Use this skill whenever the task is to add a new feature to Warpnet that crosses layers — anything described as a "new endpoint", "new route", "new RPC", "new event", "new path", or any feature that needs a frontend (Vue desktop or Android/Tusky) to talk to the local node and/or one node to talk to another. Triggers include phrases like "add /private/get/...", "add a handler for X", "expose Y to the UI", "make nodes exchange Z", "expose X to warpdroid", "add support for X in the Android client". Also use when the change is Android-only but touches a Warpnet protocol ID (new ProtocolIds entry, new DTO, new WarpnetRepository method). Do NOT use this skill for pure-Tusky-side UI changes that don't touch any Warpnet protocol, pure DB-schema changes that no one calls, or for fixing existing handlers without adding new routes.
Use this skill whenever the task is to add a new feature to Warpnet that crosses layers — anything described as a "new endpoint", "new route", "new RPC", "new event", "new path", or any feature that needs a frontend (Vue desktop or Android/Tusky) to talk to the local node and/or one node to talk to another. Triggers include phrases like "add /private/get/...", "add a handler for X", "expose Y to the UI", "make nodes exchange Z", "expose X to warpdroid", "add support for X in the Android client". Also use when the change is Android-only but touches a Warpnet protocol ID (new ProtocolIds entry, new DTO, new WarpnetRepository method). Do NOT use this skill for pure-Tusky-side UI changes that don't touch any Warpnet protocol, pure DB-schema changes that no one calls, or for fixing existing handlers without adding new routes.
Adding a new handler to Warpnet
A handler in Warpnet is one entry in a multi-layer pipeline. Adding one means touching all of these layers in lockstep — skipping any single layer leaves the feature dead. The pipeline is fixed; treat it as a checklist, not a suggestion.
There are now two client-side pipelines that talk to a desktop node: the in-process Vue UI (the original member app) and the Android client (warpdroid/). They share the same server-side handler — there is one canonical handler per route, registered once in member-node.go. They differ only in how the request reaches the node:
always wrap the inner payload in an event.Message envelope ({body, message_id, node_id, path, timestamp, version, signature});
use the same path string (e.g. /public/post/like/0.0.0);
hit the same Go handler on the desktop side.
What changes per route:
All routes work with the desktop UI — Vue can hit any route registered in member-node.go.
A subset of routes work with warpdroid today. Only the protocol IDs explicitly listed in warpdroid/warpnet-transport/.../ProtocolIds.kt are wired through, and only those routes whose middleware doesn't strictly verify the envelope signature are usable end-to-end (see §A4 — the gomobile binding does not yet expose the signing key).
Step 0 — Do not skip context discovery
Before writing a line of code, read these files in this order. Each is small enough to fit in a single view:
Server side (always):
cmd/node/member/app.go — see App.Call and the special-case for PRIVATE_POST_LOGIN / PRIVATE_POST_LOGOUT. Everything else hits the default: branch and goes through node.SelfStream.
event/paths.go — naming convention for routes.
core/handler/like.go — the cleanest small reference. ~215 lines. Read top to bottom.
core/handler/like_test.go — stub style used everywhere; no testify.
warpdroid/warpnet-transport/src/main/kotlin/site/warpnet/transport/WarpnetClient.kt — request() (~180) and pair() (~148) for the calling pattern.
warpdroid/warpnet-transport/src/main/kotlin/site/warpnet/transport/dto/WarpnetDtos.kt — the DTO catalogue (only routes wired through warpdroid).
warpdroid/app/src/main/java/com/keylesspalace/tusky/warpnet/WarpnetRepository.kt — how Tusky-shaped entities are produced from Warpnet DTOs.
warpdroid/app/src/main/java/com/keylesspalace/tusky/warpnet/WarpnetMapper.kt — the lossy mapping rules between Warpnet and Mastodon-shaped types.
If the feature touches a resource that already exists (replies, follows, chats…), also read the matching core/handler/<that>.go — you will reuse its repo or DTO.
Decide upfront which clients the route is for. Only the desktop UI? Only warpdroid? Both? This determines which of the steps below you must do. Server-side steps (1–5, 7) are always required. Step 6 (Vue) is required only if the desktop UI should call the route. Section §A (warpdroid) is required only if the Android client should call the route.
private = only the local UI calls this (talks to its own node only). Examples: PRIVATE_GET_TIMELINE, PRIVATE_POST_TWEET, PRIVATE_DELETE_CHAT, PRIVATE_GET_STATS.
public = nodes call this on each other over libp2p. Examples: PUBLIC_POST_LIKE, PUBLIC_GET_USER, PUBLIC_POST_FOLLOW.
