Telegram-iOS/docs/superpowers/postbox-refactor-log.md

Postbox → TelegramEngine refactor — log

This file is the historical record of the Postbox → TelegramEngine refactor. It is not loaded by default into AI sessions (only CLAUDE.md is). Read this file when you need wave-specific context, a full worked example of a pattern, or the running tally of module Postbox-freeness.

The short, actively-maintained rules and references live in CLAUDE.md under the "Postbox → TelegramEngine refactor" section. This file holds the narrative backstory, verbose example scripts, and per-wave outcomes that would otherwise bloat every AI session's context.

---

Wave-selection guidance — full versions

The following subsections are trimmed to terse bullets in CLAUDE.md. Full versions (rule + backstory + scripts + per-wave examples) live here.

Shape-selection backstory

The "leaf module, drop Postbox in isolation" approach only works for modules whose public API doesn't leak Postbox domain types. Most candidate leaf modules DO leak such types (postbox: Postbox / account: Account in public inits, Media/Message in public function parameters). Those modules need paired caller-migration waves, not isolated refactors.

Before selecting a wave's module list, grep each candidate for:

• :\s*Postbox\b, :\s*Account\b, :\s*MediaBox\b in public signatures → abandon candidate
• Media/Message as public parameter types → likely needs paired wave with callers

Inventory at execution time, not just planning time

Inventory at execution time, not just planning time. Wave 2's SaveToCameraRoll task was planned from a narrow grep that only matched MediaResource/TelegramMediaResource and missed three postbox: Postbox public-function leaks plus multiple postbox.mediaBox.* bodies. Planning-time inventory should grep the full set \b(postbox|mediaBox|transaction|PostboxView|combinedView|MediaResource|PostboxDecoder|PostboxEncoder|MemoryBuffer)\b|^import Postbox over the module's Sources, not just the tokens specific to that wave's goal. If the planning inventory under-counts, the executor should re-inventory at Task-1 time and abandon early before editing code.

Two feasible wave shapes

Two feasible wave shapes. Wave 1 tried "per-module Postbox drop". Wave 2 tried "per-engine-facade-API migrate MediaResource to EngineMediaResource (modify in place, update all call sites in one commit)". The second shape worked well: narrow, clean commits, no abandonment cascade. Prefer it when the refactor target is an API surface that multiple consumer modules depend on.

Enum-payload migrations need a full case-site grep

Enum-payload migrations need a full case-site grep, not just a facade call-site grep. If a wave changes the payload type of a public enum (wave 4 changed UploadStickerStatus.complete's payload from CloudDocumentMediaResource to EngineMediaResource), inventory ALL construction and destructure sites of the enum across TelegramCore, not just call sites of the facade that returns it. Wave 4's plan undercounted by 6 consumer sites inside ImportStickers.swift itself (3 shortcut .complete(...) constructions in guard branches, 3 destructure+field-access sites using CloudDocumentMediaResource-specific members). For enum-payload waves, grep case \.|let \.|\.<caseName>\( over the enum's defining module before execution and add those sites to the plan.

Unused-import sweeps are a valid wave shape

Unused-import sweeps are a valid wave shape. After a round of facade migrations, consumer files accumulate import Postbox lines whose last semantic use was removed. Periodically sweep these:

