Merge pull request #167 from AreevAI/release/v1.7.2 #362

Workflow file for this run

	name: CI

	on:
	push:
	branches: [main]
	pull_request:
	workflow_dispatch:

	concurrency:
	group: ci-${{ github.ref }}
	cancel-in-progress: true

	env:
	CARGO_TERM_COLOR: always
	RUST_BACKTRACE: 1
	# `cargo test --workspace` links ~113 separate test binaries, and each one
	# emits full debug info. On Windows that is a large share of both link time
	# and disk I/O (measured: ~6 min between the last `Compiling` line and the
	# first test). Line tables keep the file:line that RUST_BACKTRACE=1 above
	# exists to produce, at a fraction of the size. CI-only, set here rather
	# than in Cargo.toml so local `cargo test` debugging is untouched.
	CARGO_PROFILE_TEST_DEBUG: line-tables-only
	CARGO_PROFILE_DEV_DEBUG: line-tables-only

	jobs:
	test:
	name: test (${{ matrix.os }})
	runs-on: ${{ matrix.os }}
	strategy:
	fail-fast: false
	matrix:
	os: [ubuntu-latest, macos-latest, windows-latest]
	steps:
	- uses: actions/checkout@v4
	- uses: dtolnay/rust-toolchain@stable
	- uses: actions/setup-python@v7
	with:
	python-version: "3.11"
	# Defender scans every file cargo writes, and this job writes a lot of
	# them. Best-effort: a future runner image that refuses the exclusion
	# must slow the build down, not fail it.
	- name: Exclude the build tree from Defender (windows)
	if: runner.os == 'Windows'
	continue-on-error: true
	shell: pwsh
	run: \|
	Add-MpPreference -ExclusionPath "${{ github.workspace }}"
	Add-MpPreference -ExclusionPath "$env:USERPROFILE\.cargo"
	foreach ($p in 'rustc.exe','cargo.exe','link.exe') {
	Add-MpPreference -ExclusionProcess $p
	}
	- uses: Swatinem/rust-cache@v2
	- uses: taiki-e/install-action@nextest
	# `cargo test` runs test BINARIES one at a time, parallelising only
	# within each; with ~113 binaries, most of them small, the cores sit
	# idle. Measured on windows-latest: 11.1 min of test execution out of a
	# 22 min step. nextest schedules across binaries instead (locally:
	# 55s -> 30s for the same already-built suite, 2,535 tests, same result).
	#
	# Sharding this job was considered and rejected on the same numbers:
	# only ~11 of the 22 min is test execution, so every shard would repeat
	# the ~11 min of compile+link to split the other half — near-zero
	# wall-clock gain for 3x the runner minutes. It would also have to keep
	# reporting under the exact name `test (windows-latest)`, which the
	# protect-main ruleset requires.
	- name: Test workspace
	run: cargo nextest run --workspace
	# nextest deliberately does not run doctests, so they stay on cargo.
	- name: Doctests
	run: cargo test --workspace --doc