If unsure: does a remote peer ever need to invoke this? Yes → public. No → private.
verb is get, post, delete. (No put, no patch — Warpnet does not use them.)
resource is short, lowercase, one word if possible. Plural for collections (tweets, users), singular for items (tweet, user).
version is 0.0.0 for new routes. Bump only when you make a breaking wire change to an existing route — do not bump just because the file changed.
The same constant is declared in two or three places, depending on which clients use it. All declarations must hold the same string byte-for-byte:
Go: event/paths.go — always. Alphabetical within its visibility block.
JS: frontend/src/service/service.js — required if the Vue desktop UI calls the route. Top of file.
Kotlin: warpdroid/warpnet-transport/.../ProtocolIds.kt — required if warpdroid calls the route. Grouped under one of the existing // admin / pairing, // public reads, // public writes, // private reads, // private writes headers.
If any pair of these strings drifts, the call silently goes to the default: of App.Call and fails with not attached server node or with handler-not-found. There is no compile-time link between the three.
Step 2 — Define the DTOs
In event/event.go, alphabetically among the other types:
// FooEvent defines model for FooEvent.type FooEvent struct {
TweetId domain.ID `json:"tweet_id"`
UserId domain.ID `json:"user_id"`
}
// FooResponse defines model for FooResponse.type FooResponse struct {
Count uint64`json:"count"`
}
Rules:
IDs use domain.ID, not string. Other primitive fields use plain Go types.
JSON tags are snake_case. Field names in Go are CamelCase.
Reuse existing types where possible. event.UnlikeEvent = LikeEvent is the precedent — if your delete event is shape-identical to the create event, alias it.
If the response is just a count, reuse LikesCountResponse. Don't invent a parallel type.
Put the godoc one-liner // FooEvent defines model for FooEvent. — every existing type has it; linters may rely on it.
Step 3 — Add or extend the repo
If a repo for this resource already exists (database/<resource>-repo.go), add a method to it. If not, create a new file modelled on like-repo.go.
Validate empty IDs at the top with local_store.DBError("empty ..."). Same wording style as the other repos — it ends up in error messages users see in logs.
Always defer txn.Rollback() immediately after NewTxn. Rollback after a successful Commit is a no-op, this is intentional.
Always end the success path with txn.Commit(). Any branch that returns without committing is a bug.
Use local_store.IsNotFoundError(err) to test for not-found, never err == ... or errors.Is against datastore errors.
repo.statsDb is optional — guard with if repo.statsDb == nil { return ... } exactly like in like-repo.go. Stats failures are warnings, not errors: log.Warnf(...) and continue.
Do not introduce a new repo dependency on something other than LikeStorer-shaped interfaces. Repos depend on storage, not on each other.
Define a sentinel error if callers need to distinguish "not found" from real errors:
var ErrFooNotFound = local_store.DBError("foo not found")
Add a _test.go next to it. Check the existing like-repo_test.go for the test harness — they use a real BadgerDB in a temp dir, no mocks.
Step 4 — Write the handler
File: core/handler/<feature>.go. Use like.go as the template literally — copy the AGPL header from any existing handler file (the exact 22-line block ending with // SPDX-License-Identifier: AGPL-3.0-or-later, then package handler).
Skeleton:
// Storer interfaces are declared at the top of the file, named after the handler:type FooStorer interface {
DoSomething(id, userId string) (uint64, error)
// ...minimal surface only — do not pass *FooRepo around
}
type FooStreamer interface {
GenericStream(nodeId string, path stream.WarpRoute, data any) (_ []byte, err error)
NodeInfo() warpnet.NodeInfo
}
funcStreamFooHandler(
repo FooStorer,
streamer FooStreamer,
) warpnet.WarpHandlerFunc {
returnfunc(buf []byte, s warpnet.WarpStream) (any, error) {
var ev event.FooEvent
if err := json.Unmarshal(buf, &ev); err != nil {
returnnil, err
}
if ev.TweetId == "" {
returnnil, warpnet.WarpError("foo: empty tweet id")
}
// ...more validation, in the order the fields appear in the struct...
