feat: rewrite the Epic and Amazon store modules (#802)

Both store modules are rewritten from a protocol-level spec: new Rust engines behind a small JNI surface (nativeCall/nativeInstall) and thin Kotlin facades, with the launcher's own user agent.

Epic: sign-in carries over, library, manifest download with byte-exact verification, assembly, uninstall, launch arguments, achievements.
Amazon: sign-in carries over, library, manifest download with per-file hash checks, path checks, executable detection from fuel.json, uninstall.
On-disk install layout is unchanged, so existing installs keep working.

Store game pages now take controller focus on their lead hero action (Install, Uninstall, Pause, Resume, Get, Buy), fall back to Back, and keep D-pad focus inside the page while it covers the library.
This commit is contained in:
The412Banner authored and GitHub committed 2026-10-14 07:20:06 -04:00
1 parent a849556f3f
commit 7ad3c60efe
105 files changed
+10975 -13233

No files matched your search

+175
View File
@@ -0,0 +1,175 @@
# Amazon Games module: clean-room rewrite
The Amazon Games integration (the Rust `amazon` module in `libdroiddeckstores.so` and the Kotlin
package `com.droiddeck.launcher.stores.amazon`) was rewritten so that it carries no code from the
earlier implementation. This page records how that was done and every judgement call the new code
makes.
## Process
1. **Spec, written separately.** A functional specification (six documents: summary and unknowns,
service protocol, manifest format, install behaviour, on-disk compatibility, and the app-facing
interface) was written from the old code by a separate pass. It describes behaviour and data
formats only, not code structure.
2. **Implementation from the spec only.** The implementer worked from those documents and from
code that is not Amazon-specific:
- read: the spec; the crate's `Cargo.toml`, `Cargo.lock`, `lib.rs`, `stores_jni.rs`,
`priority.rs`, `fetch_tuning.rs` and the shared `fetch_core.rs`; the clean-room Epic module
(Rust and Kotlin) as the structural template; `JNI.md`; the shared Kotlin store code (account
store, install records, download queue, login activity, catalog and state); the signatures and
call sites of `StoreSizes`' Amazon branch (to keep them compiling); build files and CI;
- not read: anything under `stores/amazon/` (Kotlin, Java, tests) or `src/amazon*` (Rust) of the
old tree, other branches or checkouts, or any Amazon client source. The old files were deleted
with `git rm` without being opened.
3. **Fresh code.** Every file has its own structure, names and comments, with a plain description
and `SPDX-License-Identifier: GPL-3.0-or-later` (the crate's licence). Where the spec left a fact
open, the choice was made from general knowledge of Amazon's public services or from
unauthenticated probes of the live endpoints (below), and is listed here.
## Layout
Rust, `app/src/main/rust/stores/src/amazon/`:
| File | Role |
|---|---|
| `net.rs` | blocking HTTP (reqwest + rustls), the three User-Agent identities, hex / percent / base64url helpers, URL scrubbing for logs |
| `auth.rs` | per-attempt device serial, client id, PKCE verifier and challenge; sign-in URL; code from the return URL; registration, refresh, deregistration |
| `distribution.rs` | the distribution service call shape (`X-Amz-Target`, `x-amzn-token`, `Content-Encoding: amz-1.0`), refusal detection, the download lookup |
| `library.rs` | entitlement pages, field fallbacks, de-duplication, add-on filter, sort |
| `cdn.rs` | path insertion before the signed query, file URLs, the manifest GET |
| `codec.rs` | manifest framing (big-endian header length) and xz / LZMA-alone decompression (`lzma-rs`) |
| `proto.rs` | a strict protobuf wire reader (and test writers) |
| `manifest.rs` | packages → files, path normalisation and safety, duplicate handling |
| `install/` | `mod.rs` (the run), `check.rs` (completeness, stray `.tmp`, free space), `writer.rs` (streamed download, SHA-256, atomic placement), `tests.rs` (loopback end-to-end) |
| `launch.rs` | `fuel.json`, exe scoring, the FuelPump environment |
| `service.rs`, `ffi.rs` | named operations and the JNI boundary |
Kotlin, `app/src/main/java/com/droiddeck/launcher/stores/amazon/`: `AmazonBackend` (the
`StoreBackend`, installs the login flow), `AmazonLogin` (the WebView flow), `AmazonCredentialStore`
(the account through `StoreAccounts`, single-flight refresh), `AmazonPrefs`, `AmazonLibrary` (sync,
cache, posters, publishing), `AmazonInstaller` (the queue job), `AmazonNative` (JNI), and two thin
wrappers kept for `StoreSizes`: `AmazonApiClient` (`getGameDownload`, `getBytes`, `appendPath`) and
`AmazonManifest` (`parse(bytes).totalInstallSize`).
The JNI surface (`nativeConfigure`, `nativeCall`, `nativeFetch`, `nativeInspect`, `nativeInstall`,
`nativeCancel`, `nativeRelease`) is documented in `app/src/main/rust/stores/JNI.md`. The old
`nativeStart` symbol is gone. The Java xz library stays in `app/build.gradle`: commons-compress
needs it for the `.tar.xz` Windows components, so only Amazon stopped using it.
