fix: session lifecycle, harbor cache, console verdicts and paging, no generation caps or trimmed output by abhishekgahlot2 · Pull Request #104 · DeepGym/deepgym

fix: key harbor clones by url digest and keep full test output

abhishekgahlot2

committed

07294aa

fix: harbor reward errors keep the whole test output

abhishekgahlot2

committed

69bbf03

fix: refuse session calls after close and hand failed starts to the closing pipeline

abhishekgahlot2

committed

b337d3e

fix: a request racing a session delete gets 409, not 500

abhishekgahlot2

committed

dccf079

fix: send no generation caps and keep truncated outputs out of difficulty

abhishekgahlot2

committed

ee0878d

fix: claude runs at its maximum output, no read timeout on z.ai replies, lm-eval takes the model's maximum

abhishekgahlot2

committed

95fa5e4

fix: show no-verdict console episodes as unscored and page the full episode history

abhishekgahlot2

committed

2ee6b89

merge fix/api-session-lifecycle

abhishekgahlot2

committed

be6fa52

merge fix/harbor-cache-and-output

abhishekgahlot2

committed

a3faea4

merge fix/console-verdict-and-paging

abhishekgahlot2

committed

b5c00dc

merge fix/no-generation-caps

abhishekgahlot2

committed

47c98bb

fix: local results record their isolation, unsandboxed runs warn once, a closed session still returns its fixed result

abhishekgahlot2

committed

445e244

fix: no trimmed output in errors, reports or qa details; verifier ids are digests, not a task prefix

abhishekgahlot2

committed

a148fe8

docs: changelog for the 0.6.2 fixes

abhishekgahlot2

committed

ad52fe4

fix: page the full run history in the console with a total instead of capping the runs list

abhishekgahlot2

committed

14899c1

merge fix/console-runs-paging

abhishekgahlot2

committed

cb36e53

fix: claude replies must reach message_stop; audit ids are environment identity and their hash covers all grading inputs

abhishekgahlot2

committed

7f85372

fix: local session results record their unsandboxed actions, which warn once

abhishekgahlot2

committed

60965a4

fix: move console runs and episodes back to the last page when a poll shrinks the total

abhishekgahlot2

committed

c59b34c

fix: cybergym scripts keep completion status, average only whole judged answers, drop the answer-key fallback and generation deadlines

abhishekgahlot2

committed

e12da6d

merge fix/console-page-index

abhishekgahlot2

committed

d35029c

docs: changelog for the review fixes

abhishekgahlot2

committed

31bed70

fix: scan scripts use the shared grading fingerprint and report every problem

abhishekgahlot2

committed

ebf4a92

test: wait for the solution's path before cancelling, and give the late-retry case room on busy runners

abhishekgahlot2

committed

02731fe

fix: only a natural stop is a whole answer, cached completions never parse legacy text, the audit version covers the whole environment

abhishekgahlot2

committed

60792dd

Navigation Menu

Repository navigation

fix: session lifecycle, harbor cache, console verdicts and paging, no generation caps or trimmed output - #104

fix: session lifecycle, harbor cache, console verdicts and paging, no generation caps or trimmed output#104
abhishekgahlot2 merged 25 commits into
mainfrom
fix/0.6.2

Commits on Oct 23, 2026

fix: key harbor clones by url digest and keep full test output

fix: harbor reward errors keep the whole test output

fix: refuse session calls after close and hand failed starts to the closing pipeline

fix: a request racing a session delete gets 409, not 500

fix: send no generation caps and keep truncated outputs out of difficulty

fix: claude runs at its maximum output, no read timeout on z.ai replies, lm-eval takes the model's maximum

fix: show no-verdict console episodes as unscored and page the full episode history

merge fix/api-session-lifecycle

merge fix/harbor-cache-and-output

merge fix/console-verdict-and-paging

merge fix/no-generation-caps

fix: local results record their isolation, unsandboxed runs warn once, a closed session still returns its fixed result

fix: no trimmed output in errors, reports or qa details; verifier ids are digests, not a task prefix

docs: changelog for the 0.6.2 fixes

fix: page the full run history in the console with a total instead of capping the runs list

merge fix/console-runs-paging

fix: claude replies must reach message_stop; audit ids are environment identity and their hash covers all grading inputs

fix: local session results record their unsandboxed actions, which warn once

fix: move console runs and episodes back to the last page when a poll shrinks the total

fix: cybergym scripts keep completion status, average only whole judged answers, drop the answer-key fallback and generation deadlines

merge fix/console-page-index

docs: changelog for the review fixes

fix: scan scripts use the shared grading fingerprint and report every problem

test: wait for the solution's path before cancelling, and give the late-retry case room on busy runners

fix: only a natural stop is a whole answer, cached completions never parse legacy text, the audit version covers the whole environment