pi-go

Five coding agents, one protocol

Sat, 29 Aug 2026 00:00:00 +0200

pi-go runs other coding agents as subagents. It already spoke to Claude, Gemini, Cursor, and Copilot over the Agent Client Protocol (ACP) — a stdio protocol that lets one agent drive another. Adding a fifth, Google Antigravity, sounded like a one-afternoon job. It wasn’t, and the reason is a good lesson about what “supports ACP” actually means.

The twist: the vendor CLI is not the ACP endpoint

For Claude, Gemini, Cursor, and Copilot, the ACP adapter wraps the vendor’s own interactive CLI. You launch claude, gemini, or cursor and it speaks ACP over stdio.

Memory Palace: four layers of context

Sat, 29 Aug 2026 00:00:00 +0200

The Memory Palace is pi-go’s persistent contextual memory — SQLite facts, semantic embeddings, a temporal knowledge graph, and background project miners that turn your codebase into structured recall the agent can query on demand.

What the Memory Palace is

Most coding agents start every session with the same blank slate: whatever fits in the system prompt plus whatever they re-read from disk. pi-go keeps a second brain on disk. Every tool call the agent makes is captured, compressed, and stored; every file in your project can be mined into searchable embeddings; relationships between entities are tracked as temporal triples that can be invalidated when the world changes. The result is the Memory Palace — a four-layer contextual memory that lives under internal/palace/ and internal/memory/ and is queried through native ADK tools and the pi memory command.

Running Mistral in a multi-provider CLI

Sat, 29 Aug 2026 00:00:00 +0200

Adding a new LLM provider to a tool that already speaks to Claude, OpenAI, Gemini, and Grok sounds mechanical: implement the request, map the response, ship it. The Mistral integration in pi-go turned out to be a series of small, sharp lessons — each one a bug that only shows up against the live API, never in a unit test.

Lesson 1: `prompt_mode` is not `reasoning_effort`

pi-go’s default thinking level is “high”. When the Mistral provider was first written, it translated that into prompt_mode: "reasoning" on the wire for the magistral family — because the reference implementation the plan was drawn from uses Mistral’s conversations API, which takes prompt_mode.

Subagents: fanning out work without losing the thread

Sat, 29 Aug 2026 00:00:00 +0200

When a task is too big for one loop, pi-go spawns typed child agents — some in isolated git worktrees, some over the Agent Client Protocol — and orchestrates them with a bounded pool, retries, and clean cancellation so you never lose track of what’s running.

What a subagent actually is

A subagent is a separate process pi-go launches to do a focused job and report back. The main agent’s tool loop can call a subagent tool with one of three modes — single, parallel, or chain — and the orchestrator handles the rest: it resolves the agent’s role and model, acquires a pool slot, optionally creates a worktree, launches the process, and streams events back as JSONL.

The DeepSeek and GLM-5.3 effect

Sat, 29 Aug 2026 00:00:00 +0200

August 2026 — a data-driven look at my pi-go agent session history

TL;DR

Switching my coding agent’s default model to DeepSeek (via local Ollama) and adding GLM-5.3 to the rotation produced the single biggest jump in usage I’ve ever recorded. Monthly session volume went from a mid-year slump of 326 to an all-time high of 1,868 — a 5.7× increase — and the new model stack now accounts for ~54% of all sessions.

Your coding agent should draw diagrams

Sat, 29 Aug 2026 00:00:00 +0200

Coding agents talk in text. That’s fine for a diff, but it’s a terrible medium for architecture. When I ask an agent “how does auth flow through this service?”, the honest answer is a diagram — boxes, arrows, a fork where a request can go two ways. A wall of prose is the wrong shape for that.

So I taught pi-go to draw Mermaid diagrams in the terminal, and the interesting part turned out to be everything that isn’t the diagram.

OpenRouter arrives in Pi-Go

Mon, 24 Aug 2026 00:00:00 +0200

One provider, many model families

OpenRouter exposes models from multiple labs behind one OpenAI-compatible API. In Pi-Go, select one by adding the openrouter/ provider prefix to its OpenRouter model ID. The prefix is detected automatically, so no separate provider flag is needed.

# Configure your OpenRouter key
$ export OPENROUTER_API_KEY="sk-or-v1-..."

# Use a specific model
$ pi --model openrouter/google/gemini-3.7-flash

# Or let OpenRouter choose
$ pi --model openrouter/auto

Pi-Go uses OpenRouter’s standard endpoint by default. Self-hosted gateways and compatible proxies can override it with OPENROUTER_BASE_URL.

Introducing Pi-Go

Sat, 18 Apr 2026 00:00:00 +0200

Why another coding agent?

The browser-tab and IDE-plugin coding agents are powerful, but they pull you out of the terminal — where real work happens. pi-go is opinionated about staying there. It’s a TUI that speaks JSON-RPC to language servers, talks to whatever LLM you prefer, and executes tools inside a strict sandbox so you don’t have to trust it beyond what it needs to see.

What’s in the box

Multi-provider LLMs. Anthropic, OpenAI, Gemini, and local Ollama models. Switch providers mid-conversation with a slash command.
Sandboxed tools. File read/write/edit, shell execution, grep, find, tree, and git — all confined to the project directory via os.Root. No accidental rm -rf ~.
Memory Palace. A four-layer contextual memory combining SQLite, semantic embeddings, a temporal knowledge graph, and background project miners that learn your codebase while you work.
LSP integration. A JSON-RPC client for Go, TypeScript, Python, and Rust with auto-format and diagnostics hooks wired into the agent loop.
Subagent system. Process-based orchestration with typed roles — explore, plan, designer, reviewer, task runner — so big tasks fan out cleanly.
Session branching. JSONL append-only event logs with branching, compaction, and resume. Fork a conversation, try a different approach, merge back.

Get started in one command

# Linux and macOS
$ curl -fsSL https://raw.githubusercontent.com/dimetron/pi-go/main/scripts/install.sh | bash

# Windows PowerShell
PS> powershell -NoProfile -Command "iwr https://raw.githubusercontent.com/dimetron/pi-go/main/scripts/install.ps1 -UseBasicParsing | iex"

# Or with go install
$ go install github.com/dimetron/pi-go/cmd/pi@latest

# Run it
$ pi

What’s next

This is the first post and the project is very much in motion. Upcoming: deeper dives on the Memory Palace, the sandbox threat model, and how the subagent system dispatches work. Subscribe to the RSS feed or watch the repo on GitHub to follow along.

pi-go

Five coding agents, one protocol

The twist: the vendor CLI is not the ACP endpoint

Memory Palace: four layers of context

What the Memory Palace is

Running Mistral in a multi-provider CLI

Lesson 1: prompt_mode is not reasoning_effort

Subagents: fanning out work without losing the thread

What a subagent actually is

The DeepSeek and GLM-5.3 effect

TL;DR

Your coding agent should draw diagrams

OpenRouter arrives in Pi-Go

One provider, many model families

Introducing Pi-Go

Why another coding agent?

What’s in the box

Get started in one command

What’s next

Lesson 1: `prompt_mode` is not `reasoning_effort`