## Compatibility kept exactly
- Credentials (`stores/amazon/credentials.json` through `StoreAccounts`): the same keys
(`access_token`, `refresh_token`, `device_serial`, `client_id`, `expires_at`, `display_name`). A
refresh merges `access_token`, `device_serial`, `client_id`, `expires_at` into the stored record
and keeps every other key; a new sign-in writes the whole record. Registration and refresh keep
the launcher's device constants, so existing refresh tokens keep working.
- Preferences `stores.amazon`: `library_cache` (same element keys), `library_synced_at`,
`vcover_<id>`, `size_<id>`, `version_<id>`; sign-out removes only the first two.
- Install folders, the sidecar keys (`extra.entitlementId`, `extra.sku`, no `cloud` key), the
launcher name `.droiddeck-launch.bat`, the `.tmp` naming and the two marker files.
- The catalog item's `extra` (`entitlementId`, `sku`, `version` = `""`) so the shared size cache and
the saved queue keep working.
## User-Agents and what the probes showed
Unauthenticated probes of the live services (2026-10-14):
- **Distribution service** (`gaming.amazon.com`): `GetEntitlements`, `GetGameDownload` and
`GetLiveVersionIds` answered identically for `AGSLauncher/1.0.0`, `DroidDeck/…`, a curl UA, with
and without the `UserAgent` header. A bad token gives HTTP 401 `<NotAuthorizedException/>`; no
token gives HTTP 400 with `com.amazon.coral.service#NotAuthorizedException`. The probes cannot
show whether the UA matters for an authorised call, so the launcher's `User-Agent` and
`UserAgent` values are kept (Amazon's own client identifiers, one constant each in `net.rs`).
- **Identity API** (`api.amazon.com`): register, token and deregister gave the same validation
errors whatever the UA (Firefox, `DroidDeck/…`, none). The desktop Firefox string is kept for
these calls, the same as the WebView: it is a generic browser label, not a client name, and
existing refresh tokens are known to work with it.
- **Sign-in page** (`www.amazon.com/ap/signin`): every UA got Amazon's bot check from the probe
host, so nothing could be learned; the WebView keeps the desktop Firefox string (a sign-in page
that misbehaves would lock every user out).
- **CDN**: no URL is reachable without a token. Manifest and file requests now send
`DroidDeck/<versionName>`, the only third-party label the old integration sent; this needs the
device test below.
Refusal detection follows the probes: 401, 403, and 400 whose body names `NotAuthorizedException`
mean "sign in again".
## Judgement calls
Recommended fixes from the spec, adopted:
1. **Complete = size and SHA-256.** One test for the space check and the download plan; a
same-size corrupt file is repaired. Files without a SHA-256 (another algorithm) are judged by
size. No cross-run "verified" cache: every present file is hashed on each run.
2. **A failed library page fails the sync** and keeps the old cache; a refused token is reported
with the shared "session expired: sign in again" line, also when a cache is shown.
3. **Single-flight refresh** (`AmazonCredentialStore`); a forced refresh after a failed library call
skips the refresh when another caller already replaced the token.
4. **Mark updating** (`beforeReplace` → `StoreInstalls.markUpdating`) once, before the first file
is put in place (new or replacing; the shared step only acts on an installed record).
5. **Manifest path safety**: absolute paths, any `:` (drive letters and stream names; Windows file
names cannot contain it), NUL, `.` / `..` and empty segments, and DroidDeck's own root files
(`.droiddeck-store.json` and its `.tmp`, `.droiddeck-launch.bat`, the two markers, compared
ignoring case) fail the install naming the path. A manifest that lists both `X` and `X.tmp` is
refused (the second would collide with the first's temporary file). The protobuf reader rejects
truncation, over-long varints, groups and unknown wire types, field number 0, wrong wire types
on the fields used, and a non-empty file without a hash or a SHA-256 that is not 32 bytes.
6. **Exe scoring on relative paths.** The Unreal-folder rule is tested against `"/" + relative
path` so a top-level `Binaries/Win64/…` still matches. The "first five letters" rule compares the
title with the file name **without** `.exe` (the extension's letters would otherwise count). An
empty title never matches the "path contains the title" rule. Ties: larger file, then the earlier
path. Only regular files are candidates; symlinks are not followed; depth is capped at 16.
7. **CDN User-Agent** is `DroidDeck/<versionName>` (see above).
8. **Empty files** whose hash is absent or the SHA-256 of empty input are created locally; stray
`.tmp` siblings of manifest files are removed at run start (other `.tmp` files are untouched).
9. **Markers** `.amazon_download_in_progress` / `.amazon_download_complete` are still written.
10. **Duplicate paths**: the later entry wins with a log warning; files sharing one hash are fetched
once per path (no copy optimisation).
Other choices:
- **Install size** is the sum over distinct paths (equal to the spec's "duplicates included" sum
for every manifest without duplicate paths, which none have been seen to have); it is the
progress total, the disk size and `size_<id>`.
- **Free space** is measured in Rust (`statvfs` on the folder or its nearest existing parent),
which is what the shared probe measures; unknown never blocks, a measured zero does.
- **Pre-download check** is reported in the queue's Manifest stage ("Checking files…" with byte
progress), like GOG and Epic: the Downloads UI orders stages Manifest → Download → Verify →
Install, so a Verify stage before Download would read backwards.
