5aab279b98a76a44afc48f6173c2d22c5c4120d6
2
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
  pastilhasandClaude Opus 5
|
d7f3223b6a |
qwen-local: the first provider that runs on our own GPU
sd-server (stable-diffusion.cpp, Vulkan) serving Qwen-Image 2.1 on this machine. It already speaks OpenAI's image API, which is the vocabulary `ImageRequest` is written in, so there is no translation layer — only the parts that are true of a local GPU and not of a hosted one. ── keyless is a category, not an exception ── `credentials.ts` decided a provider was usable by asking whether a key existed, so a service that takes none read as unconfigured forever. The fix is an adapter declaring `keyless`, not a dummy secret: a dummy would record a credential that does not exist and make the secret store lie about what is configured, and that store is the one place meant to be the truth about them. A keyless adapter never reaches `keyFor` at all. That function's own comment says why it uses `getSecret` rather than `getKey` — so a missing FOREIGN credential is never minted — and a provider with no credential by design has no business in that path. whisper, OCR and TTS already run on this machine, so this will not be the last. ── measured, not assumed ── The integration doc said "tens of seconds, not two". The server's own log says 512x512 completes in ~185s and 1024x1024 samples at ~41.4s/step over 20 steps, about 14 minutes. And it serialises strictly: a second caller waits for the first to finish entirely, so a queued 512 can wait 17 minutes before a byte comes back. So: 512 is the default and the substitute for `auto`, 1024 is offered with the cost in its label, and `timeout: false` is a requirement rather than caution — every default anyone would reach for is far under 17 minutes. The caller's `signal` still aborts, because a user cancelling is not a client giving up. ── the model id moved, because only it could ── `GET /v1/models` returns `sd-cpp-local`, hardcoded in sd-server with no override. The catalogue now agrees with the server and carries "Qwen-Image 2.1 (local)" in the label, so requests name a model the server knows while the UI still says something meaningful. ── not shipped: edit() ── Same GPU, same queue. It doubles the ways generation can be blocked while the generate path is unproven in anger, and `ImageAdapter` makes absence the way to say so — the route answers 400 instead of the queue finding out late. Verified: /v1/models answers, provider statuses report qwen-local as configured/keyless/usable, and the shipped entry survives the merge with the stored catalogue.json. NOT verified end to end — a generation is occupying the GPU and I left it alone. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Officer-Session: de3ed4e6-f88a-429c-94e6-a760336ba83b |
||
|
pastilhas
|
0de38f0b13 |
feat(catalogue): a curated file of providers and models, refreshed daily
The five connections in the database were never configured by anyone — a migration minted one per model from the adapter's own list, so the table held five decisions nobody made. What that list actually is, is a CATALOGUE: what you can choose from, which is not the same kind of thing as what you chose. So it is a file. `sidecar/catalogue/shipped.ts` is the curated baseline, in git, reviewed; `$OFFICER_ROOT/data/atelier/catalogue.json` is what a refresh writes, and it wins where it has something to say. Data dir, not the plugin dir — a plugin directory is a clone an update replaces, which is the rule env.ts already states for images. Refresh runs in the sidecar, daily, and only on boot when the file is actually stale — this process has restarted twenty times in an afternoon and should not hit OpenAI twenty times for it. Two sources, unequal, and the code says so. `GET /v1/models` is authoritative and thin: ids and nothing else, so a model matching /^(gpt-image|dall-e)/ appears the day it ships without a release. Prices come off an HTML page because there is no pricing endpoint for any plan, and that parser is written to give up rather than guess — a guess gets written to disk and then believed. When it gives up it says so in the log; the durable replacement is handing the page to a model for structured output, which Officer can already do locally. Nothing a refresh cannot establish overwrites a curated value: fields merge one by one, not objects wholesale, so an id-only listing cannot wipe the sizes and prices that were written by hand. Retiring is decided by the merge, which can tell a complete listing from a request that failed halfway — and it marks `deprecated` rather than deleting, because the gallery labels 88 historical rows from this list. Per-model sizes, qualities and transparency now travel with each entry, which is what will let the form stop offering dall-e-2 a transparent background it cannot do. The catalogue is advisory throughout: validate.ts still refuses to enforce per-model rules, exactly as before. |