Skip to content

opencode_cli.md

Latest commit

 

History

History
68 lines (54 loc) · 3.78 KB

opencode_cli.md

File metadata and controls

68 lines (54 loc) · 3.78 KB
Raw

OpenCode (coding agent — experimental alternative)

OpenCode is an open-source coding agent. The client is free; you connect it to a model provider of your choice. It reads AGENTS.md from the repository root, so the assistant's instructions, skills and wiki all apply, and it can install PyAutoLens, run fits and inspect results if the model you connect can drive the workflow.

This page is here because OpenCode is the route people ask about when they have no paid Claude Code or Codex access. Read the caveats before spending time on it.

What "free" does and does not mean

  • The client is free; the model is not OpenCode's. Cost, capability, rate limits and availability all belong to the provider you configure. OpenCode's own docs recommend new users start with its hosted provider (OpenCode Zen), which is pay-as-you-go with a handful of models offered at no cost — and those free models are described by OpenCode as available for a limited time while feedback is collected. Do not plan a project around a free model staying free.
  • Not every model can run the assistant. The workflow needs a model that (a) follows multi-step tool use reliably over a long session, (b) does not hallucinate PyAutoLens API from training data faster than the self-enforced audit can catch it, and (c) can look at figures — the real-data gate and every fit check involve inspecting a PNG. Many free models have no vision input, which rules them out for the full workflow even if they can write code.
  • Nothing here is validated yet. As of 2026-09-10 no provider/model configuration has been tested against this assistant's benchmarks, so there is no default free model to recommend. When one is validated it will be recorded here with its test date, the model and provider, and its limitations; until then treat OpenCode as compatible (it loads the instructions and runs code) rather than supported. The evaluation protocol maintainers use is in ../evaluation/agent_evaluation.md.

Setup

  1. Install OpenCode and configure a model provider — follow the official instructions at github.com/sst/opencode. Pick a model that supports tool use and image input; check the provider's model card, not the model's name.
  2. Clone this repository and start the agent inside it (running from the repository root is what lets it discover the assistant's instructions in AGENTS.md):
git clone https://github.com/PyAutoLabs/autolens_assistant.git
cd autolens_assistant
opencode
  1. OpenCode has no PreToolUse hook, so the code gate is self-enforced: the assistant should run python autoassistant/audit_skill_apis.py --file <script.py> on generated PyAutoLens code before executing it. If the model skips this, ask it to.

Your first prompt

Example Natural Language Prompt for Claude Code, Codex or other AI coding agent

Find the data on the Cosmos-Web ring, give me a short script to plot it in PyAutoLens
and then given that I'm a new user give me an overview of the different ways we can
perform strong lens modeling of this system.

Then a quick capability check before trusting it with real work: ask it to plot the dataset and describe what it sees in the image. A model that cannot view the figure will describe the code instead — that is the signal it cannot run the full workflow.

Reporting back

If you run the assistant on OpenCode, please open an issue with the provider, model, date, and what worked or failed (the three checks in the evaluation protocol are a good template). User reports are how a configuration graduates from "compatible" to "tested".