- **Concurrency**: one streamed GET per file through the shared fetch core in stream mode (the
speed tier's window clamped to 1–128, per-host cap `max(6, ceil(window / hosts))`, 120 s to the
response headers, the core's 30 s idle cap on the body, its five attempts with back-off and its
stall and 429 handling). A piece at offset 0 restarts the file (a retry); an out-of-order piece,
a body longer than the manifest size, a short body or a hash mismatch deletes the `.tmp` and is
retried; a write error, a refused `beforeReplace` or a failed rename fails the run.
- **Decompression** uses the `lzma-rs` crate (pure Rust; xz with any check type and LZMA-alone),
with a 256 MiB cap on the decompressed manifest.
- **Live version lookup** (`GetLiveVersionIds`) is not implemented; nothing used it.
- **DLC filter**: add-ons are hidden only when they also name a parent product (as the spec's sync
contract says), unless that would hide everything; the Rust call returns the filtered, sorted list
plus the unfiltered count.
- **Sign-in code**: the authorization code is read from the return URL's query (`%XX` decoded, `+`
kept literally; codes are URL-safe). The WebView flow passes the URL to Rust and never logs it.
- **Randomness** for the serial and verifier comes from `/dev/urandom`.
- **`fuel.json`** is found by exact name, else a case-insensitive match at the folder root; a UTF-8
BOM is accepted. `Main.Command` is accepted after stripping leading `/` and `./`, and must be a
safe relative path naming a regular file.
- **Logs**: the Rust side strips the query string and fragment of every URL in a line before it
reaches `onLog`; the shared redaction runs after that as before.
- **No engine, no store**: without the native library, sign-in, sync and install report
"Amazon Games: the store engine is not available"; there is no Java fallback.
## What needs device testing
- An existing signed-in user stays signed in after the update: the library shows from cache at once,
and a forced sync refreshes the token with the stored refresh token and lists the games.
- A fresh sign-in through the WebView (registration, display name), and sign-out (deregistration
fired, library cache cleared).
- A library with many pages; add-ons hidden; posters.
- Install, pause, resume, cancel (folder removed); repair of an installed game (nothing downloaded
when intact; a corrupted file fetched again; "Resume install" offered when a repair is
interrupted); an SD-card install; out-of-space refusal.
- Every endpoint with the chosen User-Agents, the CDN with `DroidDeck/…` in particular.
- Launch of an existing install through its existing Steam shortcut (same shortcut, same prefix)
and of a fresh install (exe choice, `fuel.json` arguments, the five FuelPump variables in
`.droiddeck-launch.bat`); uninstall.
+143
View File
@@ -0,0 +1,143 @@
# Epic Games Store module: clean-room rewrite
The Epic integration (the Rust `epic` module in `libdroiddeckstores.so` and the Kotlin package
`com.droiddeck.launcher.stores.epic`) was rewritten so that it carries no code from the earlier
implementation, whose Rust module stated derivation from third-party Epic clients. This page
records how that was done and every judgement call the new code makes.
## Process
1. **Spec, written separately.** A functional specification (six documents: summary and unknowns,
service protocol, manifest and chunk formats, install behaviour, on-disk compatibility, and the
app-facing interface) was written from the old code by a separate pass. It describes behaviour
and data formats only, not code structure.
2. **Implementation from the spec only.** The implementer worked from those documents and from
code that is not Epic-specific:
- read: the spec; the crate's `Cargo.toml`, `lib.rs`, `stores_jni.rs`, `priority.rs`,
`fetch_tuning.rs`, the shared `fetch_core.rs` and the GOG module (for crate, threading and JNI
conventions); `JNI.md`; the Kotlin code outside `stores/epic/` that calls Epic (to keep its API);
build files and CI;
- not read: anything under `stores/epic/` (Kotlin, Java, tests) or `src/epic*` (Rust) of the old
tree, the body of the old `EpicCloud` class and of `CloudSavePaths.epic`, other branches or
checkouts, or any Epic client source. The old files were deleted with `git rm` without being
opened. (A search over shared files incidentally printed two lines of the old `EpicCloud`, its
base URL and User-Agent constant; both values are in the spec and only the URL is used.)
3. **Fresh code.** Every file of the new module has its own structure, names and comments, with a
plain description and `SPDX-License-Identifier: GPL-3.0-or-later` (the crate's licence). Where
the spec left a fact open, the choice was made from general knowledge of Epic's public services
and is listed below.
