Sem

Part of the Ataraxy Labs stack: agent-native infrastructure for software development. See also: weave (entity-level git merge driver) · inspect (semantic code review) · opensessions (tmux sidebar for coding agents).

Read the manifesto: https://ataraxy-labs.com/#thesis · Essays: https://ataraxy-labs.com/blogs · LLMs: https://ataraxy-labs.com/llms.txt

[sem]

[Ataraxy-Labs%2Fsem | Trendshift]

Semantic version control built on Git.
Instead of lines changed, sem tells you what entities changed: functions, methods, classes.

Why sem? · Using sem with agents · Install · Commands · Agents (MCP) · Cloud consent · Releases

[Release] [Rust] [Tests] [License] [Languages]

sem is a semantic version control tool that works on top of Git. It parses your code with tree-sitter, extracts every function, class, and method as an entity, and diffs at the entity level instead of lines. This means you see "function blahh was modified" instead of "lines x-y changed."

It works in any Git repo with no setup.

Cloud-backed queries are opt-in per repo: logging in does not upload a repo or send a query. See the cloud consent flow for the public/private repo states, preview screen, local audit log, and forget controls.

[sem diff]

Using sem with agents

Four questions, four verbs. Every verb takes --json.

Where is it? Find a function, class or method by name:

sem find parse_config

function parse_config src/config.py:1

Who calls it (--callers), what it uses (--refs), or its code with its callers and callees (--context):

sem find parse_config --callers

function parse_config src/config.py:1  function load src/config.py:6  function test_parse_config tests/test_config.py:4

Search text, like rg:

sem grep splitlines

src/config.py:2:    pairs = [line.split("=", 1) for line in text.splitlines() if line]

What does my change touch? Everything that depends on an entity, and the tests to run:

sem impact parse_config --tests

⚡ 1 tests affected:    tests/test_config.py      function test_parse_config (L4–5)

sem impact --diff HEAD --tests answers the same for your whole uncommitted change.

Is it correct? Run the project's own compiler, type checker, linter and tests, rechecking only what the change can affect when that gives the same answer. Exit 0 means pass, 1 fail, 2 could not decide:

sem check

cmd    FAIL      full        sh (0 rechecked, 279 ms)  `python3 -m pytest -q tests` exited 1:  E       AssertionError: assert {'a': '1'} == {'a': '2'}

What should a human review? A review certificate for a commit range: what changed, which callers the change leaves behind, which tests reach it:

sem certify main..HEAD

# sem certificate 94f4bd74e26e..d73d4d092b431 files, 2 entities changed (1 added, 1 modified, 0 deleted, 0 renamed, 0 moved).- modified `parse_config` (src/config.py:1): 2 static caller(s) in 2 file(s) that this change does not modify

For agents that speak MCP, sem mcp serves the same verbs as tools (see Use with AI agents).

Install

curl -fsSL https://raw.githubusercontent.com/Ataraxy-Labs/sem/main/install.sh | sh

Or via Homebrew:

brew install sem-cli

Or via winget on Windows:

winget install AtaraxyLabs.sem

Or via Scoop on Windows:

scoop install sem

Or install the npm wrapper into node_modules:

npm install --save-dev @ataraxy-labs/sem

With Bun, trust the package so its postinstall script can download the binary:

bun add -d @ataraxy-labs/sembun pm trust @ataraxy-labs/sem

Once installed, update to the latest release any time:

sem config update

Or via cargo, from crates.io:

cargo install sem-cli

Or build the latest main from source (requires Rust):

cargo install --git https://github.com/Ataraxy-Labs/sem sem-cli

Or grab a binary from GitHub Releases.

Or run via Docker:

docker build -t sem .docker run --rm -it -u "$(id -u):$(id -g)" -v "$(pwd):/repo" sem diff

Name conflict with GNU Parallel

GNU Parallel ships a sem binary (/usr/bin/sem) as a symlink to parallel. If you have both installed, they'll collide. Run sem --version to check which one you're using. (#77)

Quick fixes:

# Option 1: alias in your shell profile (~/.bashrc, ~/.zshrc)alias sem="$HOME/.cargo/bin/sem"
# Option 2: make sure cargo bin comes first in PATHexport PATH="$HOME/.cargo/bin:$PATH"
# Option 3: if installed via Homebrewexport PATH="$(brew --prefix)/bin:$PATH"

If you installed via npm/bun, the binary lives in node_modules/.bin/sem and is invoked through npx sem or bunx sem, which avoids the conflict entirely.