1. grep -rl "^import Postbox$" submodules --include="*.swift" | grep -vE "/(TelegramCore|Postbox|TelegramApi)/" generates the candidate list.
2. sed -i '' '/^import Postbox$/d' <file> (BSD sed) speculatively drops the import from every candidate.
3. Run the full build with --continueOnError — without --keep_going, bazel stops at the first failing target and surfaces only a few errors per iteration. Make.py forwards --continueOnError to --keep_going; always use it.
4. Each iteration: extract failing files via grep -E "^submodules/.*\.swift:[0-9]+:[0-9]+: error:" <build-out> | awk -F: '{print $1}' | sort -u, restore via git checkout -- <file>, rebuild.
5. The dependency graph has many layers (wave 6 needed ~18 rebuilds to reach a clean build). Per-iteration failures shrink roughly: 18 → 4 → 5 → 3 → 12 → 4 → 13 → 9 → 11 → ... Accelerate by doing pattern-based preemptive restores after the first few iterations: scan still-dropped files for tokens that are definitively Postbox-only (MediaBox, PostboxCoding, PostboxDecoder, PostboxEncoder, TempBoxFile, ValueBoxKey, Postbox\b, PeerId, MessageId, MediaId, MessageIndex, MessageAndThreadId, PeerNameIndex, etc. — note that CLAUDE.md's "engine typealias cheat sheet" arrows are migration targets, not typealiases in TelegramCore — PeerId etc. are still raw Postbox types and files using them need import Postbox) and restore those files in bulk.
6. Only restore files from the candidate set. If errors surface in TelegramCore, Postbox, or TelegramApi, halt — the sweep has cascaded beyond its scope.
7. Commit the surviving drops as one atomic commit.

Tally impact from a sweep: dozens of consumer modules can become Postbox-free in a single commit. First run (wave 6): 782 candidates, 18 iterations, 183 survivors, 189 modules newly Postbox-free. Re-run after every 2-3 facade-migration waves.

Public-Postbox-type inventory (wave-11-pattern planning)

Public-Postbox-type inventory (wave-11-pattern planning). For wave-11-shape candidates (modules whose public init takes postbox: Postbox, network: Network purely for avatar/setPeer forwarding), grepping only for Postbox/Network tokens undercounts — public-surface types defined in Postbox can leak without ever naming "Postbox" literally. Wave 16 hit this: the plan missed EngineMessageHistoryThread.Info? and PeerStoryStats?, both Postbox-defined types whose names don't include "Postbox". Mitigation: build a Postbox-defined-public-types allowlist once, then grep the candidate module against it.

# Build allowlist once (or re-run if Postbox sources change):
grep -rhE "^public\s+(class|struct|enum|protocol|typealias)\s+\w+" submodules/Postbox/Sources/ \
  | awk '{print $3}' | sed 's/[(:<].*//' | sort -u > /tmp/postbox-public-types.txt

# Then, for each candidate module, grep its sources for any of those names:
grep -rhoE "\b($(cat /tmp/postbox-public-types.txt | tr '\n' '|' | sed 's/|$//'))\b" \
  submodules/<CandidateModule>/Sources/ | sort -u

Any hit in a public-surface position (field type, init param type, enum payload type, generic arg) that isn't already a documented typealias is a blocker. "Engine"-prefixed types can still be Postbox-defined — don't trust naming conventions, grep for the defining module. If the module hits only Postbox itself (i.e., literal Postbox/Network pair), it's a clean wave-11 candidate. Otherwise, decide per leak: (a) move the type to TelegramCore if it's a namespace-only class (wave 16a pattern — prototype: EngineMessageHistoryThread), (b) accept that the module can't become Postbox-free and ship a partial engine:/stateManager: collapse that keeps import Postbox (wave 16b pattern — PeerStoryStats is too baked into Postbox views to move cleanly), (c) abandon the candidate.

Wave-shape G: facade addition + consumer sweep in one commit

Wave-shape G: facade addition + consumer sweep in one commit. Validated at scale across waves 19-26. Six consecutive sessions migrated ~95 consumer sites and added ~15 mediaBox facades, all with clean first-pass builds (exception: wave 26 needed a second pass to add import RangeSet). Shape recipe:

