exploring

name	exploring
description	Autonomous scientific literature exploration. Given a theme, explores the literature using subagents until saturation or paper limit, then generates a rich synthesis report.
triggers	["exploring","explore literature","explorar literatura","literature exploration","autonomous exploration"]

Exploring Skill

Autonomous scientific literature exploration. You are the orchestrator.

Trigger

User says: exploring: {theme}

Example: exploring: multi-agent systems for scientific discovery

Your Role

You are a fully autonomous orchestrator. You:

1. Create a workspace
2. Execute search/condense/tag cycles via subagents
3. Generate rich synthesis via writer subagents
4. Deliver final report

CRITICAL CONSTRAINTS:

• You are 100% AUTONOMOUS - NEVER ask user for confirmation
• ONLY use scimesh CLI commands via Bash
• Use Read to check workspace state (log.yaml, papers.yaml, synthesis/*.md)
• Use Write/Edit ONLY for log.yaml updates
• Delegate paper-level operations (reading PDFs, updating paper index.yaml) to subagents
• NO git operations until explicitly told
• Dispatch subagents for all file operations

Parameters

Parameter	Default	Description
limit	500	Maximum papers to collect
providers	openalex,semantic_scholar	Search providers
saturation_threshold	3	Rounds with <5% new papers to stop

User can override: exploring: {theme} --limit 100 --providers openalex

Execution Flow

Phase 0: Ensure Permissions

Before dispatching background subagents, use the Skill tool to grant Write permission for the workspace path:

Skill(skill="update-config", args="allow Write to {workspace_path}/** in project settings")

This ensures background subagents can write condensed.md and other files without permission being auto-denied.

Phase 1: Initialize

uvx scimesh workspace init {workspace_path} --type exploration --question "{theme}"

Use workspace path: explorations/{slugified-theme}

Phase 2: Exploration Loop

WHILE total_papers < limit AND saturation_count < saturation_threshold:

    1. READ STATE
       - Read log.yaml for previous queries, subtopics, suggested_queries
       - Read papers.yaml for total count
       - Use Glob to find papers without condensed.md

    2. DECIDE NEXT QUERY
       - If first round: use theme directly
       - Else: pick from suggested_queries or explore new subtopic
       - Avoid repeating similar queries

    3. EXECUTE SEARCH
       ```bash
       uvx scimesh workspace search {path} "{decided_query}" -p {providers} -n 50
       ```
       Then read `{path}/papers.yaml` to identify new papers.

    4. DISPATCH CONDENSE + TAG SUBAGENTS
       For each new paper (no condensed.md file):

         Step A - Condense (parallel across papers):
         Agent(subagent_type="scimesh:paper-briefer", prompt="""
           Extract from: {workspace_path}/{paper_path}/fulltext.pdf
           Write to: {workspace_path}/{paper_path}/condensed.md
         """)

         Step B - Tag (after condenser completes for each paper):
         Agent(subagent_type="scimesh:paper-tagger", prompt="""
           Condensed: {workspace_path}/{paper_path}/condensed.md
           Protocol: {workspace_path}/index.yaml
           Output: {workspace_path}/{paper_path}/condensed.md (prepend frontmatter)
         """)

       Condensers run in parallel. Each tagger waits for its condenser.

    5. UPDATE LOG
       After condense+tag completes, read frontmatter from condensed.md files.
       Update log.yaml entry with:
       - notes: observations from this round
       - subtopics: newly discovered subtopics (from method_category and tags)
       - suggested_queries: promising directions
       - saturation: true if <5% new papers

    6. CHECK STOPPING CRITERIA
       - total_papers >= limit? → EXIT LOOP
       - saturation_count >= 3? → EXIT LOOP

Phase 3: Synthesis

1. IDENTIFY SUBTOPICS
   Read condensed.md frontmatter for each paper (method_category, tags, relevance.score)
   Group papers by method_category or dominant tag clusters
   Select top 3-5 groupings by paper count as subtopics

2. DISPATCH SCIENTIFIC WRITERS (3-5 in parallel)

   For each major subtopic:
     Agent(subagent_type="scimesh:scientific-writer", prompt="""
       workspace_path: {path}
       task: Synthesize the subtopic '{subtopic}' with depth and critical analysis
       papers: [{list of paper paths in this subtopic}]
       context:
         theme: {theme}
         all_subtopics: {list}
       output_file: synthesis/{subtopic-slug}.md
     """)

   Cross-analysis:
     Agent(subagent_type="scimesh:scientific-writer", prompt="""
       workspace_path: {path}
       task: Analyze patterns, connections, and contradictions across all subtopics
       papers: [{all papers with relevance.score >= 4}]
       context:
         theme: {theme}
         subtopics: {list}
       output_file: synthesis/cross-analysis.md
     """)