n, err := repo.DoSomething(ev.TweetId, ev.UserId)
if err != nil {
log.Errorf("foo handler failed: %v", err)
returnnil, err
}
// For PUBLIC routes that propagate to the resource owner's node:// 1. resolve target node id via userRepo// 2. streamer.GenericStream(targetNodeId, event.PUBLIC_POST_FOO, ev)// 3. swallow ErrNodeIsOffline — return local result anyway// 4. unmarshal possible ResponseError from the remote and log itreturn event.FooResponse{Count: n}, nil
}
}
Hard rules:
Use the project's json package: import "github.com/Warp-net/warpnet/json" (NOT encoding/json). The aliased jsoniter is fine inside app.go because it's the dispatcher; in handlers use json.
Logger: log "github.com/sirupsen/logrus". Use log.Errorf for handler-level failures, log.Warnf for non-fatal degradation, log.Infof only for one-shot lifecycle events.
Error wrapping: handler-internal errors are prefixed with the handler name in lowercase: warpnet.WarpError("foo: empty tweet id"). This is so log greps work.
Handlers receive warpnet.WarpStream as the second arg. Most handlers ignore it. Do not close it manually — the framework owns its lifecycle.
Return (any, error). Returning a typed event struct is the norm; the framework marshals it.
For PUBLIC handlers that call out to another node, follow the like.go propagation pattern verbatim:
errors.Is(err, warpnet.ErrNodeIsOffline) is not an error from the caller's perspective — return the local result.
After a remote call, try to unmarshal event.ResponseError from the response and log if it's non-empty. Don't surface it as an error to the local caller.
The Storer interfaces declared at the top of the handler file are handler-local — each handler declares only the methods it needs from the repo. This is intentional. Don't replace them with a shared interface.
Step 5 — Register the handler
In cmd/node/member/node/member-node.go, find the m.node.SetStreamHandlers([]warpnet.WarpStreamHandler{...}) block (~line 385). Add an entry in the same group as related routes — like is near unlike, follow is near unfollow. Don't sort alphabetically; sort by feature affinity. The existing block uses positional struct literals ({event.PUBLIC_POST_LIKE, handler.StreamLikeHandler(...)}), preceded by //nolint:govet. Match that style.
If your handler needs a new repo, instantiate it in the same function, near where the other repos are constructed (~line 370):
fooRepo := database.NewFooRepo(db, statsDB)
statsDB is only relevant if you increment counters; otherwise just db.
Step 6 — Wire the Vue frontend
Skip this step if the route is warpdroid-only. Otherwise, two edits in frontend/src/service/service.js:
Body keys are snake_case and must match the Go struct's json tags exactly. sendToNode adds message_id, node_id, timestamp automatically — do not set them in request.
The Vue component imports warpnetService and calls await warpnetService.fooSomething(...). No backend awareness in components.
§A — Wire the warpdroid Android client
Skip this section if the route is desktop-only. Otherwise, the work falls in three submodules and (sometimes) the Go binding:
The thin client connects to one desktop peer (the user's paired fat node) over a single libp2p connection, and tunnels every request through binding.stream(protocolId, json). There is no DHT lookup per call, no multi-peer fanout, no local Badger storage on the device — Android is purely a client of the user's own desktop node.
A0. Performance and battery budget — warpdroid targets low-end devices
warpdroid is designed for budget Android phones (4-core ARM, 2-3 GB RAM, 60 Hz screens, no JIT headroom) where the user expects the app to coexist with the rest of their day on a single charge. Frame budget is 16 ms (>50 ms on main = visible jank). Battery budget is implicit but strict: every architectural decision that costs CPU, radio, or wakelock time compounds in BatteryStats. The codebase carries Tusky's XML-era legacy plus a Compose migration, and several patterns that look fine on a flagship are not OK on the target hardware.
Frame / responsiveness rules. Watch for these when adding or modifying UI/glue code:
No ConstraintLayout in Compose for trivial layouts. Solver + double-measure pass is expensive on low-end CPUs and gets multiplied per LazyList item. Use Row / Column / Box with Arrangement and Alignment. Reserve ConstraintLayout for genuine constraint graphs (barriers, guidelines, chains across nested groups) — and even then, prefer custom Layout {} if it's hot in a list.