1. Target: a MediaBox (or similar Postbox type) method where Postbox's signature uses clean leaf types (MediaResourceId, Data, String, Bool) and the return type is either non-Postbox or has an existing Engine* wrapper.
2. Pre-flight inventory: grep context\.account\.postbox\.mediaBox\.<methodName> over submodules/ (excluding TelegramCore/Postbox/TelegramApi). Classify each hit:
   • Shape A: context.account.postbox.mediaBox.X(...) → migratable.
   • Shape B: context.account.postbox.mediaBox.X(id: ...) (different overload) → migratable with identical pattern.
   • Shape C: account.postbox.mediaBox.X(...) where account: Account is a local (not AccountContext) → skip this wave (needs per-module rework).
   • Shape D: self.postbox.mediaBox.X(...) where postbox: Postbox is a stored field → skip this wave.
   • Plus: check for accountManager.mediaBox.X(...) which is Account-manager-scoped, a different migration path entirely. Never migrate via TelegramEngine.Resources.*.
3. Facade design rules:
   • Signatures take EngineMediaResource.Id (MediaResourceId aliased at call site via EngineMediaResource.Id(x.id)) or EngineMediaResource (wraps resource when the Postbox overload takes a resource with members accessed via .id).
   • Parameters with Bool defaults (synchronous: Bool = false) preserve defaults on the facade.
   • Return types: prefer Void, String, String?, Signal<T, NoError> where T is a non-Postbox type or an Engine* wrapper. Where Postbox return types are wrapped (e.g., Signal<MediaResourceData, NoError> → Signal<EngineMediaResource.ResourceData, NoError>), confirm the Engine* wrapper exists and decide whether consumer-side field-access rewrites are acceptable for the wave.
4. WIP interference check: before starting, git status --short | grep -v "^??" to list modified files. If any Shape-A site is in a WIP file, either skip those sites (document the skip in the outcome) or wait for the WIP to commit. Wave 23 hit this in ChatMessageInteractiveMediaNode.swift.
5. Name collision check: if a facade return type names a Swift stdlib type that has availability restrictions (e.g., RangeSet — iOS 18+), verify the third-party module import is present in TelegramEngineResources.swift. Wave 26 needed import RangeSet.
6. Replace_all usage: for files with duplicate identical call text, replace_all=true on the exact call expression (without leading whitespace) batches the migration. When leading whitespace varies across identical-call sites within a file, the tool still matches if the unchanged prefix (context.account.postbox.mediaBox.X(...)) is unique enough — but verify via post-edit grep.
7. Cheapness: ~5-50 sites per wave, single atomic commit, expected first-pass-clean build. If post-migration grep for context\.account\.postbox\.mediaBox\.<methodName> returns empty (exclude Shape-C/D) and build is green, commit.

---

Wave 1 outcome (2026-04-16)

4 modules done: ChatInterfaceState, ChatSendMessageActionUI, ContactListUI, DrawingUI. 6 modules abandoned with recorded reasons in the wave-1 plan: ActionSheetPeerItem, ChatListSearchRecentPeersNode, DirectMediaImageCache, FetchManagerImpl, GalleryData, ICloudResources.

Wave 2 outcome (2026-04-17)

5 TelegramEngine facades migrated to EngineMediaResource (signatures changed in place; _internal_* Postbox layer unchanged):

• TelegramEngine.Peers.uploadedPeerPhoto, uploadedPeerVideo, updatePeerPhoto
• TelegramEngine.AccountData.updateAccountPhoto, updateFallbackPhoto
• TelegramEngine.Contacts.updateContactPhoto
• TelegramEngine.Auth.uploadedPeerVideo

1 consumer submodule fully de-Postboxed: MapResourceToAvatarSizes (signature changed from (postbox: Postbox, resource: MediaResource, …) to (engine: TelegramEngine, resource: EngineMediaResource, …); 27 call sites migrated).

1 consumer signal type swapped: AuthorizationUI/AuthorizationSequenceController.swift (Signal<TelegramMediaResource?> → Signal<EngineMediaResource?>).

1 task abandoned with recorded reason in the wave-2 plan: SaveToCameraRoll (full-module Postbox coupling, needs its own wave).