	postgres:
	name: postgres backend
	runs-on: ubuntu-latest
	services:
	postgres:
	image: pgvector/pgvector:pg16
	env:
	POSTGRES_PASSWORD: postgres
	POSTGRES_DB: areev_test
	ports: ["5432:5432"]
	# Mandatory: the service starts concurrently with the job and the
	# first connection races it without a health gate.
	options: >-
	--health-cmd "pg_isready -U postgres"
	--health-interval 5s
	--health-timeout 5s
	--health-retries 10
	env:
	DATABASE_URL: postgres://postgres:postgres@127.0.0.1:5432/areev_test
	steps:
	- uses: actions/checkout@v4
	- uses: dtolnay/rust-toolchain@stable
	with:
	components: clippy
	- uses: actions/setup-python@v7
	with:
	python-version: "3.11"
	- uses: actions/setup-node@v4
	with:
	node-version: 22
	- uses: Swatinem/rust-cache@v2
	# The conformance runner hard-fails (rather than skipping) when
	# CI=true and DATABASE_URL is missing, so a broken service can never
	# masquerade as a skipped suite.
	- name: Conformance + CAL smoke (postgres)
	run: cargo test -p areev-conformance --features postgres
	- name: Store unit tests with the postgres feature
	run: cargo test -p areev-store --features postgres --lib
	# `postgres-tls` is a SEPARATE build, not a superset in test terms: the
	# refusal path (STO-E003, "asks for encryption, gets none") only exists
	# in the build above, and the verifiers only exist in this one. Its
	# handshake tests need no database — they stand up a throwaway CA and a
	# fake Postgres on loopback — so they belong wherever the feature
	# compiles.
	- name: Store unit tests with the postgres-tls feature
	run: cargo test -p areev-store --features postgres-tls --lib
	# The workspace clippy job lints default features only; without this
	# the feature-gated backend would accumulate warnings invisibly (and
	# docs.rs builds all-features, so it must at least compile clean).
	- name: Clippy with the postgres feature
	run: cargo clippy -p areev-store -p areev-conformance --features postgres --all-targets -- -D warnings
	- name: Clippy with the postgres-tls feature
	run: cargo clippy -p areev-store -p areev-conformance --features postgres-tls --all-targets -- -D warnings
	# The bindings against a REAL postgres — the deployment surface the
	# server tier exists for. (The python/node jobs have no database, so
	# their postgres tests skip there; this is where they run.)
	- name: Python bindings against postgres
	run: \|
	python -m venv .venv
	.venv/bin/pip install maturin pytest
	VIRTUAL_ENV=$PWD/.venv .venv/bin/maturin develop -m crates/areev-py/Cargo.toml
	.venv/bin/pytest crates/areev-py/tests/ -q
	- name: Node bindings against postgres
	working-directory: crates/areev-js
	run: \|
	npm install
	npx napi build --platform
	node --test __test__/smoke.mjs __test__/pg_smoke.mjs

	clippy:
	name: clippy
	runs-on: ubuntu-latest
	steps:
	- uses: actions/checkout@v4
	- uses: dtolnay/rust-toolchain@stable
	with:
	components: clippy
	- uses: actions/setup-python@v7
	with:
	python-version: "3.11"
	- uses: Swatinem/rust-cache@v2
	- name: Clippy
	run: cargo clippy --workspace --all-targets -- -D warnings
	# The scheduler core's purity is enforced mechanically (governed-agents
	# §4): its dependency tree may not contain a clock, randomness, or IO
	# crate — one SystemTime::now() silently invalidates deterministic
	# replay. `cargo deny` cannot ban std, so: (a) the dep tree is grepped
	# for the crates that would smuggle ambient authority in, and (b) the
	# crate-local clippy.toml disallowing SystemTime::now/Instant::now may
	# never be silenced with an allow().
	- name: areev-run-core purity (no clock/rand/net in the dep tree)
	run: \|
	set -e
	BAD=$(cargo tree -p areev-run-core -e normal --prefix none \
	\| awk '{print $1}' \| sort -u \
	\| grep -E '^(rand\|rand_core\|getrandom\|chrono\|time\|tokio\|ureq\|reqwest\|hyper\|mio)$' \|\| true)
	if [ -n "$BAD" ]; then
	echo "areev-run-core dependency tree contains ambient-authority crates:"; echo "$BAD"; exit 1
	fi
	- name: areev-run-core purity (disallowed_methods never silenced)
	run: \|
	if grep -rn "allow(clippy::disallowed_methods)" crates/areev-run-core/src; then
	echo "the purity lint may not be silenced in the scheduler core"; exit 1
	fi

	# NOTE: `cargo fmt --check` is intentionally NOT gated — the tree is not
	# uniformly rustfmt-formatted by design (see CONTRIBUTING.md). Contributors
	# format only the lines they touch.

	msrv:
	name: msrv
	runs-on: ubuntu-latest
	steps:
	- uses: actions/checkout@v4
	- uses: dtolnay/rust-toolchain@1.90.0
	- uses: actions/setup-python@v7
	with:
	python-version: "3.11"
	- uses: Swatinem/rust-cache@v2
	- name: Build on minimum supported Rust version
	run: cargo build --workspace --locked

	docs:
	name: doc
	runs-on: ubuntu-latest
	steps:
	- uses: actions/checkout@v4
	- uses: dtolnay/rust-toolchain@stable
	- uses: actions/setup-python@v7
	with:
	python-version: "3.11"
	- uses: Swatinem/rust-cache@v2
	- name: Build API docs
	run: cargo doc --workspace --no-deps

