Skip to content

ci(search): automate domain-credibility list refresh #143

ci(search): automate domain-credibility list refresh

ci(search): automate domain-credibility list refresh #143

name: Engine Regression Gate
# Proves the bundled engine still loads a real model and runs real inference
# correctly. Runs on every PR and is the mandatory gate on engine-bump PRs.
#
# This is the same shape the field uses to gate a llama.cpp bump (Ollama builds
# the runtime on a version change; Jan/cortex build + run a real-model E2E smoke):
# build it, load a real model, generate, assert. It exercises the real
# scripts/ensure-llama-server.ts path — build/verify/prune/re-sign — so a changed
# dylib closure or a signing breakage fails here, not in a user's hands. It then
# verifies the signature, loads the pinned gate model, drives it through
# llama-server's /v1 chat endpoint (Thuki's actual runtime path, so a broken chat
# template is caught) with greedy decoding, and asserts the answers are
# semantically correct rather than byte-identical to a frozen golden (greedy
# output is only reproducible on the same build + hardware). Generation
# throughput is reported in the job summary as an informational number only: it
# does not gate. The hosted runner's tok/s for this model swings wide enough that
# any floor worth setting fired on runner noise rather than on real breakage, and
# it never caught a CPU fallback (Metal was live in every failing run). Perf
# regressions are caught by the mandatory human review on engine-bump PRs.
on:
pull_request:
push:
branches: [main]
permissions:
contents: read
jobs:
gate:
name: Engine loads + inference
runs-on: macos-15
timeout-minutes: 60
steps:
- uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
- uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # v2.2.0
with:
bun-version: 1.3.11
- run: bun install --frozen-lockfile
- name: Cache llama.cpp sidecar
uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
with:
path: src-tauri/binaries
key: llama-cpp-${{ runner.os }}-${{ hashFiles('scripts/ensure-llama-server.ts') }}
- name: Build the shipped engine (real build/verify/prune/re-sign)
run: bun run engine:ensure
- name: Verify code signatures
run: |
set -uo pipefail
bin=src-tauri/binaries/llama-server-aarch64-apple-darwin
pass=true
codesign -vv "$bin" 2>&1 || pass=false
n=0
for dylib in src-tauri/binaries/*.dylib; do
codesign -vv "$dylib" 2>&1 || pass=false
n=$((n + 1))
done
printf '{"pass":%s,"detail":"binary + %d dylibs verified"}' "$pass" "$n" > codesign.json
cat codesign.json
- name: Cache gate model
uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
with:
path: .gate-model
# Keyed by the pinned model's sha256 prefix; keep in sync with GATE_MODEL
# in scripts/engine-gate/config.ts.
key: gate-model-6a1a2eb6
- name: Download + verify gate model
run: |
set -euo pipefail
eval "$(bun scripts/engine-gate/model-info.ts)"
mkdir -p .gate-model
if [ ! -f ".gate-model/${GATE_MODEL_FILE}" ]; then
curl -fL --retry 3 -o ".gate-model/${GATE_MODEL_FILE}" "${GATE_MODEL_URL}"
fi
echo "${GATE_MODEL_SHA256} .gate-model/${GATE_MODEL_FILE}" | shasum -a 256 -c -
- name: Run inference + measure throughput
run: |
set -euo pipefail
eval "$(bun scripts/engine-gate/model-info.ts)"
GATE_MODEL_PATH=".gate-model/${GATE_MODEL_FILE}" \
GATE_RESULT=result.json \
bun scripts/engine-gate/collect.ts
- name: Decide
run: |
set -euo pipefail
GATE_RESULT=result.json \
GATE_CODESIGN=codesign.json \
bun scripts/engine-gate/decide.ts