Wave 3 outcome (2026-04-18)

3 thin forwarders added on TelegramEngine.Resources over MediaBox:

• fetch(reference:userLocation:userContentType:) → Signal<FetchResourceSourceType, FetchResourceError> (Postbox return types remain a documented accepted leak)
• status(resource: EngineMediaResource) → Signal<EngineMediaResource.FetchStatus, NoError>
• data(resource: EngineMediaResource, pathExtension:, waitUntilFetchStatus:) → Signal<EngineMediaResource.ResourceData, NoError> (takes a Bool rather than exposing ResourceDataRequestOption, per YAGNI)

1 consumer submodule fully de-Postboxed: SaveToCameraRoll. Public signatures changed from (context:, postbox: Postbox, userLocation:, …) to (context:, userLocation:, …); FetchMediaDataState.data payload changed from MediaResourceData to EngineMediaResource.ResourceData; internals rewired through context.engine.resources.*. 23 call sites across 14 files migrated atomically with the module.

Pre-flight verified that ShareController.swift:2406's self.currentContext.stateManager.postbox is equivalent to context.account.postbox in the ShareControllerAppAccountContext path (because AccountStateManager is constructed with the account's own postbox), so the postbox: argument could be dropped without behavior change.

No tasks abandoned. Shape validated: "per-engine-facade-API migration + full consumer module rewrite" (the wave-2 shape, scaled up to a full module drop).

Plan: docs/superpowers/plans/2026-04-18-postbox-to-telegramengine-wave-3.md

Wave 4 outcome (2026-04-18)

1 TelegramEngine facade migrated in place to EnginePeer + EngineMediaResource (signature changed; _internal_uploadSticker keeps its raw Peer/MediaResource parameter list):

• TelegramEngine.Stickers.uploadSticker(peer: Peer → EnginePeer, resource: MediaResource → EngineMediaResource, thumbnail: MediaResource? → EngineMediaResource?, …)

1 public enum payload migrated: UploadStickerStatus.complete(CloudDocumentMediaResource, String) → .complete(EngineMediaResource, String). _internal_uploadSticker wraps EngineMediaResource(uploadedResource) at its one .complete(...) result-construction site — a narrow, spec-allowed one-line deviation from "internal Postbox-facing stays raw", taken to keep UploadStickerStatus as a single public enum.

Plan-time inventory undercount — worth recording as a lesson. The spec and plan enumerated 2 external call sites and 1 internal construction site. Execution uncovered 6 additional consumer sites inside ImportStickers.swift itself that also needed adapting: 3 shortcut .complete(...) construction sites (lines 204, 371, 492, each emitting .complete(CloudDocumentMediaResource, String) directly from as? CloudDocumentMediaResource guards) and 3 destructure sites (lines 216, 384, 505) that accessed CloudDocumentMediaResource-specific fields. Each construction site now wraps via EngineMediaResource(resource); each destructure site unwraps with let rawResource = resource._asResource() as? CloudDocumentMediaResource. MediaEditorScreen's two stickerFile(resource:) calls also needed as! TelegramMediaResource casts because _asResource() returns the Postbox MediaResource protocol while stickerFile takes the TelegramCore TelegramMediaResource sub-protocol. Future planning-time inventory for enum-payload migrations should grep not only call-sites of the facade but every case .complete / case let .complete of the migrated enum across the whole TelegramCore source tree.

2 external call sites migrated atomically with the facade:

• submodules/ImportStickerPackUI/Sources/ImportStickerPackController.swift:91 (plus a peer: Peer → EnginePeer(peer) wrap, since the local peer comes from postbox.loadedPeerWithId(...) which returns raw Peer)
• submodules/TelegramUI/Components/MediaEditorScreen/Sources/MediaEditorScreen.swift:8099 (plus 6 cascading sites inside the enclosing block for the new UploadStickerStatus.complete payload)

No module becomes Postbox-free in this wave (both caller files import Postbox for unrelated reasons).