	coverage:
	name: coverage
	runs-on: ubuntu-latest
	steps:
	- uses: actions/checkout@v4
	- uses: dtolnay/rust-toolchain@stable
	with:
	components: llvm-tools-preview
	- uses: actions/setup-python@v7
	with:
	python-version: "3.11"
	- uses: Swatinem/rust-cache@v2
	- uses: taiki-e/install-action@cargo-llvm-cov
	- name: Measure coverage
	run: cargo llvm-cov --workspace --lcov --output-path lcov.info
	# Three gates: the committed figure has not drifted (the README quotes
	# it), the scored set is above its global floor, and no individual crate
	# has slipped below its own floor. Per-crate is the point — one workspace
	# number lets a regression in one crate hide behind a gain in another.
	- name: Coverage figure is current and every crate is above its floor
	run: python3 scripts/coverage.py --lcov lcov.info --check
	- name: Upload coverage report
	uses: actions/upload-artifact@v4
	with:
	name: coverage-lcov
	path: lcov.info
	if-no-files-found: error

	python:
	name: python bindings
	runs-on: ${{ matrix.os }}
	strategy:
	fail-fast: false
	matrix:
	os: [ubuntu-latest, macos-latest]
	steps:
	- uses: actions/checkout@v4
	- uses: dtolnay/rust-toolchain@stable
	- uses: actions/setup-python@v7
	with:
	python-version: "3.11"
	- uses: Swatinem/rust-cache@v2
	- name: Build the extension module + run pytest
	run: \|
	python -m venv .venv
	. .venv/bin/activate
	python -m pip install --upgrade pip maturin pytest
	maturin develop -m crates/areev-py/Cargo.toml
	pytest crates/areev-py/tests/ -v
	# The Hermes MemoryProvider plugin needs exactly what this job already
	# built — a maturin `areev` in .venv — and nothing else. It used to
	# ride the `adapters` job's venv for that reason; that job left with
	# the adapters (AreevAI/areev-adapters), the plugin did not.
	- name: Hermes provider smoke
	run: \|
	. .venv/bin/activate
	python examples/hermes/test_provider.py

	agent-example:
	# The keyless floor, enforced. `examples/agents/` promises every agent
	# example runs end to end with no credential, no network and no model
	# key — both acts, smoke then improve — and that where an agent ships
	# more than one language stack, all of them mint the identical
	# content-addressed plan (created_at is pinned; run-smokes.sh compares
	# the hashes). `invoice-to-accounting` ships Python + TypeScript + Rust;
	# the other agents are Python, and run-smokes.sh skips a stack an agent
	# does not have before it ever consults REQUIRE, so adding a language is
	# a wrapper plus an agent file and needs no edit here.
	# An untested example in the flagship repo is wrong at exactly the
	# moment somebody is evaluating, so the promise is a job, not a README
	# claim — the SAME entry point a developer runs locally before pushing
	# (docs: examples/agents/docs/testing.md).
	name: agent examples (keyless, every stack)
	runs-on: ubuntu-latest
	steps:
	- uses: actions/checkout@v4
	- uses: dtolnay/rust-toolchain@stable
	- uses: actions/setup-python@v7
	with:
	python-version: "3.11"
	- uses: actions/setup-node@v7
	with:
	node-version: "22"
	- uses: Swatinem/rust-cache@v2
	- name: Build the Python binding into a venv (maturin)
	run: \|
	python -m venv .venv
	. .venv/bin/activate
	python -m pip install --upgrade pip maturin
	maturin develop --release -m crates/areev-py/Cargo.toml
	- name: Build the Node binding (napi)
	run: cd crates/areev-js && npm ci && npm run build
	# Both acts of every agent, in all three stacks, plus the
	# cross-language plan-hash comparison. CARGO_TARGET_DIR points the
	# detached Rust stack at the cached workspace target so its path deps
	# do not recompile from scratch.
	- name: Every agent, every language — smoke + improve
	run: \|
	. .venv/bin/activate
	CARGO_TARGET_DIR="$PWD/target" REQUIRE="python typescript rust" \
	examples/agents/run-smokes.sh

