Skip to content

breakdown-configuration.md

Latest commit

 

History

History
201 lines (140 loc) · 6.43 KB

breakdown-configuration.md

File metadata and controls

201 lines (140 loc) · 6.43 KB
Raw

Magic Context - Breakdown & Configuration

Installation

Proper Install - Global Scope (recommended)

opencode plugin @cortexkit/opencode-magic-context --global

This:

  1. Runs npm install --ignore-scripts into the opencode cache
  2. Adds the plugin to opencode.jsonc plugin[] list
  3. Registers server and tui targets

Package Cache Locations

Location Purpose
~/.cache/opencode/packages/@cortexkit/ Proper install by opencode plugin
~/.cache/opencode/node_modules/@cortexkit/ Manual npm install (not canonical)
~/.npm/_npx/ Cached from npx (not persistent)

Backend Data

Location Contents
~/.local/share/cortexkit/magic-context/ SQLite DB, model limits, ONNX model, RPC sockets
~/.local/share/opencode/opencode.db OpenCode conversation database

Verify

grep 'magic-context' ~/.config/opencode/opencode.jsonc
ls ~/.cache/opencode/packages/@cortexkit/
ls ~/.local/share/cortexkit/magic-context/

Plugin Schema

opencode.jsonc Entry

{
  "plugin": [
    "@cortexkit/opencode-magic-context"
  ]
}

No @latest tag - opencode plugin resolves and strips it on install.

Agent Model Binding

All agents reference "model": "opencode/big-pickle" (160K context).

Available opencode/ model IDs:

  • big-pickle - 160,000 context
  • deepseek-v4-flash - 1,000,000 context
  • deepseek-v4-flash-free - 200,000 context
  • deepseek-v4-pro - 1,000,000 context
  • claude-sonnet-4 - 1,000,000 context
  • claude-opus-4-7 - 1,000,000 context
  • gpt-5.4 - 922,000 context
  • gemini-3.5-flash - 1,048,576 context

Compaction Settings

{
  "compaction": {
    "auto": false,
    "prune": false,
    "reserved": 20000
  }
}

When Magic Context is active, compaction is managed by the plugin, not OpenCode's native engine.

Backend Architecture

Data Directory: ~/.local/share/cortexkit/magic-context/ (~96 MB)

File Purpose
context.db (+SHM, +WAL) SQLite - session state, compacted history
model-context-limits-opencode.json Per-model context limits (2,400+ entries)
models/Xenova/all-MiniLM-L6-v2/ ONNX embedding model
rpc/ RPC communication ports
last_announced_version Version announcement tracker

Model Context Limits

The model-context-limits-opencode.json caches context window limits populated from OpenCode's config.providers() SDK endpoint. Magic Context reads the live resolved config (never reads models.json directly) to avoid torn-read issues.

{
  "opencode/big-pickle": 160000,
  "opencode/deepseek-v4-flash": 1000000,
  // ... 2400+ entries
}

Context Resolution Flow

Plugin startup -> warm apiCache from last-known-good file
OpenCode SDK -> config.providers() -> merged models
  (live + compiled-in + auth-plugin caps)
Magic Context -> consumes merged result
Result bounded to [20k, 3M] range

Process Architecture

OpenCode (main process)
  +-- MCP servers (git, mempalace, hf-mcp, etc.)
  +-- Magic Context (child process)
        +-- SQLite (context.db)
        +-- ONNX model (all-MiniLM-L6-v2)
        +-- RPC socket

Troubleshooting

"Model not found: opencode/small-pickle. Did you mean: big-pickle?"

Cause: OpenCode's runtime tries to resolve opencode/small-pickle which doesn't exist in the model catalog. Appears during context compaction cycles.

Solution: Harmless - no functional impact. big-pickle (160K context) is the correct model already configured for all agents. This is an upstream model alias issue - cannot fix via config. Ignore or report to OpenCode.

Duplicate Plugin Entries

Solution: Check both config files:

grep 'magic-context' ~/.config/opencode/opencode.jsonc

# Check if local scope files exist and what's in them:
cat ~/.opencode/opencode.json 2>/dev/null
cat ~/.opencode/tui.json 2>/dev/null

To remove only the plugin entry (not the whole file):

# Targeted removal — removes just the magic-context entry from the plugin array,
# never deletes the entire file even if it contains other config.
if [ -f ~/.opencode/opencode.json ]; then
  # Always filter first, then del only if nothing remains.
jq '(.plugin |= map(select(. != "@cortexkit/opencode-magic-context"))) | if (.plugin | length) == 0 then del(.plugin) else . end' \
  ~/.opencode/opencode.json > /tmp/opencode_tmp.json && mv /tmp/opencode_tmp.json ~/.opencode/opencode.json
  echo "Removed magic-context entry from plugin array"
fi

Keep plugin in only one scope.

Package Missing (Plugin Registered But Not Installed)

Symptom: Plugin listed in config but OpenCode shows loading errors.

Solution: Re-install with the proper command:

opencode plugin @cortexkit/opencode-magic-context --global

This installs to the canonical cache path (~/.cache/opencode/packages/@cortexkit/). Only use manual npm install inside ~/.cache/opencode as a last resort — it installs to node_modules/ which the package cache table above notes is not canonical and may not be discovered by OpenCode.

FAQ

Q: Why does Magic Context use its own SQLite database?

The plugin needs persistent session state for context management that survives OpenCode restarts. It cannot rely on OpenCode's in-memory state. The database at context.db stores compacted history, configuration overrides, and session bookkeeping.

Q: Where does the ONNX model come from?

all-MiniLM-L6-v2 is a sentence-transformer embedding model downloaded from Hugging Face (Xenova/onnx conversion). It's used by Magic Context's sidebar for semantic compression features.

Q: Does Magic Context read ~/.cache/opencode/models.json?

No. Magic Context explicitly does NOT read models.json due to torn-read issues. It consumes the merged, live-resolved model config from OpenCode's config.providers() API endpoint, which already combines the live cache, compiled-in snapshot, opencode.json overrides, and auth-plugin caps.

Q: What happens if both local and global scope have the plugin?

OpenCode loads both entries, potentially causing duplicate initialization. Remove one scope (preferably the local ~/.opencode/ files) to keep only the global scope.

Q: Is the opencode/small-pickle error breaking anything?

No. It's a console warning during compaction with zero functional impact. The model routing still works correctly with the configured opencode/big-pickle.