The root barrel is now core-only and side-effect free: types,
createModels/createProvider, auth substrate, lazyStream/lazyApi, faux,
utils. Generated catalogs, api-registry, env-api-keys, images, global
stream functions, and per-API lazy wrappers leave the root.
New @earendil-works/pi-ai/compat preserves the old surface verbatim as
a strict superset of the root: api-dispatch stream/complete with env
key injection, the builtin registration side effect (skip-if-present so
it cannot clobber earlier overrides), deprecated getModel/getModels/
getProviders aliases of the new getBuiltin* reads in providers/all,
lazy api wrappers + setBedrockProviderModule, and image generation.
Compat dies with the coding-agent ModelManager migration.
Packaging: exports map gains ./compat, ./providers/*, ./api/*;
sideEffects array lists only the effectful modules.
Old-global imports across agent/coding-agent/examples and pi-ai tests
switch to /compat (path-only; compat is a superset). The coding-agent
extension loader resolves the pi-ai ROOT specifier to compat, so
existing user extensions using the old global API keep working at
runtime until compat is removed. vitest configs alias /compat to src;
browser smoke imports old globals from /compat.
anthropic, openai-codex, and github-copilot flow modules gain OAuthAuth
exports (login/refresh/toAuth) wired to the prompt()/notify() login
callbacks, making the lazyOAuth attachments on the provider factories
functional. Copilot's modifyModels baseUrl rewriting becomes toAuth()
returning ModelAuth.baseUrl derived from the token proxy endpoint.
Callback-server flows race a manual_code prompt and abort it through
AuthPrompt.signal once the flow settles; OAuthAuth has no
usesCallbackServer flag. The old OAuthProviderInterface exports stay
unchanged until the coding-agent migration.
Auth helpers in src/auth/helpers.ts: envApiKeyAuth() (stored key wins,
then env vars in order, with secret-prompt login) and lazyOAuth()
(flow loads on first use through bundler-opaque dynamic imports in
utils/oauth/load.ts; the OAuthAuth flow exports land in phase 4).
There is no OAuth factory toggle: providers that support OAuth always
attach it, advertising costs nothing until login/refresh runs.
createProvider() in models.ts builds providers from parts: single API
implementation or a map dispatched on model.api (mixed-API providers
like opencode and github-copilot); unknown api yields a stream error.
generate-models.ts now emits one providers/<id>.models.ts catalog per
provider (35 files, biome-excluded like models.generated.ts) and
models.generated.ts becomes a generated aggregator, so importing one
provider factory pulls one catalog. Typed getModel globals unchanged.
One factory per built-in provider under src/providers/: envApiKeyAuth
for standard providers, OAuth for anthropic/openai-codex/github-copilot,
ambient ApiKeyAuth for amazon-bedrock (AWS env/profile/IAM) and
google-vertex (explicit key or ADC+project+location).
providers/all.ts: builtinProviders(), builtinModels(), getBuiltin*
re-exports. fauxProvider() factory returns a real Provider for tests;
legacy registerFauxProvider() unchanged.
On macOS tmpdir() is a symlink (/var -> /private/var) while the spawned
CLI's process.cwd() is physical, so session cwd filtering never matched
the fixtures and the fork-target rejection test failed locally.
Picks up the Claude Fable 5 thinking-off metadata (off: null in
thinkingLevelMap) that 9ccfcd7c added to the generator without
regenerating, fixing anthropic-thinking-disable and supports-xhigh
tests. Includes upstream catalog drift; OpenRouter delisted
moonshotai/kimi-k2.6:free, so the kimi compat test now pins only the
listed variant.
Stream implementations move from src/providers/ to src/api/, renamed by
API id (anthropic.ts -> anthropic-messages.ts, google.ts ->
google-generative-ai.ts, mistral.ts -> mistral-conversations.ts,
amazon-bedrock.ts -> bedrock-converse-stream.ts). Every module now
exports exactly stream/streamSimple; shared helpers move alongside.
New ProviderStreams dispatch contract in types.ts, lazyApi() wrapper in
api/lazy.ts, and one .lazy.ts wrapper per API. Bedrock's wrapper keeps
the node-only variable-specifier import and setBedrockProviderModule()
(now taking ProviderStreams).
providers/register-builtins.ts deleted; interim until the compat
entrypoint lands, builtin api-registry registration lives in stream.ts
and lazy wrappers are exported from the root barrel. Old per-API lazy
exports (streamAnthropic, ...) are gone; package.json subpaths retarget
to dist/api/.
New Models/MutableModels/createModels collection: provider map, async
model listing (best-effort aggregation), getAuth decision tree with
double-checked locked OAuth refresh, stream/complete with per-field
auth merge over lazyStream.
Auth substrate: ProviderAuth { apiKey?, oauth? }, one type-tagged
credential per provider, CredentialStore (read/modify/delete; modify
is the only write path, serialized RMW), OAuthAuth login/refresh/toAuth
split, prompt()/notify() login callbacks, browser-safe default
AuthContext.
types.ts: Provider alias renamed to ProviderId; ApiOptionsMap and
ApiStreamOptions<TApi> for typed per-API stream options; hasApi()
runtime narrowing guard.
Behind PI_EXPERIMENTAL=1, show a first-time setup dialog on interactive startup when the default agent directory is used and settings.json does not exist. The dialog preselects the detected terminal appearance with an explicit dark/light choice (live preview) and asks for opt-in analytics data sharing. Submitting writes settings.json, which serves as the completion marker; opting in stores a generated trackingId. Escape at any point skips out of setup.
When Bedrock rejects a request with "data retention mode '<mode>' is not
available for this model", append a pointer to the AWS data retention
documentation so users can configure a supported mode.
Azure Foundry deploys gpt-5.4 and gpt-5.5 with a 1,050,000 context
window, but the Azure provider cloned OpenAI-direct's 272k API limit.
Override the context window in the Azure derivation.
Also fix gpt-5-pro maxTokens, which upstream metadata set to 272000 (a
duplicate of the input sub-limit) instead of the actual 128000 max
output; corrected at the source so OpenAI-direct and Azure both match.
closes#5559
Mario Zechner
Vegard StikbakkeandGitHub
Burak Varli
haoqixu
Armin Ronacher
Sviatoslav Abakumov
Danila Poyarkov
github-actions[bot]