fish_bridge

Session-scoped knowledge graph engine for AI chat context compression.

Converts raw AI chat (40k+ tokens) into a compact typed knowledge graph (~300–800 tokens) and writes it to .github/copilot-instructions.md or CLAUDE.md — automatically included in every AI turn across all modes (ask, edit, agent). No MCP server required for the core workflow.

Raw session (40k tokens) → [fish_bridge] → Compressed graph (350 tokens)
                                              written to copilot-instructions.md
                                              picked up by every AI turn automatically

Install

Don't have uv? Get it first: curl -LsSf https://astral.sh/uv/install.sh | sh (macOS/Linux) or see uv docs. It replaces pip + pipx + pyenv in one tool — no virtualenv management needed.

Recommended — uv tool install (installs both the fish-bridge CLI and fish-bridge-mcp MCP server on your PATH):

# Local Ollama backend — free, offline (requires Ollama running)
uv tool install fish-bridge-mcp

# Gemini backend (~$0.0002/turn, ~95% quality — recommended cloud option)
uv tool install "fish-bridge-mcp[gemini]"
export GEMINI_API_KEY=...

# Claude backend (~$0.002/turn, ~97% quality)
uv tool install "fish-bridge-mcp[claude]"
export ANTHROPIC_API_KEY=sk-ant-...

# OpenAI backend (~$0.0003/turn, ~93% quality)
uv tool install "fish-bridge-mcp[openai]"
export OPENAI_API_KEY=sk-...

# Everything
uv tool install "fish-bridge-mcp[all]"

After install, two commands are available on your PATH:

• fish-bridge — the main CLI (ingest, compile, show, serve, ...)
• fish-bridge-mcp — the MCP server for VS Code agent mode

MCP config only (no permanent install needed): use uvx directly in your .vscode/mcp.json — it downloads and runs the MCP server on demand:

{ "command": "uvx", "args": ["fish-bridge-mcp"] }

See the MCP server section below for the full config.

pip install fish-bridge-mcp
pip install "fish-bridge-mcp[gemini]"   # with Gemini backend
pip install "fish-bridge-mcp[claude]"   # with Claude backend
pip install "fish-bridge-mcp[all]"      # everything

2-minute quickstart

# 1. Initialize for your project
fish-bridge init --tool copilot --project ./

# 2. Ingest the latest Copilot session (auto-discovers JSONL on macOS/Linux/Windows)
fish-bridge ingest --source copilot

# 3. View the graph
fish-bridge show

# 4. Compile to your instructions file (done automatically after ingest)
fish-bridge compile

That's it. .github/copilot-instructions.md now contains a ~350-token compressed summary of your session, replacing raw history in every future turn.

Backends

Configure with:

fish-bridge config --backend gemini
# or set GEMINI_API_KEY / ANTHROPIC_API_KEY / OPENAI_API_KEY as env vars

Full CLI reference

# --- Session init ---
fish-bridge init                          # create session for current project
fish-bridge init --tool claude            # → writes to CLAUDE.md instead

# --- Ingest chat turns ---
fish-bridge ingest --source copilot       # auto-discover latest VS Code Copilot session
fish-bridge ingest --source copilot --session <id>  # target specific session
fish-bridge ingest --source paste         # paste any chat text — opens $EDITOR (universal fallback)
fish-bridge ingest --source file --file export.json  # from a saved export file
fish-bridge watch --source copilot        # tail JSONL, auto-update on new turns

# --- Merge external knowledge ---
fish-bridge merge --source document --file HANDOVER.md
fish-bridge merge --source codebase --path ./            # git log + README
fish-bridge merge --source obsidian --vault ~/notes
fish-bridge merge --source deps --path ./                # package.json / pyproject.toml etc.
fish-bridge merge --source testout --file results.json   # jest / pytest / JUnit
fish-bridge merge --source iac --path ./                 # Terraform / CDK / CloudFormation
fish-bridge merge --source openapi --file openapi.yaml
fish-bridge merge --source session --file prior.chatgraph.json

# --- Compile & view ---
fish-bridge compile                       # update instruction file (runs after ingest by default)
fish-bridge compile --mode digest         # full handover markdown
fish-bridge compile --mode focus --query "Redis caching"
fish-bridge show                          # pretty-print active nodes
fish-bridge show --all                    # include resolved/deferred items
fish-bridge serve                         # open Cytoscape.js graph viewer at localhost:8080
fish-bridge digest                        # generate handover digest

# --- Node management ---
fish-bridge resolve "DNC caching strategy"
fish-bridge defer "v16 index validation"
fish-bridge add "Use Redis for session cache" --type decision
fish-bridge conflict show
fish-bridge conflict resolve <node-id> --keep old

# --- Export / import / diff ---
fish-bridge export                        # save .chatgraph.json
fish-bridge import prior-session.chatgraph.json
fish-bridge diff session-a.chatgraph.json session-b.chatgraph.json

# --- Config ---
fish-bridge config --show
fish-bridge config --backend gemini

MCP server (optional — agent mode only)

The MCP server adds real-time record_turn capture when using VS Code agent mode. It is not required — the file-based workflow above works in all modes without it.

Add to .vscode/mcp.json (uses uvx — no prior install needed):

{
  "servers": {
    "fish-bridge": {
      "command": "uvx",
      "args": ["fish-bridge-mcp"],
      "env": { "FISH_BRIDGE_BACKEND": "gemini", "GEMINI_API_KEY": "${env:GEMINI_API_KEY}" }
    }
  }
}

If you used uv tool install fish-bridge-mcp, you can also reference the installed binary directly:

{ "command": "fish-bridge-mcp" }

See examples/ for Claude Desktop, Cursor, and Windsurf configs.

MCP tools: record_turn, get_context, get_focus, mark_resolved, add_node, export_session, import_session, show_active, list_deferred

Ingest sources

How it works

1. Ingest — reads raw chat turns from JSONL (Copilot), paste, or any file format
2. Extract — LLM extracts typed nodes (questions, decisions, errors, tasks, skills, files) and edges
3. Dedup — semantic similarity merges near-duplicates; conflict detection flags status reversals
4. Compile — graph is compressed to ~300–800 token XML/markdown block
5. Write — block is written to .github/copilot-instructions.md (or CLAUDE.md)
6. Deliver — AI tool reads the file automatically on the next turn — no injection, no agent required

Documentation

• Quickstart
• Configuration
• Graph schema
• Compiler modes
• Ingestors

License

MIT — see LICENSE