Repository navigation
apr code: record the prompt bytes sent and consumed, and exit ≠ 0 when consumed < sent (L22 slice 1) #5190
Closed as not planned
Description
Refs #4600. This is the first slice of L22, for apr code only. run, chat and serve come in a later slice. The SHACL shape and the 200 KB cell in the #3715 suite come in another.
Defect. apr code does not record how much of its prompt reached the model. #4601 (#4599) refuses an over-budget prompt instead of dropping it. Still, a result can say Pass while nothing proves the whole prompt was consumed.
Done when:
- Measure. The agent turn measures the current prompt as it appears in the request handed to the driver:
input_bytes_sent: the prompt's bytes.input_bytes_consumed: the bytes of that prompt present in the request.prompt_sha256: the hash of the consumed bytes.prompt_tokens: the count from the same counter that sizes the window.
- Fail loudly. When consumed < sent, the turn ends with an
input_not_consumederror.apr codeexits ≠ 0, and the JSON result reports"kind": "input_not_consumed". - Record. The four fields go into the
-p --output-format jsonresult document and into the session trace'ssession_endrecord. - Prove. A falsifier plants a one-byte drop between the loop and the driver and must turn red. A prompt at the window limit must stay green with sent == consumed.
Agent: aprender-w71-e3a
Activity
noahgift commentedon Oct 12, 2026
Status: a fix is committed locally (3f4ab48d7b, off main 8c12fca) and waits for a PR slot.
input_bytes_sent,input_bytes_consumed,prompt_sha256,prompt_tokens.input_not_consumedbefore the driver is called (exit != 0; JSON"kind": "input_not_consumed").-p --output-format jsondocument and in the trace'ssession_endrecord.context_overflow.-D warningsclean; 935/936agent::tests pass, including all 9 new ones. The single failure,test_find_apr_binary, needs anapron PATH and is unrelated to this diff.Limits: the measurement is at the request boundary, not inside a backend, and the error path still writes no session trace.
noahgift commentedon Oct 14, 2026
P0 to P1 (2026-10-14 P0 triage): P0 now means it blocks the next release or carries the operator's own words on it; this issue does neither, so it stays on the 0.71.0 milestone at P1.
noahgift commentedon Oct 14, 2026
Folded into #5335 as a checklist row (TRIAGE-100, 2026-10-14): the open-ticket cap is 100, so this ticket is now tracked on that master and closed as not planned. Nothing is lost: reopen it when its work resumes, and an open PR for it reopens it.
Agent: aprender-clerk