Skip to content

fix: reuse verified context and bound incomplete review work - #54

Merged
AppSolves merged 1 commit into
mainfrom
fix/context-budget-and-review-efficiency
Nov 16, 2026
Merged

AppSolves merged 1 commit into
mainfrom
fix/context-budget-and-review-efficiency

Conversation

@AppSolves

@AppSolves AppSolves commented Nov 16, 2026 •

Copy link
Copy Markdown
Owner

Summary

Reuse verified context and avoid spending the remaining execution budget after an irrecoverable review prerequisite fails.

context request -> mounted/verified cache -> reuse without another path slot
                -> new snapshot/path -> existing 24-path/API/byte/integrity limits
terminal missing evidence -> preserve findings -> incomplete COMMENT, failed review job
complete groups -> final synthesis -> unchanged completeness and verdict checks

Keep the global 20-minute and 100-provider-request limits unchanged. Bound the context socket separately to 100 requests, including invalid requests; probes and contents for one snapshot/path share a lookup slot. Preserve primary execution failure alongside allowlisted completion flags and lookup counters. Modified old source remains hash-verified and mounted; complete patches/current or deleted source are supplied without retransmitting full modified old text.

Evidence

  • Before: deterministic regressions show cached read25 invalidating context without any new GitHub API call, loss of context diagnostics beside a timeout, and continued groups/synthesis after terminal missing evidence. Historical production diagnostics prove the global deadline; the exact historical Calendar lookup cause is not retained.
  • After: npm run check passes251 tests, strict types, lint and build. Clean-checkout, real workerd/D1, browser, catalog, Actionlint and audit pass. Both full GitHub CI runs pass, including20 real pinned-Codex Docker scenarios with scripted Responses, isolation and security scanning. No local Docker daemon was started.
  • Real provider acceptance: TopkAPI run43911493285, candidate7835b3c14e49f487ca3667e72e86c7560fac577c, gpt-6.1-sol/xhigh, four commits,93 files/8412 changed lines,89 modified text files, four distinct640x360 PNGs, unchanged transitive JSON fixtures. All6 groups plus synthesis complete, host context complete, zero unfinished reasons. Both introduced authorization and stale-response defects are published at valid changed lines in one App-owned native COMMENT, bound to the exact head. Review step9m08s, including preparation, without extending execution limits. The test used an operator-signed isolated delivery based on fresh real GitHub comment metadata; GitHub dispatch/OIDC, runner, ChatGPT inference and publication were genuine. No production webhook was rerouted.
  • Small real-provider acceptance: TopkAPI run43914086895 passed with the same model/xhigh: two changed files, one complete group with integrated whole-PR synthesis, verified host context, no findings, one native COMMENT and a successful review job. Test policies deliberately keep approval and merging disabled. Source PR54 was merged as fabe169; its Git tree exactly matches the tested candidate.

Merge Danger

Door: two-way. No migration, dependency upgrade, credential, quota, approval or merge-policy change. Rollback still requires coordinated immutable caller/Worker trust pins.

Blast Radius: runner. Existing path/object/hash/byte/API limits and essential-evidence failure remain enforced. Synthetic and controlled live fixtures do not guarantee arbitrary heterogeneous xhigh PRs finish within20minutes. Production/KMU remain unchanged pending acceptance and protected merge.

@AppSolves
AppSolves merged commit fabe169 into main Nov 16, 2026
4 checks passed
@AppSolves
AppSolves deleted the fix/context-budget-and-review-efficiency branch November 16, 2026 17:03
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant