Repository navigation

Rust

feat: vendor the Open Computer Use source so a build needs no upstream #1010

Workflow file for this run

	name: Rust

	# BR-70 — the repo's first `cargo` CI. Two orthogonal axes:
	# * cross-check — does the code we SHIP (windows-gnu / linux-gnu) COMPILE,
	# including #[cfg(test)] code, using the release's own docker
	# recipe (scripts/cross-env.sh)? This is the PR gate.
	# * test — does the platform BEHAVIOUR work on real kernels (macOS
	# Seatbelt, Linux Landlock, Windows PowerShell + taskkill)?
	# A nightly job runs the full cross BUILD — link errors, the glibc floor, and
	# the shipped packages' runtime-library contract — none of which a `cargo check`
	# can see, because a check never links and so never emits a NEEDED entry or a
	# symbol version.

	on:
	pull_request:
	push:
	branches: [main]
	schedule:
	- cron: "0 7 * * *" # nightly: the expensive full cross-BUILD + glibc floor
	# A schedule-only job is a job nobody can check. cross-build-nightly has run
	# 36 times since 2026-07-15 — its entire existence — for 33 failures and 3
	# cancellations and not one green. Part of why that stood for a month is that
	# there was no way to ASK it a question: you could read the log the morning
	# after, change something, and then wait a day to learn whether the change
	# helped. Manual dispatch turns a 24-hour edit loop into a 20-minute one. It
	# runs every job in this file, the nightly included; the three PR-gate jobs
	# are cheap next to the cross build and their signal is never unwelcome. Read
	# cross-build-nightly's OWN conclusion rather than the run's, though — the
	# matrix carries whatever is red on main at the time (windows-latest was, the
	# day this was written), and that says nothing about the cross build.
	workflow_dispatch:

	# Keyed by event as well as ref, and the event half is load-bearing. `github.ref`
	# for the nightly is refs/heads/main — the SAME group a push to main lands in —
	# so with cancel-in-progress a morning merge silently killed the night's cross
	# build. Three scheduled runs died that way (2026-08-01, 08-07, 08-09) and were
	# indistinguishable in the run list from a real failure. Adding manual dispatch
	# would have made it worse: pressing the button on main would cancel whatever
	# push-triggered gate was in flight. Separating by event keeps the behaviour that
	# was actually wanted — a newer push supersedes an older push on the same ref —
	# without letting three unrelated kinds of run compete for one slot.
	concurrency:
	# A group holds at most ONE pending run by default: a newly queued run cancels
	# the pending one, whatever cancel-in-progress says (that governs only the
	# running one). Keyed by ref, a merge burst to main cancels its own queued
	# runs and leaves those commits with no status at all. So only a PR's runs
	# share a group, where the newest should win; any other run is keyed by
	# run_id, not sha, which repeat dispatches of one commit share.
	group: rust-${{ github.event_name }}-${{ github.event_name == 'pull_request' && github.ref \|\| github.run_id }}
	cancel-in-progress: ${{ github.event_name == 'pull_request' }}

	env:
	CARGO_TERM_COLOR: always

	jobs:
	# ── 1. Cheap guards ────────────────────────────────────────────────────────
	guards:
	# ⚠ A job with no timeout burns GitHub's SIX HOUR ceiling when it hangs,
	# and this workflow has done exactly that: `test (macos-latest)` finished
	# BUILDING in ~13 minutes and then sat silent for 5h46m on three
	# `subagent_handler` tests that await `run_complete_subagent_task`
	# unbounded, until the ceiling cancelled it. Same three names on a feature
	# branch and on clean main; not reproducible on a developer's own macOS.
	#
	# These budgets are deliberately well ABOVE the measured maxima (warm
	# cache, 2026-08-27: guards 0.1m, cross-check 8.2m, test 15.0m, serve
	# 9.5m) so a merely slow run is never failed -- they exist to make a HANG
	# cheap and legible, not to police duration. A deadlock now costs 40
	# minutes of runner time instead of six hours, on every branch that
	# pushes.
	timeout-minutes: 10
	if: github.event_name != 'schedule'
	runs-on: ubuntu-latest
	steps:
	- uses: actions/checkout@v4
	- name: Cross recipes have not drifted / glibc floor pin intact
	run: ./scripts/check-no-cross-drift.sh
	- name: Version consistency (CLI = daemon = GUI)
	run: ./scripts/check-version-consistency.sh
	# The capability is "Computer Use" wherever a person reads it -- never the
	# removed "Computer Controller", never vendor-qualified. The internal
	# `computercontroller` identifier is deliberately exempt; see the script.
	- name: Computer Use naming consistency
	run: ./scripts/check-computer-use-naming.sh

