Skip to content

fix: pre-release hardening - #3

Draft
dorukozgen wants to merge 26 commits into
mainfrom
fix/pre-release-hardening
Draft

dorukozgen wants to merge 26 commits into
mainfrom
fix/pre-release-hardening

Conversation

@dorukozgen

Copy link
Copy Markdown
Contributor

Pre-release hardening before v0.2.0. Draft: more commits are coming before review.

Done so far:

  • Local API: only answers requests addressed to the app's own host. API requests carry a per-run token that the app hands to its UI.
  • Release UI: served from the app binary with a Content-Security-Policy. Preview stays on local dev servers.
  • Proxies: on Linux, loopback traffic skips environment proxies.
  • Browser panel: runs without access to the app's commands.
  • Markdown images in chat replies and workspace markdown load only after a click.
  • Git: remote credentials stay out of PR traces, AI prompts and replies. Paths are checked before they open in an editor.
  • Windows: another program can't share the backend's port.
  • Product analytics defaults to PostHog's EU region.
  • Tests: stale tests updated; every suite is green.

Still to come:

  • config write locking
  • single instance and log-folder fallback
  • startup window
  • external links
  • reveal/package handling
  • a one-time dashboard-sync question for upgraded installs
  • the final launch checks

The install route takes source_id + scope, the PR chip spells out its
state, and onboarding's Continue is always enabled; the tests still
expected the old behaviour. phase4_apis now also runs against a
temporary FLEETIFY_HOME.

Two tests raced: the python_runtime env tests now hold the crate's env
lock, and the loopback child test no longer serves on the port its
parent released for the unreachable proxy.

Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
The embedded backend now checks the Host header on every request, the
UI files, /healthz and unmatched paths included. It answers only
127.0.0.1:<bound port> and localhost:<bound port>; any other host, or
none, gets a plain-text 403. The check is the router's outermost layer.

The CORS layer is gone. The app's pages share the backend's origin, or
reach it through the Vite proxy in dev, which now sets the backend's
Host on WebSocket upgrades as it already did for HTTP requests.

serve() now takes ServeOptions (UI files, routing setup and a per-run
API token) and replaces serve_with_routing_setup(). Nothing checks the
token yet.

Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
The shell generates the backend's API token at startup, keeps it in
Tauri state and starts the backend with it, so the backend and the
renderer share one token per run.

The renderer reads it with window.fleetify.getApiToken(), backed by the
new backend_api_token command. Like every app command, it is allowed
only for the main webview, through the app-commands set.

Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
TestServer::client() builds the reqwest client the integration tests
send their requests with, from one place, with room for headers every
request should carry. The access tests use it for every request that
isn't sent over raw TCP.

Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
The access guard's tests and the telemetry client's tests each had a
copy of the same tracing capture; they now share one in test_support.
tower-http's cors feature has had no user since the CORS layer went
away, so the workspace no longer enables it.

The access tests also check that the guard's 403 is plain text, and
that a cross-origin preflight reaches the router instead of being
refused by the guard.

Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
The renderer asks the shell for this run's backend token once and sends
it with every request to the backend. Fetches carry it in the
X-Fleetify-Token header; WebSocket and EventSource URLs, which can't
set a header, carry it as the api_token query parameter.

apiFetch adds it next to the access token. The new backendFetch sends
only the token, for the routes that don't check who is signed in, and
the call sites that used a bare fetch now use one of the two. Renaming,
pinning and reordering agents go through apiFetch, since their route
checks the access token. The status stream and the Browser panel's URL
socket open once the token is in hand, and not after their effect has
been cleaned up.

A source scan in the tests keeps bare backend fetches, and sockets or
streams without the token, from coming back.

Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
The access guard now also asks API requests for this run's token. It
reads it from the X-Fleetify-Token header. A WebSocket upgrade under
/ws/ and GET /events/status may carry it in the api_token query
parameter instead, since their URLs can't have a header. A missing,
wrong or repeated token gets the same plain-text 403 as a wrong host.

No token is needed for /healthz, for GET and HEAD of the UI files, for
the /api/internal routes (their handlers check the internal secret) or
for the system browser's MCP OAuth callback.

