Magic Context — Pi extension

Cross-session memory and context management for Pi coding agent. Shares the same SQLite database as the OpenCode plugin, so memories, embeddings, dreamer state, and project knowledge follow you across both harnesses.

⚠️ Beta release. The Pi extension is newer than the OpenCode plugin and is published as beta. Core flows (tagging, historian, memories, dreamer, /ctx-aug) are validated end-to-end on interactive pi sessions and pi --print --mode json subagents. Please report issues — every Pi-side rough edge is fixable.

Requires @earendil-works/pi-coding-agent and @earendil-works/pi-tui >= 0.71.0.

What it does

Magic Context is a context engine that keeps long Pi sessions productive by:

| Feature | What it does | |---|---| | Tagging + drops | Tags every assistant/user/tool message with §N§ so you can drop specific turns later via ctx_reduce | | Historian | Background subagent compresses old conversation into compartments + facts at threshold pressure or commit boundaries | | <session-history> injection | Prepends compressed history into the system prompt every turn so the agent never loses context | | Project memories | Persistent cross-session knowledge store with embedding-based semantic search | | Dreamer | Scheduled background subagent that consolidates, verifies, archives, and improves stored memories | | /ctx-aug | On-demand sidekick that augments the next turn with relevant memories | | Auto-search hint | When user prompts mention previously-discussed topics, appends a compact memory hint | | Note nudges | Surface deferred intentions at natural work boundaries (commit, todo completion, historian publication) | | Cross-harness sharing | Memories written from OpenCode appear in Pi (and vice versa) for the same project |

Installation

The fastest path is the unified Magic Context CLI — --harness pi selects the Pi-specific setup pipeline (registers the extension with Pi, writes a sensible magic-context.jsonc, and verifies your model picks):

npx @cortexkit/magic-context@latest setup --harness pi

This handles everything for you:

1. Adds npm:@cortexkit/pi-magic-context to Pi's packages array in ~/.pi/agent/settings.json (the same place pi install writes to)
2. Creates ~/.pi/agent/magic-context.jsonc with defaults
3. Prompts you for historian, dreamer, sidekick, and embedding model choices
4. Warns about provider-specific gotchas (e.g. GitHub Copilot reasoning models need an explicit thinking_level)

If you'd rather register the Pi extension package directly with Pi (skipping the wizard), use Pi's own installer:

pi install npm:@cortexkit/pi-magic-context

This adds the extension to ~/.pi/agent/settings.json but won't write magic-context.jsonc for you — you'll need to create it manually (see Configuration below).

To check installation health later:

npx @cortexkit/magic-context@latest doctor --harness pi

Configuration

Magic Context reads two config files (in this priority order):

1. $cwd/.pi/magic-context.jsonc (project-level overrides)
2. ~/.pi/agent/magic-context.jsonc (user-level defaults)

Both are merged through a Zod schema. Invalid fields fall back to defaults — bad config never disables the plugin entirely.

Minimal config

{
  "$schema": "https://raw.githubusercontent.com/cortexkit/magic-context/master/assets/magic-context.schema.json",
  "enabled": true,
  "historian": {
    "model": "anthropic/claude-haiku-4-5"
  },
  "embedding": {
    "provider": "local"
  }
}

For the full configuration reference (including dreamer, sidekick, auto-search, and experimental features), see CONFIGURATION.md in the main repository — the schema is shared between both plugins.

Slash commands

All commands trigger triggerTurn: false (never sent to the LLM):

| Command | What it does | |---|---| | /ctx-status | Live token breakdown + queued ops + cache state | | /ctx-flush | Force-process pending ops queue | | /ctx-recomp | Rebuild compartments from raw history (heavy operation) | | /ctx-dream | Trigger a dream run on demand | | /ctx-aug | Augment your next prompt with sidekick-retrieved memories |

Storage

Magic Context stores everything in a single shared SQLite database at:

~/.local/share/cortexkit/magic-context/context.db

This is the same database the OpenCode plugin uses. Tables are scoped by:

• harness column ('pi' or 'opencode') for session-scoped data (tags, compartments, facts, notes)
• project_path (resolved git root) for project-scoped data (memories, embeddings, dreamer runs)

So memories and dreamer state are shared across both harnesses for the same project; per-session tagging stays correctly attributed.

Storage failures are fatal — Magic Context will refuse to register hooks rather than run with ephemeral state, since that would let context grow unbounded across restarts.

Cross-harness coherence

For semantic search to work across harnesses, both plugins must use the same embedding model. Magic Context detects mismatch on Pi startup and warns:

WARN embedding model mismatch detected for project ...:
stored vectors use "openai-compatible:Qwen/Qwen3-Embedding-8B" but Pi is configured with "local:Xenova/all-MiniLM-L6-v2".
Cross-harness search will return zero results until vectors are re-embedded.

Easiest fix: configure embedding once in ~/.pi/agent/magic-context.jsonc (Pi) and ~/.config/opencode/magic-context.jsonc (OpenCode) with identical settings.

Tools available to the agent

| Tool | Action set | Purpose | |---|---|---| | ctx_search | n/a | Search memories + raw session history; returns ranked results with previews | | ctx_memory | write, delete | Manage project memories explicitly (most writes happen via dreamer instead) | | ctx_note | read, write, update, dismiss | Defer intentions for later — surfaced via note nudges at work boundaries |

ctx_expand and ctx_reduce from the OpenCode plugin are intentionally not exposed on Pi — they depend on raw OpenCode message ordinals, while Pi has its own message identity model. Drops still happen automatically via threshold-driven historian; you don't need an explicit ctx_reduce to trigger reduction.

Architecture & implementation

This package is part of the magic-context monorepo. The Pi extension shares the core implementation with the OpenCode plugin via the @magic-context/core workspace dependency, exposing only the Pi-specific adapter layer:

| Pi-specific module | Responsibility | |---|---| | context-handler.ts | Pi pi.on("context", ...) adapter — tags, drops, runs nudges and auto-search | | subagent-runner.ts | Spawns pi --print --mode json --no-extensions --extension <lean-entry> ... for historian/sidekick/dreamer subagents, with a 2-second drain after the terminal message_end so child processes don't keep the parent waiting | | tools/ | Pi pi.registerTool wrappers around the shared tool implementations | | commands/ | Pi pi.registerCommand wrappers for the five /ctx-* slash commands | | dreamer/ | Pi-side adapter for the shared dreamer scheduler | | system-prompt.ts | Pi before_agent_start injector for <session-history>, <project-memory>, <project-docs> | | config/ | Pi-convention config loader ($cwd/.pi/magic-context.jsonc + ~/.pi/agent/magic-context.jsonc) |

The CLI lives in the unified @cortexkit/magic-context package — setup --harness pi and doctor --harness pi route to the Pi-specific code paths in packages/cli/src/commands/.

For deeper architectural detail, see the main repo's ARCHITECTURE.md.

License

MIT — see LICENSE.