MIMRY

Changelog

Notable user-facing changes are recorded here. MIMRY is pre-1.0; compatibility changes may still occur, but the supported release floor is tested before a release is tagged.

0.9.0

Latest

Upgrade notes

  • mimry install --platform claude-code now adds the Claude Code hooks and the prompt hook by default, at user level in ~/.claude/settings.json as well as with --project (#654). mimry upgrade and mimry install --refresh add them to an existing install that lacks them. Pass --no-hooks or --no-prompt-hook to opt out; the choice is recorded, and --no-hooks on an existing install removes the hooks. Your other settings and hooks are kept, and the hooks are skipped when the MIMRY plugin is installed, since it ships its own.
  • User-level hooks run in every folder. In a folder without .mimry/ each hook exits silently before the CLI loads, but it still starts Python, which adds about 0.15 to 0.2 seconds to each prompt and to each Bash, PowerShell, Read, Grep, Glob, Edit and Write call. Install with --project to keep the hooks in set-up repositories only, or pass --no-hooks.
  • A project install keeps its files out of git status: Claude Code hooks go to .claude/settings.local.json, and the project skill folders join MIMRY's global ignore block (#635). mimry install --refresh and mimry upgrade move MIMRY's old entries out of .claude/settings.json and leave everything else there.

New features

  • Every Claude Code session gets MIMRY's rules and a short repo map from the index when it starts, resumes, clears or compacts: a new mimry hook-session on SessionStart (#672). The map lists the most connected classes and functions of the main folders, within 2000 characters, and is cached until the index changes.
  • Subagents get MIMRY guidance: a new mimry hook-subagent on Claude Code's SubagentStart adds a short note to use MIMRY first, plus the files at the top of a fresh context pack (#663).
  • A broad search answered by the search hook gets the top ranked files with their read-first lines next to the grep result, instead of a reminder to run mimry find (#665). The grep still runs, and exact-string, piped and named-file searches stay quiet.
  • Code that runs a stored procedure or queries a view now links to it, and mimry impact lists the files that reference a SQL object in a new referencing_files group (#650).
  • Package references and version pins are symbols, so mimry symbol System.Text.Json finds the project file that pins it: MSBuild PackageReference and PackageVersion, package.json, pyproject.toml and requirements*.txt (#649).
  • mimry stats says when MIMRY has no record of agent sessions in a repository or session, instead of reporting 0 tokens saved (#666). --json adds tracked, with null totals for a scope that is not tracked.

Security

  • Secrets passed to a call on a credential assignment, such as os.environ.get("DB_PASS", "..."), typed fields, values after an Authorization placeholder, curl -u and Markdown code spans are redacted (#639). Comments and docstrings in code files are judged as prose, so a commented-out password is redacted too.
  • pypdf is now required at 6.21.4 or later, which fixes seven published pypdf advisories (#660).

Fixes

  • Documentation placeholders such as api_key: <your-api-key>, token = get_token(), masked values and password prompts no longer keep a guide out of the index (#634, #666, thanks @abhinavrastogi2627).
  • init --no-global-gitignore is remembered in .mimry/config.toml, so later runs and mimry doctor no longer undo it, and updating Git's global ignore keeps the file's bytes and line endings (#631, thanks @abhinavrastogi2627).
  • A deleted index generation, or one Windows is still deleting, rebuilds on the next query instead of being reported as damaged (#632, thanks @abhinavrastogi2627, and #638).
  • The hooks exit quietly on any payload, including deeply nested JSON (#666).
  • mimry install --refresh warns when it cannot read a settings file instead of skipping the hooks silently (#666).
  • Prose such as FROM a table no longer links to a SQL table named a; a schema-qualified dbo.a still links (#666).
  • mimry on its own prints a short guide to the main commands instead of an error that lists every command, and mimry -h no longer lists the internal hook commands (#672).
  • mimry doctor and mimry embed status suggest the embeddings install command that matches how MIMRY was installed: uv tool, pipx or pip (#653).

0.8.4

Upgrade notes

  • Upgrade with mimry upgrade now if your repository has SQL files. A SQL file with many comment lines and Windows line endings could stall mimry index for many minutes; it is read at once now (#620).
  • mimry upgrade, mimry install --refresh and mimry init add .mimry/ to Git's global ignore when it is missing, so MIMRY's folder stays out of every repository's changes (#615). Set MIMRY_NO_GLOBAL_GITIGNORE=1, or pass init --no-global-gitignore, to keep your global ignore untouched. mimry doctor reports it as git.global_ignore.

Security

  • 0.8.3 redacted a Basic user:password only up to a comma, space, $, quote or bracket in the password, so the rest of the password stayed visible (#617). The whole value, and a secret inside an encoder call such as base64(...), is now redacted to the end of its header.
  • A token after a scheme that looks like a file path, such as Bearer abc/def.key, is redacted; a file name is kept only when no scheme comes before it (#627).

Fixes

  • Code next to authorization is kept: typed fields such as token: String, header lookups such as req.get("Authorization") or headers[key], {token} placeholders and file paths (#613). Spaced and literal doc placeholders such as base64( user:pass ) or "Basic user:pass" no longer keep a Markdown file out of the index (#617).
  • Indexing a SQL file no longer stalls on a run of comment lines with Windows line endings before a statement that is not CREATE (#620).

