OpenFUT/OPENFUT_PROJECT_REPORT.md

# OpenFUT — Project Report

**Goal:** make FIFA 17 Ultimate Team fully playable offline, forever, by re-implementing
every server the game talks to.

**Status:** FUT boots, loads, and is playable. Packs, squads, the transfer market, coins
and progression all work. Two cosmetic/flow problems remain open.

**Timeline:** 2026-06-25 → 2026-08-04 · 44 commits · ~12,900 lines of Python across 39
tools · ~3,600 lines of reverse-engineering documentation.

---

## 1. What the project is

EA shut down FIFA 17's servers years ago, which kills Ultimate Team — the mode is entirely
server-driven. Your club, squads, packs, market and progression all live server-side, so
without a backend the mode is dead even though the game still installs and runs.

OpenFUT replaces that backend with local servers. The game is **unmodified retail
FIFA 17** running under Wine/Proton on Linux; nothing is patched into the game except a
single runtime tweak so it accepts our TLS certificate.

This is **clean-room work**. No EA code is copied or redistributed. The game's own
binaries are read to learn the *wire format* — which JSON keys, of which types, each
response must carry — and the servers are written from scratch against that specification.

---

## 2. Project history

The project changed target twice before finding its footing. That arc matters, because
each pivot was driven by hitting a hard wall.

**Phase 1 — FIFA 23 (June 2026).** Began as an offline FUT backend for FIFA 23: a Rust
core (`openfut-core`, Axum + SQLite), a protocol bridge (`openfut-bridge`), and a GUI
launcher. All three built and passed tests. The architecture was sound but the client
never got far enough to exercise it.

**Phase 2 — the FIFA 23 wall.** FIFA 23 refused to go online at all. Extensive reverse
engineering of the connection state machine, live-memory probing, and forcing the
"go online" gate directly all failed — the client's internal coherence checks were the
wall, not any single flag. Documented as a dead end rather than fought.

**Phase 3 — the FIFA 17 pivot (late July).** FIFA 17 turned out to be a far better target:
its network library is **unprotected and fully symboled**, exposing 979 RPC names. The
insight was to crack FIFA 17 first and port the understanding back.

That worked, quickly:
- **TLS pinning defeated** — the client's certificate verification is patched at runtime
  in memory, so a self-signed cert is accepted.
- **Origin/LSX emulation** — the local Origin client protocol was reverse engineered,
  including the repack's own crypto layer, beating "log in to Origin" and
  "title version outdated".
- **Blaze cracked end to end** — EA's binary RPC protocol: redirector, second-hop
  handshake, and the encoded pre-auth exchange. This was the big one.
- **Full FUT API mapped** — ~100 response structures reverse engineered to field level.

**Phase 4 — building the FUT backend (August).** With the protocol understood, the work
became making FUT actually *play*: card rendering, packs, the store, the transfer market,
squads, and match rewards. This is where the project stands.

---

## 3. Architecture

FIFA 17 does not talk to one backend. It talks to **four**, on different protocols, and
all four must be satisfied in sequence before FUT loads.

```
FIFA 17 (Wine/Proton)
  │
  ├─ LSX / Origin     :4216    XML/TCP — local Origin client emulation
  │                            login, entitlements, persona
  ├─ Blaze            :42127 redirector (TLS) → :42130 game, :42131 nucleus
  │                            EA's binary Fire2/TDF RPC
  │                            session, auth, and the CLIENT-CONFIG STORE
  ├─ UTAS / RS4       :8099    the FUT REST API — JSON/HTTP, ~45 endpoints
  │                            club, squads, packs, market, matches
  └─ POW / EASFC      :8094    a third HTTP API (+ :8080 content)
                               online status, level, credits, catalogue
```

Plus **roster** (:8081), serving an XML file the FUT loading screen blocks on, and
**autopatch**, which patches certificate verification in the running process.

### Redirection, in order of preference

1. **Blaze client-config keys** — the client reads its own service URLs from a key/value
   store that Blaze serves. Pointing FUT and EASFC at localhost needs **no root and no
   DNS manipulation**. This is the clean mechanism and most redirection uses it.
2. **`/etc/hosts`** — for hostnames baked into the binary.
3. **iptables DNAT** — for one hardcoded IP address.