## Layout
| Rust (`app/src/main/rust/stores/src/epic/`) | |
|---|---|
| `net.rs` | blocking HTTP helper, the one User-Agent, percent/base64 encoders |
| `account.rs` | sign-in with an authorization code, refresh, exchange codes, redirect-page parsing |
| `library.rs`, `catalog.rs` | owned items (all pages), catalog details and attributes |
| `build_info.rs` | live build lookup, CDN list, manifest download |
| `manifest.rs`, `chunk.rs` | binary and JSON manifests, chunk files |
| `locale.rs` | device language → install tags, tag selection |
| `install/` | the install engine: `staging`, `delta`, `space`, `cache`, `schedule`, `writer`, `mod` (orchestration), end-to-end `tests` against a loopback server |
| `shop.rs`, `achievements.rs`, `savesync.rs`, `launch.rs`, `timefmt.rs` | shelves, achievements, cloud saves, launch arguments, ISO dates |
| `service.rs`, `ffi.rs` | the JSON operations and the JNI entry points |
| Kotlin (`stores/epic/`) | |
|---|---|
| `EpicBackend` | `StoreBackend` + `EpicLaunchSupport`; installs the login flow |
| `EpicLogin` | the `StoreLoginActivity.Flow` |
| `EpicCredentialStore`, `EpicPrefs` | account (shared Keystore store), `stores.epic` prefs |
| `EpicLibrary`, `EpicShelves` | library sync and cache, shop shelves |
| `EpicInstaller` | the download-queue job |
| `EpicApiClient`, `EpicDownloadManager`, `EpicInstallTags`, `EpicAchievements` | API kept for callers outside the package |
| `EpicSaveSync`, `EpicSaveTemplate` | cloud-save transfers for `EpicCloud`, template parsing for `CloudSavePaths.epic` |
| `EpicNative` | JNI wrapper (see `rust/stores/JNI.md`) |
## Judgement calls
Numbers in brackets refer to the spec's list of unknowns.
**Fixed where clearly correct**
1. OAuth form values (authorization code, refresh token) are percent-encoded [9].
2. A library page that fails ends the sync with an error and keeps the cached library instead of
caching a partial one; an HTTP 401 is reported as "session expired: sign in again" rather than
an empty library [7].
**Chosen where the spec left it open**
3. **User-Agent** [1]: every HTTP request (account, library, catalog, build info, CDNs, store
GraphQL, promotions, achievements, cloud saves) sends `DroidDeck/<versionName>`, defined once in
`epic/net.rs` (`USER_AGENT_PRODUCT`). Whether every endpoint accepts it is untested; changing that
constant is the one place to adjust. The sign-in **WebView** keeps the launcher-style UA it used
before, because a sign-in page that rejects the UA would lock every user out; that string lives
in `epic/account.rs` (`WEBVIEW_USER_AGENT`).
4. **Shared fetch core** [14]: chunk downloads reuse the crate's store-agnostic `fetch_core` (also
used by GOG and Amazon; it came from Bannerlator's own crate and, unlike the old Epic module,
states no derivation from an Epic client). It already provides the adaptive window, CDN
rotation, retries with back-off, 429 and stall cooldowns the spec asks for. If its lineage is
ever in doubt, it is the one shared dependency to replace.
5. **Out-of-package Epic code** [14]: `EpicCloud` and `CloudSavePaths.epic` were rewritten to call
the new module. `StoreLaunch`'s exchange-code handling, `EpicOptions`, `WinCompSources.epic` and
`SessionPrefs.storesEpicMode` are shared code and were left as they are.
6. **Save-path template** parsing is Kotlin (`EpicSaveTemplate`), not Rust: the prefix walk it feeds
is shared Kotlin code and its existing unit test runs on the JVM without the native library. The
Epic walk matches folder names case- and punctuation-insensitively (`MyGame` finds `My-Game`), as
that test expects.
7. **One engine** [12, 13]: the hidden `storesEpicMode` switch is not read and there is no Java
fallback; without the native library the Epic store reports itself unavailable. When the full
space plan does not fit, the engine runs in a narrow mode (one file staged at a time, no chunk
cache), which needs only the net peak of the pending files.
8. **Schedule and cache**: the cache budget is enforced by the plan (eviction of the chunk needed
furthest in the future). At run time files are admitted up to 48 downloads ahead of the
furthest dispatched one, so the cache can briefly hold a few entries more than the plan's peak.
A chunk is first written into every admitted file that uses it; it touches the cache only when a
not-yet-admitted file still needs it.
9. **Space probing** uses `statvfs` (available blocks × fragment size, the same figure as Android's
`StatFs.getAvailableBytes`) from Rust; a failed probe is "unknown". A scratch directory on the same
filesystem as the game is treated as "beside the game".
10. **File hashing** is done in file order while parts land; parts that arrive early are read back
from the stage file. Files without a manifest hash are not hashed.
11. **Staging folder recovery**: an *empty* staging folder without its marker (a run stopped between
creating the folder and writing the marker) is adopted and the marker written; any other missing
or foreign marker fails the run without deleting.
12. **Unsafe manifest paths** (absolute, `..`, empty segments) fail the install.
13. **Encrypted manifests** (stored-as bit 1) are refused with an explicit message.
14. **Meta section prerequisite ids** are all read (no cap at 64), so the following fields stay aligned.
15. **Store GraphQL shelves** are requested with their arguments written into the query text
(country reduced to two letters first) instead of typed GraphQL variables, so the query does not
depend on the server's variable type names.
16. **Achievements**: when an account id is stored and the player-progress query gets no answer, the
result is "unreachable" rather than a list with everything locked.
17. **Token refresh** is single-flight (one refresh at a time; a waiting caller re-reads the stored
account first), so a rotated refresh token is never spent twice. A fresh sign-in replaces the
stored account instead of merging with it.
18. **Download progress** reported to the queue never moves backwards, although samples come from
several native threads.
19. **Network window** is the speed tier's value clamped to 1–128; the per-host cap is
`max(6, ceil(window / hosts))`.
