Conversation
Enable compositionReasoningStepDown by default while preserving explicit false values from configuration and the existing CLI overrides. Update the README and configuration/CLI regression coverage for the new default. Treat unsupported formats as normal text-mode operation instead of syntax fallback degradation. Apply this consistently to outlines, symbol reads, definition discovery, symbol mentions, packet accounting, and high-risk context warnings. Continue reporting fallback from configured parsers as degraded, preserve text/heuristic precision, and avoid claiming syntax verification for empty mention results or mixed text/syntax definitions. Cover RIDL, JSON, YAML, Markdown, SQL, unknown extensions, extensionless files, mixed search results, unavailable configured grammars, and generic packet context. Preserve real retrieval errors and context-limit reporting. Update OpenRouter DeepSeek routing to use deepseek, fireworks, and together in that order for both only/order, retaining allow_fallbacks=false and the focused automatic submit-tool compatibility path. Update payload tests and documentation to match the routing configuration. Add repository review configuration that excludes specs and eval/test fixture data. Bump the package to 0.6.1 and synchronize documented GitHub Action references with that version. Add reviewed Plan 123 and its index entry. Document reproduced OMSX search scope and result-packing defects, honest failure/completion reporting, evidence-backed question reconciliation, and consequential missing-test criteria. Plan 123 remains unimplemented for the next iteration. Validation: all 1,309 tests in 60 files, workflow checks, TypeScript checking, package build, and git diff --check pass. No live inference calls were needed for validation.
|
🧞 Codegenie review failed (
|
Implement Plan 123 and the follow-ups from the OMSX and trails-api traces. Search and evidence delivery: - Share validated repository glob semantics across discovery and search; resolve tracked revision paths into bounded literal Git pathspec batches. - Keep invalid globs/queries and backend failures distinct from empty matches. - Pack complete search entries after cache lookup for each consumer budget, preserving symbol identity and disclosing truncated or omitted evidence. - Skip oversized matching lines while retaining subsequent useful matches; reject all-omitted delivery rather than claiming an exhaustive zero result. Review outcomes and reconciliation: - Derive failed, incomplete, unresolved, and completed review health on the host, and use it consistently in reports, artifacts, CLI and Action outcomes. - Preserve completed findings and redacted failure reasons; prohibit clean conclusions when required work failed or material questions remain open. - Retain the actual pipeline stage on terminal failures and mark unmeasured coverage unavailable instead of presenting placeholder zero counts. - Preserve health diagnostics during artifact serialization and retain stage timing snapshots through the telemetry observer. - Reconcile original packet questions as well as verifier concerns using explicit supporting references, including successful complete source reads. - Admit concerns within the existing character budget independently of the display cap; deduplicate evidence and prioritize question-relevant sources. - Distinguish deliberate exclusions from incomplete work and require a material behavioral requirement before promoting missing-test findings. Defaults and CLI: - Skip SVG review by default with review.skipSvgReview and positive/negative CLI overrides; retain SVG availability for repository evidence searches. - Add codegenie use as shorthand for provider use, including help routing. - Update documentation, prompt/schema versions, and implementation records. Validation: - pnpm test: 1,362 tests passed across 62 files, including workflow checks. - pnpm run typecheck, pnpm run build, and git diff --check passed. - Fresh built-CLI Astra eval 129 passed: medium reasoning, composition step-down, prompt p10.17, 10/10 hunks, one finding, 33 successful calls, no repairs, timeouts or tool-budget refusals, 153.971s, and $3.038528 review cost. - Separate recommendation judge graded both fix and test correct. The smoke eval emitted no unresolved questions; wider model comparisons and cross-finding recommendation consistency remain follow-up work.
pkieltyka
changed the title
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
1 participant
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.
Reviews could silently miss search matches, count normal text lookup as degraded, or imply a clean result after required work failed. This 0.6.1 release improves evidence retrieval and makes failures, incomplete work, and unresolved questions explicit while preserving completed findings.
Changes
falseand--no-composition-reasoning-step-downoverrides. Skip SVG files by default throughreview.skipSvgReview, with positive/negative CLI flags and eval overrides. SVG evidence remains searchable. Investigation and verification retain their configured reasoning.allow_fallbacks: false; retain focused automatic submit-tool handling and strict validation. Addcodegenie use <model>[:<level>]as shorthand forprovider use.Validation
pnpm test: 1,362 tests passed across 62 files, including workflow checks.pnpm run typecheck,pnpm run build, andgit diff --checkpassed.Remaining follow-up
The Astra smoke run had no unresolved questions, so it does not independently establish live reconciliation quality. Other model runs still show investigation timeouts, withheld unverified advice, and occasional inconsistencies between related recommendations. These remain focused follow-up work; this release preserves their limitations rather than weakening validation. Overall budgets and repair-attempt limits are unchanged by Plan 123.