Agent Deep Research

Deep Research Skill

Perform deep research powered by Google Gemini's deep research agent. Upload documents to file search stores for RAG-grounded answers. Manage research sessions with persistent workspace state.

For AI Agents

Get a full capabilities manifest, decision trees, and output contracts:

uv run {baseDir}/scripts/onboard.py --agent

See AGENTS.md for the complete structured briefing.

Security & Transparency

Credentials: This skill requires a Google/Gemini API key (one of GOOGLE_API_KEY, GEMINI_API_KEY, or GEMINI_DEEP_RESEARCH_API_KEY). The key is read from environment variables and passed to the google-genai SDK. It is never logged, written to files, or transmitted anywhere other than the Google Gemini API.

File uploads: The --context flag uploads local files to Google's ephemeral file search stores for RAG grounding. Files are filtered by MIME type (binary files are rejected), and the ephemeral store is auto-deleted after research completes unless --keep-context is specified. Use --dry-run to preview what would be uploaded without sending anything. Only files you explicitly point --context at are uploaded -- no automatic scanning of parent directories or home folders.

Non-interactive mode: When stdin is not a TTY (agent/CI use), confirmation prompts are automatically skipped. This is by design for agent integration but means an autonomous agent with file system access could trigger uploads. Restrict the paths agents can access, or use --dry-run and --max-cost guards.

No obfuscation: All code is readable Python with PEP 723 inline metadata. No binary blobs, no minified scripts, no telemetry, no analytics. The full source is auditable at github.com/24601/agent-deep-research.

Local state: Research session state is written to .gemini-research.json in the working directory. This file contains interaction IDs, store mappings, and upload hashes -- no credentials or research content. Use state.py gc to clean up orphaned stores from crashed runs.

Prerequisites

• A Google API key (GOOGLE_API_KEY or GEMINI_API_KEY environment variable)
• uv installed (see uv install docs)

Quick Start

# Run a deep research query
uv run {baseDir}/scripts/research.py "What are the latest advances in quantum computing?"

# Check research status
uv run {baseDir}/scripts/research.py status <interaction-id>

# Save a completed report
uv run {baseDir}/scripts/research.py report <interaction-id> --output report.md

# Research grounded in local files (auto-creates store, uploads, cleans up)
uv run {baseDir}/scripts/research.py start "How does auth work?" --context ./src --output report.md

# Export as HTML or PDF
uv run {baseDir}/scripts/research.py start "Analyze the API" --context ./src --format html --output report.html

# Auto-detect prompt template based on context files
uv run {baseDir}/scripts/research.py start "How does auth work?" --context ./src --prompt-template auto --output report.md

Environment Variables

Set one of the following (checked in order of priority):

Optional model configuration:

Research Commands

Start Research

uv run {baseDir}/scripts/research.py start "your research question"

The start subcommand is the default, so research.py "question" and research.py start "question" are equivalent.

Check Status

uv run {baseDir}/scripts/research.py status <interaction-id>

Returns the current status (in_progress, completed, failed) and outputs if available.

Save Report

uv run {baseDir}/scripts/research.py report <interaction-id>

Structured Output (--output-dir)

When --output-dir is used, results are saved to a structured directory:

<output-dir>/
  research-<id>/
    report.md          # Full final report
    metadata.json      # Timing, status, output count, sizes
    interaction.json   # Full interaction data (all outputs, thinking steps)
    sources.json       # Extracted source URLs/citations

A compact JSON summary (under 500 chars) is printed to stdout:

{
  "id": "interaction-123",
  "status": "completed",
  "output_dir": "research-output/research-interaction-1/",
  "report_file": "research-output/research-interaction-1/report.md",
  "report_size_bytes": 45000,
  "duration_seconds": 154,
  "summary": "First 200 chars of the report..."
}

This is the recommended pattern for AI agent integration -- the agent receives a small JSON payload while the full report is written to disk.

Adaptive Polling

When --output or --output-dir is used, the script polls the Gemini API until research completes. By default, it uses history-adaptive polling that learns from past research completion times:

• Completion times are recorded in .gemini-research.json under researchHistory (last 50 entries, separate curves for grounded vs non-grounded research).
• When 3+ matching data points exist, the poll interval is tuned to the historical distribution:
  • Before any research has ever completed: slow polling (30s)
  • In the likely completion window (p25-p75): aggressive polling (5s)
  • In the tail (past p75): moderate polling (15-30s)
  • Unusually long runs (past 1.5x the longest ever): slow polling (60s)
• All intervals are clamped to [2s, 120s] as a fail-safe.