Plan: docs/superpowers/plans/2026-04-18-postbox-to-telegramengine-wave-4.md

Wave 5 outcome (2026-04-18)

Completes the last explicitly-named future-wave candidate from the wave-2 final review.

uploadSecureIdFile(context: SecureIdAccessContext, postbox: Postbox, network: Network, resource: MediaResource) migrated in place to (context:, engine: TelegramEngine, resource: EngineMediaResource). Function body accesses raw Postbox types via engine.account.postbox / engine.account.network (internal Postbox-facing layer stays raw per the standing rule).

1 consumer submodule fully de-Postboxed: SecureIdVerificationDocumentsContext (PassportUI/Sources). Signature changed from (postbox: Postbox, network: Network, context: SecureIdAccessContext, update: ...) to (engine: TelegramEngine, context: SecureIdAccessContext, update: ...); stored props collapsed into a single engine: TelegramEngine field. One instantiation site updated in the same commit.

After this wave, the "Known future-wave candidates" list contains only the 4 permanently-blocked classes conforming to TelegramMediaResource.

Plan: docs/superpowers/plans/2026-04-18-postbox-to-telegramengine-wave-5.md

Wave 6 outcome (2026-04-19)

First build-verified unused-import sweep. Ran the speculative-drop + build-verify methodology (see "Unused-import sweeps" under Wave-selection guidance above): dropped import Postbox from all 782 consumer files where a plain ^import Postbox$ line appeared, iterated 18 full builds with --continueOnError, restoring imports on files that failed to compile.

183 drops survived (single atomic commit 7b2b74e79b, 0 insertions / 183 deletions). 189 modules transitioned to Postbox-free status — full list is inferable by running the methodology's module-scan against HEAD. Representative additions spanning alphabetically: AccountUtils, ActivityIndicator, AdUI, AlertUI, AnimatedStickerNode, AppLock, AttachmentTextInputPanelNode, BotPaymentsUI, CalendarMessageScreen, CallListUI, Camera, ChatImportUI, etc. The running tally below preserves the per-module enumeration only for the ~10 individually-documented waves 1–5 modules. Wave 6's 189 additions are not re-enumerated here because the size would overwhelm the doc; see git show 7b2b74e79b --stat for the per-file breakdown and grep -rL "^(@_exported )?import Postbox" submodules/*/Sources --include="*.swift" for the current per-module status.

Deviation from plan: the plan capped at 3 iterations; execution needed 18 because the dependency graph is deep and each bazel build surfaces only the currently-compilable layer. Pattern-based preemptive restores (using the symbol list in the "Unused-import sweeps" guidance) were used from iteration 9 onward to accelerate convergence from iteration-by-iteration single-file restores to bulk restores. No unexpected path cascades; no abandoned state.

Plan: docs/superpowers/plans/2026-04-19-postbox-to-telegramengine-wave-6.md

Wave 7 outcome (2026-04-20)

Closed out the seven remaining raw-Postbox leaks in TelegramEngine.* public facades surfaced by a post-wave-6 scouting pass. Single atomic commit, one full build, zero abandonment.

Seven TelegramEngine facades migrated in place (all _internal_* implementations kept raw per the standing rule):

Messages (3):

• downloadMessage(messageId:) — return Signal<Message?, NoError> → Signal<EngineMessage?, NoError>. Return-side wrap via |> map { $0.flatMap(EngineMessage.init) }.
• topPeerActiveLiveLocationMessages(peerId:) — return Signal<(Peer?, [Message]), NoError> → Signal<(EnginePeer?, [EngineMessage]), NoError>. Return-side tuple wrap.
• getSynchronizeAutosaveItemOperations() — deleted. Dead facade: sole caller (StoreDownloadedMedia.swift) already bypassed it by calling _internal_getSynchronizeAutosaveItemOperations directly inside its own transaction block.

Peers (1):

• updatedRemotePeer(peer:) — return Signal<Peer, UpdatedRemotePeerError> → Signal<EnginePeer, UpdatedRemotePeerError>. PeerReference param kept as-is (no EnginePeer.Reference alias today). The sole call site in ChannelAdminsController.swift uses ignoreValues, so no caller change was needed.

