nodecode-sem
Sem + Weave: entity-level code discovery and review with the local Sem CLI, plus branch merge previews with Weave. A Nodecode rewrite of Ataraxy-Labs/sem by Ataraxy-Labs.
The original is two skills that tell an agent to type sem and weave commands at a terminal. This cell does the typing: four tools, one command and one manual topic, with the CLIs' own answers, no shell in between.
Get it
Tell the agent "add the sem cell", or run nodecode add nc://sem.
Set it up
Install Sem yourself, and Weave if you want merge previews. Tested upstream with Sem 0.27.0 and Weave 0.5.4. Check sem --version: GNU Parallel also ships an unrelated executable named sem.
The cell needs no key and no section to run: it finds sem and weave on PATH. /sem reports which of the two this machine has. When one lives elsewhere, name it:
"sem": { "sem_path": "/opt/homebrew/bin/sem", "weave_path": "weave" }
| Setting | Default | What it is |
|---|---|---|
sem_path |
sem |
the Sem executable, on PATH |
weave_path |
weave |
the Weave executable; only previews need it |
directory |
empty | the repository the CLIs run in; the asking session's directory when empty |
offline |
true |
run both CLIs with SEM_NO_NETWORK=1 SEM_NO_TELEMETRY=1 |
timeout |
120 |
seconds one call may take before its process tree is killed |
max_chars |
30000 |
how much of one answer the model is given |
A present section is an enabled one; "enabled": false vetoes it. config.example.jsonc carries all of it, commented.
Use it
The model gets four tools, and asks for them by name:
| Tool | What it answers |
|---|---|
sem_find |
where an entity is defined, who calls it, or its body with the context around it (names, callers, context, budget, in) |
sem_diff |
changed entities, not changed lines, for a range or the working tree (range) |
sem_impact |
what depends on an entity, and which tests reach it (name, tests, file) |
weave_preview |
how a branch would merge into HEAD, without merging it (branch, file) |
So "find parse_config and its callers, then review the uncommitted entity changes" is one or two calls rather than a shell command per question.
/sem with no arguments lists both programs and their versions; /sem find parse_config --json runs Sem itself, the arguments split on whitespace and never through a shell. The manual topic (help :sem) carries the guidance the original skills gave: choose the smallest query, an empty caller or test list is not proof there is none, do not invent a base branch, and a clean merge preview is not compiler or behavioral verification.
Nothing here edits, stages, commits or merges, and sem check -- which runs the repository's own build, linter and tests -- is not one of the tools. Sem may keep a local index outside the repository; with offline on, neither CLI is given the network.
The tests stub every CLI call: NODECODE_SRC=<nodecode checkout> .github/test-cell.sh tools/nodecode-sem from the cells repository.
License
MIT. The plugin it rewrites is Copyright (c) 2026 Ataraxy Labs: see LICENSE and NOTICE.