No synchronous IO from the main thread. That includes EncryptedSharedPreferences first-read, ImageView.setImageURI on custom URI schemes (falls through to Drawable.createFromPath → blocking FileInputStream.open), Drawable.createFromPath, File.exists / listFiles in onCreate. The MaterialDrawer iconUrl + warpnet:// scheme regression (~70 ms per call) is the canonical case. Move to Glide via WarpdroidAsyncImage or a registered ModelLoader; lazy-init prefs behind a coroutine.
First-compose JIT cost is real. Compose compiles ~5 KB of bytecode per composable on first invocation, blocking the main thread. TweetCard + TweetButtons + LazyList.measure on first Profile open routinely produces 1.5-2 s Davey! frames. Keep composables small, avoid deep nesting, and reuse stateless leaves so JIT amortises. Long-term fix is a Baseline Profile; until then, simplify hot composables.
SparkButton, complex Modifier.graphicsLayer, and animation-heavy widgets are not free on these devices. Audit whether the animation actually plays in the warpdroid context (the retweet/quote split path skips it because the animation never fires). Cut what doesn't run.
Don't ship Tusky's Glide+Coil mix. Glide is the single image loader (see WarpdroidAsyncImage, GlideModule, WarpnetAvatarLoader). Coil3 was migrated out for memory / cold-start reasons. Don't reintroduce a second loader.
Image / avatar URLs are warpnet://avatar/<userId>/<key>, not HTTP. Anything that goes through Uri → ImageView.setImageURI will block. Only Glide (via the registered WarpnetAvatarLoader) can resolve them, and only when the call site uses Glide.with(applicationContext) or Glide.with(imageView) — not Glide.with(activity) (crashes on activity destroy, see the WarpdroidAsyncImage crash fix).
Diagnostics workflow when something feels slow. Enable StrictMode (already configured in debug builds). Look in logcat for: Davey! duration=...ms, Choreographer: Skipped N frames, StrictMode policy violation; ~duration=N ms, Compiler allocated ... KB to compile, Quality: stackInfo. The compile-allocation line names the exact composable JIT spent time on. Pair with ANR_LOG blocks to see what message was on the looper at the time.
Comments referencing performance in MainActivity, WarpdroidAsyncImage, loadDrawerAvatar, etc. encode hard-won fixes (sync-IO escapes, Glide context lifecycle, ANIMATE_GIF_AVATARS pref split). Don't refactor them away without re-testing on a budget device.
Battery / wakelock rules. Watch for these when adding background work, network calls, or anything that fires on a timer:
No persistent foreground services for the libp2p connection. The fat node is always-on; the thin client connects on resume and tears down on pause. WarpdroidApplication.onStop calls binding.Pause(), onStart calls binding.Resume(). Holding the libp2p host alive in the background means the radio stays on the relay's keep-alive cycle — that's a constant drain. Don't add a ForegroundService "to keep the connection warm" — measure first, the user will not thank you for the warm pocket.
No polling. If the UI needs fresh data, the fat node pushes it via the existing stream or a pull on screen open. Don't Handler.postDelayed a refresh, don't LaunchedEffect { while(true) { delay(30s); refresh() } }. The one polling worker that exists (PairRefreshWorker) is governed by WorkManager's MIN_PERIODIC_INTERVAL_MILLIS (15 min), NetworkType.CONNECTED, RequiresBatteryNotLow, BackoffPolicy.LINEAR. Match that pattern: never under 15 min, always with battery + network constraints, always with backoff.
Notification fetches go through WorkManager, not a Service or AlarmManager.NotificationWorker is the template. Periodic interval ≥15 min, constrained, opportunistic — Android batches it with other apps' jobs to amortise radio wake. A custom alarm wakes the device on its own and shows up in BatteryStats as a per-app drain.
Disconnect on pause, reconnect on resume.binding.Pause() releases the libp2p host's hot loops (yamux keep-alives, peerstore TTL, relay reservations). Without it, even with no UI activity the connection costs ~3-5% per hour on a budget device because of the relay-tunneled keep-alive cycle. The lifecycle hook is already wired in WarpdroidApplication — don't bypass it from a new screen "to keep the data live".
WakeLocks: zero. This is not a media app, a navigation app, or a messenger that needs to keep the device awake. There is no legitimate reason for PowerManager.WakeLock, WifiManager.WifiLock, or setKeepScreenOn(true) anywhere in warpdroid. If you find yourself reaching for one, you're working around a broken state machine.