0.8.3

Security

  • 0.8.2 let some real Authorization values through as doc placeholders: upper-case and upper-snake tokens such as Bearer ZK_9F2K_77QX, names starting my- or your-, and values followed by ( (#604). Only names built from doc words, such as YOUR_TOKEN or your-token-here, are placeholders now.
  • A secret inside an encoder call, such as Basic base64(admin:S3cret42), is redacted with the whole call, while base64(user:pass) is kept (#611).
  • Authorization values are redacted when passed as call arguments, such as res.setHeader("Authorization", "Bearer ..."), headers.append or C# Headers.Add, after a scheme MIMRY does not know, such as Authorization: Acme ..., and as id:signature pairs such as AWS and SharedKey use (#601). A Basic user:password whose password holds punctuation is redacted (#611); 0.8.3 still left the part after a comma, space, quote or bracket visible, which 0.8.4 fixes.
  • Table cells and labels written with bold or backticks, such as | **Authorization** | Bearer ... |, are read as headers too (#604).

Fixes

  • Code next to authorization is no longer redacted: type annotations such as def check(authorization: str), header lookups such as request.headers["Authorization"], and dotted attributes in source files (#605). Bracketed placeholders such as Authorization: Bearer <token>, [TOKEN] or ${API_TOKEN} no longer keep a Markdown file out of the index (#604).
  • The hook reads heredocs whose delimiter is quoted or escaped, such as <<\EOF or <<E"OF", skips PowerShell here-strings, ends a line at a closing backtick, and leaves option values such as --limit 5 out of the logged query (#606, #610).

0.8.2

Security

  • pypdf is now required at 6.21.1 or later, which fixes five published pypdf advisories (#588).

Fixes

  • Authorization headers written in Markdown or prose are redacted: list items, quotes, bold labels, sentences and table cells such as | Authorization | Bearer <token> | (#591, #599). A credential written without a scheme is redacted when it ends its line, a comma follows it, or a period follows it at the end of the line.
  • Prose after Bearer, Basic or Token, such as Bearer is required, is kept again; 0.8.1 redacted it (#591). Doc placeholders such as Bearer YOUR_TOKEN, Token abc. or Basic base64(user:pass) no longer keep a Markdown file out of the index (#599).
  • mimry history and mimry_history answer from stored commits when the cache is read-only, and mimry_feedback there returns a cache_read_only error instead of failing (#583, thanks @abhinavrastogi2627). MIMRY only checks whether the cache is writable when a refresh is needed or a write was refused, and a current index is not called stale (#589, #598).
  • The session log keeps every word of a query that mentions a subcommand name, such as mimry find the mimry init command, and the hook stays fast on very long commands (#590).
  • The hook reads PowerShell backtick line continuations and skips heredoc bodies, so text inside a heredoc is not taken for a search or a query (#597).

0.8.1

Fixes

  • Authorization values after Bearer, Basic or Token are redacted again whatever follows them, as before 0.8.0, and a plain-word value is caught when its header ends at a comma, a semicolon or a line break, escaped or not (#578). Common words such as disabled, needed or default after Authorization: are no longer redacted.
  • A query that mentions mimry without quotes, such as mimry find how does mimry work, keeps all its words in the session log again (#577).
  • A hook that reads the session key while another hook is writing it waits and reads again, so its query is still tagged (#579).
  • On Windows, a hook no longer drops a file event when the session log's lock is busy or still being removed, and a folder MIMRY cannot write fails at once instead of waiting (#581).

0.8.0

Upgrade notes

  • Upgrade with mimry upgrade. The first query after upgrading reparses every indexed file, and the upgrade then refreshes the installed skills and hooks.
  • mimry install --refresh adds Claude Code's PowerShell tool to the search hook, so mimry queries run there are tagged with their session (#556). mimry doctor reports a hook table from 0.7.0 as incomplete and prints the repair command.
  • Installing from a uv throwaway env (uvx mimry install) now stops before writing anything when the mimry on PATH is another version, and says how to upgrade it (#545). mimry install --refresh and the refresh after mimry upgrade check the same (#565). Hooks used to run the older copy and fail.
  • The session command log keeps a keyed hash of each query, so a short secret typed as a query cannot be guessed from the file (#546). Rows written by 0.7.0 are ignored.

New features

  • An opt-in prompt hook names the files to read first for each prompt in Claude Code: mimry install --hooks --prompt-hook (#516). Session feedback is now recorded when the agent stops, not only after the next command.
  • MCP queries are tagged with their session through the PreToolUse hook, so session learning covers agents that use the MCP tools (#518).
  • Read commands work in a sandbox that cannot write MIMRY's cache: they answer from the index that exists and say once that it cannot be updated, with the folder to point MIMRY_CACHE_HOME at (#527, #549, #557). The MCP tools, mimry status and mimry doctor say so too, rather than failing or promising a refresh (#568). A cache inside the project is no longer indexed as part of the project.