---

## 4. The wire format — and why it is unforgiving

FUT responses are JSON, but the client does not use a general JSON object model. Each
response class has a hand-written SAX-style deserializer that walks tokens and dispatches
on a **hashed key id** ("atom"). Three consequences dominate the project:

**Atoms.** Every JSON key maps to a 16-bit id via FNV-1a. A recovered table of ~900
id→name pairs is the Rosetta stone. A response spec is really "which atoms does this
deserializer read, of what type".

**Type fidelity is fatal.** A scalar where an object or array is expected does not error —
it **desyncs the reader and hard-freezes the game** in a busy loop. This is the primary
failure mode of the entire project.

**Unknown keys are usually skipped — but not always.** Most deserializers route
unrecognised atoms to a skip handler, making extra fields inert. At least one does not,
so any unexpected key desyncs it.

**Working method:** locate the deserializer, extract its atom set and per-atom getter
types, build the minimal body, and omit nested members whose shape isn't known — omission
is safe, a wrong shape freezes the game.

---

## 5. What works

| Capability | State |
|---|---|
| Boot: Origin → Blaze → FUT hub | ✅ |
| Club identity, coins, W/D/L record | ✅ |
| Active squad — 11 real players, ratings, chemistry | ✅ |
| Squad building — saves and survives relaunch | ✅ |
| Squad roster ("MY SQUADS") | ✅ |
| Transfer market — browse, bid, buy-now, list, watchlist | ✅ |
| Store — buy packs | ✅ |
| Packs — cards land in the club | ✅ via workaround (§6a) |
| Quick sell — credits coins | ✅ |
| Match loop — create/ready/play/destroy + rewards | ⚠️ built, **never requested by the client** (see below) |
| Online/EASFC status bar | ✅ when enabled |
| Seasons, tournaments, leaderboards, champions | ⚠️ routed, **never requested by the client** (see below) |
| Card identity (names, faces, ratings) | ✅ resolves from the game's own local database |

### "Untested" is two different things, and the difference matters

The server log records the User-Agent of every request. The real client identifies as
`ProtoHttp`; this project's own curl and Python probes do not. Separating them shows that
several endpoints previously filed as "built but untested" have in fact **never been
requested by the game at all**, and everything recorded against them was self-inflicted
traffic:

| endpoint | client requests | project probes |
|---|---|---|
| `/leaderboards/options` | 5 | 1 |
| `/clientdata/userHubData` | 13 | 2 |
| `/user/accountinfo` | 23 | 49 |
| `/season`, `/season/user` | **0** | 2 |
| `/tournament`, `/tournament/user` | **0** | 3 |
| `/leaderboards` (bare) | **0** | 2 |
| `/champion` | **0** | 2 |
| `/match` | **0** | 2 |
| `/clubUser` | **0** | 93 |
| `/user/list` | **0** | 180 |
| `/sbs` (SBC), `/draft/mode` | **0** | 0 |

`/clubUser` and `/user/list` are the starkest: 273 requests between them, none from the
game. Work was done on both on the assumption the client wanted them.

A zero in the client column does **not** mean the client never wants that endpoint. In
most cases it means **nobody has navigated to that part of the game yet**. It does mean no
claim about those endpoints has been tested against the client, and any analysis that does
not apply this filter is misleading by default.

**Requirement:** every capture and analysis tool in this project should apply the
User-Agent split by default rather than as an afterthought.

**Design convention.** Every risky change ships behind an environment flag whose default
is whatever is live-proven. This exists because shipping "corrections" on by default broke
two working screens — once freezing the store outright.

---

## 6. Open problems

### 6a. "Send to Club" ends the FUT session

Opening a pack displays the cards correctly. Choosing **Send to Club** produces *"there
has been an error connecting to FIFA 17 Ultimate Team"* and a logout. The cards **do**
move correctly server-side; only the acknowledgement is rejected.

Eliminated by live test: the response body (a bare `{}` fails identically), a missing
field, unknown-key desync, a network failure (zero external connections during a
failure — that error string is FIFA's *generic* session error), missing service URLs,
the online layer, the reveal screen's exit path, and the club not being loaded.

**The asymmetry:** quick sell and move receive an *identical* bare `{}` from the same
server. Quick sell succeeds. Move ends the session. The deciding factor is client-side
state, not the wire.

**Workaround:** pack contents are deposited straight into the club at open time, so the
client is never offered a move. Packs are fully usable; the reveal screen just doesn't
offer assignment.

