Skip to content

CHANGELOG.md

Latest commit

 

History

History
85 lines (77 loc) · 5.13 KB

CHANGELOG.md

File metadata and controls

85 lines (77 loc) · 5.13 KB
Raw

Changelog

All notable changes to CodeWell are tracked here.

Unreleased

Added

  • Local Python workspace indexing into SQLite.
  • Incremental re-indexing that skips unchanged files and removes deleted files.
  • SQLite schema migrations and database status/overview commands.
  • SQLite FTS5 lexical search over paths, symbols, and source content.
  • JSON context packs with selected files, symbols, imports, call edges, related revision memory, and budget-aware symbol traces.
  • Per-file context selection reasons that explain path, symbol, body, import, call-edge, and fallback matches.
  • Actionable CLI guidance for missing databases, unsupported repository URLs, empty code searches, and empty revision searches.
  • Installed-user first-run examples that exercise index, status, search, trace, context, and revision memory commands on a throwaway project.
  • Stable JSON examples for CLI trace/context output and MCP search/context/revision-memory tool results.
  • Context pack graph expansion now adds a small number of related support files through imports, callee definitions, and caller files while staying budget-aware.
  • Query-aware context expansion now recognizes related test files, command entrypoints, and route files for Python projects using static path/import/call heuristics.
  • Revision memory search now re-ranks results with local applicability signals, verification-state weighting, and warnings for stale, rejected, weak, or under-documented fixes.
  • Symbol tracing with matching definitions, capped source excerpts, incoming calls, and outgoing calls.
  • Public GitHub repository indexing by URL with local archive caching and provenance metadata.
  • Revision memory for snippet usage, adaptation failures, fixed revisions, verification states, and revision search.
  • MCP server tools for indexing, search, symbol tracing, context packs, database status, revision recording, and revision memory search.
  • Demo and real-project smoke scripts for local validation.
  • Fixture evaluation script and evaluation guide for measuring usefulness beyond unit tests.
  • Project evaluation script for running JSON task lists against local Python projects.
  • Machine-readable recall, precision, budget, and latency metrics in project evaluation reports.
  • Self-evaluation task list for CodeWell's own indexing, context, revision memory, and MCP paths.
  • Search ranking now gives path and symbol matches more weight than body-only matches.
  • Natural-language self-evaluation tasks for CodeWell's core retrieval paths.
  • Query normalization with stop-word filtering, light stemming, snake_case expansion, and fallback optional matching.
  • Acceptance evaluation tasks for CLI dispatch, GitHub ingest, schema/status reporting, Python AST parsing, and real-project smoke workflows.
  • Multi-project evaluation script plus real-project manifest and task templates for running evaluations across local repositories.
  • Open-source evaluation tasks and report for pinned PyPI sources of Click, Requests, and Rich.
  • Release-readiness report that summarizes feature coverage, known limitations, current quality gate status, and remaining launch evidence.
  • Pre-release check script that runs evaluations, tests, linting, type checks, dependency checks, and diff whitespace checks in one command.
  • Package smoke script that builds a wheel, installs it into a temporary virtual environment, and validates the installed codewell CLI against the fixture project.
  • GitHub Actions CI workflow that runs the full release gate on pushes, pull requests, and manual dispatch.
  • V1 evaluation-note and freeze-candidate docs that summarize the current TS/JS agent-eval boundary, repeated-run interpretation, and snapshot guidance for freezing a stable V1 baseline.

Changed

  • Consolidated internal MVP, technology stack, and engineering protocol notes into docs/PROJECT_GUIDE.md.
  • Context pack selection now filters noisy matches more aggressively while preserving close identifier-based implementation files.
  • Search ranking now gives extra weight to qualified symbol pairs such as Console.print.
  • Project evaluation uses a wider trace window so long methods can verify later direct calls.
  • CLI MCP server import is now deferred until serve --mcp, so ordinary installed CLI commands do not import MCP server dependencies.

Notes

  • CodeWell is now V2 beta (v2.0.0-beta).
  • The default path remains local-first and does not require LLM APIs, embeddings, hosted services, GPUs, or a background daemon.
  • V2 adds ambiguity-aware retrieval with subgoal decomposition, workspace clustering, and primary/backup branch selection — all without required LLM or embedding dependencies.
  • 17 evaluation tasks across 12 domains, 5 codebases, and 2 languages (TypeScript + Python) show net-positive results on elapsed time (88%), tool calls (100%), and token usage (94%).
  • GitHub token-based auth (via GITHUB_TOKEN or GH_TOKEN env vars), multi-language parsing (Python AST + JavaScript/TypeScript regex), and a local HTML UI dashboard (codewell ui) are implemented.
  • 484 tests passing across all modules.
  • Private GitHub repositories, embeddings, and reranking are planned but not implemented.