	# The vendored upstream Computer Use source must match its manifest AND be
	# fully committable. Repository-wide .gitignore rules matched 15 of its
	# files when it was first vendored; a tree that verifies locally while
	# those files never reach a fresh clone is worse than one that fails here.
	- name: Vendored Computer Use source
	run: ./scripts/check-vendored-computer-use.sh

	# ── 2. Cross COMPILE gate — the BR-70 core. Docker, the SAME images/flags as
	# scripts/release.sh (both source scripts/cross-env.sh). ubuntu-latest is
	# just the docker HOST; the TARGET triple is what is being proven.
	cross-check:
	timeout-minutes: 30
	if: github.event_name != 'schedule'
	runs-on: ubuntu-latest
	strategy:
	fail-fast: false
	matrix:
	target: [x86_64-unknown-linux-gnu, x86_64-pc-windows-gnu]
	steps:
	- uses: actions/checkout@v4

	# Registry + per-triple target dir, persisted across runs. The per-triple
	# target dir is why this is minutes, not a quarter-hour — and it makes the
	# mac/linux build-script clobber (release.sh) structurally impossible here.
	- name: Restore cross cache
	uses: actions/cache@v4
	with:
	path: \|
	.cross-cache/registry
	.cross-cache/target/${{ matrix.target }}
	key: cross-${{ matrix.target }}-${{ hashFiles('Cargo.lock', 'rust-toolchain.toml', 'scripts/cross-env.sh') }}
	restore-keys: cross-${{ matrix.target }}-

	- name: cargo check --workspace --all-targets (${{ matrix.target }})
	run: \|
	set -euo pipefail
	mkdir -p .cross-cache/registry ".cross-cache/target/${{ matrix.target }}"
	export CROSS_REGISTRY_MOUNT="$PWD/.cross-cache/registry"
	export CROSS_TARGET_MOUNT="$PWD/.cross-cache/target/${{ matrix.target }}"
	. scripts/cross-env.sh
	case "${{ matrix.target }}" in
	x86_64-unknown-linux-gnu)
	cross_linux "cargo check --workspace --all-targets --locked" /cross-target ;;
	x86_64-pc-windows-gnu)
	cross_windows "cargo check --workspace --all-targets --locked" /cross-target ;;
	esac

	# ── 3. Native unit tests on REAL kernels. This is where platform BEHAVIOUR is
	# proven: PowerShell selection + taskkill (windows), Seatbelt (macos),
	# Landlock (ubuntu, once BR-69 slice 2 lands — this runner is its venue).
	# NOTE: windows-latest is MSVC; we SHIP mingw/gnu. Job 2 covers the gnu
	# ABI, this job covers the Windows behaviour. Both are load-bearing.
	test:
	timeout-minutes: 40
	if: github.event_name != 'schedule'
	strategy:
	fail-fast: false
	matrix:
	os: [ubuntu-latest, macos-latest, windows-latest]
	runs-on: ${{ matrix.os }}
	steps:
	- uses: actions/checkout@v4
	- uses: dtolnay/rust-toolchain@master
	with:
	toolchain: "1.92"
	components: clippy

	# MSVC's test PDBs dominate the Windows link and cache footprint (the
	# biorouter harness alone produced a ~1.1 GiB PDB). This keeps debug
	# assertions, overflow checks and panic locations while omitting debugger
	# symbols that CI never consumes. Set it before rust-cache so the action's
	# CARGO_* environment hash separates these artifacts from debug=2 builds.
	#
	# macOS is included for a different reason: it is the only runner in this
	# matrix with NO Rust cache (see the cache-on-failure note below), so it
	# pays a cold full-workspace build every run and its debug symbols are
	# what would push the repository over GitHub's 10 GiB cache ceiling once
	# it finally saves one.
	#
	# ⚠ Ubuntu was left out while it had a warm 1.7 GiB cache and no problem
	# for this to fix. It has one now: the integration-binary step below
	# links ~117 more executables, and on Linux every one of them EMBEDS the
	# DWARF of everything it links rather than leaving it in a side file.
	# Measured 2026-09-11 (run 34628149105) with debug=2: that build took the
	# runner from 70 GB free to 1.4 GB and died with `No space left on
	# device` before it finished. The variable changes rust-cache's key, so
	# ubuntu builds cold (measured: lib + bins 3.7 -> 7.5 min) until main's
	# first run with it saves the new cache. A one-time price.
	- name: Reduce CI test debuginfo
	shell: bash
	run: echo 'CARGO_PROFILE_TEST_DEBUG=0' >> "$GITHUB_ENV"