Fixes

  • Redaction catches more secrets: in comments, URLs, f-strings and subscript assignments (#512), in tuple assignments and auth headers inside strings (#528), and in chained and one-element assignments, split string literals and Authorization values with any scheme (#553). It also catches chained and backslash-continued values that run across lines (#567), and Authorization values of plain words, while prose such as basic auth is required is kept (#571).
  • A multi-line secret is redacted line by line, so the file keeps its shape and is still indexed, and a docstring that mentions a password keeps its prose (#536). A triple-quoted secret that never closes keeps its line breaks too (#570).
  • Secret scans, the prompt hook and feedback redaction run in time proportional to the input on worst-case text, such as long runs of escaped quotes or dotted names, or a command with many mimry lines (#529, #537, #551, #566).
  • The search hook stays quiet on pipes, word regexes and searches of named files (#517).
  • Session learning keeps the session's last query through Stop, matches queries exactly, and counts edits made in the same clock tick as Stop on Windows (#531, #534). A query that starts with a secret label keeps its session tag (#556), and queries are tagged on drives without hard links, such as FAT or exFAT (#569).
  • mimry install registers its own mimry-mcp rather than the first one on PATH, keeps --prompt-hook in repair hints, and prints tidier output (#508, #532, #510).
  • MIMRY's own installed skill files are no longer indexed (#507).
  • A brief file error on a busy Windows machine is reported as itself, not as a missing index pointer (#538).

Other changes

  • An accessibility statement, and issue forms for a file MIMRY missed and for feature requests (#552).

0.7.0

Upgrade notes

  • Upgrade with mimry upgrade. The first query after upgrading (find, preflight and the like) reparses every indexed file, and the upgrade then refreshes the installed skills and hooks.
  • mimry install --refresh rewrites the installed skills, and adds the new hook entries to a project that already has MIMRY's hooks. In Claude Code those are an outline for a large file that the agent reads whole, and a restore of the working set after a context compaction.
  • Embeddings are optional and off until you install them: uv tool install "mimry[embed]", then mimry embed download, which fetches a 64 MB model once.
  • A code file with secret-looking lines is now indexed with those lines redacted, instead of being left out (#470). Files that are not code, sensitive-named files, and code that cannot be redacted cleanly are still left out.

New features

  • A long find query, of 20 or more distinct words, also matches on meaning when the optional embedding model is installed (#488).
  • Session learning records the files read and edited after a find as feedback (#486). MIMRY_SESSION_FEEDBACK=0 turns it off.
  • In Claude Code, the files the agent was working on are restored after a context compaction (#487).
  • mimry stats shows the tokens MIMRY saved, estimated, for this session and all time (#489).
  • In Claude Code, a large file that the agent reads whole is outlined instead (#484).
  • A learned reranker orders the top 50 results (#468).
  • Files that define the identifiers a task names are lifted in the ranking (#467), and a keyword ranking of files is fused into find for long queries (#472). Both change the order of results, and the ranking is more accurate.

Fixes

  • Redaction no longer reads a == comparison as a secret assignment, keeps the quotes, escapes and brackets of a redacted value, and keeps := intact (#474, #485, #498). A quoted value that a line break splits is redacted line by line, more comparison and => forms are caught, and redacted code stays parseable (#502).
  • The search hook is quiet on a search of a single file, and it is capped per session (#480).
  • A tsconfig whose extends bases share a base reads that shared base once, so indexing does not repeat the work (#466).
  • An unreadable embedding title cache is rebuilt, and embed status, status and doctor report a broken model (#496). A broken embedding model is skipped instead of being loaded on every search (#502). Recording session feedback no longer makes a finished command wait for a busy index, and keeps the files the agent opened meanwhile (#496).
  • The README and site figures match the benchmarks, and the line-range saving is described first (#481, #482).

0.6.4

Upgrade notes

  • Upgrade with mimry upgrade. The first command after upgrading reparses every file, so tsconfig bases named .jsonc show up without a full reindex.

Fixes

  • On Windows and macOS, a file renamed only in case (src/a.py to Src/A.py) is marked as uncommitted (#460).
  • A tsconfig extends base named .jsonc, or a middle base that only extends another, continues the chain to the path aliases (#461).
  • A very deep extends chain no longer stops mimry index; links past 32 are not followed (#463).

0.6.3

Upgrade notes

  • Upgrade with mimry upgrade. The first command after upgrading reparses every file, so tsconfig bases named in an extends array show up without a full reindex.

New features

  • preflight, route and explain mark a file with uncommitted changes, including untracked files and files in a nested repo or submodule: (uncommitted) in the Start with list and an Uncommitted changes: yes line in the context pack, so an agent does not take work in progress as shipped (#454, #456).

Fixes

  • A tsconfig.json whose extends is an array follows each base in order, and a base with any .json name, such as ./base.json, resolves its path aliases (#453).

0.6.2

Upgrade notes

  • Upgrade with mimry upgrade. The first command after upgrading reparses every file, so MSBuild, plist and XAML files and the Groovy fixes show up without a full reindex.
  • mimry path --json lists each other candidate in ambiguous once, as {label, path, line}, instead of a list of paths (#435).

New features

  • .csproj, .fsproj, .vbproj, .props, .targets, .plist, .xaml, .resx and .nuspec files are read as XML: MSBuild properties and targets, plist keys and XAML names are symbols, and a ProjectReference links to its project (#436).
  • A tsconfig.json that extends a base such as tsconfig.base.json resolves the base's path aliases, as in Nx and other monorepos (#446).

