[Pi] Cold start: pi-magic-context module import ~3.8s (7.7MB bundle dominates startup)

[WAY29]

Short description

[Pi] @cortexkit/pi-magic-context cold-starts ~3.8s on module import (dominates Pi startup)

What happened?

On macOS arm64, starting Pi with npm:@cortexkit/pi-magic-context installed makes CLI startup ~4s+ slower. Pi's built-in PI_TIMING=1 attributes almost all of that to the Magic Context package's module import, not factory init and not other local extensions.

Expected: Plugin load should not dominate interactive CLI cold start (ideally well under a few hundred ms, or defer heavy deps).

Observed: Magic Context alone accounts for ~89% of extension load time in a typical Pi install with many other extensions present.

Repro

# Full startup timings
PI_TIMING=1 PI_OFFLINE=1 pi -p "ok" --no-session

# Baseline without any discovered extensions
PI_TIMING=1 PI_OFFLINE=1 pi --no-extensions -p "ok" --no-session

Measured (Pi 0.82.0, plugin 0.32.4, Node v24.8.0, macOS 15.3.1 arm64)

With extensions (createAgentSessionRuntime ≈ total extension load):

--- Startup Timings: main ---
  createAgentSessionRuntime: 4330ms
  TOTAL: 4336ms

--- Startup Timings: extensions ---
  ...other extensions: mostly <30ms each; a few in the ~100ms range...
  ~/.pi/agent/npm/node_modules/@cortexkit/pi-magic-context/dist/index.js module import: 3787ms
  ~/.pi/agent/npm/node_modules/@cortexkit/pi-magic-context/dist/index.js factory: 41ms
  TOTAL: 4294ms

Without discovered extensions:

--- Startup Timings: main ---
  createAgentSessionRuntime: 43ms
  TOTAL: 49ms

Summary from the same run:

item	time
@cortexkit/pi-magic-context module import	~3787ms
@cortexkit/pi-magic-context factory	~41ms
all other extensions combined	~466ms
extensions TOTAL	~4294ms
Pi startup without discovered extensions	~49ms

Likely cause

Installed package contents:

• @cortexkit/pi-magic-context@0.32.4
• dist/index.js ≈ 7.7MB single ESM bundle
• dist/subagent-entry.js ≈ 4.2MB
• Build externalizes @huggingface/transformers, but the main bundle still eagerly evaluates a large inlined quickjs payload (many quickjs symbols in the bundle)
• Declared deps also pull a heavy local-embedding stack (@huggingface/transformers, onnxruntime-node ~210MB on disk, etc.)

This matches the timing profile: cost is almost entirely module import / evaluate, not the factory body (~40ms).

Note: this report is about cold-start import time, not runtime transform lag (#224) or idle TUI CPU (#200).

Suggested directions (not prescriptions)

Any of these would help Pi users a lot:

1. Split/lazy-load quickjs and other heavy subtrees so the Pi extension factory path doesn't parse a 7–8MB bundle on every process start
2. Defer embedding/transformers/onnx work until first actual embedding use (or when embedding.provider !== "off")
3. Ship a thinner Pi entrypoint that only registers hooks/tools, with optional heavy features behind dynamic import()
4. Document a "lightweight / no local embedding" install path if one already exists

Happy to re-run timings against a canary build.

Diagnostics

## Environment
- Pi plugin: @cortexkit/pi-magic-context@0.32.4
- Pi: 0.82.0
- OS: darwin arm64 (macOS 15.3.1)
- Node: v24.8.0

## Configuration
Config from `~/.config/cortexkit/magic-context.jsonc`:

{
  "$schema": "https://raw.githubusercontent.com/cortexkit/magic-context/master/assets/magic-context.schema.json",
  "historian": {
    "model": "<redacted>"
  },
  "dreamer": {
    "model": "<redacted>"
  },
  "sidekick": {
    "disable": true
  },
  "embedding": {
    "provider": "local",
    "model": "Xenova/all-MiniLM-L6-v2"
  }
}

## Package install
- settings packages includes: `npm:@cortexkit/pi-magic-context`
- resolved path: `~/.pi/agent/npm/node_modules/@cortexkit/pi-magic-context`
- dist/index.js size: ~7.7MB

Plugin version

0.32.4

OpenCode version

N/A (Pi only)

Platform

macOS arm64

Client

Pi

Log output (optional)

# Command
PI_TIMING=1 PI_OFFLINE=1 pi -p "ok" --no-session

# Key lines
createAgentSessionRuntime: 4330ms
@cortexkit/pi-magic-context/dist/index.js module import: 3787ms
@cortexkit/pi-magic-context/dist/index.js factory: 41ms
extensions TOTAL: 4294ms

[alfonso-magic-context]

v0.33.0 (just published) carries the fix: the QuickJS sandbox no longer parses at import — code-split into a lazy chunk, entry 8.3MB → 1.25MB, warm import ~150ms measured. A PI_TIMING=1 re-measure on your machine after updating would be a welcome confirmation of the cold-start delta.

[WAY29]

Thanks for shipping v0.33.0 — re-measured after pi update npm:@cortexkit/pi-magic-context (now 0.33.0).

Result on this machine: cold start essentially unchanged

PI_TIMING=1 PI_OFFLINE=1 pi -p "ok" --no-session (macOS arm64, Pi 0.82.0, Node v24.8.0):

version	magic-context module import	factory	createAgentSessionRuntime
0.32.4	~3787ms	~41ms	~4330ms
0.33.0	~3500–3900ms (latest sample 3838ms)	~45ms	~4463ms

So the entry split helped file layout, but Pi cold-start still spends ~3.5–4s inside @cortexkit/pi-magic-context import.

What I see in the installed 0.33.0 package

• dist/index.js is indeed ~1.2MB (down from ~7.7MB) ✅
• QuickJS looks lazy: dynamic import("./index-ayd60xnw.js") (~2.3MB) ✅
• But the entry still has a static from "./index-f3jfb2ed.js" and that chunk is ~4.3MB (esprima + sqlite/shared plugin core, etc.)
• Eager graph ≈ index.js (1.2MB) + index-f3jfb2ed.js (4.3MB) ≈ 5.5MB still parsed on every process start

That lines up with why import stays multi-second here despite the smaller entry filename.

Happy to re-test another build if more of that static chunk can be deferred past extension factory load.