	# ⚠ `cache-on-failure` is what breaks a self-perpetuating deadlock, not a
	# convenience. rust-cache's post step is skipped whenever the job does not
	# succeed, so a job that cannot finish never saves a cache, is cold on the
	# next run, and therefore cannot finish. That is exactly what happened to
	# `test (macos-latest)`: it was cancelled on main at d0a177f8 after 360
	# minutes with its post step skipped, and `gh cache list` subsequently
	# showed no `v0-rust-macos-latest-*` entry at all while ubuntu and windows
	# both had one. Every macOS run since has compiled the whole workspace
	# from scratch — 56+ minutes against a 20 minute budget, versus 17.8m on
	# the last run that had a warm cache (44ac420f).
	#
	# Compiled dependencies are valid whether or not the tests passed, so
	# saving them on a red job costs nothing and means a broken test can never
	# again strand the cache.
	# ⚠ `save-if` restricts SAVING to main, and that is what keeps a cache
	# available to anyone at all. GitHub scopes a cache to the branch that
	# wrote it: a feature branch can read its own caches and main's, and
	# nothing else. So every feature-branch run was writing its own private
	# 1-1.7 GiB copy per OS that no other branch could ever use, while the
	# repository sat at 9.74 GiB against a 10 GiB ceiling — evicting main's
	# copies, which are the only ones shared. Measured on windows-latest, by
	# rust-cache's own restore time:
	#
	# restore 0.7m -> `cargo test` 9.4m (hit, branch's own earlier cache)
	# restore 0.1m -> `cargo test` 19.5m (miss)
	# restore 0.1m -> `cargo test` 17.4m (miss)
	#
	# With saving confined to main, every branch restores one authoritative,
	# continuously refreshed cache and none of them compete for the ceiling.
	# A PR that changes Cargo.lock still builds cold and cannot save; that is
	# the accepted cost, and it is much smaller than three OSes' worth of
	# write-only caches evicting the shared ones on every push.
	- uses: Swatinem/rust-cache@v2
	with:
	key: ${{ matrix.os }}
	cache-on-failure: "true"
	save-if: ${{ github.ref == 'refs/heads/main' }}

	- name: Install protoc (build-script dep)
	uses: arduino/setup-protoc@v3
	with:
	repo-token: ${{ secrets.GITHUB_TOKEN }}

	- name: Install Linux native build dependencies
	if: matrix.os == 'ubuntu-latest'
	run: \|
	# ⚠ Drop the runner image's Microsoft apt repositories FIRST.
	# `packages.microsoft.com` has repeatedly answered 403 for the
	# `azure-cli` and `prod` lists that the hosted image ships
	# preinstalled, and `apt-get update` exits 100 on any repository it
	# cannot refresh — so an outage on a third-party CDN we do not use
	# fails this job with `E: ... is no longer signed`. Nothing here
	# installs from them. Removing them makes the update depend only on
	# the archives the packages below actually come from.
	sudo rm -f /etc/apt/sources.list.d/microsoft-prod.list \
	/etc/apt/sources.list.d/azure-cli.list \
	/etc/apt/sources.list.d/microsoft-prod.sources \
	/etc/apt/sources.list.d/azure-cli.sources
	sudo apt-get update
	sudo apt-get install --yes libdbus-1-dev libxcb1-dev pkg-config