Commands

Works in any Git repo. No setup required. Also works outside Git for arbitrary file comparison.

sem <verb> --help shows every flag with an example.

sem stores its SQLite entity cache outside the repository, under the OS cache directory by default. Set SEM_CACHE_DIR=/path/to/cache to override the cache root; repo-local overrides are ignored so cache files do not dirty the working tree.

sem find

Find entity definitions by name, then ask about one: who calls it, what it uses, or its code with the code around it. Answers come from an on-disk query index (index.sem, next to the entity cache); the first call in a repo builds it, later calls read it directly, with no daemon.

sem find "function diff_command"        # where it is defined ("type name" narrows by kind)sem find diff_command load_config       # several names in one callsem find diff_command --callers         # who calls it; says when the caller set may be incompletesem find diff_command --refs            # what it calls and referencessem find diff_command --context         # its code plus callers and callees, within a token budgetsem find diff_command --context --budget 4000 --hops 1sem find --in src/auth.ts               # every entity in a file or directorysem find diff_command --in src/         # only definitions under src/sem find --text "retry budget"          # entity bodies containing a string, named by entitysem find diff_command --json            # JSON on any of the above

With --context, when the target signature itself does not fit, JSON output reports target_omitted: true.

sem grep

Text search across source files, rg-compatible file:line:text output, served from the index's trigram postings when one exists.

sem grep "TODO"sem grep -i todo src/                   # case-insensitive, under src/sem grep -e foo -e bar --json           # several patterns, hits kept apart

sem impact

What breaks if an entity changes: its dependencies, its dependents (transitively), and the tests that reach it.

sem impact authenticateUser             # full impactsem impact authenticateUser --deps      # direct dependencies onlysem impact authenticateUser --dependentssem impact authenticateUser --tests     # the tests to runsem impact authenticateUser --file src/auth.ts   # disambiguate by filesem impact --diff HEAD --tests          # tests for the uncommitted changesem impact --diff main..HEAD --json     # one impact report per changed entity

With --diff and --tests in a JS/TS workspace, the answer is the module graph's affected-test selection over the changed files (sem graph --modules affected-tests); elsewhere it is the entity graph's. --no-default-excludes includes generated, fixture, vendor, benchmark and build trees.

sem check

Runs the project's compiler, type checker, linter and tests on the working tree and prints one verdict: exit 0 pass, 1 fail, 2 could not decide. Each checker gives the verdict the real tool would give on the whole project; it rechecks only what the change can affect when that is provably the same answer, and otherwise runs the tool in full and says why. Nothing to check is 2, never a pass.

sem check                               # every checker the project hassem check --checkers ts,lint,tests      # only thesesem check --base origin/main            # against origin/main instead of HEADsem check --promises                    # also prove every promise in .sem/promises can failsem check --json                        # one JSON object with a verification certificate

Checkers are detected from the project. TypeScript, lint and JS tests (vitest, jest) follow the module graph, Go rechecks affected packages, Python type checking (pyright scoped by imports, or mypy) and pytest rerun only the files a change can reach, and C and C++ recompile only the translation units in compile_commands.json that include a changed file. Cargo, Gradle or Maven, dotnet and SwiftPM run their own build, which already rebuilds only what changed, and are skipped outright when no input changed since the base passed (a TypeScript edit never reruns the Rust build).

An edit inside the body of a Python function that declares its return type cannot change what other files see, so only that file is rechecked.

Commands of your own go in .sem/check.json, e.g. {"commands": [{"run": "make lint", "inputs": ["src/**"]}]}. A command that lists its inputs is skipped when none of them changed since it passed; one without runs every time.

sem certify

A review certificate for a commit range: entities touched, signature changes and the callers they leave behind, callee deltas, promises kept or broken (with witnesses), module reachability deltas (JS/TS), affected tests, and the static reference cone. --arch gives the architecture view instead: new or removed data paths, side effects, dependencies, cycles, and what did not change, ranked.

sem certify main..HEAD                  # markdown certificatesem certify main..HEAD --json           # the full certificatesem certify main..HEAD --arch           # architecture deltasem certify main..HEAD --arch --view    # at most 10 ranked itemssem certify main..HEAD --html > view.html   # one self-contained page

<base>...<head> uses the merge base; one ref means <ref>..HEAD.

sem diff

Entity-level diff with rename detection, structural hashing, and word-level inline highlights.

