What is true right now. Rewritten whole when it changes, never appended to — a status file that grows is a changelog, and git already is one.
Read this before opening a PR or filing an issue: most of what is already known-broken is below, and half of it already has a fix in flight.
Last updated: 2026-09-23
The four pillars run: the loop, memory (semantic + episodic + procedural with
a retrieval gate), tools, and both eval tiers. prometheus, prometheus dashboard,
prometheus voice, prometheus telegram, prometheus discord, prometheus brief and
prometheus connect google all start.
950 deterministic evals pass offline, with no API key; 60 more are live
evals that skip without one. CI runs the offline tier on every PR along with
ruff, the skills validator, and a check that .env.example still matches the
integrations registry.
0.1.8 is on PyPI and on GitHub Releases. Pushing a v* tag publishes to
both, so the repo's "Latest" release always matches pip install prometheus.
The dashboard uses the Prometheus Memory design system, and
test_design_system.py keeps it from drifting. See
context/design-system.md.
Nothing here is a surprise. If you hit one of these, the issue exists.
| What | Where | Fix in flight |
|---|---|---|
| The model picker offers OpenAI models that 404 on use | #137 | #178 |
| GPT-5.6 tool calls fail on Chat Completions | — | #146 |
| OpenCode Zen fails with a rate-limit error | #112 | #113 |
Google Calendar sign-in has no bundled OAuth client, soprometheus connect google needs your own ~/.prometheus/credentials.json |
— | — |
Providers are the recurring theme. Three of the items above are one provider or another, and there is no single place that says which providers are known-good today. Until there is, treat the model picker as a list of things that might work.
Not a framework, not multi-agent, not production — true of prometheus/, see
architecture.md. hosted/ runs that same code as a
deployment instead; spec 001 designs it, and none of it exists in this repo
yet.
Additionally, and worth stating because people ask:
- No Windows CI. The Windows bugs so far (#140, #141, both fixed) were found by contributors, not by us. Every Windows claim in this repo is untested.
- The judge evals are not in CI.
make gateruns deterministic evals at 100% plus a judge threshold, and CI runs only the first half. The judge tier needs an API key, which CI does not have. - No provider smoke check. Nothing verifies that a model in the picker resolves, which is why #137 reached a user.
- Where the memory pillar ends and Prometheus Memory begins. This repo's memory
is local, single-machine, and yours. Prometheus Memory is the same memory across
several agents, and it is a paid hosted service. Both are true and the
README does not yet say either plainly, so a reader has to work out the
difference alone. Hosted prometheus (
hosted/, spec 001) adds a third case: a tenant's memory lives on the operator's VM, not the person's own machine, but it is still that one tenant's own directory, not Prometheus Memory. What hosted prometheus keeps there is not written down yet; a later task in spec 001 records it oncehosted/exists.