxxxxx",
  "goal": "The user's research goal",
  "paperDir": "/absolute/path/to/papers",
  "baselinePaper": "baseline_paper_name.pdf",
  "papers": [
    {"filename": "paper1.pdf", "isBaseline": false},
    {"filename": "baseline_method.pdf", "isBaseline": true}
  ],
  "relatedPapersDir": "{paperDir}/related_papers",
  "status": "in_progress",
  "currentPhase": "analysis",
  "stages": [
    {
      "id": 1,
      "name": "paper-review",
      "phase": "analysis",
      "parallel": false,
      "description": "Review papers and identify weaknesses",
      "status": "pending",
      "outputFile": "stage-1-paper-review.md",
      "startedAt": null,
      "completedAt": null
    },
    {
      "id": 2,
      "name": "gap-search",
      "phase": "analysis",
      "parallel": false,
      "description": "Search for papers compensating weaknesses",
      "status": "pending",
      "outputFile": "stage-2-gap-search.md",
      "startedAt": null,
      "completedAt": null
    },
    {
      "id": 3,
      "name": "summarization",
      "phase": "analysis",
      "parallel": false,
      "description": "Create structured summaries of all papers",
      "status": "pending",
      "outputFile": "stage-3-summaries.md",
      "startedAt": null,
      "completedAt": null
    },
    {
      "id": 4,
      "name": "idea-blending",
      "phase": "proposal",
      "parallel": true,
      "parallelGroup": "proposal",
      "description": "Propose creative idea combinations",
      "status": "pending",
      "outputFile": "stage-4-idea-blending.md",
      "startedAt": null,
      "completedAt": null
    },
    {
      "id": 5,
      "name": "math-formulation",
      "phase": "proposal",
      "parallel": true,
      "parallelGroup": "proposal",
      "description": "Independent mathematical method formulation",
      "status": "pending",
      "outputFile": "stage-5-math-formulation.md",
      "startedAt": null,
      "completedAt": null
    },
    {
      "id": 6,
      "name": "architecture",
      "phase": "proposal",
      "parallel": true,
      "parallelGroup": "proposal",
      "description": "Independent model/system architecture design",
      "status": "pending",
      "outputFile": "stage-6-architecture.md",
      "startedAt": null,
      "completedAt": null
    },
    {
      "id": 7,
      "name": "synthesis",
      "phase": "convergence",
      "parallel": false,
      "description": "Synthesize parallel proposals into unified method",
      "status": "pending",
      "outputFile": "stage-7-synthesis.md",
      "startedAt": null,
      "completedAt": null
    },
    {
      "id": 8,
      "name": "experiments",
      "phase": "convergence",
      "parallel": false,
      "description": "Baseline experiments and evaluation plan",
      "status": "pending",
      "outputFile": "stage-8-experiments.md",
      "startedAt": null,
      "completedAt": null
    }
  ],
  "createdAt": "ISO timestamp",
  "updatedAt": "ISO timestamp"
}

Resume Behavior

When the skill is re-invoked:

1. Check for existing pipeline.json in .omc/research-pipeline/
2. If found, read it and determine current phase:
   • If in analysis phase (1-3): resume from first incomplete sequential stage
   • If in proposal phase (4-6): check which parallel stages are incomplete, re-fire them
   • If in convergence phase (7-8): resume from first incomplete sequential stage
3. If not found, start fresh from initialization </State_Management>

<Output_Formats>

Stage 1 Output: Paper Review

# Paper Review Report

## Research Goal
{goal}

## Baseline Paper: {name}
### Key Contributions
- ...
### Methodology
- ...
### Results
- ...
### Weaknesses (Detailed)
1. **[W-B1] {short name}**: {weakness} — {evidence/reasoning}
   - Severity: HIGH/MEDIUM/LOW
2. **[W-B2]**: ...
### Open Questions
- ...

