uncoded 0.7.0

Symbol indexes and language-server setup for AI-assisted code navigation

uncoded

AI coding agents navigate codebases poorly. They grep for guessed keywords, skim the first few lines of files, and fill gaps from pretraining rather than reading the actual code. The result is plausible-looking output built on a hallucinated understanding of code that's sitting right there, unread.

uncoded builds a static navigation index so agents can orient themselves at the start of a task and navigate deterministically to what they need — no guessing, no grep.

It also wires up a language server so the agent can find references, rename, and edit symbols by name — no grep, no manual find-and-replace.

What it generates

Running uncoded sync produces:

.uncoded/namespace.yaml — a hierarchical YAML file listing every symbol: directories, files, classes (with attributes and methods), functions. Covers all configured source roots. An agent can load this at the start of a task and immediately know the full vocabulary of the codebase.

.uncoded/stubs/ — one .pyi stub per source file, with imports, full signatures (parameter names, types, return types), module constants, and class attributes.

.claude/skills/coherence-review/SKILL.md and .agents/skills/coherence-review/SKILL.md — a coherence review skill, written to both Claude Code and Codex skill directories (see Coherence review below).

uncoded also injects a navigation protocol into CLAUDE.md/AGENTS.md, so agents working in the repo pick up the instructions automatically.

Install

pip install uncoded

Or with uv:

uv add uncoded

Configure

Add a [tool.uncoded] section to your pyproject.toml:

[tool.uncoded]
source-roots = ["src", "tests"]

Use

uncoded sync

Run it from the repo root. It reads pyproject.toml to find your source roots, builds the index, and updates CLAUDE.md/AGENTS.md.

Commit the generated .uncoded/ directory so agents working in the repo always have a current index.

Set up the language server

uncoded setup

Generates the MCP and Claude Code configuration that wires up Serena (language-server bridge) and ty (Python backend), so agents can find references, rename, and edit symbols by name. Safe to re-run. See Using uncoded with a language server below for the generated files and notes on non-Claude-Code agents.

Keep it current with pre-commit

Add uncoded sync as a pre-commit hook so the index stays in sync automatically:

- repo: local
  hooks:
    - id: uncoded
      name: uncoded
      entry: uncoded sync
      language: system
      pass_filenames: false

Like ruff format: if uncoded sync modifies any files, the commit fails and you stage the updated index before committing again.

You can also set up your CI to run pre-commit run --all-files to verify the index is up to date.

Verify the index is fresh

For CI or scripted checks that must not modify the working tree, use the check subcommand:

uncoded check

It runs the same pipeline but writes nothing. Exits 0 if every generated file is byte-identical to what a rebuild would produce, and 1 otherwise — printing which files would change. A stale index is a silent failure mode (agents read misleading names and signatures), so gating on this in CI is worthwhile even alongside a pre-commit hook.

How agents use it

When uncoded is set up, a navigation section is automatically maintained in the configured instruction files (by default, CLAUDE.md and AGENTS.md). Agents following that protocol:

1. Read .uncoded/namespace.yaml to orient — every symbol, at a glance.
2. Read the relevant .pyi stubs to understand imports, signatures, constants, and class members.
3. Use Serena's find_symbol(..., include_body=True) when they need implementation detail for a specific symbol.

The split is deliberate: uncoded provides a stable map and signature index; Serena resolves the current source body. No grep, no stale line-number coordinates, no offset arithmetic.

Coherence review

AI coding agents tend to leave codebases in an incoherent state: names that no longer match behaviour, docstrings that describe stale signatures, dead symbols, pattern changes applied in some places but not others. uncoded sync installs a /coherence-review skill that runs a structured diagnostic sweep to find these problems.

Invoke it in Claude Code:

/coherence-review

The review works in four sweeps:

1. Orient — loads namespace.yaml and forms a vocabulary map.
2. Lexical — scans the namespace for naming inconsistency: concept duplication, qualifier accretion (_v2, _legacy, _final), vocabulary islands, name collision with drift.
3. Promissory — checks each public symbol's name / signature / docstring triple for internal disagreement. Names and signatures come from the stub; docstrings come from Serena's find_symbol(include_body=True).
4. Structural — checks for boundary violations (private symbols imported across modules), overgrown public surfaces, cross-domain imports, and zero-caller public symbols.

Output is a timestamped Markdown report saved to .uncoded/reviews/, with verbatim evidence and a confidence level (high / medium / low) for each finding. The review only reports — it proposes no fixes. The human decides what to follow up.

Using uncoded with a language server

Symbol-level operations — finding callers, reading or editing a single symbol, renaming, safe deletion — are better served by a language server than by grep and freeform text edits. Uncoded's map supplies the name_path and relative_path these tools take as input.

The recommended setup is oraios/serena as the MCP bridge with astral-sh/ty as the Python language-server backend. Serena launches via uvx, so there's nothing to install globally; ty is downloaded by Serena on first use.

Setup

uv run uncoded setup

Generates three files, tailored for Claude Code:

• .mcp.json — registers Serena as an MCP server, launched via uvx.
• .serena/project.yml — picks ty as the backend, ignores .uncoded/, and narrows Serena's tool surface (drops execute_shell_command, the memory tools, onboarding helpers, and the dashboard opener).
• .claude/settings.json — enables the Serena server and allowlists its navigation and edit tools.

Safe to re-run: JSON files merge into existing content (so pre-existing MCP servers and permissions are preserved), and the Serena project YAML is left alone once present. Restart your agent afterwards so the new MCP server is picked up.

If you're not using Claude Code, the generated .serena/project.yml is MCP-client-agnostic, and .mcp.json can serve as a starting point — replace claude-code with your client's context name.

Dev setup

Clone, install dependencies, and wire up the pre-commit hooks:

git clone https://github.com/alimanfoo/uncoded
cd uncoded
uv sync --extra dev
uv run pre-commit install

Run the tests:

uv run pytest

Run all checks (the same suite CI runs):

uv run pre-commit run --all-files

Note for Windows contributors

CLAUDE.md is a symlink to AGENTS.md (single source). macOS and Linux handle this transparently; on Windows, git's core.symlinks setting must be enabled, or the checkout writes CLAUDE.md as a plain file containing the literal string AGENTS.md and the navigation section drifts.

Releasing

GitHub releases publish to PyPI through .github/workflows/publish.yml. The workflow uses PyPI Trusted Publishing, so it does not need a PYPI_TOKEN or any other long-lived publishing secret.

The PyPI trusted publisher is configured for:

• PyPI project: uncoded
• Owner: alimanfoo
• Repository: uncoded
• Workflow: publish.yml
• Environment: pypi

To release, create and publish a GitHub release from the release tag. The published release event builds the source distribution and wheel, then uploads them to PyPI.