**Kept at parity (known weak spots)**
- Chunk requests do not carry the CDN's signing query; only the manifest request does; the
`cloudflare.epicgamescdn.com` CDN stays excluded [2].
- The manifest file name comes from the first manifest location and is requested from every CDN [3].
- The manifest header SHA-1 and the chunk header GUID/hash are not checked (chunk size and chunk
SHA-1 are) [4].
- Per-file flags, symlink targets and custom fields are ignored [5]; JSON manifests carry no hashes
as read, so their files are rewritten on every run [6].
- The refresh token's own expiry is not tracked; a failed refresh still returns the stored token [8].
- The cloud `hash` is compared as if it were an MD5 [10].
- `AdditionalCommandLine` is one argument, frozen at install [11].
- Shop search, an unused "wants an exchange code" helper and per-game uninstall work are not
ported; DLC stay hidden and are never installed separately; optional (non-language) install tags
are never selected; resume restarts a partially written file from zero [15–18].
- Library details are fetched one game at a time, as before.
## What needs device testing
- Sign-in through the WebView, and that an **existing** saved account (sealed `credentials.json`)
stays signed in after the update.
- Every endpoint with the DroidDeck User-Agent: library, catalog, build info, CDN downloads, store
GraphQL and promotions, achievements, cloud saves.
- Library sync and shelves; install, pause/resume, cancel, update and repair; a card install with
the internal scratch cache; launch with an exchange code; cloud-save up/down.
+86 -136
View File
@@ -35,8 +35,8 @@ to manifest checks of the files still needing work (GOG MD5, Amazon SHA-256 wher
Free-space display probes the nearest existing ancestor without creating the target folder; if
the probe cannot determine space, it reports **Free space unknown** rather than blocking an install.
`StoreInstallRoot` uses `-1` for unknown and preserves a measured `StatFs` zero as **0 B free**.
Fresh GOG/Amazon preflight and all three Java backend space checks reject known zero when bytes
remain. The read-only probe walks existing ancestors without creating directories; a failed
Fresh GOG/Amazon preflight and the backends' own space checks (GOG in Kotlin, Epic and Amazon in
the Rust engine) reject known zero when bytes remain. The read-only probe walks existing ancestors without creating directories; a failed
`StatFs` probe can use a positive `File.usableSpace` fallback, but that API's ambiguous zero is
unknown, not a measurement of a full disk.
Preflight failures remain failed Downloads rows with Clear and Retry. Retry reruns preflight
@@ -45,143 +45,80 @@ with a placeholder job. Failures are kept in the Downloads UI, notification and
duplicated as transient failure toasts. Active queue persistence and ordinary backend pause/retry
are unchanged.
Epic installs have four writers, all with the same per-file safety rules (below). The hidden
`storesEpicMode` string in the `session` preferences picks one: `adaptive` (default), `lookahead`,
`batch` or `java`. There is no UI for it; set it on a test device to compare the modes on the same
game. Every mode logs an `epic: summary mode=<mode>` line (files, bytes written, elapsed, average
MB/s). When `libdroiddeckstores.so` does not load, every mode runs the Java writer. Once it is
loaded, an engine failure (refused plan, ABI mismatch, network, hash or I/O failure, unsupported
atomic swap) fails the install; it never silently reruns as Java downloads.
Epic installs run in the Rust `epic` module of `libdroiddeckstores.so` (`epic/install/`), driven by
`EpicInstaller` through `EpicNative.install`. There is one engine; the old hidden `storesEpicMode`
switch is no longer read. Without the native library the Epic store cannot install (the facade
reports "the store engine is not available"). The design, and how it was written from a
functional spec only, is in [epic-cleanroom.md](epic-cleanroom.md).
**Adaptive (default)** - `EpicNative.stream`, `epic/stream.rs`. Measured on device: about
83 MB/s on a 71 GB install. The chunks go through the shared fetch core in the order their files
need them, with its full adaptivity (window ramp, per-host caps, in-flight byte budget, retries,
host rotation, stall handoff). A chunk only one file part uses is verified and written straight
into that file's staging file at the part's offset, in whatever order it arrives; it is never
cached, so each byte is written once. A chunk several parts use is written into every part whose
file the download has reached and kept, decompressed, in the cache until its last part is written.
The cache has a byte budget; the plan simulates the install in file order and, when the budget
would be exceeded, drops the cached chunk needed furthest ahead and downloads it again later.
Files are opened only once the download reaches them, so the set of half-written files stays
within the fetch window. All writes happen in the fetch core's process pool, whose completion
frees the in-flight byte budget: a slow card throttles the download instead of growing memory.
Each file's SHA-1 is computed as its parts are written (parts that arrived out of order are read
back once the bytes before them are in). Stage files go into the owned staging namespace below.
The run logs `stream-plan` (direct vs shared chunks, planned refetches, budget) and
`stream-summary mode=adaptive` (peak cache, peak staging, refetches, bytes written from the
download vs from the cache).