Fixes

  • A malformed JSON file nested very deeply no longer stops mimry index, a JSON file with a byte order mark and comments keeps its keys, and tsconfig.json with comments no longer shows a parse error (#432).
  • UTF-16 source files such as plists, project files, Python and Markdown parse, a byte order mark no longer hides a Markdown heading on the first line, and SVG icons without text are no longer listed as empty documents (#444, #451).
  • A Groovy class right after a wildcard import is found again, methods with nested generic bounds such as <T extends Comparable<T>> are indexed, and import static a.b.C.* links to the class file while import a.b.C.* links to the package, or to the class file when a.b.C is a class (#433, #445, #451).
  • A Groovy line with many generic types no longer makes indexing hang (#445).
  • Feedback records the files a task changed as missed when its context pack suggested nothing, and a folder passed as --context is not taken for a pack (#434, #447).
  • mimry path prefers a name with the exact case, names the symbol and line it used, and says when no path exists within 6 steps (#435, #447).
  • A small .map file is no longer treated as a minified bundle (#435).
  • After an upgrade, index says which files it reparsed instead of "No file contents changed", and feedback stats says "No commits yet." in a new repository (#447).
  • The README gives the data file symbol limit as 256 KiB, the site shows the current mimry path output, and the README says how to install without uv and which folders are always skipped (#436, #448, #451).

0.6.1

Upgrade notes

  • Upgrade with mimry upgrade. The first command after upgrading reparses every file, so Groovy files, Haskell class methods and Helm template keys show up without a full reindex.
  • mimry path prints the path as a chain, each relation between the two nodes it joins; and --json adds an ambiguous field when a name matches several symbols (#417).
  • A data file with a syntax error ({"a": 1,) now counts as "could not be parsed" and keeps no keys, like broken code; JSON with comments and trailing commas, as editor settings allow, still reads fine (#420).

New features

  • MIMRY parses 41 languages: Groovy and Gradle join, with classes, methods, Spock feature methods, Gradle tasks, imports that link to the exact file, and call and inheritance edges (#406, #415, #422).
  • Haskell class method signatures and defaults are symbols, and Haskell and Fortran module imports link to the exact file (#401, #403, #416).
  • index, preflight and status -v say which documents are searchable by name only (damaged, encrypted, empty, compressed too much) and which were cut at a cap (#400, #420).
  • A query that is a file's name ranks that file first, and find says when a file named in the query is not indexed (#404, #418).
  • mimry path says when a name matched several symbols and which one it used (#417).

Fixes

  • Minified bundles found by their content no longer crowd out source files, while hand-written files with long lines keep their ranking (#405, #419).
  • A file name's extension alone no longer counts as a name match, so find "session.ts" stops listing every .ts file (#418).
  • Helm templates keep their keys after the first template action (#402).
  • YAML nested too deeply to read gets its own warning instead of "Fix the syntax" (#420).
  • Size limits print as 20 MiB, status -v lines its document lists up, --language scope lines show the display name, and find -v no longer repeats config evidence (#421).

Benchmarks

  • The token benchmark README records a trained embedding model for --semantic: it located fewer tasks in every mode, so --semantic keeps its hashed vectors (#423).

0.6.0

Upgrade notes

  • Upgrade with mimry upgrade; it refreshes the skills, always-on rules and hooks you installed. The first command after upgrading reparses every file, so the new languages and whole-document search take effect, and index says how many unchanged files it reparsed (#397).
  • .html, .css and .scss files no longer take the boost source files get for edit tasks (#365).
  • .pp files are Pascal and .m files Objective-C only when their content says so; Puppet and MATLAB files become searchable text (#396).
  • A JS/TS const holding a function or arrow is now kind function (or component with JSX), and calls inside a local's initializer count for the enclosing function (#360).
  • Fortran and Pascal program blocks are kind program, not module (#397).

New features

  • MIMRY parses 40 languages, up from 21: Elixir, Haskell, OCaml, R, Julia, Solidity, Fortran, Pascal, CSS, SCSS, HTML, Vue, Svelte, YAML, JSON, TOML, XML, Terraform and SQL (#355, #357, #363, #366, #372). Data files give their keys as dotted names, Kubernetes documents give Deployment/web, Terraform blocks keep their reference names and kinds, and secret values are never indexed (#392).
  • PDF, Word, Excel and PowerPoint files are searched in full, and a result names the page, section, sheet or slide to read (#367, #390).
  • mimry install covers Windsurf, Cline, Roo Code, Zed, JetBrains Junie and Grok Build (#369, #395).
  • A Docker image, ghcr.io/jbacalso24/mimry, for amd64 and arm64 with the grammars baked in (#371).
  • mimry preflight --semantic and mimry_preflight(semantic=true) add semantic matches to the ranking; it stays opt-in, since it located the same tasks on the token benchmark (#375, #376).
  • --language accepts the names users see, such as Terraform, TypeScript or Objective-C (#397).

Fixes

  • A YAML file nested past 200 levels no longer crashes index; it is reported as unparsed and stays searchable (#391).
  • The secret scan runs in linear time over long runs of spaces; a deeply indented file indexed in 40 seconds now takes under 2 (#394).
  • A cache folder MIMRY refuses to write to gets one plain error from every command and MCP tool (#373).
  • uninstall keeps a skill file another installed platform still uses, Zed with a commented settings.json says to paste the entry by hand, and uninstall with nothing installed prints only that (#395, #397).
  • callers and callees rows say line 4 instead of a bare :4 (#397).
  • llms.txt lists all 9 default MCP tools (#389).

