Skip to content

TODO.md

Latest commit

 

History

History
33 lines (28 loc) · 4.09 KB

TODO.md

File metadata and controls

33 lines (28 loc) · 4.09 KB
Raw

TODO

Fixes

  • Fix hardcoded max turns in prompt
  • Merge overlapping findings on same line range before posting
  • Pin GetFile/ListFiles to iteration commit SHA instead of branch tip
  • CodeGraphIndexer was registered as singleton with non-keyed IGitConnector — broke DI validation locally after keyed services were introduced. Temp-fixed by resolving via IServiceProvider at runtime. Should align with IndexFunction's provider from IndexRequest.
  • Split IGitConnector — interface is too broad (pipeline + tools + chat). Consider splitting into focused interfaces (e.g. core pipeline, code browsing, chat) while keeping shared implementations.
  • QueryCodeGraph tool is registered unconditionally in CoreStrategy (reviewer + explorer) and Reviewer (chat) — LLM can call it even when EnableCodeGraph is false. Should only add the tool when codeGraph != null. (done)
  • Code graph indexing only runs on repo registration. If a repo was registered before the feature existed, or EnableCodeGraph was false at the time, the graph is never built. Trigger indexing from the review pipeline — if a review runs and the code graph is empty, enqueue an IndexRequest to index-queue so the next review benefits. Don't block the current review.

Features

  • Verification pass — after the reviewer produces findings, run a second pass that tries to disprove each finding before posting. Goal: reduce false positives. The verifier should attempt to refute each finding by checking context, reading referenced code, and confirming the issue is real. Discard findings it can disprove.

  • Build repo map (AST structural index) for compact codebase awareness (done: code graph)

  • LSP integration for type-resolved call graphs (fixes cross-file type resolution limitations in tree-sitter; requires local repo cloning)

  • Externalize prompts (load from blob/config instead of compiled Prompts.cs)

  • Switch auth from PAT-only to DefaultAzureCredential with PAT fallback

  • Semantic search — vector embeddings for similar-pattern detection, consistency checks, related tests

  • Learnings from feedback — inject past review learnings into reviewer instructions

  • Work item integration — fetch linked work items as additional review context (done)

  • Cross-provider work item enrichment — parse AB# references from GitHub PR descriptions/commits and fetch ADO work items (requires dual auth: GitHub + ADO credentials for the same review)

  • Local repo cloning support for tool-based reviews

  • Auto-detect per-repo metadata (language version, framework) to reduce false positives

  • Feature-level config system — per-repo feature toggles on Repository entity (app-owner controls) and .aporia.json ProjectConfig overrides (project-owner controls). Three-tier precedence: global AporiaOptions > per-repo Repository > .aporia.json. Applies to incremental reviews, code graph, chat, and future flags.

  • GitHub webhook: subscribe to pull_request.ready_for_review (draft→ready transition) to trigger reviews when PRs leave draft state (done)

  • GitHub chat: subscribe to pull_request_review_comment.edited for cache invalidation when users edit comments in review threads

  • GitHub issue context — scan PR body/title for #N references, fetch issue details (title, body, comments) via GitHub API, inject into reviewer context for better understanding of why the change was made

  • CodeFix fuzzy anchoring — instead of trusting LLM line numbers for suggestions, match the original code snippet against actual file content (exact → whitespace-trimmed → anchor-based) and reposition the suggestion to the real location. Reduces misplaced suggestions when line numbers drift.

  • "Review skipped" feedback — when a PR is skipped (no matching files, too large, etc.), post a comment explaining why with actionable metrics so the author knows what to fix

  • Claude Code C# LSP agent — create a dedicated agent with cclsp MCP server bridging Roslyn LSP (csharp-ls) for type-aware navigation (go-to-definition, find-references, diagnostics). Blocked on Claude Code LSP stability (tracking: anthropics/claude-code#15619, #15168)