fix(benchmark+cognition): the run pulses, empty completions fault, bench acts land in their room (#371 step 1) - #2335

Merged

joelteply merged 13 commits into

canaryfrom

fix/run-lifecycle-pulse-and-empty-completion

Aug 17, 2026

Commits on Aug 17, 2026

fix(benchmark): a dispatched run's acts happen IN ITS ROOM — agent/solve was hardcoded roomless (#329/#243)

joelteply
and
claude
committed
c9ba5f9
fix(cognition): an EMPTY completion is not a chosen silence — the Err arm's missing twin (#181/#390)

joelteply
and
claude
committed
ce82f00
docs(architecture): the round lifecycle as a recipe-owned state machine (#371 design)

joelteply
and
claude
committed
a4a60f9
feat(benchmark): the run PULSES while it runs — liveness stops being a terminal artifact (#371 law 2, #374)

joelteply
and
claude
committed
dd87643
fix(swe): astropy env builds — a 1990s vendored zlib stubs out fdopen on every modern Apple SDK (#383)

joelteply
and
claude
committed
5e6c2fd
fix(governor): shed CAPABILITY, never starve the window — model choice was floored at bare survival (#438-class)

joelteply
and
claude
committed
03c890b
docs(architecture): context is a capability axis — the governor learns the window but not the CHOICE (Joel 8/17)

joelteply
and
claude
committed
a5ad69c
fix(cognition): my empty-completion guard was standing in front of the #181 recovery — reasoning-bearing empties are not faults

joelteply
and
claude
committed
fc8964b
docs(architecture): the reserve must YIELD to the incompressible prompt floor — why raising window/4 is a refactor, not a constant

joelteply
and
claude
committed
7de2db7
feat(bench): ENFORCE the gold gate — an env that cannot score its own gold patch disqualifies itself (#380 keystone)

joelteply
and
claude
committed
943b8a2
fix(bench): grade django from CANONICAL test ids, not from prose (#383 root cause)

joelteply
and
claude
committed
13fd357
fix(bench): select django's JSON runner through SETTINGS — runtests.py has no --testrunner (#383)

joelteply
and
claude
committed
cb29884
fix(bench): django's dataset ids include DOCSTRINGS — the runner must emit its own (#383)

joelteply
and
claude
committed
2c5b09b

Navigation Menu

Uh oh!

fix(benchmark+cognition): the run pulses, empty completions fault, bench acts land in their room (#371 step 1) - #2335

fix(benchmark+cognition): the run pulses, empty completions fault, bench acts land in their room (#371 step 1)#2335
joelteply merged 13 commits into
canaryfrom
fix/run-lifecycle-pulse-and-empty-completion

Commits on Aug 17, 2026

fix(benchmark): a dispatched run's acts happen IN ITS ROOM — agent/solve was hardcoded roomless (#329/#243)

fix(cognition): an EMPTY completion is not a chosen silence — the Err arm's missing twin (#181/#390)

docs(architecture): the round lifecycle as a recipe-owned state machine (#371 design)

feat(benchmark): the run PULSES while it runs — liveness stops being a terminal artifact (#371 law 2, #374)

fix(swe): astropy env builds — a 1990s vendored zlib stubs out fdopen on every modern Apple SDK (#383)

fix(governor): shed CAPABILITY, never starve the window — model choice was floored at bare survival (#438-class)

docs(architecture): context is a capability axis — the governor learns the window but not the CHOICE (Joel 8/17)

fix(cognition): my empty-completion guard was standing in front of the #181 recovery — reasoning-bearing empties are not faults

docs(architecture): the reserve must YIELD to the incompressible prompt floor — why raising window/4 is a refactor, not a constant

feat(bench): ENFORCE the gold gate — an env that cannot score its own gold patch disqualifies itself (#380 keystone)

fix(bench): grade django from CANONICAL test ids, not from prose (#383 root cause)

fix(bench): select django's JSON runner through SETTINGS — runtests.py has no --testrunner (#383)

fix(bench): django's dataset ids include DOCSTRINGS — the runner must emit its own (#383)