Resources (4):

• renderStorageUsageStatsMessages(…existingMessages:) — [EngineMessage.Id: Message] → [EngineMessage.Id: EngineMessage] on both sides. Facade unwraps input via .mapValues { $0._asMessage() }, wraps output via .mapValues(EngineMessage.init).
• clearStorage(peerId:categories:includeMessages:excludeMessages:) — [Message] → [EngineMessage]. Facade unwraps via .map { $0._asMessage() }.
• clearStorage(peerIds:includeMessages:excludeMessages:) — same shape.
• clearStorage(messages:) — same shape. No external callers; migrated for overload-set consistency.

Consumer call-site updates (5 files):

• ChatListUI/Sources/ChatListSearchListPaneNode.swift: dropped now-redundant .flatMap(EngineMessage.init) wrap at the downloadMessage call site.
• LocationUI/Sources/LocationViewControllerNode.swift: dropped now-redundant .map(EngineMessage.init) at the topPeerActiveLiveLocationMessages call site.
• LiveLocationManager/Sources/LiveLocationSummaryManager.swift: dropped redundant EnginePeer(author) / EngineMessage(message) construction (author, message are now already EnginePeer / EngineMessage).
• TelegramUI/Components/StorageUsageScreen/Sources/StorageUsageScreen.swift: bridged at 4 facade-call points (1 renderStorageUsageStatsMessages, 2 clearStorage overloads with message arrays; the includeMessages: [], excludeMessages: [] site at line 3038 needed no change as empty arrays infer to [EngineMessage] just as well).

Minimal-scope bridging. StorageUsageScreen.swift still has 43 raw Message/MessageId references inside its AggregatedData helper class and surrounding logic — not touched in this wave. A future "StorageUsageScreen full de-Postbox" wave would drop those (migrate AggregatedData.messages: [MessageId: Message] → [EngineMessage.Id: EngineMessage], clearIncludeMessages: [Message] → [EngineMessage], etc.) and potentially drop import Postbox. Out of scope here.

No modules became Postbox-free in this wave — all five touched consumer files still import Postbox for reasons unrelated to the migrated facades.

Plan / record: docs/superpowers/plans/2026-04-20-postbox-to-telegramengine-wave-7.md.

After this wave, the "Known future-wave candidates" list contains only the 4 permanently-blocked classes conforming to TelegramMediaResource. The full public TelegramEngine.* facade surface is now engine-typed (modulo those four types).

Wave 8 outcome (2026-04-20)

StorageUsageScreen consumer-module migration of raw Message domain types to EngineMessage. Scope explicitly narrower than a full de-Postbox: two files touched, module remains import Postbox due to two out-of-scope site clusters.

Types migrated:

• StorageFileListPanelComponent.Item.message: Message → EngineMessage (item type co-located with the panel component).
• StorageUsageScreen.Component.AggregatedData.messages: [MessageId: Message] → [EngineMessage.Id: EngineMessage]; .clearIncludeMessages / .clearExcludeMessages: [Message] → [EngineMessage]. Init param updated to match.
• StorageUsageScreen.Component.SelectionState.togglePeer(availableMessages:) param: [EngineMessage.Id: Message] → [EngineMessage.Id: EngineMessage].
• StorageUsageScreen.Component.RenderResult.messages: [MessageId: Message] → [EngineMessage.Id: EngineMessage].
• openMessage(message: Message) → openMessage(message: EngineMessage) (external OpenChatMessageParams.message / chatMediaListPreviewControllerData(message:) calls unwrap via message._asMessage() at the two call sites — those APIs still take raw Message).

