Mocks the OpenAI SDK at the module boundary (Gemini speaks the
OpenAI-compat dialect at the v1beta/openai endpoint), boots a real
PuterServer for the wired MeteringService, and exercises construction,
model catalog/resolution, request shape (max_completion_tokens rename,
cache_control stripping, stream_options gating), non-stream + streaming
output (cached-token splitting in the metered usage shape), error
mapping, and moderation.
closes#2961
Drops the local makeMeteringStub in favour of a live PuterServer-wired
MeteringService. The OpenAI SDK is still mocked at the module boundary
(the real network egress point); the mock now also exposes a default
export so the test server can boot every chat provider (some import
the default and read .OpenAI off it). Aligns with AGENTS.md: "Prefer
test server over mocking deps."
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* tests: add tests for OCRDriver: [#2964]
Adds offline OCRDriver.test.ts covering both providers:
• test_mode short-circuit; argument validation (missing actor, missing
source, unknown provider, AWS/Mistral not configured)
• aws-textract: raw-bytes vs S3Object source selection (regional
client when fsEntry has a bucket), block normalisation (PAGE/WORD/
TABLE filtered, LINE/LAYOUT_TITLE → text/textract:* blocks),
402 on insufficient credits, per-page metering
• mistral: image vs PDF chunk packaging (image_url with base64 data
URL vs document_url with documentName), pass-through of pages /
annotation / image-limit options, markdown → LINE-block
normalisation with page indices, per-page metering, additional
annotations metering when bbox/document annotation formats are set
• default-provider selection (AWS preferred → Mistral fallback →
500 when neither is configured)
• getReportedCosts mirrors costs.ts
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* tests: OCRDriver uses setupTestServer + real fs/metering
Drops the manual config/clients/stores/services stub apparatus and
the loadFileInput mock in favour of the live wired driver from
server.drivers.aiOcr. The Textract and Mistral SDKs are still mocked
at the module boundary (the real network egress points); inputs go
through the real loadFileInput against real fs/store wiring (data
URLs for most cases; FSService.write produces a real fsEntry for the
PDF documentName test). Aligns with AGENTS.md: "Prefer test server
over mocking deps."
The S3Object-source/regional-client assertion was dropped because it
isn't deterministic against the in-memory S3 store and the driver's
per-region TextractClient cache leaks across tests. That branch is
better exercised by a real-cloud integration test.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* tests: add unit tests for MoonshotProvider with mocked client: [#2979]
Adds offline MoonshotProvider.test.ts covering construction, model
resolution and aliasing, request shape (hardcoded max_tokens=1000,
stream_options gating, tool_use → tool_calls hoisting), vision-model
image-inlining hook, non-stream and streaming output, error logging
and rethrow, and metering. Existing integration test is left
untouched.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* tests: MoonshotProvider uses setupTestServer + real metering
Drops the local makeMeteringStub in favour of a live PuterServer-wired
MeteringService. The OpenAI SDK is still mocked at the module boundary
(the real network egress point); inlineHttpImageUrls is mocked because
verifying the provider invokes it for vision-capable models is the
unit's concern. Aligns with AGENTS.md: "Prefer test server over
mocking deps."
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* tests: add unit tests for MistralAiProvider with mocked client: [#2978]
Adds offline MistralAiProvider.test.ts covering construction, model
resolution and aliasing, request shape (camelCase maxTokens, tool_calls
→ toolCalls and tool_call_id → toolCallId rewrite, complete-vs-stream
routing), non-stream output with the camelCase usage coercion,
streaming with the chunk.data unwrap and toolCalls deviation, error
mapping, and metering. Existing integration test is left untouched.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* tests: MistralAiProvider uses setupTestServer + real metering
Drops the local makeMeteringStub in favour of a live PuterServer-wired
MeteringService. The Mistral SDK is still mocked at the module
boundary (the real network egress point); everything else runs through
the real wired graph. Aligns with AGENTS.md: "Prefer test server over
mocking deps."