When history is insufficient (<3 data points) or --no-adaptive-poll is passed, a fixed escalating curve is used: 5s (first 30s), 10s (30s-2min), 30s (2-10min), 60s (10min+).

Cost Estimation (--dry-run)

Preview estimated costs before running research:

uv run {baseDir}/scripts/research.py start "Analyze security architecture" --context ./src --dry-run

Outputs a JSON cost estimate to stdout with context upload costs, research query costs, and a total. Estimates are heuristic-based (the Gemini API does not return token counts or billing data) and clearly labeled as such.

After research completes with --output-dir, the metadata.json file includes a usage key with post-run cost estimates based on actual output size and duration.

File Search Store Commands

Manage file search stores for RAG-grounded research and Q&A.

Create a Store

uv run {baseDir}/scripts/store.py create "My Project Docs"

List Stores

uv run {baseDir}/scripts/store.py list

Query a Store

uv run {baseDir}/scripts/store.py query <store-name> "What does the auth module do?"

Delete a Store

uv run {baseDir}/scripts/store.py delete <store-name>

Use --force to skip the confirmation prompt. When stdin is not a TTY (e.g., called by an AI agent), the prompt is automatically skipped.

File Upload

Upload files or entire directories to a file search store.

uv run {baseDir}/scripts/upload.py ./src fileSearchStores/abc123

Hash caches are always saved on successful upload, so a subsequent --smart-sync run will correctly skip unchanged files even if the first upload did not use --smart-sync.

MIME Type Support

36 file extensions are natively supported by the Gemini File Search API. Common programming files (JS, TS, JSON, CSS, YAML, etc.) are automatically uploaded as text/plain via a fallback mechanism. Binary files are rejected. See references/file_search_guide.md for the full list.

File size limit: 100 MB per file.

Session Management

Research IDs and store mappings are cached in .gemini-research.json in the current working directory.

Show Session State

uv run {baseDir}/scripts/state.py show

Show Research Sessions Only

uv run {baseDir}/scripts/state.py research

Show Stores Only

uv run {baseDir}/scripts/state.py stores

JSON Output for Agents

Add --json to any state subcommand to output structured JSON to stdout:

uv run {baseDir}/scripts/state.py --json show
uv run {baseDir}/scripts/state.py --json research
uv run {baseDir}/scripts/state.py --json stores

Clear Session State

uv run {baseDir}/scripts/state.py clear

Use -y to skip the confirmation prompt. When stdin is not a TTY (e.g., called by an AI agent), the prompt is automatically skipped.

Non-Interactive Mode

All confirmation prompts (store.py delete, state.py clear) are automatically skipped when stdin is not a TTY. This allows AI agents and CI pipelines to call these commands without hanging on interactive prompts.

Workflow Example

A typical grounded research workflow:

# 1. Create a file search store
STORE_JSON=$(uv run {baseDir}/scripts/store.py create "Project Codebase")
STORE_NAME=$(echo "$STORE_JSON" | python3 -c "import sys,json; print(json.load(sys.stdin)['name'])")

# 2. Upload your documents
uv run {baseDir}/scripts/upload.py ./docs "$STORE_NAME" --smart-sync

# 3. Query the store directly
uv run {baseDir}/scripts/store.py query "$STORE_NAME" "How is authentication handled?"

# 4. Start grounded deep research (blocking, saves to directory)
uv run {baseDir}/scripts/research.py start "Analyze the security architecture" \
  --store "$STORE_NAME" --output-dir ./research-output --timeout 3600

# 5. Or start non-blocking and check later
RESEARCH_JSON=$(uv run {baseDir}/scripts/research.py start "Analyze the security architecture" --store "$STORE_NAME")
RESEARCH_ID=$(echo "$RESEARCH_JSON" | python3 -c "import sys,json; print(json.load(sys.stdin)['id'])")

# 6. Check progress
uv run {baseDir}/scripts/research.py status "$RESEARCH_ID"

# 7. Save the report when completed
uv run {baseDir}/scripts/research.py report "$RESEARCH_ID" --output-dir ./research-output

Output Convention

All scripts follow a dual-output pattern:

• stderr: Rich-formatted human-readable output (tables, panels, progress bars)
• stdout: Machine-readable JSON for programmatic consumption

This means 2>/dev/null hides the human output, and piping stdout gives clean JSON.