Repository-scale review
Include the system instructions, selected files, current diff, tool schemas, and the output you need to reserve. Test the same review job you plan to run later.
A documented context window is not the whole request path.
A large request can fit the published window and still fail on output reserve, cache shape, timeout, rate policy, or account state. Test one representative payload before you commit another long run.
Use the workload that would hurt to repeat.
A tiny prompt proves the client can connect. It does not prove the path your application depends on.
One representative payload, one saved resultCoding-agent history and tool traces
Carry the conversation, tool calls, tool results, and current task into the request. Check what is retained, what is trimmed, and what reaches the answer.
Large document set
Send the documents, retrieval context, instructions, and expected citation shape together. A window fit is not enough if the final answer loses the evidence you need.
A context number settles only one part of the request.
Check these five boundaries before you call a route ready for a long workload.
Budget, shape, runtime, policy, accountContext plus reserved output
Add the complete input and the output you must leave available. A request can fit the documented context window and still exceed a separate output limit.
Cache shape
Keep reusable instructions and context in a stable prefix, then inspect cache-hit and cache-miss fields. Similar token totals can produce a different charge shape.
Timeout
The client, proxy, gateway, and upstream can each have a timeout boundary. Record which layer ended the request before changing the payload or retrying it.
429 and rate policy
Treat a 429 as a pacing or policy result, not proof that the context is too large. Use bounded backoff and decide when the application must stop.
Balance and effective group
Confirm available balance and the effective request group in Console. The group controls the multiplier applied to the dated base rate.
Put a dated price beside the payload.
These AIWave gateway rows separate input, cache-hit input, and output. Use them to estimate the representative request, then compare the estimate with its request record.
AIWave rate snapshot · 2026-09-10 · listed base rate| Route | Input / 1M | Cache hit / 1M | Output / 1M |
|---|---|---|---|
| deepseek-v4-flash | $0.638 | $0.0202884 | $1.914 |
| deepseek-v4-pro | $1.914 | $0.0637362 | $5.742 |
The listed rate uses default ×1. Select the VIP billing group when you create an API key to pay 10% less, then confirm the effective group before funding a production workload.
Send one request you can explain afterward.
Keep the OpenAI-compatible client, pin the tested model ID, and save the response usage with the result.
Base URL, model ID, representative payloadPython request
import os
from openai import OpenAI
client = OpenAI(
base_url="https://aiwave.live/v1",
api_key=os.environ["AIWAVE_API_KEY"],
)
response = client.chat.completions.create(
model="deepseek-v4-pro",
messages=[{"role": "user", "content": "Summarize the repository context."}],
max_tokens=2048,
)
print(response.usage)
Before production
Write down what a passing request must prove.
The checklist belongs beside the payload, not in a separate launch document nobody reads during an incident.
- Record the full input budget and reserved output.
- Set a client timeout and identify the other timeout layers.
- Bound 429 retries and preserve the first error response.
- Confirm balance, effective group, and the dated rate.
- Save model ID, request ID, status, token fields, and charge.
Before the next long run
Test one representative payload and keep the record.
If it passes, you have evidence for this route and workload. If it fails, you know which boundary to investigate before you repeat the work.