hermes-agent/tests/install

Install and Update E2E Tests

These tests answer one question: can a user on a released version get to this commit?

Each leg installs a released version and updates it through the real user surface. Source legs target the selected checkout revision; packaged legs use explicit old/new signed artifacts. The harness can use a mock LLM provider for onboarding and interaction. It does not substitute a mock installer, updater, or headless proxy for a native GUI flow.

The layers

The test family has four layers. Each layer has one job.

1. scripts/sandbox/generate-e2e-matrix.mjs declares the support matrix. It lists every {os, install-method, update-method} pair. It expands the pairs against the sampled release tags. It knows nothing about which pairs CI can run.
2. .github/workflows/install-e2e.yml is the primary workflow. It picks the release tags, runs the generator, and fans out one matrix job per OS. It also writes the plan chart and the result chart on the run summary.
3. The run workflows own the capability knowledge. install-e2e-run.yml serves linux. install-e2e-windows-run.yml serves windows. install-e2e-macos-run.yml selects either the shared script driver or the macOS GUI driver. Job-level if: gates select the supported pairs. All other pairs skip natively and show as grey.
4. The drivers do the work. tests/install/installer-script-e2e.sh handles POSIX script installs, tests/install/macos-desktop-e2e.sh handles macOS dmg installs, and tests/install/windows-e2e.ps1 handles Windows installs. Install and update methods are separate axes, subject to each workflow's capability gates.

To declare a new method, edit the generator. To implement a method, flip the gate in the run workflow and extend a driver.

The isolation trick

Source drivers redirect canonical Hermes Git URLs to a local bare clone at serve.git, using a driver-owned GIT_CONFIG_GLOBAL rewrite. This controls the source install/update boundary, not all network access: tool/dependency downloads and published bootstrap artifacts can still use the network. Packaged-update legs instead use verified signed downloads and a temporary feed.

The driver parks the main branch of serve.git at the old release. The installer runs and lands on the old release. Then the driver moves main to HEAD. An update becomes available in the same way that it does for a real user.

For script-install legs, the installer comes from the old Git ref. A script-reinstall update uses the target revision's script. A hermes-update leg starts the old release's updater, and app-update legs start its app flow. These paths are intentionally different.

The HEAD -> NEXT column

Every combination also runs from HEAD itself. The driver installs HEAD, then mints NEXT: a synthetic child of HEAD that adds one marker file (.hermes-e2e-next). NEXT exists only in the object store, with no ref, and the bare clone carries it into serve.git. The driver then moves main to NEXT and applies the update method.

This column tests two things that no tag leg tests:

• HEAD's installer on an empty machine. Tag legs run an old installer, or run HEAD's installer over an existing install.
• HEAD's own updater. Tag legs start the old release's updater.

On Windows, the HEAD leg also takes every git.exe directory off PATH and installs no remote get-url shim. install.ps1 uses any git that it finds on PATH, so without this step pinned-git staging never runs. The driver's own git plumbing uses the git path that it captured before the strip. The drivers take NEXT as --update-ref NEXT (Windows: -UpdateRef NEXT). The run workflows take it as the update-ref input.

What one leg does

Each leg with the script drivers has these phases:

1. Stage: make the bare clone, park main at the old release.
2. Install: run the old release's own installer script. Make sure that the checkout is at the old commit and that hermes --version works.
3. Update: move main to HEAD. Apply one update method. An app update must produce a new successful receipt or handoff result; a changed checkout alone is not completion.
4. Verify the installed command and products before running hermes --version. Select the command under the installation's .hermes/bin; use the old venv only for a source tree without PM. Check PM currency and compiler receipts where supported. Preserve the no-desktop scenario for a plain install. Do not rebuild, remove dependencies, or force-stop an updater during verification. Historical installs without these receipts get artifact-presence checks, not a freshness claim.

The cheap fixture checks are tests/scripts/test_source_driver.py and tests-js/source-update-observer.test.mjs. They do not run installers or prove native GUI relaunch. The native packaged-update drivers own automatic-relaunch acceptance. A later driver-owned launch checks chat only after the original update/relaunch assertions pass; it cannot repair a failed handoff. The observer does not change source files, products, dependency selections, or facts.

