Dikaya Pro 2: code review built around reproducible tests

Minimal API example

Illustrative Python, not an executed model response. The key comes from the environment and automatic retries are disabled. A lost connection can leave both outcome and cost unknown: do not retry blindly.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://dikaya.cc/v1",
    api_key=os.environ["DIKAYA_API_KEY"],
    max_retries=0,
)
response = client.chat.completions.create(
    model="seller-pro-2",
    max_tokens=512,
    messages=[{"role": "user", "content": "Create a checklist for my task. Separate source facts from assumptions."}],
)
print(response.choices[0].message.content)

Start with the contract behind the code

A reminder service stores a due date as a calendar day, but a new helper converts it through a timestamp. Users in some time zones see the previous day. Before submitting code, decide whether the value represents an instant, a local date or a date in a named business time zone. Without that contract, a model can produce a plausible patch that preserves the wrong assumption and merely moves the failure elsewhere.

Describe the runtime, language version, input format and expected output. Include an example spanning midnight and explain whether daylight-saving changes matter to the feature. Ask Pro 2 to identify assumptions that cannot be resolved from the supplied code. The modified GLM-5.3 attribution identifies this branded model’s approved foundation; it does not prove correctness on date arithmetic or establish equivalence with every other GLM-based catalogue identity.

Send the smallest useful repository slice

Provide the changed function, its caller and the relevant tests, keeping filenames and line references. Add the package manifest excerpt only when dependency behaviour matters. A whole repository dump can obscure the contract and exceed the client’s request bounds. Conversely, a function without its caller may conceal that the input was already normalized. Tell the assistant which files are authoritative and which are proposed changes.

Remove secrets and unrelated personal data from the packet. Mark comments and fixture contents as source material, not new instructions. If the behaviour depends on an unavailable library, ask for the missing interface rather than an invented method name. The model’s declared context does not mean the browser transmits unlimited files or the full stored conversation. Restate the critical date contract in the current review request.

Ask for a regression case before implementation

Request one concrete input for which the current helper violates the contract, along with the expected result and why it should fail. The example should exercise the date conversion rather than merely assert that a mocked helper returned a value. Run that test yourself in the intended environment before considering a fix. If it passes unexpectedly, investigate the fixture or original hypothesis instead of asking for a larger patch.

Then ask for the smallest implementation change and any additional boundary tests. Keep parsing failures, valid leap days and invalid calendar dates distinct. Require the assistant to explain how the patch preserves the existing caller contract, but not to invent execution results. Dikaya does not natively execute your shell commands through a text answer. An IDE or other client may provide tools, but their availability, permissions and actual results belong to that client.

Separate the Pro 2 demo from paid development

The current demo is centred on Pro 2 and text Chat Completions, with account-specific request snapshots and a separate output limit. Older grants retain their own conditions. It is not unrestricted access to every model or an agentic repository session. Use a small illustrative function when exploring the workflow, and check the current first-party limits instead of assuming a large declared model context is available through the demo.

A reasoning-related setting in the demo does not guarantee zero reasoning usage or a nonempty final answer. Do not infer that an empty response proves nothing was processed. For paid work, images sent to Pro 2 use a separately paid text bridge; actual code text and logs usually preserve details more directly than screenshots. Low/high/max are not declared controls for this primary branded tier, so do not import the separate GLM model’s effort menu into a Pro 2 request.

Review the API wrapper as carefully as the patch

Use seller-pro-2 explicitly and choose an output cap appropriate to a small diff plus tests. When integrating the OpenAI Python SDK, use Dikaya’s base URL and max_retries=0 to prevent hidden automatic repetitions. The SDK README is a source for connection configuration and errors, not evidence that a generated patch is valid. Do not assume every tool, cache or thinking extension from another API is supported without checking Dikaya documentation.

If a request returns a definite HTTP rejection, preserve that response for diagnosis. If the connection disappears, keep the outcome uncertain rather than announcing a free failed attempt. A stream can fail after some processing occurred. Check the available request history and accounting before deciding on a new submission, and do not build a public API resume or idempotency promise around an arbitrary local label.

Finish with evidence that belongs to your environment

Read the proposed diff before running it. Confirm that it does not change unrelated formatting, dependency versions or date semantics outside the stated contract. Run the regression test, related existing tests and the project’s normal checks as appropriate. Save actual output separately from the model’s explanation. If a suggested command cannot run, report the blocker honestly; an assistant’s confident statement that the suite should pass is not a substitute for execution.

To begin from this article, open the prepared draft and inspect its contents. Stored draft text may take precedence, and model selection depends on account permissions. Confirm seller-pro-2 in the selector before sending; no article action automatically executes code or submits an inference request. If documentation needs checking, paid browser search is explicit and separate, while public API access does not grant an internet tool. Finish by recording the tested contract and any unresolved cases alongside the patch.

Questions and answers

Is Pro 2 identical to seller-glm-5-3?

No. Pro 2 has the public ID seller-pro-2 and the approved description of a modified version based on GLM-5.3. seller-glm-5-3 is a separate catalogue identity with its own declared settings and conditions. Shared foundation wording is not a reason to substitute IDs or transfer effort controls between them.

Can I treat a generated test as proof of a fix?

Only after inspecting and executing it in the relevant environment. First establish that a regression case fails for the intended reason, then apply the patch and run the checks again. A test that merely repeats a mocked result may miss the bug, and a model-written success message is not an execution receipt.

Does the demo provide the full paid coding workflow?

No. Pro 2 demo access is text-based and bounded by its own account snapshot, request allowance and output conditions. It does not grant all models, file workflows or external agent tools. Keep the example small and check current conditions; older grants may retain different allowances without being reset by a new guide.

Can Pro 2 run the commands it suggests?

A normal model response provides text, not native command execution. You can inspect and run appropriate commands locally, or use a separate client that exposes tools under its own permissions. Distinguish proposed commands, actual tool invocations and verified results when deciding whether a patch is ready to merge.

Sources

AI-assisted editorial material. Examples are illustrative, not captured model responses or comparative tests. External sources do not replace the service terms.

• Dikaya documentation: Pro 2 and development clients
  dikaya.cc
• Dikaya catalogue: model identities and conditions
  dikaya.cc
• OpenAI Python SDK: configuring automatic retries
  github.com
• MDN Fetch: responses, failures and aborted reads
  developer.mozilla.org