## Paper: {name}
### Key Contributions
...
### Weaknesses
1. **[W-{id}1]**: ...
...

## Summary of Key Weaknesses Across All Papers
| Paper | ID | Weakness | Severity |
|-------|----|----------|----------|
| ... | ... | ... | ... |

Stage 2 Output: Gap Search

# Gap Search Report

## Weakness → Compensatory Paper Mapping

### [W-B1]: {weakness description}
**Found Papers:**
1. **{Title}** ({Authors}, {Year}, {Venue})
   - URL: {url}
   - PDF: `[DOWNLOADED]` → `related_papers/{FirstAuthor}_{Year}_{ShortTitle}.pdf`
   - Relevance: {how it addresses the weakness}
   - Key technique: {what technique it offers}

## Cross-Disciplinary Findings
{Papers from unexpected domains that offer relevant techniques}

## Download Summary
| Paper | Filename | Status |
|-------|----------|--------|
| {Title} | `{filename}.pdf` | DOWNLOADED |
| {Title} | — | DOWNLOAD_FAILED: {reason} |

- Total downloaded: {count} / {total found}
- Download directory: `{paper-directory}/related_papers/`

Stage 7 Output: Proposal Synthesis (KEY OUTPUT)

# Unified Method Proposal

## Method Name: {name}

## One-Paragraph Summary
{What the method does, why it's novel, what it achieves}

## Convergence Analysis
### Where proposals agree:
- {convergence point 1}
- {convergence point 2}

### Conflicts resolved:
- {conflict}: chose {approach} because {justification}

## Core Idea
{From Idea Blender, refined with Math/Arch insights}

## Mathematical Formulation
{From Math Expert, refined to match chosen architecture}

## Architecture Design
{From Arch Expert, refined to implement chosen math}

## Novelty Statement
{What is genuinely new — not incremental combination}

## Comparison with Baseline
| Aspect | Baseline | Proposed | Improvement |
|--------|----------|----------|-------------|
| ... | ... | ... | ... |

Stage 8 Output: Experiment Plan

# Experiment Plan

## Datasets
| Dataset | Size | Task | Source |
|---------|------|------|--------|

## Baselines
| Method | Paper | Why Include |
|--------|-------|-------------|
| {Baseline paper method} | {cite} | PRIMARY comparison target |

## Main Experiments
| Experiment | Goal | Metrics |
|------------|------|---------|

## Ablation Studies
| Ablation | What's Removed | Tests |
|----------|----------------|-------|

## Expected Results & Compute Budget

</Output_Formats>

<Tool_Usage>

• Use Read tool to read PDF papers (Claude Code supports PDF reading natively)
• Use Read tool to read agent prompt templates from agents/ directory
• Use Agent tool with subagent_type="general-purpose" for each stage's agent
• For parallel phase (Stages 4-6): Fire THREE Agent calls in a SINGLE message block
• Use Write tool to save stage outputs and state
• Use Edit tool to update pipeline.json status
• Use WebSearch (via agent) for Stage 2 paper searching
• Use WebFetch (via agent) to access found paper abstracts/details
• Use Glob to list PDF files in the paper directory </Tool_Usage>

User: "/claude-research:research-pipeline ~/papers/transformers --goal 'Improve transformer efficiency for long sequences'" Pipeline detects 5 PDFs including `baseline_attention_is_all_you_need.pdf`. Stage 1 identifies O(n²) attention as key weakness. (sequential) Stage 2 finds linear attention, sparse attention, and state-space model papers. (sequential) Stage 3 summarizes all 5 original + 4 found papers. (sequential) --- Parallel phase fires 3 agents simultaneously --- Stage 4 (Idea Blender): proposes blending sparse attention with state-space models Stage 5 (Math Expert): formalizes a sub-quadratic attention mechanism with provable bounds Stage 6 (Arch Expert): designs a transformer variant with local-global attention hierarchy --- Parallel complete, synthesis begins --- Stage 7 (Synthesis): unifies into "SSM-Sparse Transformer" — takes the local-global hierarchy from Arch, the provable bounds from Math, and the SSM integration from Idea Blender Stage 8: plans experiments on LRA benchmark comparing against vanilla transformer, Linformer, S4 Why good: Parallel proposals provide diverse perspectives, synthesis creates a stronger unified method. Parallel diversity scenario: Stage 4 proposes "attention + graph neural networks" Stage 5 proposes "kernel-based O(n) attention formulation" Stage 6 proposes "mixture-of-experts with adaptive routing" Stage 7 finds convergence on "adaptive computation" theme, synthesizes into a kernel-attention method with MoE routing — mathematically grounded with practical architecture. Why good: Independent proposals led to a richer synthesis than any single sequential path would. Running stages 4-6 sequentially: "Let me first blend ideas, then formalize the math based on the blend, then design architecture based on the math." Why bad: Stages 4-6 MUST run in parallel. They should be INDEPENDENT proposals. Stage 5 should NOT read Stage 4's output. Stage 6 should NOT read Stage 5's output. The synthesis (Stage 7) handles integration. No baseline identified: Starting the pipeline without confirming which paper is the baseline. Why bad: Must stop and ask if no baseline_* prefix is found. The baseline is critical for experiment design.