A run, in order: CDN list from the build information (`EpicApiClient.getManifestApiJson`) →
manifest from the first CDN that serves it → parse (binary or JSON) → files selected by the device
language's install tags (`EpicInstallTags`) → staging recovery → the delta pass (a selected file is
up to date when it has the manifest size and SHA-1; a file without a manifest hash is always
rewritten) → space plan → chunk cache preparation → fetch schedule → download and write → the
installed record (`StoreInstalls.complete`). Resume, update and repair are the same rerun.
On an update, a staged file that replaces an existing one sits beside its old copy until its swap,
so the engine paces replacements with an overlap budget: files are admitted in order, a replaced
file only while the planned sizes of admitted, not-yet-swapped replaced files fit the budget (one
is always admitted, so the run progresses), and a chunk is fetched only once its file is admitted.
Fresh files are never held back, and admission only delays files, never reorders them, so the
cache plan stays exact. The run logs `overlap_budget=` in `stream-plan` and
`peak_replace_overlap_bytes=` in `stream-summary`. With room to spare the budget is large and
updates run as wide as fresh installs; on a nearly full volume, replaced files are staged one
after another and the fetch pauses at each replaced-file boundary until the previous one is
swapped in.
- **Schedule** (`epic/install/schedule.rs`). The pending files' parts form one timeline. A chunk is
downloaded at its first use and feeds every later use in the same file directly. If it is used
again in a later file it is kept, decompressed and verified, in the chunk cache, which has a byte
budget; when keeping a chunk would exceed it, the cached chunk needed furthest in the future is
dropped and downloaded again when needed. Chunks a previous run left in the cache are reused,
soonest-needed first, as far as the budget allows.
- **Writing** (`epic/install/writer.rs`). Files are admitted strictly in order, a little ahead of
the download, and a download is dispatched only once the file of its first use is admitted, so
the number of half-written files stays small. Each admitted file is written into a unique
`epic-<pid>-<seq>-<nanos>.stage` file in `<game>/.droiddeck-epic-staging/`; the file SHA-1 is
computed in file order as parts land. A complete file is synced, size- and hash-checked, and
renamed over its destination. Right before the first rename of the run `StoreInstalls.markUpdating`
switches an installed record to `updating`; a refusal fails the run with nothing replaced.
- **Fetching** uses the shared `fetch_core` (adaptive window, CDN rotation, retries with back-off,
stall and 429 cooldowns) with the speed tier's window and decode workers. Every chunk is checked
(header, zlib stream end, exact size, SHA-1) before any byte reaches a file or the cache.
- **Space** (`epic/install/space.rs`). Free space is probed per volume (`statvfs`; unknown is never
"no space"). The plan: net peak of the pending files in order (each old copy credited at its own
swap), a cache budget of half the remaining space (at most 1 GiB, 256 MiB when unknown, 0 is
allowed), an overlap budget that limits how many bytes of files replacing existing data are
staged at once on an update (at least the largest replaced file), and a 64 MiB transient
allowance. A card install keeps its cache in `cacheDir/stores/epic/<id>/` when internal storage
can hold a useful cache and the game has no non-empty `.chunks` yet; otherwise the cache is
`<game>/.chunks`. When the full plan does not fit, the run falls back to one file at a time with
no cache (needs only the net peak); when even that does not fit the install is refused before any
chunk request, naming the storage, the need, the free space and the shortfall.
- **Progress**: the bar counts pending-file bytes written (not compressed chunk bytes), at most every
50 ms plus once per finished file. The run logs `epic: plan …` and `epic: summary mode=adaptive|narrow
files=… bytes_written=… avg_mbps=… fetches=… refetches=… peak_cache_mib=… hosts=…`.
- **Cancel / pause / failure**: the cancel flag is forwarded within 250 ms; the run stops its
requests, joins its threads, removes every unfinished stage file and `.part` cache entry, and only
then returns. Swapped files stay; unswapped destinations keep their old content; the chunk cache
stays for the next run. Success removes the cache and the staging folder (when only its marker is
left).
Its space plan (`EpicDownloadManager.planAdaptiveSpace`, unit-tested in `EpicAdaptivePlanTest`)
checks each volume for what lands on it, with the lookahead planner's rule (the need must not
exceed the free reading). The game volume: the net growth at its worst point (`settledNetPeak`,
the running Σ(new − old) with each old copy released at its swap), plus the overlap budget, plus
the cache when it is beside the game. The overlap budget is whatever the game volume has left
after the net growth and the cache, never less than the largest replaced file; when even that
does not fit, the plan hands over to the lookahead engine. The cache volume: the budget plus 64 MiB of
chunks in transit. The cache budget is half the cache volume's free space after those needs, at most
1 GiB (256 MiB when free space is unknown; 0 is allowed and means every reuse is a refetch). On a
card install the cache stays on internal scratch unless internal storage cannot hold at least
64 MiB of it; then the cache goes beside the game (still released chunk by chunk) instead of
failing. A plan that does not fit logs the short storage and the shortfall and hands over to the
lookahead engine, which narrows its window to the space there is (or fails its own preflight with
an explicit error). Running out of space in the middle of an adaptive run is not destructive: a
failed write fails the run, unfinished staging files are removed, the old files stay.
The staging folder is recognised by its `owner` marker (`DroidDeck Epic staging v1` + LF). Before
the delta pass the run checks that the folder is a real directory at the expected canonical path,
that the marker matches and that every other entry is a plain `.stage` file directly inside it; only
then are those orphans deleted. Anything unexpected fails the run without deleting. A manifest path
naming the staging folder, an absolute path or a `..` segment is refused. `.stage` / `.part` files
elsewhere in the game folder are never touched.
