Repository navigation
fix: separate optional context discovery from required evidence - #44
Merged
Merged
Conversation
This was referenced Nov 15, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
A speculative filename lookup must not turn an otherwise complete review into an incomplete one. Separate optional discovery from mandatory content retrieval: the new
/probeoperation returns only existence from verified immutable Git tree metadata;/filecontinues to require verified contents and permanently invalidates completeness if it fails. Both share the existing 24-request and GitHub API budgets. There is no model-controlledrequired: falseescape hatch, additional inference pass or policy relaxation.Verified absence is distinct from API failures, special Git objects, size limits and integrity failures. Probes never materialize source or satisfy content evidence; API/tree verification failures remain fail-closed. Neither probes nor later successful reads can clear missing changed-file, trusted-instruction or dependency evidence.
Preserve the original final-completion correction: close/drain lookups before evaluating the trusted host gate, retain validated findings in non-approving submissions, and report only safe completion fields and approved failure codes. Source, prompts, paths, credentials and model output are not diagnostic fields.
Regression coverage includes optional absence, mandatory absence, initially missing dependencies, API 404/403/429/500, truncated/inconsistent trees, special objects, oversized/corrupt blobs, shared lookup limits and concurrent lookups. The actual runner test exercises both paths through a real context socket and preserves findings on mandatory failure. Docker acceptance uses the actual pinned Codex binary with scripted Responses and verifies the probe/file socket contract and existing isolation.
Validation:
npm run check: typecheck, lint, all 224 tests and build passed.npm run test:clean: fresh dependency installation and full check passed without generated Env declarations.npm run test:workers,npm run test:web,npm run catalog:check: passed.npm run test:sandbox: passed with synthetic providers.npm audit --audit-level=moderate: zero vulnerabilities; Actionlint and Wrangler deployment dry-run passed.No merge, deployment, downstream repository changes, trusted-pin changes, live inference or authorization/credential changes are part of this task. Metadata probes cannot prove exhaustive model reasoning; failed required evidence still prohibits approval and merging.