Other changes

  • The README and site lead with what MIMRY gets you (#368).
  • The adapter modules live in one adapters package, and CI runs the tests in parallel (#354, #358, #359, #362).

0.5.0

Upgrade notes

  • Upgrade with mimry upgrade; it refreshes the skills, always-on rules and hooks you installed. The first command after upgrading rebuilds the index, and it now indexes source files the secret scanner used to skip by mistake.
  • Install and init now add only .mimry/ to ignore files, and init removes the mimry-out/ line older versions added (#316, #351).
  • mimry_why and mimry_explain return fields instead of CLI text, so MCP clients that read their stdout should read the fields (#348).
  • The post-commit hook records feedback for the last context pack only when the commit is about that task (#340).

New features

  • mimry history searches commit messages and changed files, preflight names the commits behind a removed or renamed path, and the mimry_history MCP tool does the same (#315, #347).
  • Suggested checks come from Makefile and justfile targets, tox and nox sessions, CI workflow steps and the scripts they run; nested packages run from their own folder with their own package manager, and checks guessed from tool names appear only when the project names none (#313, #323, #350).
  • find matches Chinese, Japanese and Korean words in content and file names (#322).
  • callers, callees and impact pick one of two same-named symbols with --line N or --file path:N (#319).
  • Preflight, the context pack, route and explain list the files to edit first, and tests first only when the task is about tests (#317).
  • explain ranks its key symbols by the task and shows how they connect (#320, #352).
  • Plan commands accept the plan name shown in plan list (#349).

Fixes

  • The secret scanner no longer skips source files over a JSON array, a typeof check or a short number under a credential-looking name, and status -v names the files it skips (#337).
  • route matches role words whole, so "helper" no longer routes to docs or "latest" to QA (#338).
  • The grep hook answers with a symbol only when the whole pattern names it (#339).
  • impact leaves local variables out of its targets and credits their calls to the enclosing function (#342).
  • doctor reports a missing grammar cache folder the same way index does, without a reinstall hint (#343).
  • bench passes over release and bump commits that start with an emoji, and pre-commit-only changes (#318, #346).
  • install --status keeps --no-mcp in its repair hint, --file misses name the file and suggest close paths, and a failed why lookup prints only the miss (#349).
  • Help for every option, mimry update --help, local times in status -v and the pack, and readable small size limits (#321, #351).

Other changes

  • The README walks through one task before and after (#314).
  • The AUR build fetches and verifies the sdist once (#311), and the CI job limit is 45 minutes (#345).

0.4.1

Upgrade notes

  • Upgrade with mimry upgrade; it refreshes the skills, always-on rules and hooks you installed. The first command after upgrading rebuilds the index, so imports resolve more precisely and the lines to read first are recomputed.
  • Commands now exit 1 when they find nothing and 2 on a blank task or query, instead of 0 (#298). Scripts that treated any non-zero exit as an error should allow exit 1 for find, related and semantic.
  • mimry_impact is now one of the default MCP tools (#300).

Fixes

  • Every query command indexes or builds the graph first when it is missing, instead of answering from nothing (#295).
  • Output piped to an agent or a file is UTF-8 on every platform (#296).
  • A Python standard-library import no longer links to a same-named module inside a package, and TypeScript from '.' and from '..' imports resolve to the folder's index file (#297).
  • A missing name or file, or a search with no matches, exits 1; a blank task or query exits 2 and writes nothing; MCP marks real failures with isError, and path no longer guesses for a file that is not indexed (#298).
  • The hook finds names that start with an underscore, stays quiet for globs like **/*.py, and never answers with a Markdown heading or SQL table (#299).
  • The installed skill names only MCP tools the server lists, install --status respects --no-mcp, and uninstall removes only what MIMRY created and says what it did; uninstall --git-ignore removes MIMRY's global ignore lines (#300).
  • Suggested verification commands come only from the project's own manifests, never from install snippets in README files (#303).
  • Files over the 1 MB limit are counted and named, grammar failures show their real cause, and status totals stay right after a rename (#304).
  • impact --changed includes the tests that reach a changed file when the changed function has no callers (#305).
  • Feedback notes redact AWS, GitHub and Slack tokens and private key headers without a key= label (#306).
  • The lines to read first cover the function that holds the code, instead of a stray piece next to it (#308).

Other changes

  • The MIMRY logo in the README, on the website and as the social preview (#307).

0.4.0

Upgrade notes

  • Upgrade with mimry upgrade, then run mimry install --refresh once to bring the skills, always-on rules and hooks you installed up to date (#267). From the next release on, mimry upgrade does this for you.
  • The first command after upgrading rebuilds the index, so calls MIMRY resolved through an exact import are no longer marked inferred.

New features

  • mimry impact <symbol|file> lists callers by depth, the files that import it and the tests to run; --changed does the same for your uncommitted changes (#263).
  • mimry bench measures how well MIMRY finds the files your own merged pull requests changed, locally and without model calls (#264).
  • When an agent greps for a function or class by name, the hook answers with its definition and callers (#260).
  • path, callers and callees mark calls MIMRY could not resolve exactly as (inferred), and mimry path --json reports each step's confidence (#261).
  • Feedback counts changed files more than opened ones, and past tasks more like the current one count more (#259).
  • A troubleshooting guide, linked from mimry doctor --help (#265).

