Talicode

A taste-first AI coding instrument for the terminal. Not chat. Not autocomplete. A controlled engineering companion.

Talicode is a fast, inspectable CLI that helps you understand, plan, and modify codebases with surgical precision — using both local models and cloud APIs.

It focuses on:

• evidence-first planning
• minimal blast-radius changes
• replayable sessions
• repo memory
• rituals (structured workflows)
• transparent routing decisions

---

Why Talicode exists

Most AI coding CLIs:

• blindly call a model
• mutate files immediately
• provide no reasoning structure
• forget everything next run

Talicode instead:

inspect → plan → propose → review → execute

You always see:

• which files will change
• why they were chosen
• blast radius
• confidence
• review checklist

---

Features

Surgical patch planning

talicode patch auth

Outputs:

Surgical Patch Proposal
Query: auth
Blast radius: low

Files:
- internal/auth/session.go
- internal/auth/middleware.go

Reason: Prefer smallest touch set first.

---

Repo memory graph

Talicode remembers architecture decisions and relationships.

talicode memory build
talicode recall auth

Example:

Recall: auth

Files:
- internal/auth/session.go
- internal/auth/middleware.go

Past decisions:
- Preferred middleware boundary for auth checks

---

Replayable sessions

Every run is stored and reproducible.

talicode replay last
talicode compare A B

This allows:

• debugging agent decisions
• comparing routes
• reproducing fixes
• auditing changes

---

Ritual workflows

Reusable engineering workflows.

talicode ritual list
talicode ritual run bugfix

Built-in rituals:

• bugfix
• pr-review
• migration (coming)
• refactor (coming)

---

Local + cloud model routing

Talicode automatically chooses:

• local models → repo scan
• cloud models → reasoning
• hybrid routing when needed

Transparent routing:

talicode route "fix auth bug"

---

Trust ladder

Talicode enforces safety levels:

Level	Behavior
L0	read only
L1	safe writes
L2	dev writes
L3	shell allowed
L4	unrestricted

---

Commands

Core

talicode run "summarize repo"
talicode chat
talicode route <prompt>

Planning

talicode map
talicode impact <query>
talicode patch <query>

Memory

talicode memory build
talicode recall <query>
talicode note add "text"

Replay

talicode replay last
talicode compare A B

Rituals

talicode ritual list
talicode ritual run bugfix

Stats

talicode stats

---

Installation

Build from source

git clone https://github.com/moswek/talicode
cd talicode
go build ./cmd/talicode

Run:

./talicode

---

Configuration

Create talicode.json

{
  "default_mode": "craft",
  "default_model": "talicode-auto",
  "trust_level": "safe-write",
  "providers": [
    {
      "name": "openai",
      "type": "cloud",
      "base_url": "https://api.openai.com/v1",
      "api_key_env": "OPENAI_API_KEY",
      "models": ["gpt-5-mini"]
    },
    {
      "name": "ollama",
      "type": "local",
      "base_url": "http://localhost:11434",
      "models": ["qwen2.5-coder:7b"]
    }
  ]
}

---

Philosophy

Talicode is designed around five principles:

1. Inspect before mutate

Never blindly edit code.

2. Small blast radius

Prefer minimal change sets.

3. Explain decisions

Routing and planning must be visible.

4. Memory matters

Agent should learn repo structure.

5. Taste over automation

Precision beats chaos.

---

Roadmap

Phase 1

• CLI engine
• patch planner
• routing

Phase 2

• memory graph
• recall
• notes

Phase 3

• replay + compare
• session timeline

Phase 4

• rituals system
• plugin architecture

Phase 5

• streaming models
• premium TUI

Phase 6

• benchmarking + auto routing

---

Comparison

Feature	Talicode	opencode	crush
Surgical patch preview	yes	no	no
Replay sessions	yes	no	no
Repo memory	yes	no	no
Trust ladder	yes	no	partial
Ritual workflows	yes	no	no
Explainable routing	yes	no	no
Local + cloud hybrid	yes	partial	partial

---

Example workflow

talicode map
talicode impact auth
talicode patch auth
talicode run "fix session expiration"
talicode replay last

---

Contributing

Talicode is early but stable.

Areas to help:

• provider adapters
• analyzers
• rituals
• TUI layer
• memory graph improvements

---

Vision

Talicode is not trying to replace editors.

It aims to become:

a persistent engineering companion for real codebases

Fast. Calm. Inspectable. Powerful.

---

Latest release

v0.1.0 — First public release

• Surgical patch planning
• Repo memory (v1)
• Replayable sessions
• Ritual workflows
• Hybrid local/cloud routing

---

License

MIT