@forecastx/deep-research

Deep research tool for Pi-compatible agent CLIs with full streaming output

Packages

Package details

extension

Install @forecastx/deep-research from npm and Pi will load the resources declared by the package manifest.

npm report
$ pi install npm:@forecastx/deep-research
Package
@forecastx/deep-research
Version
0.1.1
Published
May 26, 2026
Downloads
not available
Author
yanjin734
License
MIT
Types
extension
Size
28.3 KB
Dependencies
0 dependencies · 0 peers
Pi manifest JSON
{
  "extensions": [
    "./src/index.ts"
  ]
}

Security note

Pi packages can execute code and influence agent behavior. Review the source before installing third-party packages.

README

@forecastx/deep-research

Deep research tool for Pi-compatible agent CLIs with full streaming output.

Provides two tools that the LLM can call:

  • deep_research(query) — 5+ web searches, structured markdown report (1-3 min)
  • super_deep_research(query) — multi-phase research with planning, 8-15 searches, 3000-8000 char report (2-5 min)

All intermediate steps (search calls, planning text, token usage) are streamed in real-time via the Pi extension onUpdate callback.

Install

pi install npm:@forecastx/deep-research

Or project-local:

pi install -l npm:@forecastx/deep-research

Or symlink for development:

ln -sf /path/to/packages/deep-research ~/.pi/agent/extensions/deep-research

Requirements

  • Pi-compatible CLI (pi, forecastx, or any pi-mono fork)
  • Web search tool available (via MCP or built-in) — the subprocess agent uses it to perform searches
  • LLM API key configured in the host CLI

Provider Override

By default, the subprocess inherits the host CLI's configured provider and model. To override:

Variable Description
RESEARCH_PROVIDER Provider name (e.g. anthropic, openai, deepseek, openrouter)
RESEARCH_MODEL Model ID (e.g. claude-sonnet-4-20250514, gpt-4o, deepseek-chat)

These map to --provider and --model flags on the subprocess.

How It Works

  1. The LLM calls super_deep_research({ query: "..." })
  2. The extension spawns a pi --mode json -p --no-session subprocess with a research prompt
  3. The subprocess performs multi-phase research using available search tools
  4. JSON events from stdout are parsed in real-time
  5. Each search call, text delta, and usage update is streamed back via onUpdate
  6. Final report is saved to ~/.forecastx-reports/<id>/ and returned to the LLM

Streaming Output

The TUI shows live progress during research:

⏳ super_deep_research "2026 FIFA World Cup prediction"
  researching (6 searches), 3 turns, 45s
  → web_search "2026 FIFA World Cup favorites odds"
  → web_search "France squad depth 2026"
  → web_search "Brazil national team form 2025-2026"

When complete:

✓ super_deep_research done
  12 searches | 8 turns | 3m 24s | $0.0847
  Report: a1b2c3d4e5f6
  研究报告：2026年世界杯冠军预测分析
  1. 摘要 (1.2k)
  2. 背景与上下文 (0.8k)
  3. 核心发现 (3.4k)
  ...

Reports

Reports are stored at ~/.forecastx-reports/ with structure:

<id>/
  report.md       — full markdown
  meta.json       — metadata (query, stats, summary, toc, sources)
  sections/
    01.md, 02.md  — individual sections

Compatible CLIs

CLI Install
pi pi install npm:@forecastx/deep-research
forecastx forecastx install npm:@forecastx/deep-research
Any pi-mono fork Same install command

License

MIT