Skip to content

Changelog

This page lists every change to the contract, including every fix that makes uncoded do what its documentation or its skills say. Refactoring, tests, CI and the documentation get no entry. The release notes on GitHub list every merged pull request.

Each release after 3.0.1 groups its entries under these headings, in this order, and leaves out any heading that would be empty. Each heading names the type of release that its entries make, as Compatibility describes the types:

  • Breaking for a change that makes the release major;
  • Added for an addition that makes the release minor;
  • Changed for any other change that makes the release minor; and
  • Fixed for a fix that makes the release a patch.

A change stays under Unreleased, with the same headings, until a release includes it. Each entry links the pull request that made the change.

Unreleased

6.0.1 — 2026-10-29

This release makes no change to the contract. uncoded init now separates the [tool.uncoded] table from the rest of pyproject.toml with a blank line. (#422)

6.0.0 — 2026-10-29

Before upgrading, follow Upgrading from version 5 to version 6.

Breaking

  • The skills run uncoded from the project's uv environment, so a project must use uv, require Python 3.12 or later, and add uncoded as a development dependency. .uncoded.toml is no longer read, and uncoded init now appends the configuration to pyproject.toml. (#415)

5.0.0 — 2026-10-29

Before upgrading, follow Upgrading from version 4 to version 5.

Breaking

  • uncoded no longer reads the required-version setting, and uncoded init no longer writes it. uncoded runs at any version, and a project pins its version in the entry of its pre-commit hook. (#380)

Changed

  • The navigation and consistency review skills tell the agent to run uvx uncoded sync and uvx uncoded body, in place of the project's configured command, which nothing configures. (#407)
  • The docs map lists every CommonMark ATX heading, including one indented by up to three spaces, and leaves out a line that starts with # inside an HTML block. (#403)

Fixed

  • uncoded sync now reads a file under overlapping source or documentation roots once, so it warns about and counts a skipped file once. (#408)
  • The docs map follows CommonMark's fence-closing rules, so a shorter fence or one with an info string no longer exposes headings inside a fenced code block. (#403)
  • uncoded refs now finds every reference when valid whitespace separates a def, async def or class keyword from its name. (#401)
  • uncoded refs retains the host from a file URI, so it reports the correct path for a project on a Windows network share. (#398)
  • uncoded refs ignores a request from ty while waiting for a response, even when its id equals its own request's. It could fail with a traceback. (#395)

4.0.0 — 2026-10-28

Before upgrading, follow Upgrading from version 3 to version 4.

Breaking

  • On Windows, the paths that uncoded refs prints and the path comment in each stub use / to separate directories, as the namespace map and the docs map already do. So does every path in a message or a warning. (#312)
  • uncoded body and uncoded refs find the project directory from the current directory, as the other commands do, and stop outside a project. They also stop when the --in file is outside the project directory. uncoded refs no longer looks for a project from the --in file's directory, and no longer falls back to the nearest pyproject.toml. The code navigation skill tells the agent to run them from the project directory. (#325)
  • Stubs are laid out in the standard stub style that ruff and Black write for .pyi files. A function, a method, or a class with no members is one line that ends with .... No blank line separates consecutive functions, the members of a class, or consecutive classes with no members. (#366)
  • A stub lists the symbols of each module body and class body in the order that the Python file defines them, not grouped by kind, so it shows which of two definitions of a name comes later. (#367)

Added

  • uncoded --version prints the installed version. (#312)
  • The required-version setting names the one version of uncoded that a project uses. At any other version, every command stops, writes nothing, and gives the command that runs the required version. A value that is not one exact version, such as "3.1", stops every command with a message that names the configuration file. (#322, #326)
  • uncoded init sets uncoded up in a project. It writes .uncoded.toml with the roots that exist among src, tests, README.md, AGENTS.md and docs, and with the running version as the required version. It then runs uncoded sync, says which files to commit, and prints the lines that tell agents to load the navigation skills. (#323)

Changed

  • uncoded sync no longer removes skills under the names that earlier versions of uncoded gave them: coherence-review, uncoded-coherence-review and uncoded-review. It removed them on every run, so it deleted a skill of the user's own with one of those names. It removes only the skills that the running version generates, when the roots they belong with are gone. A project that still has a skill under an old name removes it as Upgrading says. (#337)
  • uncoded body and uncoded refs report an --in file that does not exist with the reason that the system gives, as they do for every other file that cannot be read: Error: nope.py: No such file or directory. They said Error: nope.py: file not found. (#354)