<Escalation_And_Stop_Conditions>

• STOP if no PDF files found in the specified directory
• STOP if no baseline paper identified and user hasn't specified one
• STOP if paper reading fails (corrupted PDF, too large)
• STOP if Stage 2 web search finds zero relevant papers (unlikely but possible)
• If one parallel agent (4/5/6) fails but others succeed, proceed to synthesis with available outputs — note the gap
• Continue if a stage produces lower-quality output -- flag it and proceed
• If the same stage fails twice, report the issue and ask for user guidance </Escalation_And_Stop_Conditions>

<Final_Checklist>

• Paper directory validated with PDFs found
• Baseline paper identified
• All 8 stages completed (status = complete in pipeline.json)
• Stages 4, 5, 6 were fired in parallel (not sequentially)
• Stage 7 synthesis references all three parallel proposals
• Stage 7 produces internally consistent unified method (math matches architecture)
• Each stage output file exists and is non-empty
• Final report generated with all sections
• User informed of completion with summary of proposed research direction </Final_Checklist>

## Configuration

The pipeline behavior can be customized via arguments:

--goal "<goal>"           # Required: research objective
--baseline "<filename>"   # Override baseline paper detection
--stages "1-3"           # Run only specific stages (for partial re-runs)
--skip-search            # Skip Stage 2 (if you already have all papers)
--top-k 3               # Number of blending strategies in Stage 4 (default: 5)

Integration with OMC

If oh-my-claudecode is installed, the pipeline can use specialized agents:

• Stage 1: oh-my-claudecode:scientist (opus) for deep paper analysis
• Stage 2: oh-my-claudecode:document-specialist (sonnet) for web search
• Stage 4: oh-my-claudecode:scientist (opus) for creative synthesis
• Stage 5: oh-my-claudecode:scientist (opus) for mathematical formulation
• Stage 6: oh-my-claudecode:architect (opus) for architecture design
• Stage 7: oh-my-claudecode:scientist (opus) for proposal synthesis

Without OMC, all stages use general-purpose agents with detailed prompts.

Why Parallel Stages 4-6?

Running idea blending, mathematical formulation, and architecture design in parallel provides:

1. Diversity: Each expert approaches the problem from a different angle without being anchored by the others
2. Independence: Math expert proposes what's theoretically elegant; Arch expert proposes what's practically implementable; Idea Blender proposes what's creatively novel
3. Better synthesis: Stage 7 has three independent perspectives to merge, avoiding the tunnel vision of a single sequential path
4. Speed: Three Opus agents running in parallel is ~3x faster than sequential

Troubleshooting

PDF reading fails?

• Ensure PDFs are not password-protected
• Large PDFs may need page-range reading: Read(file, pages="1-20")
• If a paper is too large, the agent will summarize available pages

One parallel agent failed?

• Stage 7 can synthesize from 2/3 proposals — note the gap in the synthesis
• Consider re-running just the failed stage before synthesis

Synthesis is incoherent?

• Check if the three parallel proposals are too divergent
• Re-run Stage 7 with explicit instructions to prioritize one proposal as the backbone

Stage 2 finds irrelevant papers?

• Provide more specific research goal
• Cross-disciplinary results are intentional (area doesn't matter for technique relevance)

Original task: {{PROMPT}}