| name | subzeroclaw-contribute |
| description | Develop SubZeroClaw — the single-file ~550-line C agentic runtime. Load this before changing src/subzeroclaw.c or src/test.c: it carries the anti-framework thesis (the goal is NOT to grow), the code map (the loop, config, the shell tool, async compaction), what will and won't be merged, and the build/test loop. To merely run or configure SubZeroClaw, use subzeroclaw-use. |
Contributing to SubZeroClaw
SubZeroClaw is an anti-framework: the goal is not to grow but to stay minimal,
readable and correct. The whole runtime is one file, src/subzeroclaw.c
(~550 lines). Before you add anything, read CONTRIBUTING.md — a PR that adds
surface has to clear a high bar, and several categories are refused outright.
The two contributions that are welcome
- Reduce complexity — same behaviour with less code, fewer moving parts, or
clearer logic: remove a function that doesn't earn its place, flatten a control
path, drop a dependency or an allocation, replace a hand-rolled pattern with
libc. A PR that removes lines while keeping tests green is probably good.
- Prove a runtime limitation — a task where "skill + shell + LLM loop" genuinely
breaks down in the runtime, not the skill. The honest test: could a better
.md skill fix it? If yes, it's a skill problem, not a SubZeroClaw one. If the
limitation is structural (e.g. binary tool output can't round-trip through the
JSON string protocol), that's worth discussing. Open an issue with the task, the
skill you wrote, what happened, and why no skill can work around it.
Won't be merged: new tools (the shell is the only tool — the model runs git,
tee, curl itself); plugin systems, hooks, event buses, middleware (multi-user
platform problems SubZeroClaw doesn't have); backward-compatibility shims.
Code map (src/subzeroclaw.c)
The loop, top to bottom:
| Function | Role |
|---|
config_parse_line / config_load | parse the config keys; env overrides; scrub the four provider secrets (SUBZEROCLAW_API_KEY/_ENDPOINT/_REQUEST_EXTRA/_COMPACT_EXTRA) from the environment (wipe-in-place + unsetenv) so the model's shell never sees the key. SUBZEROCLAW_SKILLS is preserved on purpose. |
main | read config + skills, build the system prompt, seed the message array, drive the loop. |
agent_run | the turn loop: POST to endpoint, parse, dispatch tools, until a stop or max_turns. |
parse_response | pull assistant text / tool calls / errors out of the completion JSON. |
process_tool_calls | the one tool: run each shell command via popen() (stderr merged), append results as turns. |
round_has_command | guard: did this round actually request a shell command? |
compact_fire / compact_splice / compact_url | async context sealing: on the router's x_router.compact signal, fire /v1/compact in the background and splice the sealed block in ahead of intervening turns — no turn is ever blocked. |
log_write, write_temp, mkdirp | session logging + scratch files. |
cJSON is vendored (src/cJSON.c/.h) and used automatically when system
libcjson is absent (the Makefile detects it via pkg-config).
Build & test
make
make test
make install
make clean
src/test.c #includes src/subzeroclaw.c directly and defines SZC_TEST to
exclude main, so tests link against the real functions. All 31 tests must
pass, and if you change subzeroclaw.c you must update test.c to match — the
suite covers the shell tool, turn framing, request building + request_extra
merge, response/error parsing, the compact flag + splice, tool dispatch, the
system prompt, skill loading, and config (including the env scrub).
The bar, restated
Every line justifies its existence. The code that remains is the code that can't
be removed. If your change grows the file, the burden is on you to show the
runtime — not a skill — needed it. See CONTRIBUTING.md for the full policy.