	# Start narrow (lib + bins on every OS) so the matrix is green on arrival;
	# widen as it proves stable. This selector leaves out `tests/`; those
	# binaries run in their own step below, ubuntu only, with no network.
	# (They were long assumed to need cassettes and credentials the runners
	# lack. Measured, they do not: `mcp_integration_test` REPLAYS its
	# cassettes offline and only records under BIOROUTER_RECORD_MCP, and the
	# tests that need a real provider or vendor CLI are `#[ignore]`d.)
	#
	# ⚠ It does NOT follow that nothing in this step touches the network. The
	# selector filters by TARGET, not by behaviour, and a test that lives in
	# `src/` is a unit test as far as cargo is concerned however far it dials.
	# `biorouter-server/src/tunnel/lapstone_test.rs` did exactly that against a
	# third-party Cloudflare worker, and because `biorouter-server/src/main.rs`
	# re-declares `mod tunnel;` rather than using the library, both `--lib` and
	# `--bins` carried it: 2 tests x 2 targets x 3 platforms, twelve live
	# requests per run, red whenever that worker had a bad day. It is now
	# opt-in behind BIOROUTER_TEST_LAPSTONE_TUNNEL. Gate the test, not the
	# selector, if another one appears.
	# ⚠ `--no-fail-fast` is load-bearing on this matrix, not a nicety. Without
	# it cargo stops at the FIRST test binary that fails, and everything
	# ordered after it never runs. On windows-latest that was `biorouter_mcp`,
	# so `biorouter_sandbox`, `biorouter_server`, `biorouterd` and
	# `biorouter_test` were unreported for a hundred consecutive runs: each
	# batch of Windows fixes made the job fail one binary further along, which
	# read as "a new rotating set of failures" and was really one hidden queue
	# being drained one CI round-trip at a time. One run now names them all.
	# ⚠ **The OS credential store must be off, or macOS hangs until the job
	# is killed.** `Config::read` resolves a secret through `keyring`, and on
	# a runner with no unlocked login keychain `SecKeychainFindGenericPassword`
	# blocks on an authorization prompt nothing is there to answer. The
	# secrets tests in `routes::action_required` sit on that call, so the
	# whole binary stops: `test (macos-latest)` ran SIX HOURS on main
	# (run 32997530845) and was cancelled. Reproduced locally and sampled —
	# `Config::read` -> `keyring::Entry::get_password` ->
	# `SecKeychainFindGenericPassword` -> `CSSM_DecryptData`, blocked. With
	# this set the same binary finishes in seconds.
	#
	# Windows reaches the same code through the Credential Manager. Whether
	# that is what has made windows-latest slow is NOT measured, so this is
	# not offered as the explanation there — it is set on the whole matrix
	# because the store is unwanted on every runner, not as a Windows fix.
	#
	# It costs no coverage: nothing in the workspace asserts keyring
	# behaviour — the store is incidental to these tests, which are about
	# routes. `serve` below and `release-artifact-smoke.yml` already set it
	# for the same reason; this is the one job that was missing it.
	- name: cargo test (workspace, lib + bins)
	env:
	BIOROUTER_DISABLE_KEYRING: "true"
	run: cargo test --workspace --lib --bins --locked --no-fail-fast --timings

	- name: Upload Windows Cargo timings
	if: matrix.os == 'windows-latest' && always()
	uses: actions/upload-artifact@v4
	with:
	name: windows-cargo-timings
	path: target/cargo-timings/cargo-timing.html
	if-no-files-found: warn
	retention-days: 14

	# The cross-platform-sensitive suites BR-70 exists to prove, run on every
	# OS they apply to. The workspace step above already runs the sandbox,
	# developer shell/background and security unit tests through their library
	# targets. The sandbox's integration target is the one `tests/` binary
	# that runs on every OS (the rest follow in the next step, ubuntu only);
	# name it directly rather than executing those unit tests a second time.
	- name: cargo test (sandbox integration)
	run: cargo test -p biorouter-sandbox --test sandbox --locked --no-fail-fast