Fixed

  • What uncoded prints is UTF-8 on every system. It printed in the system's output encoding, so uncoded body printed a definition that holds é as bytes that are not UTF-8, such as on Windows, and stopped with a UnicodeEncodeError where that encoding lacks the character, such as ASCII. (#368)
  • The uncoded skills use one word for each concept, such as "code index", "definition" and "project". They also stop listing directories and files among the symbols in the namespace map. (#344)
  • The documentation navigation skill and the docs map's header say which heading a " (N)" suffix in the docs map marks: the Nth heading with the same text under one parent. The skill said that every heading in the docs map is literal text, and the header said to take the Nth match of the heading's text, which finds the wrong heading when the same text comes earlier elsewhere in the file. (#344)
  • uncoded body prints a definition with the line endings that it has in the file. It printed every line ending as LF, and on Windows as CRLF, so an edit tool given the printed text could fail to find it in the file. (#342)
  • uncoded body prints the definition when a form feed or a Unicode line separator, such as \u2028 in a string, comes before the definition's end. It counted those characters as line breaks, which Python does not, so it printed the wrong lines. (#342)
  • When a name is defined twice in one module body, the namespace map shows the later symbol, and uncoded body and uncoded refs find no member of a class that a later function, assignment or type statement redefines. The map showed a function over a class, and a class over an assignment, whatever the order. uncoded body and uncoded refs used the class. (#341)
  • uncoded refs prints the symbol's definition and each import of it, as well as each use of it. It printed only the uses, so a symbol whose only reference was an import looked unused. The code navigation skill tells the agent to delete a symbol only when it is unused. (#334)
  • When ty fails, uncoded refs says that ty failed and passes on what ty and uvx wrote about why, such as that ty could not be fetched. It discarded that output, and a ty that stopped early, or a uvx that could not start, could be reported as a fault in the --in file. (#335)
  • When a source root is removed from the configuration while another source root stays, uncoded sync removes the removed root's stubs, and uncoded check counts them as files that would change. Both left the stubs alone, so uncoded check said that the index was current while an agent could still read the stub of a file that was no longer indexed. (#339)
  • uncoded sync and uncoded check end their warnings about the files they skipped with a line that says how many files they skipped. A skip was easy to miss among the other lines, such as in the output of a pre-commit hook. The exit status is still 0. (#329)
  • When a configuration file is not valid TOML or a setting has the wrong type, every command names the file and the problem in one line, exits with status 1 and writes nothing. These gave a Python traceback, or a message that misled. A pyproject.toml on the way up to the configuration file that is not valid TOML stops the command too. (#326)
  • uncoded sync and uncoded check check every root before they write anything. A doc root that could not be used was found only after the code index and the installed skill files were written. The message names the configuration file, the setting and the root, and says whether the root does not exist, is the wrong kind, or is outside the project directory. A root that did not exist was reported as "not a directory". (#328)
  • When uncoded sync or uncoded check cannot read, write or remove a generated file, or uncoded init cannot write the configuration file, the command says in one line which file, what it could not do, and why, and exits with status 1. This gave a Python traceback. (#331)
  • uncoded body and uncoded refs stop when two files in one directory both configure uncoded, as the other commands already do. (#322)
  • uncoded refs runs ty 0.0.92, in place of 0.0.53, and finds references that the older ty missed. (#308)
  • uncoded refs on Windows prints each reference's path relative to the current directory, in place of a path that names no file. (#312)
  • The search for the configuration file passes over a pyproject.toml without a [tool.uncoded] table, so a Python package's own pyproject.toml inside the project no longer hides the configuration file above it. (#321)
  • The navigation skills tell the agent to run uncoded sync before it reads the namespace map or the docs map, and whenever it notices that something else has changed files, such as a pull or a checkout. They told it to run uncoded sync only when a map was missing, so an agent could read a map that was stale. (#333)

3.0.1 — 2026-10-09

  • uncoded refs finds the project directory from a .uncoded.toml configuration file, as well as from a pyproject.toml one. (#251)
  • uncoded-code-navigation tells the agent to start with the namespace map and to read a stub or source only when the task needs that detail, in place of reading the stub of every file it touches. (#256)

3.0.0 — 2026-09-19

Before upgrading, follow Upgrading from version 2 to version 3.

  • uncoded sync writes .uncoded/.gitignore, which keeps the index out of Git. The navigation skills tell the agent to build a missing index, and to refresh it after editing indexed files. (#241)
  • uncoded-coherence-review is renamed uncoded-consistency-review, and reports only disagreements between two claims about the same concept. uncoded sync removes the skill under its earlier names. (#245)
  • The navigation skills tell the agent to load each of them once per session. (#240)

2.1.1 — 2026-06-26

  • A Markdown file that starts with a byte-order mark keeps its first heading in the docs map, and every stub carries the provenance marker. (#201)

2.1.0 — 2026-06-24

  • uncoded reads a Python file in the encoding that the file declares, and every installed skill file carries the provenance marker. (#189)
  • The skills are reworded for clarity. (#190)

2.0.0 — 2026-06-23

Before upgrading, follow Upgrading from version 1 to version 2.

  • uncoded generates its navigation guidance as the skills uncoded-code-navigation and uncoded-doc-navigation, and no longer writes a section into AGENTS.md or CLAUDE.md. The instruction-files configuration key is removed, and the coherence-review skill is renamed uncoded-coherence-review. (#184)
  • uncoded sync skips a Markdown file that it cannot read, with a warning, instead of stopping. (#177)

1.2.0 — 2026-06-19

  • uncoded sync writes the coherence review skill only when source roots are configured, and leaves its section of AGENTS.md or CLAUDE.md alone after a Markdown formatter rewraps it. (#172)

1.1.0 — 2026-06-18

  • The doc-roots setting adds the docs map, and a .uncoded.toml file can hold the configuration in a project without a pyproject.toml. (#165)

1.0.1 — 2026-05-19

  • The guidance that uncoded writes into AGENTS.md or CLAUDE.md, and the coherence review skill, describe uncoded refs as finding references, not only callers. The skill runs uncoded through uvx. (#160)

1.0.0 — 2026-05-18

  • uncoded body prints a symbol's source, and uncoded refs finds its references with the ty language server. They replace Serena, and the uncoded setup command that configured it is removed. (#132, #145, #151)

0.x

Releases 0.1.0 to 0.7.1, from 2026-04-21 to 2026-05-14, came before uncoded's first stable release. Their changes are in the GitHub releases.