Two corrections, the first to my own work. The `real` tasks shipped without a `target` field, so the cascade ran with an empty document path and its single-file tiers had nothing to act on. The 0/6 I published was therefore partly a handicap I had introduced. `target` is now set on the four tasks whose fix really is single-file; the other two span two files, which is why they are in the suite. The result is unchanged — still 0/6 — so the finding was right for a reason I had not actually established, and now is. Second: an L4-enabled two-pass run over this suite does not amortise. All six resolved at L4 on pass one and all six paid full price again on pass two; nothing closed at L0. At $0.3274 per resolved task the cascade is not cheaper than raw claude at $0.2809 — slightly worse, for the local overhead on top of the same cloud call. That contradicts the `mix` result, where the cache fired perfectly. The likely cause is the same missing target: with no single file to identify a close by, the L0 key has nothing stable to hang on. That run predates the fix and re-running it costs about $4, so it is stated as the likely cause rather than the confirmed one. What no fix repairs is the two genuinely multi-file tasks. They have no single target by construction, so if L0 keys on one file's content a multi-file change cannot be cached at all — and multi-file is 53-66% of real commits. That limit is stated now rather than waiting, because `mix` never tested the case. Cost note: this run was estimated at ~$1.80 on the assumption pass two would be free. It was not, and came to ~$3.93. Co-Authored-By: Claude Opus 5 <[email protected]>
5 lines
353 B
TOML
5 lines
353 B
TOML
name = "retry-backoff"
|
|
prompt = "call_with_retry retries every failure including ones that cannot succeed. Add a Permanent exception that is raised immediately without retrying, and wait 1, 2, 4... seconds between transient attempts via an injectable sleep callback."
|
|
test_cmd = "python3 -m pytest -q"
|
|
test_files = ["test_retry.py"]
|
|
target = "retry.py"
|