**Lookahead** - `EpicNative.installAdaptive`, `epic/adaptive.rs`; also the adaptive engine's
fallback when its space plan does not fit. The Java delta pass sends a versioned snapshot of
pending files, chunk metadata, verified cache lengths, old output lengths and both measured
free-space probes to Rust. The native planner and execution share one ordered moving-window
policy. Before any chunk request, Rust chooses the largest safe lookahead (128, 64, 32, 16, 8, 4,
2 or 1 part references) and reusable-cache budget (64 MiB, 16 MiB or zero) whose simulated
staged-file and cache peaks fit. Native device IDs also detect an explicitly named cache on the
game volume, so two separate paths cannot incorrectly turn one filesystem's requirement into two
independent checks.
Amazon installs run in the Rust `amazon` module (`amazon/install/`), driven by `AmazonInstaller`
through `AmazonNative.install`; sign-in, the library sync and the size lookup call the same module
through `AmazonNative.call`. The design, and how it was written from a functional spec only, is in
[amazon-cleanroom.md](amazon-cleanroom.md).
| Storage situation | Lookahead policy |
| --- | --- |
| Plenty of room | Up to 128 references ahead, up to 32 network requests, 64 MiB reusable data |
| Tight | Smaller lookahead and/or reusable budget; parallel fetching remains where space permits |
| Barely fits | One-reference lookahead, one request, no reusable retention if needed |
| Does not fit even the minimal schedule | Explicit preflight error before requests or replacement |
| Free space unknown | Conservative bounded lookahead (at most eight), not an invented zero |
The lookahead fetch coordinator reuses `fetch_core`'s host scheduling, retries, adaptive network
window and decode pool. Each chunk is cached, then its parts are copied into staged files by a
writer that runs concurrently, waiting only for its next required chunk, through the native
assembler's streaming copy primitive. GUIDs in the lookahead are reserved once before submission,
at their full decoded size. Chunks are released at last use on **both** cache layouts; distant
reusable data is oldest-first bounded and may be refetched. At most the selected reusable budget
plus distinct lookahead decoded sizes is resident/reserved. New request admission re-probes both
volumes and can shrink unsent lookahead or retention if headroom disappears; already in-flight
reservations remain protected. If even existing reservations cannot fit, the run fails and joins
its workers rather than waiting indefinitely or deleting old game files. An install to SD retains
its internal scratch cache, so assembled output is the only game data written to the card.
**Java** - without the native library, or with `storesEpicMode=java`. Eight Java workers prefetch
the next eight chunk-part references, across file boundaries, while one writer assembles and
verifies files in manifest order. Shared GUIDs within that window are fetched only once. Chunks
are deleted after their last use; other reusable chunks are retained up to 64 MiB. At most eight
additional decoded chunks (complete or in flight) coexist with that retained cache. The size
bound is 64 MiB plus the decoded sizes of the window's distinct GUIDs, not a fixed eight-MiB
allowance: manifests can supply larger chunk windows. Evicted distant shared chunks may be
downloaded again. No worker downloads outside the moving window; cancellation and failure
disconnect and join all workers before returning or removing the cache. Its summary line is
`mode=java-bounded`.
**Batch** - the earlier native design, kept for A/B benchmarks: every chunk fetched into the
cache in parallel, then the files assembled from it (`EpicNative.run` + `EpicNative.assemble`).
It needs the cache and the files at once (about twice the game) and checks exactly that before
starting; it writes files in place, so an existing install is marked `updating` before assembly.
`epic/assemble.rs::build_assemble_plan` enables `scratch_cache` only for a nonempty cache path:
`run_assemble` then removes chunks at their last use; the default beside-game cache is removed at
the end (`default_cache_assembles_like_the_kotlin_loop_and_removes_the_cache_at_the_end` and the
scratch-cache assembly tests cover that distinction).
The Epic download bar, measured rate and ETA (adaptive, lookahead and Java) count pending file bytes written, not unique
compressed chunks fetched. Shared chunks therefore cannot complete the bar before later files
are written, and an evicted chunk's refetch is reflected in the time spent doing that work.
Streaming stays in the Download stage until all files are written and verified, then enters
Install for registration; it does not bounce between stages for each file.
For the lookahead and Java writers, the space check follows the manifest delta pass and uses the
same pure moving-window policy as execution, reserving all incoming decoded chunk sizes before writing parts and simulating
last-use releases and unpinned oldest-first evictions. Existing verified cache bytes count against
current free space and are credited only at their scheduled release. This avoids adding the entire
cache to the entire game without undercounting concurrent downloads. Each file is written to a
unique file in an owned staging namespace on its output volume (never the literal `<filename>.part`, which can
be a legitimate manifest file), size/hash verified, synced and atomically renamed over its output.