Release builds always require the token. A debug build running with
FLEETIFY_NOAUTH=1 doesn't, unless FLEETIFY_REQUIRE_API_TOKEN=1 is set;
both are read once at startup.

The integration tests that sign their own JWTs send the token through
the shared test client, which now sets it by default.

Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
The file tree's create, delete and rename now send what the workspace
file API reads. Create sends { path, is_dir }, so "New folder" makes a
folder instead of an empty file. Delete sends its path as a JSON body.
Rename sends { from, to }, where to is the new name in the same folder,
joined with the separator the path already uses; a name that is empty,
"." or "..", or contains a separator is refused without a request.

All three take success from the reply's ok field, since the API also
answers 200 when an action fails.

Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
…y policy

A release build now embeds the built UI in the executable, and the
embedded backend serves exactly those files, each with a
Content-Security-Policy, X-Content-Type-Options and Referrer-Policy.
A path that is not one of the files gets an empty 404, and hashed
assets may be cached for a year. The bundle no longer ships a dist
folder, Tauri's own CSP setting is off, and `tauri build` stops when
the UI has not been built.

Browsers report what the policy blocks to POST /api/csp-report, which
needs no API token; each report is logged on the fleetify.csp target.

Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
…ating the window

The Preview pane pins only local dev servers (localhost or 127.0.0.1),
the hosts the content security policy lets it frame, and says so when
it refuses an address. The marketplace dialogs already cancel the
browser's own form submission; tests now keep it that way.

Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
…click

Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
…replies

Create PR's trace is logged and shown when a step fails. Its lines now
go without the user name and password a remote URL can hold, the
origin URL it records included, and they are removed before a line is
cut short. The failure diagnosis sends the AI provider `git remote -v`
without them, and adding the origin remote answers with the existing
URL without them. A remote URL that doesn't parse loses them as well.

parse_origin reads the host after dropping them, so Create PR through
the GitHub link accepts an origin URL that holds them.

Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
Open in editor now checks the path the way Reveal does before anything
looks at it: it must be absolute and exist, and a Windows network or
device path is refused. The editor is started with the resolved path.
A refused path gets the usual { ok: false, error } answer.

Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
…indows

Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
The backend binds 127.0.0.1 with the default socket options again. On Windows the default bind on a specific address already refuses a second bind to the same address, and an exclusive bind can keep a quick relaunch from binding the port while a peer still holds its end of an old connection open.

Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
Both tests failed on every Windows CI run. Because the test step stops
at the first failing test binary, no backend integration test ran there.

run_command_times_out_cleanly asked cmd.exe to find `ping` on PATH, and
in the CI test process that lookup fails ("'ping' is not recognized").
Naming %SystemRoot%\System32\ping.exe skips the lookup. The test still
checks that a command running past its timeout comes back as a
"timed out" error.

build_ssh_args_expands_tilde_in_identity_file read the home directory
from USERPROFILE, which earlier tests leave pointing at a temp dir. On
the runner that dir is under C:\Users\RUNNER~1, so the expanded path
contained a `~` and `!path.contains('~')` failed although the expansion
was right. build_ssh_args now takes the home directory from its caller;
the test passes one that contains a `~` and checks the exact -i value.

Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
Without --no-fail-fast, cargo test stops after the first test binary
that has a failure. On the Windows runner that meant a lib-test failure
in fleetify-backend kept every backend integration test from running.
The step stays advisory.

Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
run_command spawns every command through `cmd /C` with the PATH the app
inherited. cmd.exe truncates a PATH longer than 8191 characters, and then
not even System32 programs are found; the dev launch chain grows PATH
well past that. The CLI-agent bridges already give their children the
user's canonical PATH through apply_clean_path, and run_command now does
the same. It is a no-op off Windows and keeps the inherited PATH when
the registry can't be read.

A Windows-only test runs a bare `where cmd` with the process PATH padded
past the limit. The timeout test still names ping.exe by full path, so
it stays about the timeout alone; its comment now says why.

Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
The comment gave a count of failing tests and a cause that no longer
hold. It now says only what the step is: advisory, and run on the
Windows runner.

Signed-off-by: Doruk Özgen <140007919+dorukozgen@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant