Update 22 — Bring your own key

Metadata

Update number: 22
Version: 2.0.0 → 2.0.1
Theme: the assistant for people who do not have a subscription CLI, and three things that were quietly not being watched

The short version

Update 21 shipped the ✦ Ask panel and said, at the bottom, “No API-key assistant yet. If you have an Anthropic or OpenAI key but no CLI, the panel cannot use it. That is the next release.”

This is that release. Under Settings → AI → Assistant, Run it with now offers an API key of your own alongside the claude command. Anthropic, OpenAI, Google, xAI, Together, DeepSeek, Alibaba, or a custom OpenAI-compatible endpoint.

It is the same panel, the same tools, the same confirmation card, and the same rules on the same system channel. What changed is who runs the loop: resmon does it itself instead of handing it to a command.

Three things are different, and resmon says all three

It reports tokens, not money. A provider API tells you how many tokens a call used. It does not tell you what they cost, and resmon does not maintain anyone’s price list. So a turn on this route shows a token count and says cost not reported, rather than a dollar figure computed from a table this app would have to keep correct for eight providers. The claude route still shows a real cost, because the command reports one.

A turn stops after eight tool steps or 100,000 tokens. The command enforces a spending ceiling on itself; here the ceiling has to be on something resmon can count. Both numbers are twice the largest of the ten requests measured for 2.0.0 — four tool steps and about 53,000 tokens at the top end — so an ordinary answer is nowhere near them and a loop is stopped.

It remembers what was said, not what the tools returned. A provider API keeps no session, so resmon replays the conversation itself: your messages and its replies. The raw output of an earlier tool call is not re-sent. That means a long conversation does not pay to re-send a page of papers on every turn, and text lifted from a paper’s abstract does not follow you from turn to turn. The assistant can see what it told you; it cannot see the evidence it told you from.

Which providers can do this at all

The same three-state answer the embedding lane uses, with the evidence that established each one:

OpenAI and Anthropic — established from the SDKs resmon ships, which is better evidence than a probe and is checked by a test rather than quoted in a comment.
xAI and Google — established by a probe. Sending a deliberately malformed tools field gets a complaint naming tools; sending a made-up field gets “cannot find field”. The difference is the answer.
Together, DeepSeek, Alibaba — unknown, and it says so. Those APIs check your key before they look at the body, so an invalid key gets the same 401 whatever you send and there is nothing to observe from outside. resmon will send the tools and show you exactly what comes back, rather than recording a guess with a citation attached.
A custom endpoint is unknowable in advance, by definition.

A first screen that asks for nothing

A brand-new install now opens on a Getting started card: three optional things — a command you already subscribe to, a provider key, a key for one of the five sources that take one — each with a link to where it is set and a mark for what resmon can already see.

None of it is required. resmon searches all 25 sources with no AI and no keys at all, and the card says so in its first line. It reports found and configured and never working, because resmon cannot tell whether a command is signed in or a key accepted until the first paper goes through it. Skip puts it away permanently, and it retires itself the moment you have run anything.

When the CLI has lost your conversation

claude keeps its conversations in its own storage. If you clear its history, or restore resmon’s database onto a machine whose command has never seen that conversation, resmon used to answer “exit code 1” and lose your message.

Now it recognises what the command actually said, answers your message in a fresh session, and tells you in place: your earlier messages are still on screen, and the assistant can no longer see them.

The part nobody was watching

resmon has a suite of tests that talk to the real internet — every source client, the whole tool surface, the search extension loading inside a running backend. They are excluded from the ordinary test run because they are slow and because a provider having an outage is not a bug in resmon. Nothing ran them on a schedule.

In 1.9 one of them started asserting the opposite of what the search contract says, and it sat red for two releases with nobody hearing about it. That was found in 2.0.0 by someone happening to type the command.

There is now a weekly job. It runs every live test that needs only a network — 56 of the 72 — and its summary names, by test, every one it could not run and why. The sixteen it skips need an agent command installed and signed in, which no build machine has.

It found something on its first run. 2.0.0 added an assistant settings group to the app after the tool surface’s allowlist was frozen earlier in the same release, so the assistant could not change its own model — and nobody had decided that. It can now, behind the same confirmation card as everything else. Contract v2.1.

Also in this release

The assistant’s settings section hides the effort control on the API-key route, because none of the provider APIs takes one. A control that silently did nothing would be worse than no control — the same rule the summary lanes follow.
Both routes now go through one permission function. The command reaches it as a tool in another process and the API-key loop reaches it as a function call, so “the same card” is a fact about the code rather than two things that resemble each other.

What is still not here

No voice. The packaging spike came back with numbers: bundling local transcription adds 80–90 MB to every installer, the redistribution questions for its dependencies are unanswered, and there is no signing story for microphone access on macOS. It returns as its own piece of work when the model can be downloaded on first use instead of shipped, not before.

No Ollama assistant. Ollama does support tool calling — resmon checked, and got a real tool call back from a local model — but driving it as an assistant is its own piece of work and is recorded rather than rushed.