Don't burn CPU on background JSON / crypto. Envelope signing, when it lands (see §A4), runs on Dispatchers.Default, not main, and only on routes that actually require it. Don't pre-sign in advance, don't loop over all cached payloads on app start. Heavy crypto (Ed25519IdentityStore.derive) blocks one core for ~600 ms on a budget device — currently it runs once on the auto-pair path. Don't add more callers without measuring.
Image / network budget on cellular. Glide already has a memory cache; the disk cache is also on. Don't add a "warm the cache on launch" pass that fetches every avatar the user has ever seen — that's bytes on cellular for nothing. Lazy-load on first view, let Glide's LRU handle eviction.
Animation loops.SparkButton, ripple effects, etc. allocate frames per draw. The retweet animation runs ~600 ms at 60 fps when fired — that's bounded and fine. Continuous looping animations (InfiniteTransition, repeating Animatable) keep the render thread awake indefinitely and prevent the GPU from idling. Don't add ambient animations to feed cards or avatars.
GPS, Bluetooth, camera, mic. warpdroid does not use any of these. There is no permission for GPS, BT, mic. The camera is requested only on the PairingActivity for the QR scan and unbinds immediately after a successful scan. Don't add features that need new permissions without an explicit design discussion — every new permission is a battery vector (BT scanning especially) and an attack surface.
Diagnostics. OEM logcat has Battery AppStats lines (com.oplus.persist.system on OPPO, similar on other vendors) emitted every few minutes — they break down CPU / Wifi / Camera / WakeLock time per package. After any background-work change, leave the device idle for an hour and compare CpuFgTime / CpuBgTime / WifiRxBytes / WifiTxBytes / WakeLockBgTime against a baseline run on the same device. If CpuBgTime rises by more than a few seconds per hour, you've introduced a wakeup loop somewhere — find it.
When you add a new UI screen or hot path that touches the main thread, mentally check it against the frame list before committing. When you add anything that runs in the background, on a timer, or holds a system resource, check it against the battery list. The Pre-flight checklist in Step 8 reflects both.
A1. Decide if the binding needs to change
The Go binding (warpdroid/node/) is a separate Go module (module github.com/Warp-net/android-binding) with its own go.mod and its own dependency tree. It does not import anything from the parent warpnet module. It exposes only six entry points (Initialize, Connect, Stream, PeerID, IsConnected, Disconnect, Pause, Resume, Shutdown) — all string-in, string-out, because gomobile-bind cannot translate complex types across the JNI.
The binding needs to change only if:
you need a new lifecycle entry point (e.g. exposing the private key for signing — see §A4);
your route requires per-stream timeouts different from the binding's hard-coded 30s.
For a routine "add new Warpnet route" task, the binding does not change. You don't recompile warpnet.aar. You add the route in Kotlin, mirror DTOs, write a WarpnetRepository method.
A2. Add the path constant in Kotlin
In warpdroid/warpnet-transport/src/main/kotlin/site/warpnet/transport/ProtocolIds.kt, add the constant under the matching group header (// public reads, // private writes, etc.). Same string as the Go constant.
A3. Mirror the DTOs in Moshi
In warpdroid/warpnet-transport/src/main/kotlin/site/warpnet/transport/dto/WarpnetDtos.kt, add @JsonClass(generateAdapter = true) data classes for the request payload and the response shape. Rules:
Field names in Kotlin are camelCase. JSON wire names use @Json(name = "snake_case_name") when they differ.
The wire names must match the Go struct's json:"..." tags exactly.
Default values where the Go side treats a missing field as zero-value (string "", number 0, boolean false). This avoids Moshi throwing on responses that happen to omit optional fields.
Strings for IDs (String, not custom types) — the Go side uses domain.ID which is itself a string alias.
Time fields are RFC3339 strings on the wire — keep them as String in Kotlin DTOs and convert at the mapper boundary if needed.
Only add DTOs the Tusky side actually uses. The DTO file's docstring states this rule explicitly: don't bulk-mirror event.go — add exactly what WarpnetRepository needs.
If the response is a paged collection, mirror both the wrapper ({ items, cursor }) and the inner items.
A4. Decide signing
Every request goes through WarpnetClient.request(protocolId, bodyJson), which builds a WarpnetEnvelope and asks the injected EnvelopeSigner to sign the body. Today only NoOpSigner exists — it throws WarpnetException.SigningUnavailable because the gomobile binding does not export the Ed25519 private key.
