OpenFUT/OPENFUT_HANDOFF.md

# OpenFUT — project handoff

Self-contained briefing for an assistant with **no access to this repo or machine**.
Everything needed to understand the project and reason about its open problems is here.

---

## 1. What this is

**OpenFUT** is a clean-room, fully offline re-implementation of the server backend for
**FIFA 17 Ultimate Team (FUT)**. EA's servers for FIFA 17 are long dead. The goal is to
make the retail game's FUT mode fully playable again — open packs, build squads, use the
transfer market, play matches and earn rewards — by emulating every server the client
talks to, on localhost.

Nothing is decompiled *into* the project. The game's binaries are read to learn the
**wire format** (which JSON keys, of which types, each response must contain), and the
servers are written from scratch in Python against that spec.

The client is unmodified retail FIFA 17 running under Wine/Proton on Linux.

---

## 2. Architecture — four independent servers

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 over TCP. Local Origin client emulation.
  │                             Login, entitlements, persona.
  ├─ Blaze              :42127 redirector (TLS) → :42130 game server, :42131 nucleus
  │                             EA's binary "Fire2/TDF" RPC protocol. Session, auth,
  │                             and — critically — the CLIENT-CONFIG STORE.
  ├─ UTAS / RS4         :8099   The FUT REST API. JSON over HTTP. ~45 endpoints.
  │                             Club, squads, packs, market, matches. The bulk of it.
  └─ POW / EASFC        :8094 (+ :8080 content)   A third HTTP API, discovered late.
                                Online status bar, level, EASFC credits, catalogue.
```

Plus two helpers: a **roster** server (:8081) serving a roster-update XML the FUT
loading screen blocks on, and **autopatch**, which patches the running process's
ProtoSSL certificate verification so the client accepts our self-signed TLS.

### How the client is redirected

Three mechanisms, in decreasing order of preference:

1. **Blaze client-config keys** — the cleanest. Blaze serves a key/value config store
   and the client reads its own service URLs from it. `FUT_RS4_APIURL_<MODULE>` and
   `FUT_RS4_URL_<CALL>` point the FUT API at `127.0.0.1:8099`; `FIFA_POW_URL` points
   the EASFC layer at `127.0.0.1:8094`. **No root, no DNS games.**
2. **`/etc/hosts`** — for hosts baked into the binary (`easw.easports.com`,
   `gosredirector.ea.com`).
3. **iptables DNAT** — for a hardcoded IP (`159.153.51.20` → the Blaze redirector).

---

## 3. 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** (an "atom"). This has three consequences that dominate the project:

**Atoms.** Every JSON key name maps to a 16-bit atom id via FNV-1a. There is a recovered
table of ~900 (id → name). A response is really "which atoms does deserializer X read,
and of what type".

**Type fidelity is fatal.** Feeding a scalar where the parser expects an object or array
does not error — it **desyncs the token reader and the game hard-freezes** in a busy loop
at `0x1801c7f1a`. This is the single most common way to break the game, and it has bitten
this project repeatedly. Arrays must be arrays; nested objects must be objects.

**Unknown keys are usually skipped safely** — most deserializers route an unrecognised
atom to a value-skip handler (`FUN_180135ff0`), so extra fields are inert. **But not
always**: at least one deserializer (`FutMoveCard`, `0x180128600`) has *no* skip handler
at all, so any unexpected key leaves its value unconsumed and desyncs the reader.

### Working method

For any endpoint: find the response class's deserializer, extract the atom ids it
compares against, map them to names, note the getter used for each (int / string / bool /
nested), and build the minimal body. Omit nested members unless their shape is known —
omission is skip-safe, a wrong shape freezes the game.

---

## 4. What works today (live-verified)

| Area | Status |
|---|---|
| Boot to the FUT hub | ✅ |
| Club identity, coins, W/D/L record | ✅ |
