resmon Update 10 — August 30, 2026
Update 10 — AI On Your Terms
Metadata
- Update number: 10
- Version: 1.7.0 → 1.8.0
- Theme: one provider failing should not cost you the run — and when something does fail, the app should be able to say what
The short version
Until this release, AI summarization in resmon had exactly one way to succeed and one way to fail. You picked a provider; if it worked, you got summaries; if it didn’t, you got a line in the activity log and a run that completed looking perfectly healthy.
Two things change in 1.8.0.
You can now configure a chain of providers. Anthropic first, a cheaper model second, Ollama on your own machine as the floor — any providers, in any order. When one cannot produce a summary, the next one tries.
Failures became facts the app can reason about. Every attempt is now recorded: which provider, which model, how many papers it summarized, and the classified reason it stopped. That closes the largest blind spot the 1.7 watchdog had — a run where every single AI call failed used to finish and look fine.
This release also adds four new sources, bringing the catalog from 15 to 19.
Not every failure means the same thing
This is the idea the whole feature rests on, and getting it wrong is expensive in both directions.
A rejected API key is not going to start working on the next paper. Neither is an exhausted quota, a model that does not exist, or a provider your machine cannot reach. Presenting a dead credential once per paper burns an entire run rediscovering the same fact two hundred times. resmon calls these lane-fatal and retires that provider for the rest of the run.
An abstract longer than the model’s context window is a different animal entirely. So is content a provider declines, or a one-off server error. The provider is fine; this paper is the problem. Abandoning a working provider over one difficult abstract would silently downgrade every summary that came after it. resmon calls these document-local, falls through for that one paper, and keeps the provider primary.
| Failure | Examples | What happens |
|---|---|---|
| Lane-fatal | rejected key, exhausted quota, unknown model, unreachable provider | that provider is retired for the run |
| Document-local | context window exceeded, content declined, one-off 5xx |
that paper falls through; the provider stays |
If this sounds like the distinction the watchdog draws between broken — a recorded
fact — and unusual — an inference from your own history, it is the same discipline
applied to a different problem. Two things that look alike in a log are worth separating
when they demand opposite responses.
What gets recorded
Each execution now stores one row per provider it tried: the provider and model, the alias of the credential slot used (never the key itself), how many papers it attempted, how many it summarized, and the classified reason it stopped.
A provider that was never reached is recorded as skipped rather than left out — because “we never needed it” and “it was not configured” are different facts and you are entitled to both. A provider that got some papers and not others is partial, held deliberately apart from failed: partial is a normal day with one awkward abstract, failed is a provider that did not work.
And the report header now names the provider that actually produced the summaries. A report saying “anthropic” when Ollama did the work would be precisely the kind of quiet lie the last release was about.
Two things this release refuses to do
There is no summarizer of last resort. If every provider in your chain fails, those papers have no AI summary, and the execution says why. resmon has never had a keyless extractive fallback underneath the AI — our own planning documents claimed it did, and that was wrong. It has been corrected rather than quietly built to match.
The subscription lane is not here yet. Driving the Claude Code or Codex CLI you have already installed and logged into — so AI usage draws on a plan you already pay for — is the next release, not this one. The lane exists in the data model and is honestly reported as unimplemented rather than silently skipped.
Four new sources
The catalog goes from 15 to 19:
- Zenodo — CERN’s general repository: preprints, datasets and software, all with DOIs
- INSPIRE-HEP — the curated high-energy physics literature database
- OpenAIRE — the EU-wide open-science aggregator
- medRxiv — which resmon has been able to search since early on, but never made selectable. The client supported it; the registry never registered it. Fixed.
Each carries an honest description of how its upstream actually combines your keywords. OpenAIRE’s now reads “Combination semantics undocumented”, because OpenAIRE documents that behavior as undefined — and a catalog field that renders to you as a statement about someone else’s search engine should not guess.
Under the hood
- New
execution_aitable (schema 9), the sibling of the per-source record 1.7 added. - Provider configuration resolves into an ordered list of lanes. An existing single-provider setup is read as a one-lane chain at load time — no data moves, nothing migrates, and downgrading to an earlier build still works.
- The MCP tool surface is now frozen as a written contract in
docs/api-contract/mcp.md, ahead of the server itself. Writing it surfaced a genuine gap: resmon has no way to run a routine on demand at all. That is being fixed. - 485 → 570 backend tests, 111 → 121 renderer tests, 15 live-network tests against real scholarly APIs.
Upgrading
Nothing to do. Your existing AI configuration keeps working exactly as it did — the fallback section starts empty, and an empty chain is the old behavior. Add a fallback only if you want one.