	tls:
	# The non-default native-TLS feature (the recorded rustls exception):
	# round-trip + no-plaintext-downgrade, plus a clean feature build.
	name: tls feature
	runs-on: ubuntu-latest
	steps:
	- uses: actions/checkout@v4
	- uses: dtolnay/rust-toolchain@stable
	- uses: Swatinem/rust-cache@v2
	- run: cargo test -p areev-server --features tls
	- run: cargo check -p areev --features tls

	node:
	name: node binding (${{ matrix.os }})
	runs-on: ${{ matrix.os }}
	strategy:
	fail-fast: false
	matrix:
	os: [ubuntu-latest, macos-latest]
	steps:
	- uses: actions/checkout@v4
	- uses: dtolnay/rust-toolchain@stable
	- uses: actions/setup-node@v7
	with:
	node-version: "22"
	- uses: Swatinem/rust-cache@v2
	# areev-js is a DETACHED cargo workspace with its own Cargo.lock, so a
	# dependency added to any workspace crate it depends on (areev-run,
	# areev-store, …) does not reach this lockfile — and Dependabot's cargo
	# entry for "/" does not either. It drifts silently, and since
	# release-npm.yml builds `--locked`, the drift surfaces as a failed
	# RELEASE rather than a failed build. Assert it here instead.
	# Fix: `cd crates/areev-js && cargo metadata >/dev/null` and commit.
	- name: Assert the areev-js lockfile is current
	working-directory: crates/areev-js
	run: cargo metadata --locked --format-version 1 > /dev/null

	- name: Build the native addon + run smoke test
	working-directory: crates/areev-js
	run: \|
	npm ci
	npx napi build --platform --release -- --locked
	node --test __test__/smoke.mjs __test__/pg_smoke.mjs

	# `index.js` / `index.d.ts` are generated by `napi build`, but they are
	# ALSO the files npm ships (package.json `main`/`types`/`files`) — and
	# release-npm.yml's publish job never regenerates them: it publishes them
	# straight from the checkout, since the build legs upload only `*.node`.
	# A stale committed copy therefore ships a loader that does not match the
	# toolchain that built the binaries, and nothing else would catch it.
	# (This is why they are tracked rather than gitignored — gitignoring them
	# would publish a package whose `main` does not exist.)
	# Regenerate with `npx napi build --platform --release` and commit.
	# napi's output is host-independent, so this holds on every matrix leg.
	- name: Assert the generated binding loader is committed
	working-directory: crates/areev-js
	run: git diff --exit-code -- index.js index.d.ts

	# areev-sandbox is a standalone package, deliberately outside the workspace
	# (it carries wasmi). `cargo test --workspace` never touches it, so it needs
	# its own job or a change here can break it silently — the same reason
	# areev-js has one.
	sandbox:
	name: sandbox (wasm tier C)
	runs-on: ubuntu-latest
	steps:
	- uses: actions/checkout@v4
	- uses: dtolnay/rust-toolchain@stable
	- name: build
	run: cargo build --manifest-path areev-sandbox/Cargo.toml
	- name: test
	run: cargo test --manifest-path areev-sandbox/Cargo.toml
	- name: clippy
	run: cargo clippy --manifest-path areev-sandbox/Cargo.toml --all-targets -- -D warnings