| Active squad — renders 11 real players, chemistry links | ✅ |
| Squad building — `PUT /squad/<id>` fires, persists across relaunch | ✅ |
| Squad roster ("MY SQUADS") | ✅ |
| Transfer market — browse, bid, buy-now, sell, watchlist | ✅ |
| Packs — buy, cards land in club | ✅ (via a workaround, see §5) |
| Quick sell — destroys cards, credits coins | ✅ |
| Match loop — create/ready/play/destroy + coin rewards | ✅ implemented, **never played in-game** |
| Online/EASFC status bar (no "servers unreachable") | ✅ when enabled |
| Seasons / tournaments / leaderboards / champions | ⚠️ routed with reversed schemas, never live-tested |

Current save state: 99 club items, 8,400 coins, 0-0-0 record.

**Design convention:** every risky change ships behind an environment flag, default set
to whatever is live-proven (`FUT_MASSINFO`, `FUT_USERINFO`, `FUT_MODES`,
`FUT_PACK_AUTOCLUB`, `FUT_POW`, `FUT_STORE_GROUPS`, …). This exists because two working
screens were broken by shipping "corrections" on by default.

---

## 5. Open problems

### 5a. "Send to Club" kills the FUT session — UNSOLVED

Opening a pack shows the cards correctly. Choosing **Send to Club** results in:

> *"We are sorry but there has been an error connecting to FIFA 17 Ultimate Team. You
> will be returned to the FIFA 17 Main Menu."*

The client then POSTs `ut/delete/auth` (logout) about a second later. The cards **do**
move correctly server-side every time — only the acknowledgement is rejected.

**Eliminated by live test:**

| Hypothesis | Result |
|---|---|
| Missing `chemistry` field | Failed without it too |
| Extra keys + no skip handler in the deserializer | Real finding; sending *only* the parsed key still failed |
| The response body at all | **A bare `{}` also fails** |
| A network failure ("error connecting") | Zero non-loopback connections during a failure — that string is FIFA's *generic* session-error text |
| Missing per-call service URLs | 146 were genuinely missing and are now served; unchanged |
| The POW layer saturating HTTP | Same failure with POW fully disabled |
| The reveal screen's exit path | **"Quick Sell All" from the same screen works** |
| The club not being loaded | Loading MY CLUB first, then moving, still failed |

**The key asymmetry:** quick sell (`POST ut/delete/<sku>/item`) and move
(`PUT ut/<sku>/item`) both receive an identical bare `{}` from the same server, with the
same headers and status. Quick sell succeeds; move kills the session. Whatever decides
this is **client-side state, not the wire**.

**Workaround in place:** pack contents are deposited straight into the club at open time
and the pending pile is kept empty, so the client is never offered a move. Packs are
fully usable; the cost is that the reveal screen shows nothing to assign.

### 5b. "MY CLUB" counter always reads 0 — UNSOLVED

The hub tab bar shows `MY CLUB  0` despite the club holding 99 items.

**Eliminated by live test:**
- **Not the item list** — the user opened MY CLUB, the client fetched the club and
  **displayed all 99 players correctly**, and the counter still read 0.
- **Not `pileSizeClientData`** — the massinfo member that carries pile sizes as
  `{"entries":[{"key":int,"value":int}]}`. A probe sent 16 entries with uniquely
  identifiable values; the counter stayed 0.
- **Not lazy loading** — the club endpoint was fetched 6 times that session.

**Unexplored contrast:** the `ACTIVE SQUAD` tab in the same bar correctly shows `11/23`.
So some counters work. Whatever differs between that one and the club one is likely the
answer.

### 5c. Store tiles render "unknown" — DIAGNOSED, NOT FIXED

Pack tiles show `unknown` with zero item counts. The `"unknown"` string is an
unconditional **default** in a string constructor — the field simply never gets written.
The store renders *display groups*, and `displayGroup` is parsed **recursively by the same
element parser**. Sending it populated **froze the store** (the type-desync busy loop), so
it is behind a flag, default off. Doing it properly needs the group's own field set worked
out rather than a self-referential copy of the pack.

