Every silent-stop fix so far was a detector for a shape somebody had already
found in the field, which is why the class kept coming back wearing a new hat.
turn_spoke.py moves the question down a tier: at the one exit every terminal
path passes through, ask whether anything readable appeared since the user last
spoke, and if not, say the honest line. Cause no longer has to be enumerated for
the user to be answered.
proactive_prune.py is the hermes trigger we were missing. Their own tests say
our bug out loud: on a large window, a percentage-of-window compaction check
almost never fires, so aged tool output rides in history and is re-sent verbatim
every turn. Measured here, our shaping cut 0.0% at every session size; with a
fixed 60K-token cost trigger it cuts 88% at 12 turns, 93% at 30, 94% at 60. The
prompt-cache contract is load-bearing rather than optional, because our prune is
a rebuild: it commits only when it reclaims enough to pay for the busted prefix,
then disarms until history has regrown a full runway.
Also: lane preflight now treats only 401/403 as a dead credential, since
testStatus=="unavailable" conflated a throttled lane with a revoked one and told
users to reconnect a merely rate-limited Gemini; and awaiting_reconnect is
cleared when the retry budget is spent, so a stale flag can no longer muzzle the
floor and end an ask in total silence.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014wtspwSFzZmjCx9UNPAorQ