# Semantic diff of working changessem diff
# Staged changes onlysem diff --staged
# Specific commitsem diff --commit abc1234
# Commit rangesem diff --from HEAD~5 --to HEAD
# Verbose mode (word-level inline diffs for each entity)sem diff -v
# Plain text output (git status style)sem diff --format plain
# JSON output (for AI agents, CI pipelines)sem diff --format json
# Markdown output (for PRs, reports)sem diff --format markdown
# Compare any two files (no git repo needed)sem diff file1.ts file2.ts
# Read file changes from stdin (no git repo needed)echo '[{"filePath":"src/main.rs","status":"modified","beforeContent":"...","afterContent":"..."}]' \  | sem diff --stdin --format json
# Only specific file typessem diff --file-exts .py .rs

sem graph

How the code is connected. With no flag, the entity dependency graph (the graph sem impact and sem find --context are built on); the flags select other layers.

sem graph --json                        # entities and the calls/references between themsem graph --modules metrics             # JS/TS module graph: density, cycles, depth, centralitysem graph --modules blast-radius pkg-a  # who breaks if pkg-a changessem graph --modules path src/a.ts src/b.tssem graph --dataflow                    # reads, writes and source -> sink pathssem graph --dataflow --witness          # experimental: witness tasks per static flowsem graph --system                      # exact dependency versions from every lockfilesem graph --system build --out sys/     # the layered whole-system graph

sem history

How an entity changed through git history, logic changes told apart from cosmetic ones.

sem history authenticateUsersem history authenticateUser -v         # with the content diff of each versionsem history authenticateUser --limit 20sem history                             # repo hotspots and co-change pairssem history --blame src/auth.ts         # who last changed each entity in the file

With no entity, sem history analyzes recent repo history at the entity level: hotspots (most-changed functions/classes, with author counts) and co-change pairs (entities that repeatedly change in the same commits: "if you touch one, don't forget the other").

sem cloud, sem config

sem cloud login                         # API key, or GitHub when omittedsem cloud enable                        # cloud queries for this public repo (shows what is sent, asks first)sem cloud disable                       # stop cloud for this reposem cloud review listen <diff-id>       # attach an agent to a cloud code reviewsem cloud xref --json                   # cross-repo dependenciessem cloud repos                         # where your code is stored
sem config setup                        # make `git diff` show sem's entity diffsem config telemetry offsem config completions zsh              # shell completionssem config updatesem config stats                        # local diff counters; nothing leaves your machine

Old command names

Every earlier command name and flag still works, with the same output. At a terminal, sem prints a one-line note on stderr with the new spelling; in --json mode or when output is piped, it prints nothing extra.

Promises

A promise is something an agent (or a person) claims about the codebase, written down as a check that either passes or fails. "Done" means the check passes. Promises live in .sem/promises/*.json; each law may carry a human "promise". Code-shape laws are tree-sitter queries that work in any language sem parses; every capture whose name does not start with _ counts as a violation, and nested captures are reported once, at the outermost node. For example, .sem/promises/jsx-no-logic.json:

{ "laws": [  { "id": "jsx-no-logic/conditionals",    "promise": "JSX contains no conditional rendering",    "forbidPattern": { "from": ["**/*.tsx", "**/*.jsx"], "within": ["jsx_expression"],                       "query": "(ternary_expression) @conditional" } },  { "id": "jsx-no-logic/inline-handler-bodies",    "promise": "JSX event handlers are references, not block bodies",    "forbidPattern": { "from": ["**/*.tsx", "**/*.jsx"], "within": ["jsx_attribute"],                       "query": "(arrow_function body: (statement_block (_))) @handler_body" } }] }

sem promises check                        # KEPT / BROKEN n per promise, then file:line:col  capture  snippet; exit 1 if any brokensem promises check --changed src/App.tsx  # only violations in these files (what an edit hook runs)sem promises check --since main           # only files changed since a ref, uncommitted and untracked includedsem promises status --json                # id, kept, violation count per promisesem check --promises                      # prove every promise can fail (each needs a mutation)

sem promises is not listed in sem --help; sem certify reports promises kept or broken for a range. The same file also works with sem graph --modules check --laws <file>, which adds graph and import laws (forbid, only, acyclic, layers, forbidImport, allowImports) for JS/TS workspaces. With --changed or --since, those laws still run over the whole graph, but only violations that involve a changed file (or the package containing it) are reported.

To have an agent see a broken promise right after it makes an edit, run scripts/promises-hook.sh after each edit. The script exits 2 and prints the violations to stderr when a promise breaks, and stays silent otherwise. In Claude Code, add it to your settings as a PostToolUse hook:

{ "hooks": { "PostToolUse": [ { "matcher": "Edit|Write|MultiEdit",    "hooks": [ { "type": "command", "command": "sh /path/to/sem/scripts/promises-hook.sh" } ] } ] } }

The script reads tool_input.file_path from the hook's stdin. Other harnesses can pass the edited paths as arguments instead. In pi, for example, an extension can do this from its tool_result handler for the edit and write tools: run sh promises-hook.sh <path> and append the stderr to the tool result when the script exits 2.

