Repository navigation

benchmark: integrate recovery fixes for continued EdgeBench work - #6624

Merged

songoow merged 29 commits into

mainfrom

edgebenchtest

Oct 17, 2026

Merged

benchmark: integrate recovery fixes for continued EdgeBench work#6624
songoow merged 29 commits into
mainfrom
edgebenchtest

Commits on Oct 16, 2026
- fix(benchmark): budget complete controller capture preparation
  loopx-agent
  committed
  47247ad
- docs(benchmark): explain bounded full checkpoint capture window
  loopx-agent
  committed
  0721781
- fix(planning): preserve initial quality in shared delivery guidance
  songoow
  committed
  da01f99
- fix(turn): reuse the canonical frozen-request digest matcher
  loopx-agent
  committed
  a9dfe8c
- fix(benchmark): keep a refused checkpoint handoff from killing the source wake
  
  songoow
  and
  claude
  committed
  34f1c99
- Merge PR #6621: preserve initial quality in shared delivery guidance
  
  songoow
  and
  claude
  committed
  05de5f7
- Merge PR #6618: preserve full checkpoint preparation within trial phase
  
  songoow
  and
  claude
  committed
  c1eb303
- test(planning): qualify bounded first-delivery guidance growth
  
  songoow
  and
  claude
  committed
  311126b
- fix(research): bind stride outcomes to complete source evidence
  songoow
  committed
  fae4371
- docs(research): clarify outcome evidence replay and planning guidance
  songoow
  committed
  cae3257
- Merge PR #6623 into the EdgeBench experiment branch
  songoow
  committed
  0f9663a
- fix(testing): recognize planning guidance without dynamic regex
  songoow
  committed
  f7756ff
- fix(testing): keep the CLI output probe importable against a base checkout
  
  songoow
  and
  claude
  committed
  fb6f060
- fix(benchmark): keep the uploaded native prompt readable under any launcher umask
  
  songoow
  and
  claude
  committed
  a0fa279
- fix(benchmark): retry a dropped planning wake within the trial budget
  
  songoow
  and
  claude
  committed
  b7c47c9
- refactor(planning): keep quality message rendering with its owner
  songoow
  committed
  495ceae
- test(architecture): reconcile the retired governance module ratchet
  songoow
  committed
  aa12459
Commits on Oct 17, 2026

Navigation Menu

Repository navigation

benchmark: integrate recovery fixes for continued EdgeBench work - #6624

benchmark: integrate recovery fixes for continued EdgeBench work#6624
songoow merged 29 commits into
mainfrom
edgebenchtest

Commits on Oct 16, 2026

fix(benchmark): budget complete controller capture preparation

docs(benchmark): explain bounded full checkpoint capture window

fix(planning): preserve initial quality in shared delivery guidance

fix(turn): reuse the canonical frozen-request digest matcher

fix(benchmark): keep a refused checkpoint handoff from killing the source wake

Merge PR #6621: preserve initial quality in shared delivery guidance

Merge PR #6618: preserve full checkpoint preparation within trial phase

test(planning): qualify bounded first-delivery guidance growth

fix(research): bind stride outcomes to complete source evidence

docs(research): clarify outcome evidence replay and planning guidance

Merge PR #6623 into the EdgeBench experiment branch

fix(testing): recognize planning guidance without dynamic regex

fix(testing): keep the CLI output probe importable against a base checkout

fix(benchmark): keep the uploaded native prompt readable under any launcher umask

fix(benchmark): retry a dropped planning wake within the trial budget

refactor(planning): keep quality message rendering with its owner

test(architecture): reconcile the retired governance module ratchet

Commits on Oct 17, 2026

Merge current main recovery fixes into the EdgeBench experiment branch

fix(benchmark): consume refused handoffs before checking capture roots

docs(recovery): align continuation guidance with clock-free planning

fix(benchmark): keep accept failures outside packet refusal recovery

test(planning): qualify the observed complete-guidance upgrade

fix(replan): deliver ordinary declared-work evidence to followthrough

docs(replan): define ordinary evidence readback and qualification limits

feat(replan): retain unresolved work evidence through recorded followthrough

Merge pull request #6702 from loopx-project/codex/research-result-adoption-20261017

fix: scope research planning guidance to explicit benchmark entry

test(benchmark): bind capture assertions to one phase deadline

chore(benchmark): retain merged analysis delivery in research branch