---

## 6. Notable reverse-engineering findings

- **Class → deserializer resolution.** A response class's name literal is preceded by a
  **4-byte header**, and the constructing factory's `lea` points at *the header*, not the
  text. Lookups must use `name_address - 4`. Six attempts failed on this off-by-four;
  four of them returned zero results and nearly got recorded as "this class has no
  deserializer".
- **The request-template table is a floor, not a ceiling.** Several real endpoints are
  built by the caller appending a suffix and therefore never appear in the binary's URL
  table: `squad/list`, `user/club`, `club/stats/*`, `clientdata/<key>`. Only live traffic
  reveals them. This has caught the project three separate times.
- **The documentation lies.** The project's own `ENDPOINT_MAP.md` (~1,360 lines, ~100
  reversed structs) has been wrong repeatedly: it claimed `FutMoveCard` parses
  `chemistry` (it does not); it called seven store pack fields "skipped no-ops" (all are
  parsed); it gave price-object keys as `amount`/`currency` (the parsers read
  `externalPriceId`). **Verify against the decompiler before relying on any row.**
- **The online layer was hiding in an unpacked DLL.** "EA FC servers are unreachable"
  comes from a third HTTP API implemented in a *loose, unpacked, string-rich* library —
  not the packed executable, and not any layer previously emulated. It is redirectable
  purely through the Blaze config store.
- **The main executable is Denuvo-packed.** Its code exists only in a live process. Live
  memory is readable via `/proc/PID/mem` (the PE is mapped flat), which is how a crash
  site was disassembled. Any logic living there cannot be reversed statically.

---

## 7. Tooling built

- A **PyGhidra harness** with helpers for decompiling, xrefs, vtables, byte scanning and
  class→deserializer resolution. (Ghidra's Java/OSGi scripting is broken on this machine;
  PyGhidra bypasses it entirely.)
- A **minidump reader** — exception record, fault-time registers, module map, and a stack
  walk that recovers a usable backtrace.
- A **live code grabber** that reads and disassembles unpacked code out of a running
  process.
- A **read-only live probe** pattern for polling client model state while playing.
- **Two test suites**: 380 live contract checks (freeze-safety: asserts every response
  field's type against the reversed schema) and 51 pure unit checks for match rewards.
- A **traffic-replay audit** that diffs current responses against a known-good session —
  this is what proves a change did not alter what the client sees.

---

## 8. Documentation in-repo

| File | Contents |
|---|---|
| `ENDPOINT_MAP.md` | ~100 reversed response structs, atoms, types, freeze risks |
| `FUT_RESPONSE_REBUILD_PLAN.md` | Squad family, massinfo, the service-layer call graph |
| `REBUILD_RESEARCH.md` | Complete API surface, gap analysis, POW layer, and every eliminated hypothesis with its evidence |
| `CARD_SYSTEM.md` | How card identity resolves locally from the game's own database |
| `REPACK_INTEL.md` | Origin/LSX emulation notes |

---

## 9. Where help would be most valuable

1. **The move-vs-quick-sell asymmetry (§5a).** Two sibling endpoints, identical responses,
   one works. What client-side precondition could a "move item between piles" operation
   have that a "discard item" operation does not — a loaded destination collection, a
   known pile capacity, a valid target index?
2. **The MY CLUB counter (§5b).** Given the client demonstrably *has* the items and
   *renders* them, what else could a tab counter read from? Note that a sibling counter in
   the same bar works correctly.
3. **Whether either is fixable server-side at all**, or whether the honest answer is that
   the deciding logic lives in the packed executable and only live instrumentation can
   settle it.

Useful framing: this project's failures have almost always come from proposing a fix
before testing the assumption under it. Hypotheses that come with a cheap way to
disconfirm them are worth far more than plausible ones.