Repository navigation

[Improve] Check finished coding turns against the request in about a second instead of a minutes-long judge pass - #3071

Merged

mrubens merged 20 commits into

developfrom

feat/realtime-judge-gate

Sep 21, 2026

Commits on Sep 21, 2026

[Improve] Check finished coding turns against the request in about a second instead of a minutes-long judge pass
mrubens
committed
e43b82f
Run the completion check only with a hosted judgment model and address review feedback
mrubens
committed
430e0e3
Keep late follow-ups, mid-check requests, and untracked noise files from slipping past the completion check
mrubens
committed
cbad542
Hold a follow-up that lands mid-check until the closing turn completes
mrubens
committed
6eb4f0b
Hold follow-ups for the whole closing path, not just the check itself
mrubens
committed
c225920
Cover what the judge pass weighed: checklist, validation evidence, proof claims, evident defects, and clipped diffs
mrubens
committed
c03eee3
Count earlier validation runs only while the code they covered is unchanged
mrubens
committed
9995d78
Treat a validation run as stale once source is edited after it
mrubens
committed
32e842b
Count shell rewrites of source as edits that make an earlier validation run stale
mrubens
committed
4b28144
Only file-rewriting forms of git checkout, restore, and reset make a validation run stale
mrubens
committed
e4607c0
Moving to an existing branch makes a validation run stale; creating a branch does not
mrubens
committed
2d1b787
Decide from checkout and switch arguments whether the working tree was replaced
mrubens
committed
918fb46
See through git global options when deciding whether a command replaced the working tree
mrubens
committed
971b6bf
Drop validation runs that predate the last source edit instead of asking the model to discount them
mrubens
committed
cf725fb
Catch an unsupported validation claim reliably with its own measured threshold
mrubens
committed
6f15a40
Decide whether a validation run is stale from the code itself, not from the command text
mrubens
committed
5b21bae
Fingerprint every changed path, falling back to size and mtime past the content caps
mrubens
committed
b13e475
Keep parentheses in the code fingerprint
mrubens
committed
91249ec
Fingerprint files past the caps by their bytes rather than size and mtime
mrubens
committed
c0f73ae
Stream large files into the fingerprint instead of buffering them
mrubens
committed
8559bef

Navigation Menu

Repository navigation

[Improve] Check finished coding turns against the request in about a second instead of a minutes-long judge pass - #3071

[Improve] Check finished coding turns against the request in about a second instead of a minutes-long judge pass#3071
mrubens merged 20 commits into
developfrom
feat/realtime-judge-gate

Commits on Sep 21, 2026

[Improve] Check finished coding turns against the request in about a second instead of a minutes-long judge pass

Run the completion check only with a hosted judgment model and address review feedback

Keep late follow-ups, mid-check requests, and untracked noise files from slipping past the completion check

Hold a follow-up that lands mid-check until the closing turn completes

Hold follow-ups for the whole closing path, not just the check itself

Cover what the judge pass weighed: checklist, validation evidence, proof claims, evident defects, and clipped diffs

Count earlier validation runs only while the code they covered is unchanged

Treat a validation run as stale once source is edited after it

Count shell rewrites of source as edits that make an earlier validation run stale

Only file-rewriting forms of git checkout, restore, and reset make a validation run stale

Moving to an existing branch makes a validation run stale; creating a branch does not

Decide from checkout and switch arguments whether the working tree was replaced

See through git global options when deciding whether a command replaced the working tree

Drop validation runs that predate the last source edit instead of asking the model to discount them

Catch an unsupported validation claim reliably with its own measured threshold

Decide whether a validation run is stale from the code itself, not from the command text

Fingerprint every changed path, falling back to size and mtime past the content caps

Keep parentheses in the code fingerprint

Fingerprint files past the caps by their bytes rather than size and mtime

Stream large files into the fingerprint instead of buffering them