Testing the web server (local, no external CI)
The web-server-technology test suite is a set of self-contained, runnable
scripts under native/. There is deliberately no external CI vendor and no
GitHub Actions: the "CI brand" is offloaded to local scripts you (or any
runner, hook, or cron) invoke directly. Each script exits non-zero on failure,
so they compose into any harness without special plumbing.
Build + unit tests
cd web-server-technology/native
make # builds build/sleela-web-server (links the SQL + slweb cores)
make check # hardening unit tests + the CPU keep-warm convergence test
End-to-end scripts
| Script | What it proves |
|---|---|
smoke-test.sh [ro|rw] | Side-1 static + /db SQL (both dialects) + read-only policy |
trav-test.sh | path-traversal containment (..%2f.. โ 403) |
envelopment-test.sh | Side-2 Scientist gate + warrant routing (meme/static/sql) |
hardening-live-test.sh | rate limit (503), write-token (401/ok), trusted-proxy anti-spoof |
config-hardening-test.sh | the hardening features driven from a config file |
warm-live-test.sh | CPU keep-warm status, /__sleela/warm per-core readings, --no-warm |
stress-fuzz-test.sh [CLIENTS] [REQS] | concurrency soak + SQL fuzz + HTTP framing fuzz |
Run one:
sh stress-fuzz-test.sh # defaults: 8 clients x 25 requests
sh stress-fuzz-test.sh 16 100 # heavier soak
What the stress + fuzz harness checks
stress-fuzz-test.sh starts a fresh server (writable /db, a write token) and
runs three phases, failing if any regress:
- A. Concurrency. Multiple clients hammer
/dbwith mixed reads and token-authenticated writes (reads driven viacurl --parallelso load is real without forking a process per request). Every request the server accepts must return 200, and the server must still be healthy afterward. Confirms no torn CSV / no races on the shared engine under concurrent load. - B. SQL fuzz. Gibberish, empty, multi-statement, unknown-table, 16 KiB,
and binary-ish SQL must each yield a clean 4xx (e.g. 400/401/413/422) โ
never a 5xx and never a dropped connection. Confirms the
/dbsurface degrades safely on bad input. - C. HTTP framing fuzz. Malformed requests on raw sockets (bad method, no
HTTP version, missing
Host, a garbage first line, a 70 KiB header) must be rejected with 400/405 or a clean close, and a good request must still succeed afterward โ i.e. the process survived every malformed frame.
The harness is intentionally conservative about load defaults so it finishes
quickly even where process spawning is expensive; raise CLIENTS/REQS for a
real soak. It self-reports a clear startup health gate so a port-in-use or
bind failure is diagnosed as such rather than looking like a crash.
A note on honesty
If a run reports failures, read the per-line output: phases B and C print the
status for each case, and phase A prints ok/non-200 counts plus the server's
post-run health. A non-200 under phase A is a genuine server defect; a 4xx in
phase B/C is the expected safe rejection (those lines are labelled ok).