| name | _summarize-problem-context |
| description | Create research document with problem statement and context from multiple sources (repositories, papers, files) |
| user-invocable | false |
| allowed-tools | ["Read","Write","Glob","Grep","Task","WebFetch","litellm_web_search","AskUserQuestion","Bash(gh *)"] |
ARGUMENTS
[PROBLEM_FILE_PATH] (required): Path to a file containing the problem statement
Optional pre-configured sources (if provided, skip asking user):
[INCLUDE_CURRENT_REPO] (optional): true/false - include current repository[LOCAL_REPO_PATHS] (optional): List of local repository paths[GITHUB_REPO_URLS] (optional): List of GitHub repository URLs[REMOTE_URLS] (optional): List of paper/documentation URLs[WEB_SEARCH_QUERIES] (optional): List of web search queries[SKIP_BACKGROUND] (optional): true to skip background entirely[OUTPUT_FILE] (optional): Path for the output file. Defaults to [PROBLEM_DIR]/research.md. When set, use this path instead of the default.[DOCUMENT_TITLE] (optional): Top-level heading for the output document. Defaults to "Research Document". Example: "Problem Context".
If any optional source arguments are provided, skip Steps 2-3 and proceed directly to Step 4 (Launch Agents).
DERIVED PATHS
[PROBLEM_DIR] = directory containing the problem file[OUTPUT_FILE] = argument value if provided, otherwise [PROBLEM_DIR]/research.md
TASK
Read the problem statement from [PROBLEM_FILE_PATH], gather context from multiple sources (current repository, other repositories, online papers, etc.), then create [OUTPUT_FILE] containing the problem statement and all relevant background context.
STEPS
Step 1: Read the Problem Statement
Read and understand the problem statement from [PROBLEM_FILE_PATH]. Identify:
- The core problem being addressed
- Key technical domains involved
- Specific components, modules, or areas that may be relevant
Step 2: Ask About Context Sources (Skip if pre-configured)
If source arguments were provided (any of [INCLUDE_CURRENT_REPO], [LOCAL_REPO_PATHS], [GITHUB_REPO_URLS], [REMOTE_URLS], [WEB_SEARCH_QUERIES]), skip Steps 2-3 and proceed directly to Step 4.
If [SKIP_BACKGROUND] = true, skip to Step 5 and create a minimal document with just the problem statement.
Otherwise, current repository is included by default. Ask about additional sources:
AskUserQuestion:
questions:
- question: "The current repository will be analyzed by default. Do you want to add more context sources?"
header: "Context"
multiSelect: true
options:
- label: "Add other local repositories"
description: "Include related code from other local directories"
- label: "Add GitHub repositories"
description: "Include remote GitHub repos (will be fetched via API)"
- label: "Add papers or URLs"
description: "Include arXiv papers, docs, blog posts, or other web content"
- label: "Search the web for relevant content"
description: "Use web search to find papers and resources related to the problem"
Store results:
[INCLUDE_CURRENT_REPO] = true (always default)[ADD_LOCAL_REPOS] = true if first option selected[ADD_GITHUB_REPOS] = true if second option selected[ADD_PAPERS] = true if third option selected[ADD_WEB_SEARCH] = true if fourth option selected
Step 3: Gather Source Details
3.1: Collect Local Repository Paths (if [ADD_LOCAL_REPOS] is true)
AskUserQuestion:
questions:
- question: "Enter the paths to local repositories (use 'Other' to type paths, one per line or comma-separated):"
header: "Local Repos"
multiSelect: false
options:
- label: "I'll provide the paths"
description: "Enter absolute paths like /Users/me/projects/my-repo"
- label: "Let me browse first"
description: "I need to check which local repos are relevant"
If user selects "I'll provide the paths":
- The user will type paths in the "Other" text field
- Parse the input (split by newlines or commas)
- Validate each path exists and is a directory
- Store as
[LOCAL_REPO_PATHS] (list of valid paths)
- Report any invalid paths to the user
3.2: Collect GitHub Repository URLs (if [ADD_GITHUB_REPOS] is true)
AskUserQuestion:
questions:
- question: "Enter GitHub repository URLs (use 'Other' to type URLs, one per line or comma-separated):"
header: "GitHub Repos"
multiSelect: false
options:
- label: "I'll provide the URLs"
description: "Enter URLs like https://github.com/owner/repo or owner/repo shorthand"
- label: "Let me look them up first"
description: "I need to find the relevant GitHub repos"
If user selects "I'll provide the URLs":
- The user will type URLs in the "Other" text field
- Parse the input (split by newlines or commas)
- Normalize URLs (e.g., "owner/repo" → "https://github.com/owner/repo")
- Store as
[GITHUB_REPO_URLS] (list of normalized URLs)
3.3: Collect Paper/Documentation URLs (if [ADD_PAPERS] is true)
AskUserQuestion:
questions:
- question: "Enter URLs to papers or documentation (use 'Other' to type URLs, one per line or comma-separated):"
header: "Paper URLs"
multiSelect: false
options:
- label: "I'll provide the URLs"
description: "Enter URLs to arXiv, papers, docs, blog posts, etc."
