Documentation

Known limitations

This page lists what was not verified and what is not there. It is kept honest on purpose.

Not verified

Only Anthropic has been called for real. On 2026-09-30 the Anthropic adapter passed pnpm smoke:provider and the console's connection test with a real key (claude-sonnet-5-5 and claude-opus-5-5), and ran a task end to end. OpenAI and OpenAI-compatible local endpoints were exercised only through the official SDK against local stand-ins that speak the same wire format, which checks request shapes, response parsing and error mapping but not a real model's behaviour or rate limits. Run pnpm smoke:provider with your key, and the connection test in Settings, before relying on either. See Tests and verification.
Anthropic refusal fallback (the opt-in beta option) was implemented from documentation and has not been run against the API.
The console image has not been built. The server image (API and worker) and the deploy/ layout with Caddy were built and run on a real Ubuntu server on 2026-09-30, and the split layout in Console on Vercel is what runs at poietescmd.xyz. The web service of the Compose file, which builds the console image, has not been built or started anywhere yet.
Browsers. The console and the site were checked in headless Chrome at desktop and phone widths, not in Firefox or Safari.

Single owner. One account, no roles, no teams.
No shell or code execution tool. The agent cannot run commands on the host. Adding one needs a sandbox with its own limits.
No tool writes into the workspace. Generated files go to the artifact store.
Text files only. No PDF, Word or spreadsheet parsing. Those are refused with an explicit error.
Keyword memory retrieval. PostgreSQL full-text search with English stemming. No embeddings. Entries in other languages match on exact words only.
Literal file search. No regular expressions.
OpenAI through Chat Completions. The Responses API is not used.
One request at a time per turn. Tool calls in a turn run one after another, in order.
No automatic retry of a failed task. The SDKs retry transient errors twice inside a request. After that the task stops as failed and you choose whether to retry.

Context estimates are estimates. The check that a saved conversation fits a model uses characters divided by three. A provider may still reject a request as too long.
Token budgets need usage reports. Some local servers do not report token usage. The console then says "usage not reported" and the token budget is not enforced for that task.
Time-of-check gap on files. A path is validated, then opened. Someone with write access to the workspace folder on the host could swap a file for a link in between. The agent itself has no tool that creates links.
Per-address limits need a reverse proxy. See Security model.
robots.txt is not consulted by fetch_url.
Exactly-once is not claimed for external actions. See Tasks and recovery.
Rotating the encryption key means entering stored provider keys again.
No licence file is included. The site does not describe the project as open source, and shows a source link only when you configure a repository URL. Choose a licence before publishing the code.