Skip to content

Releases: Viraj465/CodeTrace-ai

Release list

Codetrace AI v1.0.3

Codetrace AI v1.0.3 Latest
Latest

Choose a tag to compare

View all tags
@Viraj465 Viraj465 released this 05 Mar 06:15
v1.0
2460519
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

CodeTrace AI v1.0.3 — The Governance Release

v1.0.3 changes how CodeTrace answers. Every structural claim now carries a live file:line citation and an evidence grade. When the evidence isn't there, the agent says so instead of guessing.

This release also:

  • connects to any OpenAI- or Anthropic-compatible endpoint
  • makes Ollama hardware-aware
  • closes a path-containment gap
  • fixes several bugs that silently degraded results
pip install -U codetrace-ai

Requires Python 3.10–3.12.

⚠️ Upgrade notes (read first)

  • Re-register MCP in each project. Previous versions registered the MCP server against the indexed project instead of the installed package, so your IDE integration may be broken. Run codetrace register-mcp in each project. Registration now writes project-local .mcp.json, .cursor/mcp.json and .vscode/mcp.json. These files contain absolute paths for your machine, so you'll probably want them in .gitignore.
  • Rebuild your index. Re-indexing a file used to drop inbound call edges from other files, so graphs updated incrementally on ≤1.0.2 can understate blast radius. Delta sync alone won't repair this, because unchanged files are skipped. Run a full rebuild: delete .codetrace/graph_metadata.db and .codetrace/sync_metadata.db, then run codetrace index ..
  • Gemini now always runs at temperature 1.0.
  • Ollama now uses the native Ollama API and sizes num_ctx automatically. If you tuned context manually before, pin it again with codetrace set-ctx <tokens>, or use 0 for auto.

✨ Highlights

🏛️ Governed Pipeline Protocol

Every structural claim is graded:

  • CONFIRMED: established by a tool call in this session
  • INFERRED: reasoned from confirmed evidence, and labelled as such
  • UNRESOLVED: not enough evidence

Claims need a file:line citation from live tool output. Memory from compressed history or prior sessions counts as context, not evidence. When something is missing, the agent names what's missing and what re-indexing would fix it.

Investigation follows the narrowest path the question needs:
search_codebase → get_symbol_relations → read_file → analyze_impact → report

🔌 Any LLM endpoint via custom

Point CodeTrace at any OpenAI- or Anthropic-compatible API: DeepSeek, vLLM, LM Studio, Together, Fireworks, or a corporate gateway. You provide the API style, base URL and model name. API keys are optional for local endpoints. Model lists are fetched automatically when the endpoint exposes them.

🖥️ Hardware-aware Ollama

  • Detects GPU memory on NVIDIA and Apple Silicon and picks a safe context window.
  • Lowers the window after the current turn if the model spills to CPU.
  • On a hard out-of-memory error, stops and asks before retrying smaller.

Tune it with:

codetrace set-ctx 16384           # pin a window
codetrace set-ctx --backoff 0.75  # keep 75% on each auto-lower

🧩 MCP from the CLI

  • codetrace mcp [PATH] starts the stdio MCP server directly.
  • codetrace register-mcp [PATH] re-registers it for Claude Code, Cursor and VS Code. Existing entries are preserved, and non-JSON config files are left untouched and reported.

Also new

  • codetrace set-default-ctx and codetrace set-model-limits for cloud models LiteLLM doesn't recognise.
  • codetrace index <GitHub/GitLab URL> keeps the clone under ~/.codetrace/repos/, so re-indexing is incremental and you can codetrace chat against it.
  • Output Normalizer: headings, bullet styles, code-fence tags and spacing render the same regardless of provider.

🔒 Security

  • Path containment fix. read_file and write_file previously used a prefix check, so a sibling directory such as project-secret passed as inside project. Checks now use Path.is_relative_to, and out-of-project writes are refused before the approval prompt appears. Upgrading is recommended.
  • Config hardening. ~/.codetrace/config.json, which holds your API key, is now written atomically with owner-only permissions.

🐛 Fixes

  • git_diff always returned an empty diff, because the revision was passed after -- and git treated HEAD as a filename.
  • Re-indexing a file dropped inbound call edges. Also, calls no longer resolve across unrelated languages.
  • NameError: logger crashed every Ollama session at startup.
  • MCP auto-registration now registers the current interpreter and installed package.
  • Gemini 400 errors caused by stream_options.
  • Blank Architect Error: messages: timeouts and empty exceptions now show a readable cause.
  • The API-key leak detector no longer flags commit SHAs and long identifiers.

⚙️ Improvements

  • Thread-safe Tree-sitter parsing.
  • Chunked ChromaDB upserts and SQLite-safe chunked deletes.
  • Batched FlashRank re-ranking, which ends out-of-memory crashes on CPU-only machines.
  • Built-in ignore list (node_modules, virtualenvs, IDE folders) and correct handling of nested .gitignore files.
  • Case-insensitive path handling on Windows.

🧪 Tests

First automated test suite, running in CI. It covers indexing regressions, the MCP server, the output normalizer, safety and config, and the token manager.

Full changelog: v1.0.2...v1.0.3
PyPI: https://pypi.org/project/codetrace-ai/1.0.3/

Assets 2
Loading