The Windows desktop-installer@latest install downloads the published Hermes-Setup.exe and drives its GUI with AutoHotkey. The selected update method is a separate axis. App-update methods click the running app's Update control; script and CLI methods use their corresponding entry points.

Desktop chat at install and update checkpoints

Desktop-bearing routes run the same message check used by post-build bundle smoke and the desktop chat spec: tests-js/scripts/desktop-chat-smoke.ts. It types a unique prompt through the real composer, requires a new request at the existing loopback mock provider, and waits for a new completed assistant reply in the active transcript. A reply preserved from OLD cannot satisfy NEW. Only inference is mocked; Electron, the installed backend, and rendering run for real. These checks require no external model credentials.

• A desktop-bearing OLD installation must chat before its update. App-driven updates also check the actual OLD window before clicking Update now.
• After the existing completion and product checks, NEW must chat if the route expects desktop. A route that adds desktop only at update retains the OLD no-desktop assertion.
• Plain source install/update routes that never request desktop record not-applicable; the chat driver does not build desktop for them.
• Native package updates must prove automatic relaunch first. The driver then closes the verified app normally and launches the same installed binary for post-update-launch chat. That second launch is not automatic-relaunch proof.

tests/install/e2e-assets/desktop-smoke.ts runs an exact installed executable with an isolated home/userData and checks the actual backend listener's origin. Source checkpoints use that leg's installed tree; bundle checkpoints use its embedded payload. Missing products fail rather than trigger a build or repair. Driver Node/Playwright come from the current checkout, not OLD's dependencies.

Each checkpoint writes a desktop-chat-old.json, desktop-chat-new.json, or desktop-chat-installed.json result plus a screenshot. Failed chat fails the leg even when version, health, and update receipts passed. Existing plugin and user-state preservation checks remain required. The mock configuration writer preserves unrelated settings, including the controlled update feed.

Post-build artifact smoke

desktop-bundled-release.yml calls desktop-bundle-smoke.yml on fresh native Windows and macOS runners. It downloads the admitted commit's receipt-bound bytes through the public archive and installs the actual artifact before chat:

Native target	Independently installed formats
macOS arm64 and x64	DMG, ZIP
Windows arm64 and x64	MSIX, universal MSIXBUNDLE

The universal envelope runs on both Windows architectures, verifying native slice selection. Windows assembly/staging is separate from feed publication; both Windows smoke matrices must pass before publishing its canary feed. macOS publication and stable candidate acceptance also require their smoke results. Failed checks retain the downloadable build and diagnostic artifacts.

Stable candidate schema 2 binds these successful smoke groups into the pinned candidate manifest. Promotion renders that evidence, not the skipped smoke jobs from its own phase. Schema-1 candidates lack this admission and cannot be promoted by the new workflow; create a new candidate tag. They remain valid as previous-version upgrade baselines.

This covers signed downloadable commit builds, canaries with upload enabled, and stable candidates. No-upload tag builds have no cross-job download handoff. Linux bundle jobs remain disabled, but Linux source-install desktop checks run. Unsigned Store submissions are not sideloaded or re-signed for this smoke; their Store acceptance remains separate. Passing helper tests does not establish native installation or chat acceptance: those require the corresponding runner.

Old versions

A leg can install a release from months back. The driver must not assume that the old version has today's CLI surface. The rule: probe, do not assume.

• For the installer, read the flag from the old ref's own script text.
• For the installed CLI, ask the binary with --help.
• If a flag is not found, omit the flag. This is not an error.

The install methods

• packaged-app: a signed Windows MSIX bundle or macOS application ZIP. This pairs only with open-app-update, using separately pinned package inputs rather than the source-tag cross product. See bundled update acceptance for the manifest and proof contracts.
• installer-script: the platform's one-liner (curl | bash on linux and macos, irm | iex on windows).
• installer-script+desktop: the same one-liner with its desktop stage opted in (--include-desktop / -IncludeDesktop). The stage builds the desktop app during the install. On windows it also registers Start Menu and Desktop shortcuts. On linux and macos it builds the app inside the checkout and registers no OS entry point.
• desktop-installer@latest: the published GUI installer (Hermes-Setup.exe on windows, Hermes-Setup.dmg on macos), driven through the real user flow.

