semitexa / llm

2026.10.08.0620 2026-10-08 09:35 UTC

Requires

• php: ^8.4
• semitexa/cache: *
• semitexa/core: *
• semitexa/prompt: *

Requires (Dev)

• phpunit/phpunit: ^10.0

MIT

• Semitexa

• assistant
• ai
• skills
• llm
• semitexa

Published on Packagist.org on 2026-10-08 18:16 UTC

README

LLM assistant for Semitexa with console-based skill discovery and execution. Talks to a local Ollama, a remote Ollama, or Google Gemini.

Install

Included in every project created by the installer (https://semitexa.com/install.sh).

Purpose

Integrates LLM capabilities into the Semitexa CLI. Console commands marked with #[AsAiSkill] are discoverable by AI agents with structured metadata including risk assessment and execution policies.

Role in Semitexa

Depends on semitexa/core for attribute discovery and the console command system. Skills are registered via ClassDiscovery and exposed to LLM agents as a structured manifest with execution constraints.

Key Features

• #[AsAiSkill] attribute for skill discovery, #[AsAiPersona] for personas
• SkillRegistry builds the manifest from ClassDiscovery
• Execution policies: AiExecutionKind, AiRiskLevel, AiConfirmationMode
• Providers: LocalOllamaProvider, RemoteOllamaProvider, GeminiProvider, selected by LlmProviderResolver
• Console: ai (interactive assistant), ai:skills (list skills), prompt:eval (send a prompt to the configured model)

Configuration

LLM_BACKEND selects the provider: local (default), remote_ollama or gemini. An unset or unknown value falls back to local.

Backend	Keys (defaults)
local	LLM_BASE_URL (http://127.0.0.1:11434), LLM_MODEL (gemma3:4b), LLM_TIMEOUT (60), LLM_RETRIES (1), LLM_CONNECT_TIMEOUT (5)
remote_ollama	LLM_REMOTE_OLLAMA_URL (required), LLM_REMOTE_OLLAMA_MODEL (gemma4:e2b), LLM_REMOTE_OLLAMA_TIMEOUT (120), LLM_REMOTE_OLLAMA_RETRIES (2), LLM_REMOTE_OLLAMA_CONNECT_TIMEOUT (5)
gemini	GEMINI_API_KEY (a Gemini API key from Google AI Studio), GEMINI_MODEL (gemini-2.5-flash), GEMINI_BASE_URL, GEMINI_TIMEOUT (120), GEMINI_RETRIES (2), GEMINI_CONNECT_TIMEOUT (5), GEMINI_CONTEXT_CACHE (off)

LLM_DECIDER_MODEL (optional) routes silent classification and tool-picking to a cheaper model on the same provider. Only providers that can switch model honour it (Gemini today); on the Ollama backends it is ignored.

LLM_PROVIDER=ollama is not read by this package: it tells bin/semitexa server:start to add the docker-compose.ollama.yml overlay, which runs an Ollama container in the project stack and points LLM_BASE_URL at it (http://ollama:11434).

Docs: https://semitexa.com/docs/llm/providers

Notes

Skills are only useful once a provider is reachable. Execution policies ensure destructive operations require explicit confirmation.