	# ── Every other `tests/*.rs` binary, bar a table of exceptions ─────────
	#
	# Until this step no workflow ran them. The lib + bins selector above
	# leaves `tests/` out, and `clippy --all-targets` below only COMPILES
	# them, which says nothing about whether they pass. So the guards
	# CLAUDE.md treats as load-bearing (the privacy master-switch binaries,
	# the keyless-daemon binaries, `privacy_ar15_is_retired`,
	# `every_test_binary_is_sandboxed`) ran only when someone remembered to.
	# The cost is on record: from eb594ded until PR #229,
	# `the_documented_closure_is_the_one_the_code_performs` read the wrong
	# function, and would have stayed green with the AR-15 gate deleted.
	#
	# ⚠ The table is an EXCLUSION list on purpose. An allowlist would rebuild
	# the gap one file at a time: a new binary would run only if its author
	# remembered to add it. This way it runs the day it lands, and one that
	# cannot run hermetically gets a line saying why. A line naming a target
	# that no longer exists fails the step, so the table cannot outlive what
	# it excuses.
	#
	# ⚠ Never add `--ignored` or `--include-ignored`. The live tests (real
	# providers, the vendor CLIs, llama.cpp downloads, GitHub fetches) are
	# `#[ignore]`d inside binaries whose other tests belong here, so the
	# binary stays in and its live tests stay out.
	#
	# ⚠ No network, enforced rather than assumed. The tests run in a network
	# namespace holding nothing but loopback, so a test that dials out fails
	# at once with a connection error instead of passing whenever the third
	# party is up (the lapstone note above is what that looks like).
	# `setpriv` drops back to the runner user once loopback is up: as root,
	# every "permission denied" assertion would pass for the wrong reason.
	# Servers on 127.0.0.1 and ::1 work as usual.
	#
	# The table's second entry is the instructive one. `ui_example_apps`
	# bundles each example with `npx --yes esbuild` when no local esbuild
	# exists, and this job installs none. Offline it does NOT fail: measured,
	# npx sits on the unreachable registry until the bundler's 60 s timeout,
	# then the test passes on the type-stripper fallback without esbuild
	# ever running. That is a minute spent proving nothing, and a pass that
	# turns into a failure the day npm's retries give up before the timeout.
	#
	# Measured 2026-09-11, one binary at a time on hosted runners (scratch
	# runs 34628149105 and 34629544896). ubuntu, inside the namespace: all
	# 119 binaries green, 687 tests passed, 0 failed, 49 ignored. Building
	# them took 135 s on top of the lib + bins build (14.5 GiB of
	# executables, with the debuginfo setting above) and the 117 this step
	# selects spend ~135 s in their tests. macOS was green but for one
	# log-capture test that failed once in ~70 runs. windows-latest fails
	# 31 tests in 9 binaries that had never run there, which is why this is
	# ubuntu only: widening means fixing those first, and `unshare` is
	# Linux-only, so another OS also needs its own way to stay offline.
	- name: cargo test (integration binaries, no network)
	if: matrix.os == 'ubuntu-latest'
	env:
	BIOROUTER_DISABLE_KEYRING: "true"
	CARGO_NET_OFFLINE: "true"
	run: \|
	set -euo pipefail
	# One line per target, no wrapping: the first word is the name.
	# target why it does not run in this step
	excluded='
	sandbox "cargo test (sandbox integration)" above runs it, on every OS
	ui_example_apps needs the npm registry (`npx --yes esbuild`); apps-smoke.yml runs it after `npm ci`
	'
	all=$(cargo metadata --format-version 1 --no-deps --locked \
	\| jq -r '.packages[].targets[] \| select(.kind == ["test"]) \| .name' \| sort)
	# `--test` selects by name across the workspace, so a name two
	# packages share would run (or be excluded) twice over.
	dup=$(uniq -d <<<"$all")
	if [ -n "$dup" ]; then
	echo "::error::two packages have an integration target named: $dup"; exit 1
	fi
	skip=$(awk 'NF { print $1 }' <<<"$excluded" \| sort)
	stale=$(comm -13 <(echo "$all") <(echo "$skip"))
	if [ -n "$stale" ]; then
	echo "::error::the exclusion table names targets that no longer exist; delete their lines: $stale"; exit 1
	fi
	awk 'NF { t = $1; $1 = ""; printf "not run here: %s --%s\n", t, $0 }' <<<"$excluded"
	args=()
	for t in $(comm -23 <(echo "$all") <(echo "$skip")); do args+=(--test "$t"); done
	echo "running $(( ${#args[@]} / 2 )) integration binaries"
	sudo -E unshare --net -- sh -c \
	'ip link set lo up && exec setpriv --reuid="$SUDO_UID" --regid="$SUDO_GID" --init-groups -- "$@"' sh \
	env HOME="$HOME" PATH="$PATH" \
	cargo test --workspace --locked --no-fail-fast "${args[@]}"