This means at the time of writing, routes whose middleware verifies the envelope signature do not work end-to-end through warpdroid. Check core/middleware/middleware.go (around the line where signature verification happens) for which routes are gated.
What to do depends on your route:
Read-only public routes that the middleware does not gate (the bulk of PUBLIC_GET_* calls — that's why WarpnetRepository covers timeline/user/replies/stats today): no signing concerns. Proceed normally.
Write routes or signature-verified routes: the route will fail at runtime with WarpnetException.SigningUnavailable until the binding is extended. Either land your code with a TODO comment referencing EnvelopeSigner.kt (it already documents the problem in detail), or — if your task is specifically to fix this — extend mobile.go with either Sign(body) string or SetIdentityKey(pem) string, then implement BindingSigner in Kotlin and register it via Hilt.
Do not invent a parallel signing mechanism. The envelope's signature must match what the Go middleware verifies (Ed25519 over the body bytes, base64-encoded).
A5. Add a WarpnetRepository method (Tusky-facing)
In warpdroid/app/src/main/java/com/keylesspalace/tusky/warpnet/WarpnetRepository.kt, add a suspend fun modelled on getStatus/favouriteStatus/reblogStatus (~145–290). The pattern:
suspendfunfooSomething(tweetId: String, userId: String): TuskyShapedReturn {
val raw = client.request(
ProtocolIds.PUBLIC_POST_FOO,
fooEventAdapter.toJson(FooEvent(tweetId = tweetId, userId = userId)),
)
val resp = fooResponseAdapter.fromJson(raw)
?: throw IllegalStateException("foo returned empty body for $tweetId")
return resp.toTuskyShape()
}
Rules:
client.request(...) is the only allowed transport — never reach into WarpnetClient internals or call the binding directly.
Adapters are private val fooAdapter = moshi.adapter<FooEvent>() declared near the top of the class (~80–100). Add yours alongside the existing ones.
Convert to/from Tusky entities (Status, Account, Notification, Relationship, TimelineAccount) via extension functions in WarpnetMapper.kt. Don't return raw Warpnet DTOs from WarpnetRepository — the rest of the Tusky app only knows Mastodon-shaped types.
Use runCatching { ... }.getOrNull() for optional sub-fetches that should degrade gracefully (e.g. fetching stats alongside a tweet). Hard throw for anything required.
Throwing IllegalStateException for empty bodies is the established style.
A6. Extend the mapper if shapes don't fit
In WarpnetMapper.kt, add fun WarpnetX.toTuskyY(): Y extensions for any new wire→Tusky conversion. The mapper file's header docstring explains the philosophy: Mastodon types are richer than Warpnet types (rich HTML content, attachments, visibility, emojis), so when translating Warpnet→Tusky pick safe defaults (public visibility, no attachments, empty emoji list). Don't invent fields that the wire didn't send.
When translating Tusky→Warpnet (e.g. for a postStatus type call), the Tusky side carries fields the Warpnet side ignores — drop them silently, don't throw.
A7. The .aar is committed; gomobile is NOT in CI
This is the single most surprising thing about the warpdroid build. The CI workflow (.github/workflows/build-warpdroid.yml) runs:
It does not invoke gomobile bind. The compiled warpnet-transport/libs/warpnet.aar is checked into the repository. This means:
Pure-Kotlin changes (new ProtocolIds, new DTO, new WarpnetRepository method) are CI-buildable as-is.
Changes to warpdroid/node/*.go require a manual rebuild via warpdroid/node/build-native.sh, which produces warpnet.aar. The script must be run locally; CI will not catch a stale .aar.
Commits that change Go code under warpdroid/node/must also commit the regenerated warpnet-transport/libs/warpnet.aar (and warpnet-sources.jar). Reviewers should expect both. A binary-only diff for the .aar is normal.
The warpdroid/node/ Go module has its own go.mod with go 1.26.0 and pulls dependencies directly (no vendor). Building it requires Go 1.26+, gomobile, and Android NDK on the developer's machine.
A8. Where Java/Kotlin code goes — Tusky vs warpnet-transport
The Android app is a fork of Tusky with a Warpnet integration grafted on. Two rules of thumb for where new code belongs:
Tusky-namespaced code (com.keylesspalace.tusky.*) is the existing Mastodon client. Touch it only when the feature surface is UI/UX (e.g. a new screen that consumes WarpnetRepository). Stay close to existing patterns — Hilt for DI, coroutines for async, paging library for lists.
warpnet-transport-namespaced code (site.warpnet.transport.*) is the wire layer. Anything that knows about Warpnet protocol IDs, envelopes, signatures, or libp2p semantics goes here. The Tusky side must never see a ProtocolIds.X or build a WarpnetEnvelope.
The split exists so that the Tusky fork can be re-merged with upstream Tusky changes without touching Warpnet wire concerns, and so that Warpnet wire concerns can be tested independently of Android.
A9. Tests on the warpdroid side
Tests are JVM tests (src/test/kotlin/), JUnit + standard Kotlin. The transport module ships a fake WarpnetBinding that returns canned strings — use it instead of touching the real gomobile binding. Verify:
the envelope wraps the body correctly (path, version, signature placeholder, body verbatim);
the Moshi DTO round-trips with the Go-side wire format (paste a representative JSON sample from a captured Vue request to make sure tags align);
error paths surface WarpnetException.ProtocolError with the correct code/message when the binding returns an error JSON;
Do not write instrumented Android tests for new routes — the wire layer is plain JVM.
Step 7 — Server-side tests
A new handler is incomplete without a _test.go. Style is dictated by the existing tests in core/handler/:
Plain testing package, no testify.
package handler (white-box, not handler_test).
File starts with //nolint:all — the existing tests do this.
Stubs are local structs with <method>Fn func(...) (...) fields, falling back to a default if nil. See stubLikeRepo in like_test.go for the canonical shape.
Subtests with t.Run("invalid payload", func(t *testing.T) {...}). Always cover at least: invalid payload, each empty-field validation, the happy path, and remote-call branches if any.
Helper marshal(t, ...) is already defined in like_test.go and reused — if you need it and it's not in package scope, copy it.
The repo also has core/handler/notifications_e2e_test.go for end-to-end style; only add to that level if the feature is genuinely cross-handler.
Step 8 — Pre-flight checklist before commit
Before git commit, verify each of these. Missing any one is a real bug, not a style nit. Skip the warpdroid block if the route is desktop-only; skip the Vue block if the route is warpdroid-only.
Always (server side):
AGPL header present at top of every new .go file. Copy it verbatim from core/handler/like.go lines 1–26 (the 22-line block plus // Copyright ... and // SPDX-License-Identifier: AGPL-3.0-or-later).
Path constant in event/paths.go.
DTO in event/event.go with the // XxxEvent defines model for XxxEvent. godoc.
Run go test -short -p 8 -v ./core/handler/... ./database/... locally (this is what make tests does for the whole repo).
go vet ./... clean. golangci-lint run clean (CI runs it; config in .golangci.yaml).
vendor/ not modified. If go mod accidentally touched it, git checkout vendor/.
version file: do not bump manually — the pre-commit githook auto-bumps the patch on non-main branches and updates snap/snapcraft.yaml. Make sure make setup-hooks has been run (git config core.hooksPath .githooks).
If the route is callable from the Vue desktop UI:
Same path string in frontend/src/service/service.js, byte-identical with event/paths.go.
Service method on warpnetService exists, parameter list matches what components will call, body keys are snake_case matching the Go struct's json tags.
If the route is callable from warpdroid:
Same path string in warpdroid/warpnet-transport/.../ProtocolIds.kt, byte-identical with the other two.
DTOs added to warpdroid/warpnet-transport/.../dto/WarpnetDtos.kt with @Json(name = "snake_case") for any field whose Kotlin name differs.
WarpnetRepository method added, returns Tusky-shaped types (not raw Warpnet DTOs).
If the response shape doesn't fit any existing Tusky entity, mapper extension added in WarpnetMapper.kt with safe defaults for fields the wire doesn't carry.
If the route requires envelope signature verification (most write routes): added a TODO comment referencing EnvelopeSigner.kt's known limitation, OR landed the binding extension first.
If warpdroid/node/*.go was changed: regenerated warpnet-transport/libs/warpnet.aar via warpdroid/node/build-native.sh and committed both the .aar and the .jar.
JVM tests added for envelope shape and DTO round-trip; instrumented tests not needed.
./gradlew :app:assembleDebug succeeds locally (CI runs the same).