The two app-update variants

The desktop app has two launch paths, so the matrix has two app-update methods. Both click "Update now" in the running app. They differ in how the app starts:

• open-app-update: the app starts from the installed app entry point. On Windows, both the desktop installer and installer-script+desktop create shortcuts, so both support this route. On Linux and macOS, the script's opt-in desktop stage builds inside the checkout without registering an OS entry point. The macOS route therefore requires a desktop-installer install; Linux has no open-app-update leg.
• hermes-desktop-app-update: the app starts with the hermes desktop command. Every install method provides this command, on each OS that ships the desktop app. On linux this is the only app surface: no desktop installer and no packaged desktop artifact exist for linux. The driver captures the product's own launch call (argv, cwd, environment) with e2e-assets/launch-capture/sitecustomize.py and re-executes it under Playwright, which owns the app and clicks the update flow. Pre-PM console scripts load the capture hook via PYTHONPATH; PM launchers use -I, so launch-capture/pm-launch.py obtains the installed launcher's own isolated runtime command and loads the driver hook before its bootstrap.

Skips

A grey leg is normal. There are two causes:

• The method pair is declared but cannot run: either no OS entry point exists for it (open-app-update after a plain script install registers nothing to open), or no driver arm exists yet. The gate in the run workflow lists the pairs that run.
• The starting release predates the surface under test. Example: a release without apps/desktop has no window to launch. The tag annotation tag_has_desktop from the primary workflow marks these releases.

The result chart on the run summary shows each leg as passed, failed, or skipped. Confirmed historical upgrade limitations records failures that cannot be fixed in the update target, with exact release commits and CI evidence. These are not blanket skips: the original paths still run. Exact signature matches are non-red, counted separately as known failures, and linked to footnotes at the bottom of the result chart. An unrelated error on the same tag still fails.

Triggers and cost

The full matrix does not run on pull requests. One leg installs real toolchains and takes more than 10 minutes. The triggers are:

• A pull request that touches the install/update surface (the paths: list in install-e2e.yml). It runs only the pr route: installer-script -> hermes-update on Linux (newest release -> PR, and PR -> NEXT), Windows (PR -> NEXT) and macOS (newest release -> PR). The four legs run in parallel, so the wall clock is about one leg.
• A schedule, every 12 hours. This finds upstream drift. When a scheduled run goes red, install-e2e-red.yml opens one issue labelled install-e2e-red, rewrites its body in place on every later red run, and closes it on the first green one.
• A matching release tag push.
• A reusable workflow call from the stable release gate.
• Manual dispatch. You can select the route and the tag count:

gh workflow run install-e2e.yml --ref <branch> -f route=all -f tag-count=2

The generator's output is the leg-count authority. Read the workflow's plan chart before dispatching a large run. Scheduled/tag runs default to two sampled tags; manual dispatch defaults to three. tag-count accepts 1–10.

all selects every source OS. both, update, and installer select Linux source legs, not Windows plus macOS. windows-desktop and macos-desktop select those source/GUI routes. windows-bundled, macos-bundled, and bundled require their package manifests; they do not expand the source-tag cross product. install-ref selects one exact source baseline. Stable release calls also exclude the candidate tag so it cannot serve as its own old version.

Per-leg timeouts and GitHub's matrix limits remain workflow constraints, not proof that every declared combination ran. Native package acceptance is separate from the deferred desktop Playwright application suite.

Running the drivers locally: don't, except in a disposable VM. The windows driver kills every process named Hermes during teardown and the macos driver operates on /Applications/Hermes.app; on a machine with a real Hermes install they will interfere with it.

Plugin upgrade preservation

Every upgrade leg also carries a plugin-survival contract: a tagged upgrade must not delete or modify anything under the active home's plugins/** tree or any profile's profiles/<name>/plugins/** tree. Destructive flows (explicit uninstall, plugin removal, profile or user deletion) are out of contract and not exercised.