	- name: clippy (host, informational until warnings are burned down)
	if: matrix.os == 'ubuntu-latest'
	run: cargo clippy --workspace --all-targets --locked

	# ── 4. Nightly: the FULL cross BUILD. `check` cannot catch link errors
	# (aws-lc-sys / winpthread) or glibc symbol-version regressions.
	cross-build-nightly:
	timeout-minutes: 120
	# The three jobs above are gated `!= 'schedule'`, so they pick up
	# workflow_dispatch on their own; this one has to name it, or the button
	# would run everything EXCEPT the job it was added for.
	if: github.event_name == 'schedule' \|\| github.event_name == 'workflow_dispatch'
	runs-on: ubuntu-latest
	steps:
	- uses: actions/checkout@v4
	- name: Build shipped binaries for both cross targets (release command)
	run: \|
	set -euo pipefail
	. scripts/cross-env.sh
	cross_linux "cargo build --release --bin biorouterd --bin biorouter"
	cross_windows "cargo build --release --bin biorouterd --bin biorouter" "" "$WIN_DLL_STAGE"
	- name: Assert the glibc floor (2.31 / bullseye) is not raised
	run: ./scripts/check-glibc-floor.sh

	# The glibc floor check covers the libraries the binaries need from libc.
	# This covers the ones they need from everywhere ELSE — today exactly one,
	# libxcb.so.1 — and asserts that the packages we ship name it, so apt/dnf
	# put it on the box before the binary ever runs. See the script's header;
	# the short version is that a new system dependency is invisible to every
	# other gate in this file and surfaces as exit 127 on a user's machine.
	- name: The shipped binaries' runtime deps are declared by the shipped packages
	id: runtime-deps
	run: \|
	set -euo pipefail
	./scripts/check-linux-runtime-deps.sh
	echo "deb_packages=$(./scripts/check-linux-runtime-deps.sh --print-deb-packages)" >> "$GITHUB_OUTPUT"

	# This step used to `docker run` the bare ELF out of target/ in a stock
	# debian:bullseye, and had therefore failed EVERY night since it was
	# written — ~40 consecutive red runs, never once green. The failure was
	# real and reproducible and meant nothing: `biorouter` links libxcb.so.1
	# (arboard's clipboard, xcap's screen capture), the base image ships no
	# X11 libraries at all, so the loader aborted before `main`. It was
	# reporting a catastrophe that no user can reach, because no user runs the
	# naked binary — they install biorouter-cli_*.deb or the .rpm, and those
	# DECLARE libxcb1/libxcb, so the package manager installs it as a matter
	# of course. Measured on a clean debian:bullseye and a clean rockylinux:9:
	# both packages install and both binaries report their version.
	#
	# What the step proves now is the thing that was worth proving all along:
	# that the declared dependency set is SUFFICIENT. It installs exactly the
	# packages the shipped specs name — derived from the ELF by the guard
	# above rather than hardcoded here, so the list cannot go stale in one
	# place while moving in the other — and nothing else. If a future crate
	# adds a system library, the guard above fails first and names the fix; if
	# the declared set were merely incomplete, this boot fails and says so.
	# Both binaries are booted: they ship together and link the same set, and
	# only `biorouter` was ever being checked.
	#
	# bullseye is the floor pin (GLIBC_FLOOR / LINUX_RUST_IMG in
	# scripts/cross-env.sh). Its apt mirrors are what make the install step
	# work; when Debian 11 finally leaves them, this image moves forward
	# together with that pin, not on its own.
	- name: Boot both Linux binaries on the oldest supported distro
	env:
	DEB_PACKAGES: ${{ steps.runtime-deps.outputs.deb_packages }}
	run: \|
	set -euo pipefail
	docker run --rm -e DEB_PACKAGES -v "$PWD":/w debian:bullseye sh -euxc '
	apt-get update -q
	apt-get install -y --no-install-recommends $DEB_PACKAGES
	/w/target/x86_64-unknown-linux-gnu/release/biorouter --version
	/w/target/x86_64-unknown-linux-gnu/release/biorouterd --version
	'