Fixes

  • explain, path and why update a stale index before answering, and no command tells you to reindex by hand any more (#256).
  • find ranks vendored, minified and repository-template files lower (#257).
  • The server.json description fits the MCP Registry limit (#254).

Other changes

  • A feedback replay benchmark (#258) and partner-file diagnostics in the localization benchmark (#255).
  • Shared command logic moved out of commands.py (#262).

0.3.4

Upgrade notes

  • Upgrade with mimry upgrade. The first command after upgrading rebuilds the index, so Python from pkg import module imports and tsconfig path aliases link files.

New features

  • mimry status lists indexed files per language (#231).

Fixes

  • find, preflight and the context pack rank files better for long task texts such as issue bodies (#239, #240, #241). Intent is read from the task's first line, so a test word in a traceback no longer pulls test files up. A file the task names by path, module or unique file name ranks first. On SWE-bench Verified, the right file is in the top 10 for 74.6% of tasks, up from 66.2%.
  • mimry callers finds calls through from pkg import module and through tsconfig and jsconfig paths aliases, and no longer says a function is unused when it found no callers (#243).
  • Feedback keeps ordinary words after "pass" and "password" ("pass the duplicate flag", "password reset") instead of redacting them (#244).
  • Feedback with a context pack records the changed files the pack did not suggest as missed, so mimry feedback stats counts misses (#245).

Other changes

  • A SWE-bench Verified file localization benchmark, with a separate dev set for tuning (#226, #230, #233, #234, #235, #242).
  • The token benchmark reports recall@k and missed files (#229) and runs in a GitHub workflow (#246).
  • The README introduces MIMRY as a local context engine for AI coding agents (#225, #214).
  • CI jobs may run for 30 minutes (#228).

0.3.3

Upgrade notes

  • Upgrade with mimry upgrade. The first command after upgrading rebuilds the index, and the rebuild keeps nothing from the old one, so text the new privacy rules redact is gone from the cache (#206).
  • Run mimry install --platform <name> again to add .mimry/ to Git's global ignore file (#213). After that, mimry init no longer edits each repo's .gitignore. --no-global-gitignore skips it.
  • The privacy rules are stricter, so a few files may drop out of the index (#204). .conf, .properties, .tfvars, Dockerfile and other text formats MIMRY doesn't parse are now checked for credentials, and *.env files are skipped by name. Files whose only "secret" was a URL-valued setting such as TOKEN_URL are indexed again.

New features

  • mimry callers and mimry callees show who calls a function and what it calls (#163).
  • MIMRY parses Bash, PowerShell, Lua, Perl (#157), Dart, Objective-C and Zig (#166).
  • mimry install adds .mimry/ and mimry-out/ to Git's global ignore file, once per machine (#213).
  • The site has a changelog page (#158), and the site and README show contributors (#159).

Fixes

  • Commands run from a subfolder of an indexed repo answer from the repo's index, and create no .mimry/ there (#212).
  • A repo under a non-ASCII path works; --root pointing at a file or a drive root fails with one line; find on a repo that was never indexed indexes it first; why says when a path is missing or outside the repo (#212).
  • Credentials in the middle of a sentence are redacted, in queries as well as in files (#204). A placeholder after a credential label is no longer treated as a secret (#165).
  • Calls resolve by scope, through package re-exports, and within Swift, Kotlin and C# modules (#208).
  • Go, Rust and PHP imports link files (#201), and absolute Python imports resolve in src/ layout projects (#164).
  • Zig function bodies, Rust Type::method() calls and Objective-C .h headers are recorded (#207).
  • Python files that start with a BOM, subclasses of Base[T], and files with a syntax error keep their symbols; index and status name the files that could not be parsed (#211).
  • Indexing a minified JavaScript bundle takes seconds, not minutes (#211).
  • mimry symbol <file.py> matches the file name (#187, thanks @wzd8888).
  • A case-only rename is picked up, and --file accepts either slash (#210).
  • The MCP server stays up when a tool is asked about a folder that isn't set up, and rejects an invalid outcome with a clear error (#203).
  • install never overwrites a settings file it can't parse, an MCP config it can't write is a warning, and install-only warnings no longer mark a repo unhealthy in doctor (#205).
  • MIMRY says when tree-sitter grammars can't be downloaded, and tries only once (#155).
  • A copied .mimry folder is explained the same way in every command (#160).
  • The AUR build waits for the new sdist on PyPI (#154).

0.3.2

Upgrade notes

  • From 0.3.1, upgrade the old way one last time: uv tool upgrade mimry, pipx upgrade mimry or pip install --upgrade mimry. On Windows, close the agents that use MIMRY first. From 0.3.2 on, mimry upgrade does it for you.

New features

  • mimry upgrade (or mimry update) upgrades MIMRY through whatever installed it: uv, pipx or pip (#147). On Windows it refuses while MIMRY servers are running, and --stop-servers stops them first. After upgrading it tells you how to switch each open agent to the new version.
  • MIMRY installs as a Claude Code plugin: /plugin marketplace add jbacalso24/mimry, then /plugin install mimry@mimry (#148). It brings the skill, the hooks and the MCP server, and only needs uv.
  • npx skills add jbacalso24/mimry installs the MIMRY skill for other agents (#149).
  • The README lists the supported languages and the ones planned next (#151).