- label: "I'll paste content instead"
description: "I'll copy-paste the relevant content directly in a follow-up"
If user selects "I'll provide the URLs":
- The user will type URLs in the "Other" text field
- Parse the input (split by newlines or commas)
- Store as
[PAPER_URLS] (list of URLs)
If user selects "I'll paste content instead":
- Set
[PASTE_PAPER_CONTENT] = true
- After all questions, prompt user: "Please paste the paper/documentation content:"
- Store pasted content as
[PASTED_CONTENT]
3.4: Configure Web Search (if [ADD_WEB_SEARCH] is true)
AskUserQuestion:
questions:
- question: "What topics should I search for? (Enter custom queries or use auto-generated)"
header: "Web Search"
multiSelect: false
options:
- label: "Auto-generate search queries (Recommended)"
description: "I'll derive search terms from your problem statement"
- label: "I'll provide specific queries"
description: "Enter your own search terms in the text field"
If auto-generate selected:
- Generate 3-5 search queries based on problem statement keywords
- Store as
[WEB_SEARCH_QUERIES] (list of query strings)
If user provides queries:
- Parse user input (split by newlines or commas)
- Store as
[WEB_SEARCH_QUERIES]
Step 3.5: Confirm Sources Summary
Display a summary of collected sources and ask for confirmation:
Summarize what will be analyzed:
- Current repository: [Yes/No]
- Local repositories: [count] paths
- /path/to/repo1
- /path/to/repo2
- GitHub repositories: [count] URLs
- https://github.com/owner/repo1
- https://github.com/owner/repo2
- Papers/Documentation: [count] URLs
- https://arxiv.org/...
- https://...
AskUserQuestion:
questions:
- question: "Does this look correct? Ready to generate background?"
header: "Confirm"
multiSelect: false
options:
- label: "Yes, proceed"
description: "Generate background from these sources"
- label: "No, let me modify"
description: "Go back and change the source list"
If user selects "No, let me modify":
- Return to Step 2 to re-collect sources
Validation: At least one source must be selected. If no sources are selected (current repo is off, no other repos, no papers), prompt:
AskUserQuestion:
questions:
- question: "No context sources selected. What would you like to do?"
header: "No Sources"
multiSelect: false
options:
- label: "Include current repository"
description: "Add the current repo as the only context source"
- label: "Let me add sources"
description: "Go back and specify repositories or papers to include"
- label: "Skip background entirely"
description: "Generate ideas without any background context"
Step 4: Launch Parallel Background Agents
Launch independent background agents for each context source. This allows concurrent summarization for faster results.
4.1: Current Repository Agent (if [INCLUDE_CURRENT_REPO] is true)
Task tool:
description: "Summarize current repository"
subagent_type: Explore
run_in_background: true
prompt: |
Analyze the current repository to gather context relevant to this problem:
[Insert problem statement summary here]
Explore and summarize:
1. Repository Structure: Top-level directories, technology stack
2. Relevant Code Areas: Components related to the problem
3. Documentation: README, architecture docs, API docs
4. Patterns and Conventions: Code patterns relevant to the problem
Return a structured summary in markdown format with sections:
- Repository Overview
- Technology Stack
- Relevant Architecture
- Key Files and Components
- Existing Patterns
Store agent ID as [CURRENT_REPO_AGENT_ID] (or null if not included).
4.2: Local Repository Agents (if [LOCAL_REPO_PATHS] is non-empty)
For each path in [LOCAL_REPO_PATHS]:
Task tool:
description: "Summarize local repo: [REPO_NAME]"
subagent_type: Explore
run_in_background: true
prompt: |
Analyze the repository at [REPO_PATH] to gather context relevant to this problem:
[Insert problem statement summary here]
Explore and summarize:
1. Repository purpose and structure
2. Technology stack and dependencies
3. Components or patterns relevant to the problem
4. How this repository relates to the main problem
Return a structured summary in markdown format.
Store agent IDs as [LOCAL_REPO_AGENT_IDS] (list).
4.3: GitHub Repository Agents (if [GITHUB_REPO_URLS] is non-empty)
For each URL in [GITHUB_REPO_URLS]:
Task tool:
description: "Summarize GitHub repo: [REPO_NAME]"
subagent_type: general-purpose
run_in_background: true
prompt: |
Explore and summarize the GitHub repository at: [GITHUB_URL]
Use the GitHub CLI to explore the remote repository:
1. `gh repo view [OWNER/REPO] --json description,readme` - get basic info and README
2. `gh api repos/[OWNER]/[REPO]/contents` - list top-level structure
3. `gh api repos/[OWNER]/[REPO]/contents/[PATH]` - fetch specific files
4. Look for key files: README.md, package.json, requirements.txt, go.mod, pyproject.toml, etc.
Focus on extracting information relevant to this problem:
[Insert problem statement summary here]
Explore and summarize:
1. Repository purpose and description
2. Technology stack and dependencies
3. Architecture and key components
4. How this repository relates to the main problem
Return a structured summary in markdown format with sections:
- Repository Overview
- Technology Stack
- Relevant Architecture
- Key Files and Components
Store agent IDs as [GITHUB_REPO_AGENT_IDS] (list).
