@datspike/pi-runtime-info-extension

Pi package that exposes the current runtime model, thinking level, and session metadata to agents. It is built for workflows that need reproducible artifacts, model-sensitive subagent orchestration, or reviewer-facing metadata such as model_actual and thinking_actual.

Why this exists

Pi already lets the user select a model and a thinking level, but an agent writing a report cannot always prove which runtime settings were actually used. That becomes painful when you need to:

• write model_actual and thinking_actual into review, research, or handoff artifacts;
• compare requested subagent routing with the resolved runtime model;
• debug model picker and profile overrides;
• keep multi-model review output reproducible.

This extension keeps that check inside Pi, without patching Pi core.

Features

• runtime_info tool for the current Pi session.
• subagent_runtime_info tool for checking an active pi-subagents run by agent_id.
• Tested subagent integration with Tintinweb's pi-subagents: git:github.com/tintinweb/pi-subagents.
• runtime_artifact_fields tool that returns ready-to-paste YAML/JSON artifact fields.
• /runtime-info command for a quick human-readable runtime summary.
• No external service and no network calls.

Quick start

1. Install as a Pi package

pi install npm:@datspike/pi-runtime-info-extension

2. Reload Pi

/reload

3. Ask the agent to verify runtime metadata

Call runtime_info and print the JSON result.

Expected shape:

{
  "scope": "current_session",
  "model": {
    "provider": "openai",
    "id": "gpt-5.5",
    "ref": "openai/gpt-5.5"
  },
  "thinking": {
    "level": "xhigh"
  },
  "session": {
    "id": "...",
    "file": "...",
    "cwd": "/path/to/project"
  },
  "confidence": "selected_model_from_extension_context"
}

Tools

Tool	Use it when	Output
runtime_info	You need the current session runtime.	Model, thinking level, session id/file/cwd, last assistant message metadata.
subagent_runtime_info	You launched a subagent and need the resolved runtime by agent_id.	Subagent status, model, thinking level, session metadata, output file.
runtime_artifact_fields	You are about to write a report, review, plan, or handoff artifact.	Ready artifact fields plus a YAML block.

Example artifact fields:

model_requested: zai/glm-5.1
model_actual: openai/gpt-5.3-codex
thinking_requested: high
thinking_actual: xhigh
runtime_verified_at: 2026-05-04T12:00:00.000Z
runtime_info_source: pi-runtime-info
runtime_info_confidence: subagent_session_model_and_thinking
runtime_scope: subagent
runtime_agent_id: 57444cd3-fb66-4b7

Command

/runtime-info
/runtime-info <agent_id>

Without arguments, the command shows the current session model, thinking level, session id, and cwd. With an agent_id, it shows the same summary for a known subagent record.

Installation options

npm package

pi install npm:@datspike/pi-runtime-info-extension

Git package

pi install git:github.com/datspike/pi-runtime-info-extension

Local development path

{
  "packages": [
    "/absolute/path/to/pi-runtime-info-extension"
  ]
}

The package entrypoint is declared in package.json:

{
  "pi": {
    "extensions": ["./src/index.ts"]
  }
}

Compatibility notes

The current-session tools use documented Pi extension APIs:

• ctx.model;
• ctx.sessionManager;
• pi.getThinkingLevel();
• session assistant message metadata.

The subagent_runtime_info tool is intentionally narrower. It is tested against Tintinweb's pi-subagents package:

pi install git:github.com/tintinweb/pi-subagents

It reads the active manager from globalThis[Symbol.for("pi-subagents:manager")], which is a package-level integration seam rather than a Pi core API. Put pi-subagents in Pi settings before this package so its manager is available when the tool runs. If pi-subagents is not installed, not loaded, or changes that seam, current-session tools continue to work and the subagent tool reports a clear error.

Verification

npm run check
npm pack --dry-run

For a live Pi smoke test:

pi --mode json -p 'Call runtime_info and print its JSON result.'

Current limitations

• runtime_info reports the selected/effective session model. After at least one assistant response, last_assistant_message can also confirm provider-reported message metadata.
• subagent_runtime_info only sees subagents known to the active parent session and the loaded pi-subagents manager.
• If a subagent is started with extensions disabled, it cannot call runtime_info itself; the parent can still call subagent_runtime_info when the manager seam is available.

License

MIT