	# The version lives in five places and only one of them is inherited
	# (`version.workspace`). maturin reads pyproject.toml, npm reads
	# package.json, and the GENERATED areev-js/index.js hard-codes it in every
	# platform arm. Both drift modes have shipped: a workspace-only bump makes
	# the publish workflows skip-existing over the released version (a green run
	# that ships nothing), and a package.json bump without regenerating index.js
	# breaks require() for anyone with NAPI_RS_ENFORCE_VERSION_CHECK set. Catch
	# it on the PR that introduces it, not on release day.
	# The self-improvement bench's keyless floor. `--mock` swaps the live model
	# for a deterministic agent, so this asserts the PLUMBING of the A/B/A/B
	# causal claim — lessons applied raise held-out success, rollback returns it
	# to baseline, re-apply recovers it, and rollback really does empty the
	# LESSONS prompt section. It is never a learning claim (no model runs); the
	# live numbers come from `--agent-cmd` and are published with transcripts.
	# `--arms` extends the same floor to the passive-memory providers: keyless
	# throughout (m-llm summarizes with the mock summarizer, no API key), and
	# PLUMBING-ONLY by construction — the deterministic agent complies with any
	# context it is handed, so these arms prove a provider reached the prompt
	# and can never rank curation against retrieval. `--workers 4` is the
	# parallel path; it must emit byte-identical output to `--workers 1`, which
	# `cargo test -p areev-bench` pins separately.
	selfimprove:
	name: selfimprove A/B/A/B + passive arms (mock)
	runs-on: ubuntu-latest
	steps:
	- uses: actions/checkout@v4
	- uses: dtolnay/rust-toolchain@stable
	- uses: Swatinem/rust-cache@v2
	- name: A/B/A/B shape holds under the deterministic agent
	run: \|
	cargo run -p areev-bench --bin selfimprove_aba -- \
	--workdir "$RUNNER_TEMP/aba" --seed 1 --mock --assert-shape \
	--arms m-steel,m-all,m-llm --workers 4
	# The governed-only invocation is what SELFIMPROVE.md's "Live pilot"
	# block tells a reader to run, and it is a different code path from the
	# one above (no providers, no arm bookkeeping). Exercised keylessly here
	# so the documented command cannot rot.
	- name: The governed-only path still runs standalone
	run: \|
	cargo run -p areev-bench --bin selfimprove_aba -- \
	--workdir "$RUNNER_TEMP/aba-governed" --seed 1 --mock --assert-shape
	# The loop+LLM arm: the canned loop-LLM authors a lesson, the scripted
	# review applies it, rollback removes it, re-apply restores it — the
	# authored-lesson lifecycle asserted keylessly (the shape floor, never
	# a learning claim; the canned lesson deliberately cannot move the
	# mock agent's rates).
	- name: The loop+LLM arm applies and restores an authored lesson
	run: \|
	cargo run -p areev-bench --bin selfimprove_aba -- \
	--workdir "$RUNNER_TEMP/aba-llm" --seed 1 --mock --mock-llm \
	--llm-lessons --assert-shape
	# The committed runs are evidence, not decoration: every number in
	# RESULTS.md is recomputed from the transcripts that shipped with it,
	# and MANIFEST.md's checksums are re-derived, so a published file that
	# is renamed or overwritten fails the build instead of review.
	- name: Published runs match their own transcripts
	run: \|
	python3 crates/areev-bench/scripts/verify_run.py \
	crates/areev-bench/results/selfimprove-3seed-qwen3-30b-2026-08-26
	python3 crates/areev-bench/scripts/verify_run.py \
	crates/areev-bench/results/selfimprove-llmarm-3seed-qwen3-30b-2026-08-30
	python3 crates/areev-bench/scripts/verify_run.py \
	crates/areev-bench/results/selfimprove-2x2-qwen3-30b-2026-08-30
	# The cross-configuration comparison computes published numbers too, so
	# its pairing and discordant-count logic is checked on synthetic runs —
	# a silent break there would corrupt a claim, not fail a build.
	- name: The cross-configuration statistics tool agrees with itself
	run: python3 crates/areev-bench/scripts/aba_arm_stats.py --selftest

	versions:
	name: versions agree
	runs-on: ubuntu-latest
	steps:
	- uses: actions/checkout@v4
	- run: python3 scripts/check_versions.py

	# README.md quotes repository quality numbers (test vs source lines, test
	# count, error codes) from committed artifacts. Regenerate and diff, so the
	# published figures cannot quietly drift away from the tree.
	stats:
	name: repo stats current
	runs-on: ubuntu-latest
	steps:
	- uses: actions/checkout@v4
	- name: Verify the committed stats match the tree
	run: python3 scripts/repo_stats.py --check
	- uses: actions/upload-artifact@v4
	with:
	name: repo-stats
	path: \|
	docs/repo-stats.html
	docs/repo-stats.json
	docs/assets/repo-stats-*.svg

Navigation Menu

Uh oh!

Merge pull request #167 from AreevAI/release/v1.7.2 #362

Workflow file

Merge pull request #167 from AreevAI/release/v1.7.2 #362

Uh oh!

Workflow file for this run