Once large strings stopped being re-redacted on every response, cProfile
showed the cheap prefilter had become the dominant cost of the redaction
pass: 2.98s of 3.7s, spent testing 71 literal markers against each of the
~74k strings (~19MB) of a large session, i.e. ~3.1M interpreted iterations
per request — repeated by every tab and every poll for a constant verdict.
Memoize the verdict behind a bounded LRU, mirroring the existing redaction
caches: entry count and per-entry size are both capped, so retained bytes
stay bounded and oversized blobs simply run the scan uncached. Non-strings
never reach the cache, so an unhashable value cannot raise.
Measured on a real 22MB session, warm: full redaction 2.07s -> 1.07s (-49%),
output byte-identical. Combined with the large-string cache, the recurring
redaction cost drops from ~8.4s to ~1.1s.
Rejected alternative, kept in a comment: collapsing the 71 `in` scans into
one compiled alternation is the obvious-looking fix but measured 38% SLOWER
(2.60s vs 1.88s) — CPython's substring search is already a tuned C routine.
Security: the prefilter gates the entire redaction pass, so equivalence is
pinned by a differential test suite against a copy of the previous
implementation — every marker individually, case variants, embeddings,
near-misses, unicode lowercase traps, 4000 fuzz cases, plus oversized and
unhashable inputs. Verdicts are identical on all 73,658 strings of a real
session (1935 flagged before and after).
Broad selection: 22 failures before and after, identical set — all
pre-existing.
Hermes Agent
b7f30c6365