   Gaps analysis:
     Agent(subagent_type="scimesh:scientific-writer", prompt="""
       workspace_path: {path}
       task: Identify research gaps, open questions, and future directions
       papers: [{all papers}]
       context:
         theme: {theme}
         subtopics: {list}
       output_file: synthesis/gaps.md
     """)

Phase 4: Final Report

1. FINISH WORKSPACE
   uvx scimesh workspace finish {workspace_path}

2. COMPILE REPORT
   Read all synthesis/*.md files
   Compile into exploration-report.md:

Report structure:

# {Theme}: A Literature Exploration

## Summary

{Write 2-3 paragraphs integrating all synthesis sections.
Provide the "big picture" of where the field stands.}

## {Subtopic 1}

{Content from synthesis/{subtopic-1}.md}

## {Subtopic 2}

{Content from synthesis/{subtopic-2}.md}

...

## Cross-Cutting Analysis

{Content from synthesis/cross-analysis.md}

## Gaps and Opportunities

{Content from synthesis/gaps.md}

## Key Papers

{List papers with relevance.score >= 4, with one-line justification from key_contribution}

---

## Exploration Report

**Theme:** {theme}
**Period:** {started_at} → {finished_at}
**Papers collected:** {total}
**Stop reason:** {saturation reached | limit reached}

### Search Log

| Round | Query | Found | New | Saturation |
|-------|-------|-------|-----|------------|
| 1 | ... | ... | ... | false |
| 2 | ... | ... | ... | false |
...

### Subtopics Discovered

- {subtopic 1} ({n} papers)
- {subtopic 2} ({n} papers)
...

### Statistics

- Total rounds: {n}
- Providers: {list}
- Duplication rate: {%}
- High relevance papers (score >= 4): {n}

Concurrency Strategy

Avoid race conditions:

• Search: ONE at a time (sequential)
• Condense: PARALLEL (each paper writes its own condensed.md)
• Tag: SEQUENTIAL per paper (waits for its condenser), but multiple condense→tag chains run in parallel
• Writers: PARALLEL (each writes different file)
• Log updates: Only orchestrator updates log.yaml (after subagents complete)

Query Strategy

Round 1: Use theme directly Round 2+: Choose from:

1. suggested_queries from previous rounds
2. Expand discovered subtopics
3. Combine subtopics in new ways
4. Search for specific methods mentioned
5. Explore gaps identified

Avoid:

• Repeating exact queries
• Queries too similar to previous
• Overly broad queries (>100 results usually means too broad)

Example Session

User: exploring: quantum machine learning for drug discovery

Orchestrator:
[Creates workspace at explorations/quantum-ml-drug-discovery]

Round 1:
- Query: "quantum machine learning drug discovery"
- Found: 45 papers, 45 new
- Subtopics: variational quantum circuits, molecular simulation, QSPR
- Suggested: "variational quantum eigensolver molecules", "quantum kernel drug"

Round 2:
- Query: "variational quantum eigensolver molecular simulation"
- Found: 38 papers, 31 new
- Subtopics: VQE optimization, noise-resilient circuits
- Saturation: false

Round 3:
- Query: "quantum kernel methods QSPR"
- Found: 29 papers, 22 new
...

[After 8 rounds, saturation detected]

Synthesis:
- Dispatching writer for "variational quantum circuits"
- Dispatching writer for "molecular simulation"
- Dispatching writer for "quantum kernels"
- Dispatching writer for cross-analysis
- Dispatching writer for gaps

[Compiling final report]

COMPLETE:
- Workspace: explorations/quantum-ml-drug-discovery
- Report: explorations/quantum-ml-drug-discovery/exploration-report.md
- Papers: 247
- Subtopics: 5
- Stop reason: saturation reached

Error Handling & Recovery

Resuming an Interrupted Exploration

If the workspace already exists when triggered:

1. Read current state: Check log.yaml for last round, papers.yaml for total count
2. Detect uncondensed papers: Papers without condensed.md file need processing
3. Resume from last round: Continue the exploration loop from where it stopped
4. Do NOT re-initialize: Skip workspace init if workspace already has papers

Subagent Failures

• Search fails: Log the error in notes, try a different query formulation next round
• Condense/tag fails: Skip paper, log the error in notes, continue with others
• Writer fails: Retry once with same prompt. If still fails, write a placeholder noting the missing section
• All failures logged in the final exploration-report.md appendix

Rate Limiting

• If a provider returns 429 (rate limited), wait 30 seconds before retrying
• Alternate between providers across rounds to distribute load
• If all providers are rate-limited, pause for 60 seconds

Deduplication

Before dispatching condense subagents, verify the paper does NOT already have a condensed.md file. Use Glob to check:

Glob(pattern="{workspace_path}/papers/**/condensed.md")

Papers found in multiple search rounds should only be condensed once.