Package Exports

This package does not declare an exports field, so the exports above have been automatically detected and optimized by JSPM instead. If any package subpath is missing, it is recommended to post an issue to the original package (synara) to support the "exports" field. If that is not possible, create a JSPM override to customize the exports field for this package.

Readme

Synapse CLI

AI-powered coding assistant. Cloud brain, local hands.

Architecture

Local CLI (thin client)              Cloud Container (thick server)
┌─────────────────────┐              ┌──────────────────────────────┐
│ Terminal UI          │              │ Harness Engine               │
│ Tool Executor        │◄───WS────►│   System Prompt Assembly     │
│ Permission Control   │              │   Guardrails / Skills        │
│ SYNAPSE.md Reader    │              │   Context Compact (LLM)      │
└─────────────────────┘              │   Streaming LLM (Kimi K2.5)  │
                                      │ AI Gateway → Workers AI      │
                                      │ Trace Recording              │
                                      └──────────────────────────────┘

CLI only does three things: display, execute, confirm. All intelligence lives on the cloud.

Install

npm install -g synapse-cli

Usage

synapse login        # Login with username/password
synapse              # Start coding session in current directory
synapse logout       # Remove stored credentials
synapse --help       # Show help

Session Commands

Command	Description
`/compact`	Compress conversation history (LLM summarization)
`/context`	Show context usage (tokens, messages, compactions)
`/tools`	List available tools and permission levels
`/quit`	End session

Permission Model

Tool	Permission
read_file, grep, glob, list_dir	Auto (in-project only)
read outside project dir	Ask each time
write_file, edit_file	Ask (trustable via `t`)
bash	Ask every time (never trustable)
Destructive commands (rm -rf /, mkfs, etc.)	Blocked entirely

When prompted:

y — approve this operation
n — deny (AI will ask you what to do instead)
t — trust this tool type for the rest of the session (not available for bash)

Streaming

LLM responses stream token-by-token via WebSocket
Content appears in real-time as the model generates it
Reasoning (thinking) is hidden by default

Context Management

Iron rule: NEVER truncate data.

Context compression uses LLM summarization, not mechanical truncation:

autoCompact — triggers at 80% of 256k context window, summarizes older messages in batches
reactiveCompact — triggers on API 413 error, compresses and retries (max 3 attempts)
/compact — manual trigger

Each batch of messages is sent to LLM in full for summarization. The LLM decides what's important. No data is cut by character count.

Project Configuration

Place SYNAPSE.md in your project root. It's automatically read and sent to the cloud as project-specific instructions (similar to Claude Code's CLAUDE.md).

Model

Kimi K2.5 via Cloudflare AI Gateway
256k context window
temperature 1.0, top_p 0.95
Streaming with reasoning support

Tools

Tool	Description
read_file	Read a file from the project
write_file	Create or overwrite a file (shows content first)
edit_file	Precise string replacement in existing files
grep	Regex search across files
glob	Find files by name pattern
list_dir	List directory contents
bash	Execute shell commands

File Structure

synapse-cli/                    # Local client (npm package)
├── src/
│   ├── index.ts                # Entry: subcommands, WS connection, REPL
│   ├── auth.ts                 # Login/logout → JWT stored in ~/.synapse/token
│   ├── protocol.ts             # CLI ↔ Cloud message protocol
│   ├── permissions.ts          # t/y/n permission control
│   ├── tools/executor.ts       # Local tool execution
│   └── ui/terminal.ts          # Terminal output formatting
├── package.json
└── tsconfig.json

harness-agent/                  # Cloud server (Cloudflare Worker + Container)
├── src/worker/index.ts         # /cli/ws route → JWT auth → Container proxy
└── src/services/cli_session.py # CLI session: LLM loop, streaming, compact

Design References

Claude Code (March 2026 leak) — tool permission model, temperature 1.0, four-tier context compression
Kiro CLI — /compact command, streaming UX, session management
Synapse Harness — guardrails + skills injection, Dreaming pipeline