Skip to content

Releases: caffeinelabs/openai-client

openai-client 0.1.3 — serde-core 0.1.1 type-table cycle fix

27 Apr 19:02
@ggreif ggreif
v0.1.3
f912cd7

Choose a tag to compare

View all tags
openai-client 0.1.3 — serde-core 0.1.1 type-table cycle fix Latest
Latest

Fixed

  • Bumps serde-core to 0.1.1 to pull in the Candid type-table cycle fix.
    JSON.toText(to_candid(req), …) no longer traps with IC0502: Canister trapped: stack overflow when req carries a ?Map<K, V> field. Affects
    every ChatApi.createChatCompletion call in practice — the request type
    has both metadata : ?Map<Text, Text> and logit_bias : ?Map<Text, Int>,
    whose self-referential RBT type entries used to send the decoder into
    unbounded re-expansion before any byte of the value was inspected.

ChatApi.createChatCompletion is now functional end-to-end through the
connector (Wed 2026-04-29 Caffeine demo).

Assets 2
Loading

v0.1.2 — ChatApi works end-to-end (string-flatten fix)

27 Apr 16:55
@ggreif ggreif
v0.1.2
e362b39

Choose a tag to compare

View all tags
v0.1.2 — ChatApi works end-to-end (string-flatten fix)

Regenerated from caffeinelabs/openapi-generator@d009dcad56f, which adds support for anyOf / oneOf schemas whose every branch produces a JSON string. These are now emitted as a flat type X = Text alias instead of the empty record {} that 0.1.1's diagnostic flagged.

Fixed

  • ChatApi.createChatCompletion is now functional end-to-end. No more "ModelIdsShared has no synthesisable JSON form" diagnostic — the generator now produces a real type. Consumer code: model = "gpt-4o-mini". Wire: "model": "gpt-4o-mini". OpenAI: 200.
  • Symmetric fix for the other two string-flatten schemas in OpenAI's spec:
    • ModelIdsShared = Text (chat models)
    • CreateCompletionRequestModel = Text (legacy completions)
    • VoiceIdsShared = Text (TTS voice IDs)

What's still open

  • Mixed-shape oneOf / anyOf (e.g. oneOf<string, ComplexObject>) still falls through to the old empty-fallback path with the diagnostic. Future work — none on the OpenAI focusApis surface today.
  • Pre-existing template caveats from 0.1.0 (oneOf-array tag sanitisation in Completions/Embeddings, slash-in-record-fields in Moderations) unchanged. mops publish still needs --no-docs.

🤖 Generated with Claude Code

Loading

v0.1.1 — diagnostics regen

27 Apr 15:52
@ggreif ggreif
v0.1.1
5388597

Choose a tag to compare

View all tags
v0.1.1 — diagnostics regen

Regenerated from caffeinelabs/openapi-generator@383d3ae with --additional-properties=diagnostics=true.

Added

  • validate(value : T) : ?Text per model; API methods call it
    before serialisation and throw Error.reject(msg) on a non-null
    result. Surfaces the ModelIdsShared empty-fallback (oneOf<string, …> the codegen could not tag) to the caller's
    catch (e) { e.message() }, instead of letting bad JSON reach OpenAI
    as "model": {} and come back as an opaque 4xx misread upstream as
    "Invalid OpenAI API key".
  • Reply-side error context — every reply-side throw Error.reject
    (UTF-8 decode, JSON tokenise, "Failed to deserialize", "Failed to
    convert", declared error code with no body schema) appends
    — server returned: <body> (or byte count for binary). Closes the
    black-box where the server's actual response was hidden from the user
    on parse failures.

Notes

  • Same focusApis surface as 0.1.0 (Chat / Completions / Models /
    Embeddings / Images / Audio / Moderations / Files; 220 .mo files).
  • Same generator-template caveats as 0.1.0; mops publish still needs
    --no-docs until upstream sanitisation gaps are closed.

🤖 Generated with Claude Code

Loading

v0.1.0

20 Apr 16:43
@ggreif ggreif
v0.1.0
52a2b85

Choose a tag to compare

View all tags
v0.1.0

Initial release of the generated Motoko client for the OpenAI API.

Scope

Curated subset of the full OpenAI API surface — 8 APIs kept, models pruned to their transitive closure:

  • Chat, Completions, Models, Embeddings, Images, Audio, Moderations, Files

Totals: 8 API modules, 211 Model modules, 1 Config, 220 .mo files.

Status

  • mops build: green.
  • Per-file moc --check: 208 pass, 12 fail — all in Completions / Embeddings / Moderations models (two template-sanitisation gaps documented as caveats: in the generator config). The Chat surface and mops build are unaffected.

Usage

import { createChatCompletion } \"mo:openai-client/Apis/ChatApi\";
import { defaultConfig } \"mo:openai-client/Config\";

let config = { defaultConfig with auth = ?#bearer \"YOUR_API_KEY\" };
Loading