4.4: Paper/URL Agents (if [PAPER_URLS] is non-empty)
For each URL in [PAPER_URLS]:
Task tool:
description: "Summarize paper: [URL_TITLE]"
subagent_type: general-purpose
run_in_background: true
prompt: |
Fetch and summarize the content at this URL: [URL]
Use the WebFetch tool to retrieve the content.
Focus on extracting information relevant to this problem:
[Insert problem statement summary here]
Summarize:
1. Paper/document title and authors (if applicable)
2. Key concepts and findings
3. Technical approaches or methods described
4. How this relates to the problem at hand
Return a structured summary in markdown format.
Store agent IDs as [PAPER_AGENT_IDS] (list).
4.5: Web Search Agents (if [WEB_SEARCH_QUERIES] is non-empty)
For each query in [WEB_SEARCH_QUERIES]:
Task tool:
description: "Web search: [QUERY_PREVIEW]"
subagent_type: general-purpose
run_in_background: true
prompt: |
Search the web for information related to: "[QUERY]"
Use the litellm_web_search tool with this query.
Then use WebFetch to retrieve the most relevant 2-3 results.
Focus on extracting information relevant to this problem:
[Insert problem statement summary here]
Summarize:
1. Key findings from search results
2. Relevant papers, articles, or documentation found
3. Technical approaches or methods mentioned
4. How this relates to the problem at hand
Return a structured summary in markdown format.
Store agent IDs as [WEB_SEARCH_AGENT_IDS] (list).
4.6: Pasted Content Processing (if [PASTED_CONTENT] exists)
If user pasted content directly instead of providing URLs:
Task tool:
description: "Summarize pasted content"
subagent_type: general-purpose
run_in_background: true
prompt: |
Summarize the following content that was provided by the user:
---
[PASTED_CONTENT]
---
Focus on extracting information relevant to this problem:
[Insert problem statement summary here]
Summarize:
1. Document type and source (if identifiable)
2. Key concepts and findings
3. Technical approaches or methods described
4. How this relates to the problem at hand
Return a structured summary in markdown format.
Store agent ID as [PASTED_CONTENT_AGENT_ID] (or null if not applicable).
Step 5: Collect Results and Create Output File
Wait for all background agents to complete. Use TaskOutput to collect results from each agent. If an agent failed, log a warning noting which source failed and continue with results from successful agents.
Before writing: Ensure the output directory exists: Bash("mkdir -p \"$(dirname '[OUTPUT_FILE]')\"")
Create [OUTPUT_FILE] with the following structure:
# [DOCUMENT_TITLE]
## Problem Statement
[Copy the ENTIRE content from [PROBLEM_FILE_PATH] verbatim here]
---
# Background
[If [INCLUDE_CURRENT_REPO] is true:]
## Current Repository Context
[Insert summary from [CURRENT_REPO_AGENT_ID]]
[If [LOCAL_REPO_PATHS] is non-empty:]
## Local Repositories
### [Repository Name 1]
**Path:** [LOCAL_PATH]
[Insert summary from first local repo agent]
### [Repository Name 2]
**Path:** [LOCAL_PATH]
[Insert summary from second local repo agent]
...
[If [GITHUB_REPO_URLS] is non-empty:]
## GitHub Repositories
### [Repository Name 1]
**URL:** [GITHUB_URL]
[Insert summary from first GitHub repo agent]
### [Repository Name 2]
**URL:** [GITHUB_URL]
[Insert summary from second GitHub repo agent]
...
[If [PAPER_URLS] is non-empty:]
## Referenced Papers and Documentation
### [Paper/Doc Title 1]
**Source:** [URL]
[Insert summary from first paper agent]
### [Paper/Doc Title 2]
**Source:** [URL]
[Insert summary from second paper agent]
...
[If [WEB_SEARCH_QUERIES] is non-empty:]
## Web Research
### Search: "[Query 1]"
[Insert summary from first web search agent]
### Search: "[Query 2]"
[Insert summary from second web search agent]
...
[If [PASTED_CONTENT] exists:]
## User-Provided Documentation
[Insert summary from [PASTED_CONTENT_AGENT_ID]]
## Cross-Source Insights
[Brief synthesis of how the different sources relate to each other and the problem]
---
IMPORTANT: The file ends with --- followed by a blank line. This is where subsequent ideas and reviews will be appended.
GUIDELINES
- Stay focused: Only include information relevant to the problem statement. Avoid documenting unrelated content.
- Be concise: Provide enough context to understand the problem space without overwhelming detail.
- Be specific: Reference actual file paths, URLs, function names, and patterns found in the sources.
- Prioritize: Put the most important context first. Someone reading should quickly understand the relevant context.
- Parallelize: Launch all source summarization agents concurrently to minimize wait time.
- Handle failures gracefully: If a URL fetch fails or a path doesn't exist, note it and continue with available sources.
