Release 0.6.3: reliable repository tools, repair feedback, and evidence handoff - #36
Merged
Merged
Conversation
… handoff Repository investigations could exhaust local limits before reading decisive source, treat unsupported-language outlines as complete symbol inventories, and lose structured tool metadata on cache hits. Repair attempts could repeat unhelpful feedback, while successful source observations and valid repository paths could disappear from downstream verification and reporting. - Raise and document local investigation targets, expose remaining targets to models, and derive fixed 2x hard ceilings for aggregate calls, rounds, and result characters across packet review, system review, and verification. Keep per-result caps, explicit spending limits, and repair deadlines bounded; separate invalid/refused requests from executed-call accounting. - Require valid inclusive read_range bounds before SDK coercion, preserve actual delivered ranges and EOF behavior, and guide missing-file recovery through path discovery. Make unsupported-language outlines explicit and include complete small-file source when it fits. - Pack searches and file lists on complete entry boundaries, retain omission metadata, and preserve complete structured results through the tool cache. - Retain successful source observations across worker retries and failures with attempt/revision provenance; supply bounded collected source to the verifier and improve per-question reconciliation evidence selection. - Validate attention paths against repository-wide known paths, require an explicit reconciliation disposition for each supplied concern, and retain unresolved questions without accepting unsupported resolution references. - Feed repairs the latest bounded key/type diagnostics even with replacement conversations and custom prompts. Detect XML parameter syntax outside JSON strings, provide targeted schema/format guidance, and keep malformed drafts rejected until a complete submission passes strict validation. - Consolidate equivalent cross-file fallback findings without merging distinct diagnoses, preserve publication identities and source provenance, and show failed synthesis prominently without misrepresenting completed coverage. - Add Plan 124, regression coverage, and deterministic text-tool synthetics; update package version and workflow examples to 0.6.3. Validation: pnpm test (1,557 tests, 68 files, including workflow validation), make evals (39 checks, 2 files), pnpm run typecheck, pnpm run build, and git diff --check all pass. Latest OMSX runs completed without structured-output repairs or budget refusals: OpenRouter Opus 5 high in 4m26s ($4.2763), and DeepSeek 4.1 Flash max in 10m13s ($0.2581). These runs establish successful execution, not causality for the XML repair improvements, which were not exercised. Evidence selection can still omit decisive excerpts, and report severity/concision remain follow-up work; do not claim these remaining quality issues are resolved.
|
🧞 Codegenie review failed (
|
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.
Summary
Prepare 0.6.3 with more reliable repository investigation, structured-output recovery, and final reporting. Reviews could previously stop useful source discovery at a local limit, lose evidence between worker attempts and later stages, or produce a fallback report without making synthesis failure sufficiently clear. This release gives investigations bounded headroom, preserves source provenance, improves corrective feedback, and exposes reporting limitations.
Changes
read_rangebounds and preserve EOF/range metadata; guide unknown-path discovery; make unsupported-language outlines explicit and include complete small-file text when possible. Preserve structured cache results and disclose bounded search/file-list omissions.Validation
pnpm test: 1,557 tests across 68 files, including workflow validation.make evals: 39 deterministic synthetic checks across 2 files, with no model calls.pnpm run typecheck,pnpm run build, andgit diff --check: passed.Completed OMSX reviews on the same repository revision (
df1125f56a):Neither run hit a budget refusal or required composition fallback. Both still reported unresolved questions. Their success does not establish that XML repair feedback improved live recovery, because neither needed a repair.
Remaining follow-up
Reconciliation can still omit decisive, already-collected source excerpts and leave answered questions in the report. Severity calibration, preserving concrete recommendation expressions, and report concision also need further work. This release improves mechanics and disclosure; it does not claim to eliminate semantic review errors.
Merge this PR before tagging/publishing 0.6.3. This PR does not publish the release.