0.3.1

Upgrade notes

  • The MCP server now lists 8 core tools by default (#125); set MIMRY_MCP_TOOLS=all to get every tool, or a comma-separated list.
  • mimry install --status now exits 0, 1 or 2 (#128) and reports a skill written by an older MIMRY as out of date (#140). After upgrading, run mimry install --platform <name> again (with the same --project, --always-on, --hooks flags) to refresh the skill.
  • On Windows, close the agents that use MIMRY before upgrading. Their running MCP servers lock MIMRY's files and the upgrade stops halfway. uv tool install --force mimry repairs a failed upgrade (#139).

New features

  • Agent hooks now refresh the index when a session starts and after each edit (#127).
  • find, symbol, related and semantic reindex first when files changed, the same as preflight, in the CLI and MCP. When git reports the same working tree as the last index, the check is skipped, so an unchanged repository answers without walking the tree (#98).
  • mimry find is scoped to the current directory inside an indexed repository. Use --everywhere or an explicit --path-glob to search outside that directory (#92).
  • mimry doctor reports the install source, the running MCP servers and the agent registrations (#126).
  • mimry install supports Cursor and GitHub Copilot in VS Code (#122, #124).
  • MIMRY records feedback from each commit automatically (#119).
  • mimry feedback stats shows usage and ranking quality (#121).
  • Each task now gets its own context pack (#116).
  • Choose which files to index with a mimry.toml, and map file extensions to languages in it (#104, #109).
  • The search hook names mimry find on broad searches, so agents know to use it (#106).
  • A new git worktree is set up on its first lookup, with its own index (#101).
  • Results tell agents which lines of each match to read first (#27).
  • Search can return short source snippets in the CLI and MCP, on request (#23).
  • Search can filter by language and path, and pages of results stay stable across requests (#24).
  • Add mimry doctor, which checks the installation, parsers and index and says how to fix what it finds, without changing anything (#25).
  • mimry preflight shows elapsed time while it works (#96).
  • When an agent connects to the MCP server, it is told when to use MIMRY, and the server reports its version (#29).

Fixes

  • The MCP server answers its first request in about 0.1 seconds, instead of 2 to 6 seconds while a large library loads at startup (#113).
  • Keep MIMRY running on an incomplete install, and say how to repair it (#19).
  • A skill written by an older MIMRY is now reported as out of date, not up to date (#140).
  • Nested Git worktrees are excluded from their parent's index, so their files are not duplicated, and a folder named worktrees is no longer mistaken for a linked worktree (#105, #117).
  • Untracked files that .gitignore names are no longer indexed (#30).
  • An index from another cache location is explained instead of crashing, and preflight rebuilds it (#28).
  • An index from an older MIMRY is rebuilt automatically (#88).
  • Feedback keeps the leading dot in file paths (#120).
  • Every graph match gets a content score before results are cut to the limit (#26).
  • Installing the skill on Windows waits out file locks instead of failing (#81).
  • The global skill installs under the directory named by CODEX_HOME (#76).
  • Installs use the MCP library version that CI tests, and the next major version cannot arrive untested (#77).

Documentation

  • The README is reworked around what MIMRY does for coding agents, and says that each agent starts its own MCP server (#97, #86).
  • The README shows agent support status, and each agent's setup steps are on a docs page (#108, #91).
  • The project website has searchable setup guides (#22).
  • Published token benchmark results include a holdout2 run of main, and holdout3 results for whole-file and line-range reads (#100, #94).
  • The README welcomes contributions, explains how to open a pull request, and says MIMRY is free to use (#20).

0.3.0

New features

  • Stale indexes are refreshed before a task runs (#18).
  • Markdown headings and JavaScript modules are indexed (#18).
  • Code is found by what is inside files, and agents get a shorter context pack (#15).
  • PDF and SVG files are indexed, and six more languages are supported (#14).
  • mimry export writes the graph as an interactive HTML page, GraphML, a Cypher script or a notes vault (#14).
  • Git hooks refresh the index in the background after commits, merges, checkouts and rewrites, and mimry git-hooks manages them (#14).
  • mimry install registers the MCP server with the agent at user scope, so one registration serves every project (#16).

Fixes

  • Installing no longer hangs on Windows (#12).
  • Content search reads each file's whole text, not only its first 40 lines (#15).
  • Source files are no longer refused because one value only looks like a credential (#16), and source that only names a credential is kept (#11).
  • Source files rank above tests (#11).

Documentation

  • MIMRY has a security policy and the full MIT license text (#9).

0.2.1

Documentation

  • MIMRY has a project website and a listing in the MCP Registry (#6).
  • The README is readable by AI agents, and the project adds an llms.txt file (#5).

0.2.0

New features

  • Indexing is much faster: a full index of a large sample repository dropped from about 100 seconds to 16 seconds. An unchanged reindex takes about 2.7 seconds (#4).
  • Every command shares one readable output style, with a headline, indented details and aligned tables (#4).