	# `biorouter serve` end to end: the daemon serving the interface, the token
	# exchange, and the endpoints that were previously unauthenticated.
	#
	# Before this job there was NO CI coverage of browser mode at all -- `grep -rn
	# headless .github/` returned nothing while a tarball of it was a shipped
	# release asset. The one smoke test that existed asserted the served page
	# contained "<!doctype html>", which is true of a completely non-functional
	# app.
	#
	# It cannot drive a real agent turn: there are no provider credentials here,
	# and a browser session may not set a provider by design (decision SD-1 in
	# docs/deployment/serve-decisions.md). What it does assert is everything
	# between the browser and that boundary.
	serve:
	# Was 30 against a measured 9.5m. The lifecycle step below builds
	# biorouter-cli's tests, which unify its dev-dependency features and so
	# recompile part of the graph a plain `cargo build` already built (4m56s on
	# a loaded M4 Max); the budget keeps the same margin over that.
	timeout-minutes: 45
	if: github.event_name != 'schedule'
	runs-on: ubuntu-latest
	steps:
	- uses: actions/checkout@v4
	- uses: dtolnay/rust-toolchain@master
	with:
	toolchain: "1.92"
	- uses: Swatinem/rust-cache@v2
	with:
	key: serve
	- uses: actions/setup-node@v4
	with:
	node-version: 24
	- name: Install protoc (build-script dep)
	uses: arduino/setup-protoc@v3
	with:
	repo-token: ${{ secrets.GITHUB_TOKEN }}
	- name: Install Linux native build dependencies
	run: \|
	# ⚠ Drop the runner image's Microsoft apt repositories FIRST.
	# `packages.microsoft.com` has repeatedly answered 403 for the
	# `azure-cli` and `prod` lists that the hosted image ships
	# preinstalled, and `apt-get update` exits 100 on any repository it
	# cannot refresh — so an outage on a third-party CDN we do not use
	# fails this job with `E: ... is no longer signed`. Nothing here
	# installs from them. Removing them makes the update depend only on
	# the archives the packages below actually come from.
	sudo rm -f /etc/apt/sources.list.d/microsoft-prod.list \
	/etc/apt/sources.list.d/azure-cli.list \
	/etc/apt/sources.list.d/microsoft-prod.sources \
	/etc/apt/sources.list.d/azure-cli.sources
	sudo apt-get update -q
	sudo apt-get install -y --no-install-recommends libdbus-1-dev libxdo-dev

	- name: Build the interface bundle
	working-directory: ui/desktop
	run: \|
	npm ci
	npm run build:web
	# Root base, not Forge's relative base: a relative-base bundle served
	# at the root resolves its assets against the current path.
	grep -q 'src="/assets/' src/web/index.html
	! grep -q 'src="./assets/' src/web/index.html

	- name: Build the command and the daemon
	run: cargo build -p biorouter-cli -p biorouter-server

	- name: Serve, and check the whole browser contract
	run: \|
	set -euo pipefail
	export BIOROUTER_PATH_ROOT="$RUNNER_TEMP/serve-root"
	export BIOROUTER_DISABLE_KEYRING=true
	mkdir -p "$BIOROUTER_PATH_ROOT"
	./target/debug/biorouter serve \
	--port 18777 --token citoken \
	--web-dir "$PWD/ui/desktop/src/web" > "$RUNNER_TEMP/serve.log" 2>&1 &
	serve_pid=$!
	for _ in $(seq 1 90); do
	curl -fsS -o /dev/null "http://127.0.0.1:18777/?t=citoken" && break
	sleep 1
	done

	fail() { echo "FAILED: $1"; echo '--- serve log ---'; cat "$RUNNER_TEMP/serve.log"; exit 1; }

	# The document carries the daemon secret, so it is gated.
	[ "$(curl -s -o /dev/null -w '%{http_code}' http://127.0.0.1:18777/)" = 401 ] \
	\|\| fail 'the shell was served without a credential'
	[ "$(curl -s -o /dev/null -w '%{http_code}' 'http://127.0.0.1:18777/?t=wrong')" = 401 ] \
	\|\| fail 'a wrong token was accepted'

