No repo is an island. Every codebase depends on memory, intent, and shared understanding
| Package | Version | Description |
|---|---|---|
| @earendil-works/pi-coding-agent | 0.75.5 | Coding agent CLI with read, bash, edit, and write tools and session management. This is the core AI agent that powers the entire Minimum Intelligence system - it processes prompts, interacts with LLM providers, and manages conversation sessions. |
Beyond the CLI binary, GMI uses the following pi-mono feature categories:
| Feature | Location | Description |
|---|---|---|
| Session management | --session-dir, --session |
Multi-turn conversation continuity across workflow runs |
| Project settings | .pi/settings.json |
Provider, model, thinking level, compaction, and retry configuration |
| System prompt extension | .pi/APPEND_SYSTEM.md |
Behavioural guidelines appended to the default system prompt |
| Bootstrap protocol | .pi/BOOTSTRAP.md |
First-run agent identity setup |
| Context files | AGENTS.md |
Project-specific instructions loaded at startup |
| Skills | .pi/skills/ |
On-demand capability packages (memory, skill-creator) |
| Prompt templates | .pi/prompts/ |
Reusable prompts for recurring workflows (code-review, issue-triage) |
| Extensions | .pi/extensions/ |
Custom tools registered for LLM use (github_repo_context) |
These are not package dependencies but are required for the system to function:
| Dependency | Description |
|---|---|
| GitHub Actions | The sole compute runtime. Every issue event triggers a workflow that runs the AI agent. No external servers or containers are needed. |
| GitHub Issues | Used as the conversation interface. Each issue maps to a persistent AI conversation thread. |
| Git | All session state, conversation history, and agent edits are committed to the repository. Git serves as the memory and storage layer. |
| Bun | JavaScript/TypeScript runtime used to execute the agent orchestrator and install dependencies. |
| gh CLI | GitHub's official CLI tool, used by the agent lifecycle scripts to interact with the GitHub API (fetching issues, posting comments, managing reactions). |
These are referenced in .github/workflows/:
| Action | Workflow | Description |
|---|---|---|
| actions/checkout@v6 | agent | Checks out the repository so the agent can read and write files. |
| oven-sh/setup-bun@v2 | agent | Installs the Bun runtime in the GitHub Actions environment. |
| actions/cache@v5 | agent | Caches node_modules keyed on the bun.lock hash to speed up dependency installation. |
| actions/configure-pages@v5 | agent | Configures GitHub Pages deployment (auto-enables Pages when the repo is used as a template). |
| actions/upload-pages-artifact@v4 | agent | Uploads the static site artifact from .github-minimum-intelligence/public-fabric/. |
| actions/deploy-pages@v4 | agent | Deploys the uploaded artifact to GitHub Pages. |
An API key from at least one supported LLM provider is needed:
| Provider | API Key Secret | Description |
|---|---|---|
| OpenAI | OPENAI_API_KEY |
GPT models including GPT-5.4 (default provider). |
| Anthropic | ANTHROPIC_API_KEY |
Claude models. |
| Google Gemini | GEMINI_API_KEY |
Gemini 2.5 Pro and Flash models. |
| xAI | XAI_API_KEY |
Grok 3 and Grok 3 Mini models. |
| OpenRouter | OPENROUTER_API_KEY |
Access to DeepSeek, and hundreds of other models via a unified API. |
| Mistral | MISTRAL_API_KEY |
Mistral Large and other Mistral models. |
| Groq | GROQ_API_KEY |
Fast inference for open-source models like DeepSeek R1 distills. |
These are pulled in transitively by @earendil-works/pi-coding-agent:
| Package | Description |
|---|---|
@anthropic-ai/sdk |
Official Anthropic API client for Claude models. |
@aws-sdk/client-bedrock-runtime |
AWS Bedrock client for accessing models via AWS infrastructure. |
openai |
Official OpenAI API client. |
@google/genai |
Google's Generative AI SDK for Gemini models. |
fast-xml-parser |
Fast XML parser used by AWS SDK internals. |
tslib |
TypeScript runtime helpers used throughout the dependency tree. |