### 6b. "MY CLUB" counter reads 0

The hub shows `MY CLUB 0` despite 99 items. Not the item list (the client *displays* all
99), not the pile-size data (a 16-entry probe changed nothing), not lazy loading (fetched
6 times). Unexplored: the neighbouring `ACTIVE SQUAD` counter works correctly — the
difference between them is likely the answer.

### 6c. Store tiles read "unknown"

`"unknown"` is an unconditional default in a string constructor — the field is never
written. The store renders *display groups*, and that member is parsed recursively by the
same parser; sending it populated **froze the store**, so it is flagged off pending a
correct group schema.

### 6d. Large parts of the game have never been opened

Distinct from 6a to 6c, which are things that misbehave. Per the User-Agent table in
section 5, entire modes have never issued a single client request: Seasons, Tournaments,
FUT Champions, Draft, SBC, and the match loop itself. Their endpoints are routed and their
schemas are reversed, but no claim about any of them has been tested against the game.

This is not a bug list. It is unmeasured surface, and it is the cheapest information
available to the project because most of it costs nothing but navigating menus. It is
recorded here because "routed from a reversed schema" reads like a stronger claim than it
is, and section 7 warns that this project's notes have described things differently from
what is true.

*A multi-agent investigation into 6a and 6b is currently running.*

---

## 7. Notable findings

- **Class → deserializer resolution.** A response class's name literal is preceded by a
  4-byte header and the factory points at *the header*. Six attempts failed on that
  off-by-four; four returned nothing and were nearly recorded as "no deserializer exists".
- **The URL table is a floor, not a ceiling.** Several real endpoints are built by
  appending a suffix at the call site and never appear in the binary's template table.
  Only live traffic reveals them — this caught the project three separate times.
- **The project's own documentation has been wrong repeatedly** — fields described as
  inert turned out to be parsed, and documented key names didn't match the parsers.
  Verify against the decompiler, not the notes.
- **The online layer was hiding in plain sight.** "EA FC servers unreachable" comes from a
  third HTTP API in a *loose, unpacked, string-rich* library — not the protected
  executable, and not any previously emulated layer. Redirectable purely by config.
- **The main executable is Denuvo-packed**, so its code exists only in a live process.
  Live memory is readable, which is how a crash site was disassembled — but logic living
  there cannot be reverse engineered statically.

---

## 8. Tooling and quality

- **PyGhidra harness** with decompile / xref / vtable / byte-scan / class-resolution
  helpers (Ghidra's own Java scripting is broken on this machine).
- **Minidump reader** — exception record, fault-time registers, module map, stack walk.
- **Live code grabber** — reads and disassembles unpacked code from a running process.
- **Read-only live model probes** for watching client state while playing.
- **Test suites:** 380 live contract checks (type/freeze safety per reversed schema) plus
  51 pure unit checks. Both green.
- **Traffic-replay audit** — diffs current responses against a known-good session to prove
  a change didn't alter what the client sees.
- **~3,600 lines of RE documentation** across six files, including every eliminated
  hypothesis with its supporting evidence.

---

## 9. Roadmap

**Immediate (free, no code):** play a match — the reward loop is built and unit-tested but
has never run in-game. Enable the game-mode endpoints and see whether four more modes
light up.

**Near term:** close the two open problems, most likely via live instrumentation rather
than more static analysis. Implement SBC and Draft (schemas already recovered). Fix the
store display groups properly.

**Longer term:** the stated destination is porting this into the Rust `openfut-core`
behind a FIFA-17 bridge. Everything currently lives in Python prototypes; the
documentation is now good enough to write the port against.

---

## 10. Honest assessment

**What went well.** The FIFA 17 pivot was the decisive call — recognising that an
unprotected binary was worth more than persisting against a hardened one. Blaze, Origin
and the FUT API were all cracked end to end. The freeze-safety test suite has repeatedly
caught regressions before they reached the game.

**What went badly.** Progress has been slowest where fixes were proposed before the
assumption under them was tested. Both open problems absorbed many attempts built on
plausible but unverified theories; several were disproved in a single measurement that
could have been taken first. Two working screens were broken by shipping unverified
"corrections" on by default — which is precisely why the flag convention exists now.

**The most reliable technique** has been comparing a working case against a failing one:
diffing live traffic against a known-good session, and contrasting a succeeding endpoint
with its failing sibling. That has produced more answers than any amount of decompilation.