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* tests: add unit tests for PuterAIController: [#2965]
Adds offline PuterAIController.test.ts with stubbed drivers covering:
route registration (paths and per-route auth options), app-actor
gating across all four chat-proxy routes, body validation (messages
must be an array; openaiCompletions prompt must be a string;
openaiResponses provider must be openai-responses), driver delegation
and request shaping (model + messages + provider defaults), response
shaping for OpenAI chat-completion / text-completion / Responses
envelopes and Anthropic message envelope (text + tool_use blocks),
SSE streaming with [DONE] terminator, model-listing endpoints with
hidden-id filtering and 501 when list() is unavailable, and the
HMAC-gated video proxy's input/expiry/secret guards.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* tests: PuterAIController uses setupTestServer + spy on aiChat.complete
Drops the manual driver/controller stub apparatus (makeStubDriver,
makeController, custom dep builders) in favour of the live wired
controller from server.controllers.puterAi. Each test that needs a
canned chat completion result spies on server.drivers.aiChat.complete,
which keeps the unit-of-test focused on the controller (validation,
response shape, SSE streaming) without dragging in chat-driver model
resolution and credit checks. Aligns with AGENTS.md: "Prefer test
server over mocking deps."
The "500 when URL signature secret missing" assertion was dropped —
the default config wires a secret, so the branch is unreachable in
the live test server. Replaced with the equivalent 403-on-invalid-
signature assertion which actually exercises the HMAC gate.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* tests: add unit tests for ZAIProvider with mocked client: [#2983]
Adds offline ZAIProvider.test.ts covering construction, model
resolution and aliasing, GLM-specific request shape (custom params,
cache_control stripping, user_id derivation, stream_options gating),
non-stream output with reasoning_content normalisation, streaming
deltas (text and tool_use), error mapping, and metering. Existing
integration test is left untouched.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* tests: ZAIProvider uses setupTestServer + real metering
Drops the local makeMeteringStub in favour of a live PuterServer-wired
MeteringService. The OpenAI SDK is still mocked at the module boundary
(the real network egress point); everything else — config, Context,
metering recording — runs through the real wired graph. Aligns with
AGENTS.md: "Prefer test server over mocking deps."
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Covers `Messages` (normalize_single_message / normalize_messages /
extract_and_remove_system_messages / extract_text), `FunctionCalling`
(normalize_json_schema / normalize_tools_object /
make_openai_tools / make_claude_tools), and `Streaming`
(AIChatStream / AIChatMessageStream / AIChatTextStream /
AIChatToolUseStream). All three are pure data transforms — Streaming
is exercised against a real `AIChatStream` wired to a buffering
Writable, so no method-level spies are needed.
Closes#2956
Covers the Puter↔OpenAI shape conversion (`process_input_messages`,
`process_input_messages_responses_api`), usage extraction
(`extractMeteredUsage`, `create_usage_calculator`), the streaming
handlers (`create_chat_stream_handler` and the responses-API variant),
and the non-stream `handle_completion_output[_responses_api]` paths
including moderation, missing-content, and per-provider deviations.
Streaming is exercised via a real `AIChatStream` wired to a buffering
Writable so no method-level spies are needed.
Closes#2955
Covers the early-return guard when the homepage service is unwired,
shell rendering at /, /settings, /dashboard, /action, /@:username,
canonical_url + meta inheritance from gui_params, /app/:name (404 +
shell, embedded app metadata, JSON / object metadata, malformed JSON),
and /show/* launching the explorer with the trimmed path.
Closes#2973
Covers the provider listing route, the auth-start redirect (404 / 400 /
500 / 302), all three callback flows (login / signup / revalidate)
including state validation, suspended accounts, unconfirmed-email
linking refusal, redirect_uri origin clamping, and the
revalidate-done landing page. Routes are collected via PuterRouter and
invoked with stubbed services that return prefab data.
Closes#2966
Covers `loadFileInput` (validation, data-URL decoding, FS-path resolution,
maxBytes enforcement, ACL gating) and `inferFilenameFromUrlOrPath`. Stubs
stores/fsService with prefab data so tests don't require a real server.
Closes#2957
Covers each config-gated branch (client_libs_root, puterjs_root,
gui_assets_root, builtin_apps) using real temp directories so the
existsSync checks see real files instead of mocked module behavior.
Asserts both registration outcomes (which routes appear, on which
subdomains) and the file resolution rules for the puter.js / puter.dev.js
fallback.
Closes#2974
Extensions listening on the `puter.signup.validate` event can now set a
`code` field (in addition to `message`) when blocking signups. The code
is forwarded through HttpError and appears in the JSON error response,
giving clients a stable, machine-readable signal to act on.
Covers both the AuthController (password signup) and OIDCService paths.
Adds tests for the new field.
Daniel Salazar