Repository navigation
feat(ai): configurable model catalog for pickers and server/local agents - #196
Merged
Merged
Conversation
camielvs
commented
Nov 18, 2026
camielvs
left a comment
Contributor
Author
There was a problem hiding this comment.
🤖 This is an AI-generated code review comment.
9 findings, inline below (1 High, 3 Medium, 5 Low).
Notes that are not line-specific:
- The catalog block in
apps/server/src/pi/extensions/proxyProvider.tsandpackages/local-agent-host/resources/proxyProvider.tsis about 35 identical lines. Pi loads these standalone so a shared import is awkward, and the new parity test is a reasonable guard; just flagging that any fix to one now needs the other. - I could not verify from this checkout that Pi's
registerProvidermodel entries honour athinkingLevelMapfield (nonode_moduleshere). If Pi ignores unknown model fields, the DeepSeekxhigh → maxmapping never takes effect on the agent side. Worth a quick check against the Pi types before relying on it.
What's good: catalog-fetch credentials are stripped from every spawned process and the remote command carries data only, with a test proving it; redirect: "error" plus the credentials-in-URL rejection; failing startup on a misconfigured catalog instead of silently falling back; the picker keeps AVAILABLE_MODELS behaviour when no catalog is configured.
A catalog-enabled spawn onto a local host with no proxy URL used to throw from inside the PiAgent constructor, which the bridge did not catch: the server never heard about it and a zombie agent stayed in the roster. Fail it the same way a missing workspace folder fails, before spawning. Also: accept [::1] as loopback for the catalog URL, document the config router as partly installation-private now that it serves /models, stop reassigning the picker's disabled prop, import the catalog type at the top of the remote contract, and import the local-host test fixtures through the workspace package instead of a relative path. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…ides A shared resolver maps pre-catalog ids (`openai/gpt-6.1-sol`), bare ids and stale ids listed in `replacements` to the catalog model. The server uses it when spawning and falls back to the catalog default with a warning when a selection no longer exists, rather than handing Pi a provider the proxy extension no longer registers. The picker uses the same resolver so old selections show their label, highlight and thinking levels, and the remote roster records the resolved values the spawn command carries. Thinking depth is clamped to the nearest level the resolved model supports, in the picker (sent together with a model change when needed) and again at spawn for configs that predate the catalog. Pi accepts a `max` thinking level, so THINKING_LEVELS now includes it; the catalog's `max` mappings were previously dead data. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
camielvs
force-pushed
the
cvs/configurable-model-catalog
branch
from
November 18, 2026 20:42
29a8dea to
83b478b
Compare
Mbeaulne
approved these changes
Nov 19, 2026
…okie Oasis authenticates requests through Minerva headers or the MINERVA_TOKEN cookie and never reads an Authorization bearer, so the bearer-only loader would 401 and fail startup. When the catalog URL is on the TANGLE_API_URL origin the fetch now carries TANGLE_TOKEN as its Cookie header, the same credential the server's other Tangle egress uses. An explicit TANGENT_MODEL_CATALOG_TOKEN remains a bearer for catalogs served elsewhere. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
openai/..., stale ids viareplacements) on both server and picker, falling back to the catalog default with a warning; clamp thinking depth to what the resolved model supports.maxtoTHINKING_LEVELS: Pi 1.4 accepts it, so catalogmaxmappings now take effect.Validation
Test setup / rollout
See docs/model-catalog.md. Configure TANGENT_MODEL_CATALOG_FILE or TANGENT_MODEL_CATALOG_URL (optionally TANGENT_MODEL_CATALOG_TOKEN), preserve the existing inference proxy configuration, and restart the server. Local agents require the updated host package and a configured proxy connection. No catalog means existing standalone settings.
Catalog pricing is unspecified and runtime cost metadata remains zero-filled/unknown; live inference and integrated deployment tests are not yet performed.