Performance budget (§A0): new UI code does not introduce ConstraintLayout inside a LazyList/LazyColumn item, does not do synchronous IO on the main thread (setImageURI on warpnet://, SharedPreferences first-read, File.exists/listFiles in onCreate, etc.), uses Glide via WarpdroidAsyncImage for warpnet:// URIs, and keeps any new composables small enough that first-compose JIT doesn't add a visible Davey frame.
Battery budget (§A0): no new ForegroundService for the libp2p connection, no Handler.postDelayed / LaunchedEffect { while(true) } polling loops, no WakeLock / WifiLock / setKeepScreenOn(true), no new periodic worker under 15 min and without RequiresBatteryNotLow + NetworkType.CONNECTED. Lifecycle pause/resume of the binding still works after the change. No new runtime permissions added without an explicit design discussion.
Always (general):
Branch name follows IssueNumber/AmazingFeature. Commit message imperative, present tense.
PR description names the path you added so reviewers can grep it across all three sources.
Common mistakes specific to this codebase
Adding the path string only in some places. The Go, JS, and Kotlin path constants are independent strings. The compiler does not link them. A typo or omission = silent failure with not attached server node in logs (Go side) or a 30-second hang then a TransportFailure (Android side). When adding a route used by both clients, grep all three: event/paths.go, frontend/src/service/service.js, warpdroid/warpnet-transport/.../ProtocolIds.kt.
Using encoding/json instead of the project's json package. Tests will pass, runtime may pass, but you'll diverge from project-wide marshalling settings (the project uses jsoniter under the hood for performance and number handling).
Forgetting //nolint:govet on the registration entry. The block uses positional struct literals which govet doesn't like. Match neighbours.
Putting domain.ID fields where string belongs or vice versa. IDs in events are domain.ID. In handler logic they're often used as string — domain.ID is a typed string alias, the conversion is implicit.
Calling streamer.GenericStream for a private route. Private routes never propagate. Only public routes call out to other nodes. If you wrote a private handler that needs to call a peer, you've split the wrong way — that should be a public route.
Treating ErrNodeIsOffline as a real error. It isn't. The local write succeeded; the propagation failed; return the local result and move on. See like.go for the exact branch.
Bumping the version file by hand. The pre-commit hook does it. Doing it manually creates a conflict with the hook.
Modifying vendor/ directly or via go get without intent. vendor/ changes need to be a deliberate make update-deps commit, separate from feature work.
Adding a put or patch verb. Not used in the project. Use post for create-or-update, delete for delete.
Warpdroid-specific:
Forgetting that the .aar is committed. A pure-Kotlin change builds and runs in CI even with a stale .aar — meaning local Go changes won't be visible until you regenerate. Conversely, if you change warpdroid/node/*.go and forget to rebuild the .aar, your changes ship to no one. The .aar is a build artefact in source-control by convention, treat it as a deliberate commit.
Importing from the parent warpnet Go module inside warpdroid/node/. They are separate Go modules. The thin client must stay self-contained: libp2p, multiaddr, dht. Pulling in warpnet/event or warpnet/domain would balloon the .aar by tens of MB and break gomobile.
Returning Warpnet wire DTOs from WarpnetRepository. The Tusky side (com.keylesspalace.tusky.*) must only see Mastodon-shaped entities. Confine WarpnetTweet, WarpnetUser, etc. to the com.keylesspalace.tusky.warpnet package and translate at the boundary.
Building a WarpnetEnvelope outside WarpnetClient. Envelope construction (message_id, version, timestamp, signature) is centralised. If your call needs a non-standard envelope, the right move is to teach WarpnetClient about the variant — not to reach around it.
Calling the gomobile binding methods directly from Tusky code. All transport goes through WarpnetClient. The binding's global singleton state and the mutex in WarpnetClient are paired — bypassing the client risks racing pause/resume against in-flight requests.
Returning Tusky-shaped types from warpnet-transport. The transport module must not depend on Tusky entities. The split is enforced by Gradle module boundaries — if you find yourself needing a Tusky import in site.warpnet.transport.*, the conversion belongs in WarpnetMapper, not in the transport.
When the task is moderation, admin, or pair-related
These have extra ceremony — see core/handler/admin.go, core/handler/pair.go, core/handler/moderation.go. They use challenge/response signing and PSK plumbing that regular handlers don't. Read those files before treating an admin-route task as "just another handler".