Wave-7 facade-boundary bridging dropped: the renderStorageUsageStatsMessages call-site's (…).mapValues(EngineMessage.init) / .mapValues { $0._asMessage() } bridges and the two clearStorage call sites' .map(EngineMessage.init) wraps all vanish — AggregatedData.messages / .clearIncludeMessages / .clearExcludeMessages are now engine-typed and pass through the facade unchanged. Inside the AggregatedData.updateSelected... selected-messages accumulation loop, four item.message._asMessage() calls (for imageItems, which hold EngineMessage) drop back to plain item.message since the target array is now [EngineMessage]. And StorageMediaGridPanelComponent.Item(message: EngineMessage(message), …) drops the EngineMessage(…) wrap since message is already EngineMessage.

Out of scope — future-wave candidates (module still imports Postbox):

• StorageUsageScreen.swift:1047-1062 and 3131-3185: preferences-view observation of AccountSpecificCacheStorageSettings via postbox.combinedView + PreferencesView, and a postbox.transaction { transaction in transaction.getPeer / transaction.getPeerCachedData as? CachedGroupData / CachedChannelData } block classifying peer-storage-timeout exceptions. Substantial: requires EngineData-subscription rewrite for the preferences observation, plus engine-API equivalents for peer-category classification + cached-data subscriber counts.
• StorageFileListPanelComponent.swift:105: Icon.media(Media, TelegramMediaImageRepresentation) enum case, constructed only as .media(TelegramMediaFile, …) or .media(TelegramMediaImage, …) (both TelegramCore types). Trivial future wave: split into .mediaFile(TelegramMediaFile, …) / .mediaImage(TelegramMediaImage, …), drop import Postbox.

Single atomic commit. Build verified green (59s incremental build, 27 actions). Net −11 lines in StorageUsageScreen.swift (simplification).

Plan / record: docs/superpowers/plans/2026-04-20-postbox-to-telegramengine-wave-8.md.

Wave 9 outcome (2026-04-20)

Closes the first of the two "future-wave candidates" left open by wave 8: rewrites both AccountSpecificCacheStorageSettings preferences-view observation sites in StorageUsageScreen.swift using engine APIs, and drops import Postbox from that file.

Site 1 — cacheSettingsExceptionCount signal (former lines 1047–1087):

• postbox.combinedView(keys: [.preferences(keys: Set([PreferencesKeys.accountSpecificCacheStorageSettings]))]) + PreferencesView extraction → context.engine.data.subscribe(TelegramEngine.EngineData.Item.Configuration.ApplicationSpecificPreference(key: PreferencesKeys.accountSpecificCacheStorageSettings)) + preferencesEntry?.get(AccountSpecificCacheStorageSettings.self) ?? .defaultSettings.
• Downstream EngineDataMap + EnginePeer per-category counting logic unchanged (already engine-only).

Site 2 — peerExceptions signal in openKeepMediaCategory (former lines 3131–3196):

• Same preferences observation replacement as Site 1.
• postbox.transaction { transaction.getPeer / transaction.getPeerCachedData as? CachedGroupData / CachedChannelData; FoundPeer(peer:subscribers:) } → context.engine.data.get(EngineDataMap(...TelegramEngine.EngineData.Item.Peer.Peer.init(id:))) + pattern match on EnginePeer.user / .secretChat / .legacyGroup / .channel.
• Signal element type [(peer: FoundPeer, value: Int32)] → [(peer: EnginePeer, value: Int32)]. FoundPeer wrapper and its subscribers field dropped entirely — they were computed by the transaction block but never read by downstream consumers (the only consumer sites read .isEmpty, .count, and .prefix(3).map { EnginePeer($0.peer.peer) }).
• One downstream consumer updated: peerExceptions.prefix(3).map { EnginePeer($0.peer.peer) } → .prefix(3).map { $0.peer } at the MultiplePeerAvatarsContextItem construction (redundant wrap removed since $0.peer is already EnginePeer).

Typealias fixup. With import Postbox removed, var mergedMedia: [MessageId: Int64] at former line 2397 needed renaming to [EngineMessage.Id: Int64]. MessageId is the raw Postbox type name — CLAUDE.md's engine-typealias cheat sheet lists these as migration targets, not pre-existing aliases in TelegramCore. Caught by first-pass build failure (cannot find type 'MessageId' in scope).