	# The token is exchanged for a cookie and leaves the address bar. It is
	# not consumed (SD-9 in docs/deployment/serve-decisions.md): the
	# readiness loop above has already redeemed it.
	curl -s -o /dev/null -D "$RUNNER_TEMP/h" "http://127.0.0.1:18777/?t=citoken"
	grep -qi '^HTTP/1.1 303' "$RUNNER_TEMP/h" \|\| fail 'no redirect after the token exchange'
	grep -qi 'set-cookie: biorouter_session=' "$RUNNER_TEMP/h" \|\| fail 'no session cookie'
	grep -qi 'HttpOnly' "$RUNNER_TEMP/h" \|\| fail 'the session cookie is readable from script'
	grep -qi 'SameSite=Strict' "$RUNNER_TEMP/h" \|\| fail 'the session cookie is sent cross-site'

	# With the cookie, the shell and its runtime configuration.
	curl -fsS -b 'biorouter_session=citoken' http://127.0.0.1:18777/ > "$RUNNER_TEMP/shell.html"
	grep -q '__BIOROUTER_HEADLESS_CONFIG__' "$RUNNER_TEMP/shell.html" \|\| fail 'no runtime config'
	grep -qE '"secretKey":"[0-9a-f]{64}"' "$RUNNER_TEMP/shell.html" \|\| fail 'no daemon secret'
	# ABSENT, not empty: empty is falsy in the renderer and sends the
	# browser to a hardcoded 127.0.0.1:3000.
	! grep -q 'apiBaseUrl' "$RUNNER_TEMP/shell.html" \
	\|\| fail 'apiBaseUrl is present; the renderer will not use its own origin'
	grep -q 'no-store' <(curl -s -D - -o /dev/null -b 'biorouter_session=citoken' http://127.0.0.1:18777/) \
	\|\| fail 'the shell is cacheable and it carries the secret'

	# The bundle is really served.
	asset=$(grep -o '/assets/index-[A-Za-z0-9_-]*\.js' "$RUNNER_TEMP/shell.html" \| head -1)
	[ -n "$asset" ] \|\| fail 'no bundle reference in the shell'
	curl -fsS -o /dev/null "http://127.0.0.1:18777$asset" \|\| fail 'the bundle is not served'

	# The interface's own endpoints are behind the daemon's auth now. They
	# had none at all before.
	secret=$(grep -o '"secretKey":"[0-9a-f]*"' "$RUNNER_TEMP/shell.html" \| cut -d'"' -f4)
	[ "$(curl -s -o /dev/null -w '%{http_code}' http://127.0.0.1:18777/headless/health)" = 401 ] \
	\|\| fail 'an interface endpoint answered without the secret'
	curl -fsS -H "X-Secret-Key: $secret" http://127.0.0.1:18777/headless/health \
	\| grep -q '"status":"ok"' \|\| fail 'the interface endpoint did not answer with the secret'

	# And a path traversal is refused, which the retired implementation
	# allowed outright: fs_read had no validation of any kind.
	[ "$(curl -s -o /dev/null -w '%{http_code}' -H "X-Secret-Key: $secret" \
	'http://127.0.0.1:18777/headless/fs/read?path=/etc/passwd')" = 403 ] \
	\|\| fail 'the filesystem endpoint read outside its allowed roots'

	# Stopping serve by pid stops its daemon. It used to leave the daemon
	# running, still holding this port and still honouring the token.
	# `serve` reaps the daemon before it exits, so the port is closed the
	# moment `wait` returns.
	kill "$serve_pid" \|\| fail 'serve was no longer running'
	wait "$serve_pid" \|\| true
	! curl -s -o /dev/null --max-time 5 http://127.0.0.1:18777/status \
	\|\| fail 'the daemon outlived serve'
	echo 'the browser contract holds'

	# SIGTERM and SIGINT to `serve`, and SIGKILL, which leaves only the
	# daemon's own parent watch to stop it. Here rather than in the workspace
	# test job, which runs `--lib --bins` only: this needs a `biorouterd` from
	# the same tree beside `biorouter`, which the build step above provides.
	- name: Stopping serve stops its daemon
	env:
	BIOROUTER_DISABLE_KEYRING: "true"
	run: cargo test -p biorouter-cli --test serve_lifecycle

Navigation Menu

Repository navigation

feat: vendor the Open Computer Use source so a build needs no upstream #1010

Workflow file

feat: vendor the Open Computer Use source so a build needs no upstream #1010

Uh oh!

Workflow file for this run