0.1.0

Determinism

  • Add the narrow mimry plan recursive decomposer: deterministic leaf-only splits, canonical repo-local JSON storage, terminal/JSON/Markdown projections, structural validation, semantic SHA-256, deterministic inventory, and read-only MCP parity. It stores explicit tree structure only and does not execute, schedule, refine, or track work.
  • Establish an explicit determinism contract splitting canonical semantic state (file/symbol/edge/community/chunk identities, parser facts, ranking and tie-breaking, context-pack evidence ordering) from the operational envelope (timestamps, generation UUIDs, absolute paths, mtimes, feedback event IDs, SQLite byte layout). Documented in docs/determinism.md.
  • Derive canonical IDs from the schema version and the repository-relative path instead of the absolute checkout path. Two clones of the same repository at different locations now produce identical file, symbol, edge, community, and semantic-chunk identities. Semantic chunk_id no longer carries the machine-local root UUID either.
  • Refuse index generations built under the old path-dependent identity scheme (GENERATION_SCHEMA_VERSION 1-3) with an actionable rebuild message, rather than partially reusing identities nothing else in the build produces.
  • Sort filesystem traversal and every persisted collection by a single normalization rule (canonical_rel_path: POSIX separators, NFC, case preserved, codepoint-ordered and therefore locale-independent), and reject same-root native paths that collide after NFC normalization before publication.
  • Reject duplicate node, symbol, and edge IDs whose records conflict, instead of resolving them by last write. Records identical in every field are deduplicated under a documented policy.
  • Give edges a total order over every canonical field and a documented EXTRACTED-over-INFERRED conflict rule, applied at every graph assembly site rather than only inside build_graph.
  • Give symbol resolution a total, documented selection order (kind priority, relative path, line start, line end, symbol ID). Genuinely ambiguous inheritance and import targets are still declined rather than resolved by picking the first candidate.
  • Add deterministic secondary ordering before every SQL cutoff. A tied BM25 rank no longer lets physical row order decide which rows survive LIMIT, and semantic chunk reads are ordered so explanation previews and reason labels are stable.
  • Acquire each file through bounded matching reads, derive its hash, size, mtime, and parser input from that snapshot, then verify post-acquisition freshness after adapters finish. A file changed mid-index is marked unindexable; mixed-version evidence is never persisted.
  • Add mimry digest, which hashes normalized semantic state rather than raw SQLite bytes or the generation manifest.
  • Add a real end-to-end determinism gate over a frozen multi-language fixture (scripts/determinism_matrix.py) covering scanner, Python/TS/Java/PHP/Go/SQL/Markdown parsing, graph and communities, SQLite/FTS, local semantic retrieval, and context-pack generation, across creation order, checkout root, PYTHONHASHSEED, clean-cache repeats, and locale/timezone. CI runs it on Linux, macOS, and Windows and a dedicated job compares every platform's digest against a committed golden value.
  • Convert the __main__ self-checks in the graph core into committed pytest tests.

Graph engine

  • Replace the external graph dependency with MIMRY's own graph engine in src/mimry/core/. Indexing parses each file once and derives defines, imports, and calls edges from what it already read, instead of copying the tree to a temporary directory and re-parsing it in a subprocess.
  • Resolve import statements and call sites to real files and symbols. MIMRY previously collected import strings but never resolved them, so its graph had no file-to-file edges at all. Unresolvable or ambiguous references are dropped rather than guessed.
  • Extend language coverage to Go, Rust, and C# alongside Python, JavaScript, TypeScript, JSX, and TSX, at no new dependency cost.
  • Rank by following relationships, not only by matching text: query relevance now propagates two hops along edges.
  • Stop demoting full-text evidence whenever the graph has an opinion. A content-only result was previously multiplied by 0.75 if the graph returned anything, and a file found by both signals had its content score discarded; corroborating evidence now adds.
  • Measured against the recorded baselines on nine cases: ndcg@5 0.6300 -> 0.6907, recall@5 0.8542 -> 0.8958, primary-hit@3 0.5000 -> 0.7500, decoy rate 0.1351 -> 0.1081, and time to a queryable indexed graph roughly halved.
  • Rename public graph terminology for the native engine: --skip-graph, mimry status, the MCP skip_graph argument, and the status payload key graph are canonical. A compatibility window keeps --skip-graphify, mimry graphify status, MCP skip_graphify, and the graphify status key functional for existing clients. The graph engine identifier is mimry-core.

Cross-platform

  • Fix Windows portability defects in home-directory sandboxing (including APPDATA, LOCALAPPDATA, and XDG_DATA_HOME isolation), descriptor syncing, timestamp identity checks, capability-guarded file operations, and cp1252-safe CLI output. UTF-8 artifacts preserve original Unicode.
  • Test Windows, Linux, and macOS across Python 3.11-3.13 in the committed CI matrix.

Packaging and support floor

  • Restrict source distributions to reviewed source, tests, documentation, and license files; exclude vendored, generated, cache, build, and local-work bulk.
  • Add artifact contract tests and Windows/Linux/macOS CI for Python 3.11, 3.12, and 3.13; the unlocked sdist lane invokes only extracted artifact source/scripts.
  • Package determinism_aggregate.py, support standards-valid XLSX inline strings, reject empty explicit edge IDs, and fail closed on unknown future generation schemas while preserving the old-schema upgrade classification.
  • Pin the proven tree-sitter-language-pack 1.12.2 compatibility floor after 1.13.5 broke TSX symbol extraction, and test unlocked extracted artifacts in CI.
  • Define 0.1.x as an internal checkout/direct-wheel release rather than claiming compatibility with PyPI-style indexes.