The old file remains until that swap succeeds. The planner credits its bytes only after that
file's complete staging write, not by subtracting all old outputs up front. Both volume checks
must accept this safe staging peak; insufficient space is rejected explicitly, with no destructive
delete-first fallback. An unsupported/failed atomic rename fails the install and retains the old
destination. Completed files and
verified cached chunks are accounted for on resume. On a card, game and cache requirements
are checked separately. File SHA-1 hashes are checked where supplied by the manifest; failed
or cancelled staging files are removed without deleting old outputs. A successful run removes the
cache. A resume within a large incomplete file restarts that file, not its individual parts.
All three staged writers (adaptive, lookahead, Java) use `<installDir>/.droiddeck-epic-staging/`, whose `owner` marker identifies this
namespace. A successful install removes it when it holds nothing but the marker (same no-follow
and canonical-path checks); the next run recreates it. Before the delta pass and free-space probes, the manager removes orphan `.stage`
entries only from this owned namespace. A manifest naming that directory or any descendant
fails without deletion; an unowned directory or unexpected entry also fails conservatively.
The namespace must be a real directory at the expected canonical child of the install root.
Ownership markers and staging entries must be no-follow regular files; symlink directories,
markers and entries are rejected before recovery deletes anything.
Ordinary game files with `.stage` or `.part` suffixes elsewhere are never recovery targets.
Abrupt process termination is therefore recoverable on the next run without inferring ownership
from a loose filename prefix. Cancellation/failure waits for all native/Java fetch writers before
returning; completed verified cache chunks remain usable on resume.
An Amazon run, in order: download lookup for the entitlement (`GetGameDownload`: a CDN base URL with
a signed query, and the version id) → `manifest.proto` from that base (a 4-byte big-endian header
length, an opaque header, then an xz or LZMA-alone compressed protobuf) → parse (every file of every
package; paths refused when absolute, with a drive letter, `..`, empty segments, or naming
DroidDeck's own root files) → leftover `<file>.tmp` siblings removed → the check of files on disk (a
file is complete when it has the manifest size and SHA-256; the same test drives the space check and
the download plan) → space check (`Not enough free space: need …, only … free`; unknown free space
never blocks) → empty files created locally → every other pending file fetched from
`files/<sha256 hex>` under the base, one streamed GET each through the shared fetch core (the speed
tier's window; at least half the CPU cores verifying and writing), hashed as it arrives, written to
`<file>.tmp` and renamed over the destination after the size and hash checks → the installed record.
Before the first file is put in place the record is switched to `updating`, so an interrupted repair
is offered as "Resume install". There is no partial-file resume: a file is complete or fetched again
from its first byte. Files the manifest does not list are left alone.
## The sidecar
@@ -266,6 +203,15 @@ An Epic game's launch choices are the Epic card in the Games tab's launch settin
read the same thing - and "Resolve Epic sign-in", which opens Epic's account page behind its
sign-in in Android's browser.
An Amazon game's exe and arguments come from its `fuel.json` (`Main.Command`, used when it names
an existing file; `Main.Args`, used even when it does not), else from a score over every `.exe`
below the folder (Unreal shipping binaries and `Binaries/Win64` first, crash handlers, setups,
launchers and tiny stubs last; the rules see paths relative to the install folder), else the shared
picker. Its launcher sets five FuelPump variables frozen at install time: `FUEL_DIR`,
`AMAZON_GAMES_SDK_PATH`, `AMAZON_GAMES_FUEL_ENTITLEMENT_ID`, `AMAZON_GAMES_FUEL_PRODUCT_SKU` and
`AMAZON_GAMES_FUEL_DISPLAY_NAME` (the account name, `Player` without one). No Amazon Games SDK files
are deployed into the prefix and `fuel.json` `PostInstall` steps are not run.
A game's web links reach Android: `droiddeck-store-launch` sets the prefix's
`HKCU\Software\Wine\WineBrowser` `Browsers` to `/usr/local/bin/droiddeck-open-url`, which is also
`xdg-open` first on the game's PATH, and hands http(s) addresses to the app, which opens them in
@@ -306,7 +252,9 @@ the prefix's user folder appears, within a minute. Every path logs
has no cloud saves.
- **The services:** GOG `cloudstorage.gog.com/v1/<user>/<client>` with a token issued to the game's
own client (the client secret is kept at install; the Galaxy token is the fallback); Epic's
savesync data storage with signed read and write links.
savesync data storage with signed read and write links (`EpicCloud` → `EpicSaveSync` → the Rust
`epic/savesync.rs`; the template's root token and `{EpicId}` / `{AppName}` are read by
`EpicSaveTemplate`).
- **Which launches:** `droiddeck-store-launch` recognises a store game by the launcher .bat Steam runs
or, for a game without one (most GOG games), by its exe - the folders above it are searched for
the sidecar.
@@ -332,7 +280,9 @@ keeps the files; every store's install skips complete, verified files on the rer
rerun. Cancel deletes the folder. `StoreDownloadService` holds a foreground notification while
anything runs. The native engine (`libdroiddeckstores.so`, `StoresNative`, `GogNative`, `EpicNative`,
`AmazonNative`; contract in `app/src/main/rust/stores/JNI.md`) is loaded on first use; without it
each manager runs its Java fetch loop.
the GOG manager runs its Java fetch loop, and the Epic and Amazon stores (all of whose service calls
and installs run in the Rust `epic` and `amazon` modules) are unavailable: sign-in, sync and install
report "the store engine is not available".
## Settings