• e2e-assets/verify-plugin-preservation.py is the shared, stdlib-only, read-only verifier. snapshot records every entry (kind, byte size + sha256, link targets, and a recursive fingerprint of a symlink's external target) across all plugin roots, including empty directories and the roots themselves; verify diffs the live tree against that snapshot and fails on any deletion or modification. Unreadable paths are hard errors; an empty snapshot is refused as inconclusive rather than claimed as a pass.
• e2e-assets/preserve-plugins.sh is the POSIX/macOS hook pair: after the install phase it seeds controlled, non-dependency directory fixtures (a mnemosyne-wrapper plugin with its marker, a symlinked runtime, a second profile plugin tree, and the externally-owned sidecar witness outside the home — no pyproject anywhere in the scanned root, nothing downloaded) and snapshots; after the update lands it verifies and fails the leg on any violation. Both shell and Windows hooks call the same Python seed command. Existing fixtures or snapshots abort rather than masking damage by reseeding. Windows uses a junction without requiring symlink privilege.
• Unit tests live at tests/scripts/test_verify_plugin_preservation.py and exercise the verifier against a real temp filesystem (real files, real symlinks; junction fallback on Windows).
• Stable-to-stable: the drivers accept --update-ref REF (Windows: -UpdateRef), defaulting to HEAD. Pass the next release tag to target a stable→stable upgrade through the same serve.git staging; only label a leg stable-to-stable when BOTH the install ref and the target ref are release tags. NEXT is reserved for the HEAD -> NEXT column (see above).

Manual update rehearsals

There is no manual/ directory in this repo. The rehearsal scripts — pre (back up the whole HERMES_HOME, the desktop app's Electron userData, the hermes shims on PATH and the global git config, then point the install's update source at a custom repo + ref) and post (undo all of it and report how exact the restore was) — are a hand-off kit for someone with a real install, kept outside this checkout on purpose: they are not part of the repo and not wired into any lane here.

What is in the repo is the assertion layer they exist to demonstrate. The user-state preservation checks below run on every upgrade leg; the manual kit only reproduces the same upgrade on a machine you care about, where a backup and a restore are the only honest way to do it.

User-state preservation

Alongside the plugin-tree contract below, every upgrade leg also carries a user-state contract: an upgrade may add state, may rewrite config.yaml (additive config migration) and the bundled skills/ tree (the product re-syncs it), but it may not delete or modify the user's own durable state — .env, auth.json, state.db, gateway_state.json, memories/, cron/, sessions/, profiles/, photon/, desktop-plugins/, tui-widgets/, skins/, pets/, skills/.archive/ — and state.db may not lose rows.

• e2e-assets/verify-user-state.py is the shared, stdlib-only, read-only verifier. It records row counts for state.db rather than bytes (a live SQLite file changes for benign reasons), and it has no seed mode on purpose: the state it defends must be produced by the product through the ordinary user path, never hand-written by the harness.
• e2e-assets/user-state-actions.sh produces that state with real commands — hermes chat -q (a real turn → sessions/ + state.db rows), hermes auth add (→ auth.json), hermes profile create (→ profiles/<name>/) — after probing --help for each flag, per the harness's "probe, do not assume" rule. Each action asserts it actually landed, so a leg can never "pass" while testing nothing.
• e2e-assets/preserve-user-state.sh is the POSIX/macOS hook pair (snapshot before the upgrade, verify after). Windows calls the same Python snapshot/verify from windows-e2e.ps1's phases, alongside its existing plugin hooks.
• Two more invariants ride along: the redirect must stay transport-level (config --get remote.origin.url stays official while remote get-url is rewritten), and the user-visible launcher must still exist and run after the upgrade.
• Unit tests live at tests/scripts/test_verify_user_state.py and exercise the verifier against a real temp filesystem with real sqlite databases.

Artifacts

Each leg uploads its logs as an artifact. Every leg also records the screen for its whole run: the composite action .github/actions/e2e-screen-record records with the OS's capture backend (x11grab on linux, gdigrab on windows, avfoundation on macos), and fails the leg if the recording is missing or has zero frames. ffmpeg itself comes from the PM toolchain — jobs pass packages: ffmpeg to actions/setup-pm, which installs the locked, sha256-pinned build (native per-OS, including win32-arm64) and puts it on PATH; the action only verifies it is there. Linux runners have no display, so the action starts Xvfb :99 first and exports DISPLAY for every later step — the app under test and the recorder share that display. The windows GUI leg also uploads screenshots and the update result file. Get them with gh run download <run-id>.