It is tempting to expect a smaller payload to reduce model usage by the same percentage. The actual browser comparison had different changes in instruction bytes, schema bytes and input tokens. I first needed to identify the boundary measured by each number instead of calling them all traffic.
A compact request differs from the internal round trips
The compact execution record retains 1,003 bytes of input and 579 bytes of final output. Those describe the capsule handed to the runner and its last returned result, not all requests and responses exchanged while polling internally. A short summary does not prove the internal work disappeared.
The runner separately wraps client.call and counts JSON bytes for request name/arguments and returned values. Calls, requestBytes, responseBytes and statusPolls in wire metrics describe that boundary. They are not measurements of every packet header and protocol overhead. A comparison needs to state what was serialized and counted before a smaller figure can be interpreted.
Read the historical instruction/schema comparison separately
In the version 0.1.118 public-page comparison, first-call instructions changed from 16,673 to 4,990 bytes and tool schema from 22,392 to 7,485 bytes. Input tokens changed from 63,992 to 55,809. These are not the same percentage. Both sides retained the same aggregate observation context of 69,255 bytes.
Each side made two model calls and checked the extraction result, distinguishing a live historical comparison from a size-only audit. It used the internal CLI of that earlier version; that path is now retired, not newly activated here. Those figures cannot be attached to every execution of the current product merely because the underlying optimization idea remains useful.
Model context includes more than the app counted directly
PromptBytes and schemaBytes built by the application do not reveal every token the outer host sends to a model. Fixed instructions, accumulated conversation, tool definitions and cache state may contribute. I do not divide bytes by an invented conversion factor and label the result actual tokens. Returned usage is a separate measurement.
The compact runner does not trim the current Codex conversation itself. Zero model calls inside its plugin can coexist with substantial input usage outside. If the exact contribution of fixed context was not separately measured, that share cannot be asserted either. Leaving the unmeasured boundary visible is more useful than filling it with a plausible estimate.
A smaller context must still retain needed facts and references
Reduced size does not automatically prove correctness or preserved functionality. Historical observation summarization retained one fresh snapshot while reducing old bodies and actionable references. Reduced tool definitions had a discovery route for loading needed capabilities. Size and retained capability require different checks.
The historical result extracted the title and link from the same original page, with call counts and tokens recorded separately. This article reads that documentation and selected code without repeating the original live-site comparison. It does not turn one kind of verified reading into a claim of improved success across every website. The recorded scope remains one particular task and environment.
Do not merge columns with different units into one saving
The shared ledger aggregates request_bytes and response_bytes separately from token metrics. A test changes bytes substantially without automatically changing token totals. Cache-inclusive total and uncached input plus output also retain separate names. Preserving units comes before elaborate conversion.
For another service, decide which evidence measures the input capsule, internal tool traffic, model usage and completed business result. This example is neither an automatic collector nor a pricing calculator. I record the distinction because identifying exactly what became smaller is what makes the next improvement actionable, rather than producing an attractive percentage with an ambiguous denominator.