The sending console now dials the receiving one and pushes the job itself.
The receiving console hands out a one-job ticket (c2c.allow) for the sender's
key, and the sender shows it in its handshake (c2c.send). The receiver takes it
as an ordinary upload into the ticket's one root. Each side holds the other to
that one job: no calls, no other job, no other root.
The engine tries the direct route first and falls back to its relay when the
consoles can't connect, a helper is older, or the direct send stops. The job on
the receiving console is the engine's own under the same id, so either route
resumes what the other left. The queue row shows the route ("Straight between
the consoles", Beta) or "Through this computer" and why.
Verified on two PS5s (FW 13.60): 842 MB / 401 files direct in about 11 s, both
directions; a cancel then resume; checksums identical. Spec: SPEC.md §18.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ub1mez9eaVtTF8pMEFXvA9
Engine
The Rust workspace: everything between the desktop app and the PS5.
The engine is a local HTTP service. The Tauri app doesn't talk to the
console directly — it calls the engine, which speaks AVA1 to the payload.
That split is deliberate: the same API is reachable from a browser, a
script, or CI, so nothing the app can do is locked inside the app. See
the self-hosted-engine entries in ../FAQ.md.
client (Tauri/React) ──HTTP──▶ ps5upload-engine :19113
│ AVA1 (Noise XX, sealed frames)
▼
payload :9120
Crates
| Crate | What it is |
|---|---|
ava1 |
The AVA1 protocol: frames, generated message codecs, Noise handshake, sessions, lanes, pairing. Spec: protocol/ava1/SPEC.md. |
ava1-gen |
Generates the Rust and C codecs from protocol/ava1/schema/ava1.toml. Run it after editing the schema. |
ava1-ctest |
Builds the payload's AVA1 C on the host so cargo test checks C against Rust. |
ava1-chaos |
A misbehaving TCP proxy (latency, bandwidth caps, blackholes, kills) for resilience tests. |
ps5upload-core |
The bulk of the logic — socket helpers, the management seam, archive inspection and plan previews, filesystem and app RPCs, volume parsing, package install, saves, cheats, hardware, SMB, BPS patching. |
ps5upload-engine |
The Axum HTTP service (~100 routes), job tracking, SSE progress events, and the desktop + mobile entry points. |
ps5upload-pkg |
.pkg parsing — headers, entries, split-file sets. |
ps5upload-lab |
CLI for driving the console by hand over AVA1: management calls, uploads, benchmarks (bench), pairing and chaos tools. Useful when you want one call, not a workflow. |
ps5upload-tests |
Integration tests that run against in-process and loopback AVA1 peers, with no console. |
Working on it
cargo test --workspace # no PS5 required — loopback peers
cargo test -p ava1-ctest -- --test-threads=1 # payload AVA1 C vs Rust, on the host
cargo fmt --all -- --check # CI fails on this; client/src-tauri has its own
cargo build --release -p ps5upload-engine
Run it standalone and point it at a console:
PS5_ADDR=<ip> cargo run -p ps5upload-engine
curl "http://127.0.0.1:19113/api/ps5/status?addr=<ip>"
PS5UPLOAD_ALLOW_IP lets another machine reach it. The API is
unauthenticated and can read, write and delete files on the console —
keep it on a trusted LAN.
Testing
engine/crates/ps5upload-tests/tests/ runs against a loopback AVA1 console
(the Rust job host plus a scripted management node): single-file, folder,
small-file, resume-after-drop, retry classification, exclude rules,
.zip/.7z/.rar streaming, hardware commands, and volume parsing. Unit
tests live beside their modules.
Anything touching real transfer behaviour still needs hardware — see
../TESTING.md.
Things that will bite you
- The payload emits JSON; serde parses it. Field names must be
snake_caseor serde silently leaves the field at its default — a zero or an empty string, not an error. One malformed byte rejects the whole response, so anything interpolated into a payload-side JSON string has to be escaped first. ok: falseinside a 200 is a refusal, not a success. Several payload commands answer that way.client/src/api/ps5.tshas anassertOk()for action endpoints; status endpoints keepok:falseas data because there it means "unsupported on this console".- One port, one session. Everything goes to 9120 as AVA1: management calls on the session's control connection, bulk data on its lanes.
- A stream install has the PS5 fetch the package from the engine. The
engine tells the console an address to connect to; behind Docker's bridge
network, a VPN or a virtual adapter that address is one the PS5 cannot
reach.
PS5UPLOAD_PKG_HOST_IPpins the LAN IP the console is given (publish port 19113 when the engine is in a container;compose.yamluses host networking instead). Every stream-unreachable failure — the pre-install reach check, a Sony refusal before any byte was fetched, and an accepted install the console never fetched from — ends ininstall/mod.rsstream_unreachable_hint_for, which names this variable.
Where the web UI's file browser opens (PS5UPLOAD_BROWSE_ROOTS)
The browser-based UI picks local files and folders from the engine's filesystem. By default
that picker opens at the engine's home directory, which in the Docker images is /data, the
engine's own state volume, not the folder where you mounted your packages. Set
PS5UPLOAD_BROWSE_ROOTS to a comma-separated list of directories (for example /pkgs, or
/pkgs,/games) and the picker starts there instead. compose.yaml sets it to /pkgs, next to
the volume mount it describes.
- Unset or blank keeps the old behaviour (the home directory).
- An entry that is not a directory the engine can see is skipped and named in the engine log (a warning), so a typo or a volume that was never mounted is visible. If no entry is usable the picker falls back to the home directory.
- It only chooses where the picker opens. It is not a sandbox: the engine can still be asked for any path it can read.
HOMEis deliberately left alone, because FPKG conversions write under$HOME/Downloads/fpkgsand a package mount is usually read-only.
Synology and other NAS boxes (saved servers: "Permission denied")
The Docker images keep their state (saved servers, install history, artwork cache) in /data
and run as UID:GID 65532 unless you override user:. If you bind-mount a host folder that
someone else owns, writing fails and the engine says so, naming the folder, the UID:GID it runs
as, and the owner it found (#361). Fix it one of two ways:
- set
user:in the compose file to the folder's owner (ls -lnon the host shows the numbers), for exampleuser: "1234:77777"; or chown -R 65532:65532 <host folder>for the mount.
USER_ID / GROUP_ID environment variables are not read by these images; use user:.
The account must also be allowed to write that folder in the NAS's own permission settings.