Use as default Git diff

Replace git diff output with entity-level diffs. Agents and humans get sem output automatically without changing any commands.

sem config setup

Now git diff shows entity-level changes instead of line-level. No prompts, no agent configuration needed. Everything that calls git diff gets sem output automatically. Also installs a pre-commit hook that shows entity-level blast radius of staged changes.

On macOS and Linux, sem config setup also registers a Claude Code UserPromptSubmit hook (sem hook prompt-submit) for prompt-time context injection, and sem hook tool around the Edit, Write and MultiEdit tools. After each edit that hook tells the agent which sibling implementations of a changed method it did not touch, which callers still use a signature it just changed, and which tests reach the change, with a command to run them; it stays silent when there is nothing to say. It edits ~/.claude/settings.json idempotently, backs it up first, and leaves any hooks you already have untouched.

To disable and go back to normal git diff (also removes the session hooks):

sem config unsetup

Entity-level diffs on every pull request

Add the GitHub Action and every PR gets one sticky comment showing which functions, classes, and methods changed. It updates in place on each push and calls out cosmetic-only PRs (formatting/comments) explicitly:

# .github/workflows/entity-diff.ymlname: Entity diffon: pull_requestpermissions:  contents: read  pull-requests: writejobs:  entity-diff:    runs-on: ubuntu-latest    steps:      - uses: actions/checkout@v4      - uses: Ataraxy-Labs/sem/[email protected]

No config, no API keys, never fails your build. See action/ for details.

Cloud acceleration (for scale and teams)

Local is always free and always fast: the on-disk index answers day-to-day queries in single-digit milliseconds even from a cold process, so there's nothing to keep warm and no login required. You do not pay to make your laptop fast.

Cloud is for what a laptop can't do. On a very large monorepo the first local graph build can take a few seconds; a shared team graph shouldn't be rebuilt per developer; and CI wants the graph without checking anything out. sem cloud login connects those cases to sem cloud, which keeps a warm, pre-built graph for your registered repos and serves the heavy queries from it instead of rebuilding locally.

sem cloud login                        # GitHub device flow, one timesem impact myFunc --file src/foo.rs    # served from the cloud's warm graph

It is fully optional and transparent:

• Not logged in, or the cloud is unreachable? sem computes locally and prints the exact same output. No failures, no difference in results.
• SEM_LOCAL=1 forces local computation even when logged in.
• Small repos see no change, local is already fast. The win is for large codebases where rebuilding the graph each time is the bottleneck.

Related commands, all cloud-account scoped:

sem cloud logout          # log outsem cloud whoami          # show current cloud identitysem cloud enable          # turn on cloud queries for a public repo (shows what's sent first)sem cloud disable         # stop cloud for this reposem cloud xref --json     # cross-repo dependencies across your indexed repossem cloud repos           # where your code is stored: cloud-indexed repos + local cachessem cloud status          # cloud + telemetry state for this repo (offline; sends nothing)sem cloud share           # cloud queries for a private repo, with extra confirmationsem cloud forget          # delete this repo's cloud index and unregister it

status, share, forget, list, preview and log are not shown in sem cloud --help but work as listed; each one is read-only or requires explicit confirmation before it sends anything.

If your team runs code review through sem cloud, sem cloud review listen <diff-id-or-url> execs a coding agent pre-configured to join that review as a live listener that answers reviewer questions anchored to specific lines of the diff.

What it parses

32 programming languages with full entity extraction via tree-sitter:

Plus structured data formats:

Everything else falls back to chunk-based diffing.

Custom extensions and extensionless files

For files with non-standard extensions, create a .semrc in your project root:

.xyz = cpp.j = json.mypy = python

sem also reads .gitattributes patterns (diff= and linguist-language=) if you already have those set up. .semrc takes priority when both define the same extension.

For files with no extension at all, sem detects the language automatically from content (shebang lines, vim modelines, and structural heuristics like package/import/use statements). This covers 30+ languages with no config needed.

How matching works

Three-phase entity matching:

1. Exact ID match: same entity in before/after = modified or unchanged
2. Structural hash match: same AST structure, different name = renamed or moved (ignores whitespace/comments)
3. Fuzzy similarity: >80% token overlap = probable rename

This means sem detects renames and moves, not just additions and deletions. Structural hashing also distinguishes cosmetic changes (whitespace, formatting) from real logic changes.

Use with AI agents (MCP)

Need a URL-based client? The Streamable HTTP preview adds authenticated, local read-only access; stdio remains the default.

On macOS and Linux, clients in the same checkout share a warm repository daemon. Run sem mcp --status to check it. See the shared runtime contract and reproducible benchmark for session isolation, fallback behavior, and current platform limits.

sem mcp starts a Model Context Protocol server over stdin/stdout. It's not a command you run and read yourself: it's a server your coding agent launches in the background so it can ask sem questions while it works. That's the reason mcp lives alongside the normal commands. The agent gets the core verbs as tools:

Clients that call the earlier tool names (sem_entities, sem_context, sem_callers, sem_log, sem_blame) still get answers; those names are just not listed. A review-listener session (sem mcp --review, which sem cloud review listen sets up) also lists join_review, wait_for_branch, reply_to_branch and list_open_branches.

Why an agent wants these: instead of reading whole files and burning tokens, it can ask "what breaks if I change submitOrder" (sem_impact) or "give me just the context to refactor this function" (sem_find with mode context, which returns the function's source plus its callers and callees) and get a precise, deterministic answer from the dependency graph instead of a grep result that might miss a caller.

Add it once, then talk to your agent normally. It calls the tools on its own.

Claude Code:

claude mcp add sem -- sem mcp

Or one command that also installs the skill, so the agent knows when to reach for sem:

npx @ataraxy-labs/sem-skill

Cursor, Claude Desktop, or any client with an mcpServers config:

{  "mcpServers": {    "sem": {      "command": "sem",      "args": ["mcp"]    }  }}

If sem isn't on the agent's PATH, use the absolute path to the binary. No separate install is needed: sem mcp ships in the same binary as every other command.

JSON output

sem diff --format json

Real output, from a one-line logic change to a Python function:

{  "summary": {    "fileCount": 1,    "added": 0,    "modified": 1,    "deleted": 0,    "moved": 0,    "renamed": 0,    "reordered": 0,    "binary": 0,    "orphan": 0,    "total": 1  },  "changes": [    {      "entityId": "auth.py::function::authenticate_user",      "changeType": "modified",      "entityType": "function",      "entityName": "authenticate_user",      "startLine": 1,      "endLine": 6,      "oldStartLine": 1,      "oldEndLine": 4,      "oldEntityName": null,      "filePath": "auth.py",      "oldFilePath": null,      "oldParentId": null,      "beforeContent": "def authenticate_user(username, password):\n    if not username or not password:\n        return False\n    return check_credentials(username, password)",      "afterContent": "def authenticate_user(username, password):\n    if not username or not password:\n        return False\n    if not check_credentials(username, password):\n        return False\n    return True",      "commitSha": null,      "author": null,      "structuralChange": true    }  ],  "binaryChanges": []}

The named change-type buckets (added, modified, deleted, moved, renamed, reordered) always sum to total. orphan is a cross-cutting metadata count for module-level changes, and those changes are already included in the named change-type buckets. beforeContent/afterContent carry the entity's full source on either side of the change; structuralChange is false when the diff is cosmetic only (whitespace, comments).

As a library

sem-core can be used as a Rust library dependency, from crates.io:

[dependencies]sem-core = "0.23"

Used by weave (semantic merge driver) and inspect (entity-level code review).

Architecture

• tree-sitter for code parsing (native Rust, not WASM)
• git2 for Git operations
• rayon for parallel file processing
• xxhash for structural hashing
• A per-repo cache directory (SQLite entity cache + an mmap-able query index) backs find/callers/refs/grep with cold-process lookups and no background daemon
• Plugin system for adding new languages and formats (see CONTRIBUTING.md)

Telemetry

Local by default: sem counts command names (e.g. diff, impact) on your own machine only, and in that mode nothing is ever uploaded. No code, file paths, repo names, or user identity is recorded, and no network call is made.

sem config telemetry preview   # see current mode and exactly what would be sentsem config telemetry on        # opt in: also upload counts to help improve semsem config telemetry off       # record nothing at all

SEM_NO_TELEMETRY=1 or DO_NOT_TRACK=1 force the record-nothing behavior regardless of mode. Development builds (anything run out of a cargo build target/ directory) never record, so working on sem itself doesn't pollute the numbers.

Contributing

Want to add a new language? See CONTRIBUTING.md for a step-by-step guide.

Star History

[Star History Chart]

License

MIT OR Apache-2.0