Reusable pattern. TelegramEngine.EngineData.Item.Configuration.ApplicationSpecificPreference(key: ValueBoxKey) (at TelegramCore/Sources/TelegramEngine/Data/ConfigurationData.swift:356) is the general-purpose engine replacement for any postbox.combinedView(keys: [.preferences(keys: Set([key]))]) + PreferencesView idiom — takes any ValueBoxKey, returns PreferencesEntry?, decodes via .get(T.self). Crucially, callers need not import Postbox even though ValueBoxKey is a Postbox type: passing PreferencesKeys.<name> through makes ValueBoxKey an inferred-only type that never gets named in the consumer module. Use this pattern when de-Postboxing any future module that observes preferences.

StorageUsageScreen.swift is now Postbox-free. The wave 8 outcome's other candidate (StorageFileListPanelComponent.swift's Icon.media(Media, ...) enum case) remains — trivial future wave to split into .mediaFile(TelegramMediaFile, ...) / .mediaImage(TelegramMediaImage, ...) cases, at which point the StorageUsageScreen consumer module as a whole becomes Postbox-free.

Net: 1 file changed, +30 / -54. Build verified green (27 actions, cached).

Plan / record: docs/superpowers/plans/2026-04-20-postbox-to-telegramengine-wave-9.md.

Wave 10 outcome (2026-04-20)

Closes the second (and last) future-wave candidate from wave 8: eliminates StorageFileListPanelComponent.swift's Icon.media(Media, TelegramMediaImageRepresentation) enum case. StorageUsageScreen (the module as a whole) is now fully Postbox-free — the other in-module file (StorageUsageScreen.swift) landed in wave 9.

Split the enum case. Icon.media(Media, TelegramMediaImageRepresentation) → two concrete cases case mediaFile(TelegramMediaFile, TelegramMediaImageRepresentation) + case mediaImage(TelegramMediaImage, TelegramMediaImageRepresentation). Lossless split: the two construction sites already knew the concrete subtype (imageIconValue = .media(file, representation) from a as? TelegramMediaFile branch, .media(image, representation) from a as? TelegramMediaImage branch), and the consumer binding site immediately downcast via as? to pick which setSignal(...) flavor to call. New split removes the downcast; exhaustiveness-checked switch is both safer and terser.

Equatable rewritten. Old: manual outer-switch + inner if case dispatch, comparing media by media.id only. New: switch-over-tuple (lhs, rhs) with id-based equality per concrete type (lFile.fileId == rFile.fileId, lImage.imageId == rImage.imageId). Same id-based equality semantics as before.

Binding-site rewrite. Old: if case let .media(media, representation) = component.icon { ... if let file = media as? TelegramMediaFile { ... } else if let image = media as? TelegramMediaImage { ... } }. New: a compound case-binding pattern case let .mediaFile(_, representation), let .mediaImage(_, representation): lifts the shared representation variable, then an inner switch dispatches to the right setSignal branch. Works because both cases carry the same TelegramMediaImageRepresentation payload type; Swift allows compound case patterns when the bindings have identical types.

Placeholder PeerId(...) construction fixup. Second-pass build failure after dropping import Postbox surfaced a placeholder PeerId(namespace: PeerId.Namespace._internalFromInt32Value(0), id: PeerId.Id._internalFromInt64Value(0)) in the measureItem layout-measurement instance at former line 1062. Naming PeerId, PeerId.Namespace, and PeerId.Id all require import Postbox (these are raw Postbox types, not TelegramCore typealiases — consistent with wave 9's MessageId → EngineMessage.Id fixup). Replaced with component.context.account.peerId (a real EnginePeer.Id already in scope). Semantically equivalent since the measurement instance's messageId is only used for .peerId extraction inside image-fetch userLocation and for Equatable comparison (the measurement instance isn't compared to anything).

Lesson. Placeholder PeerId(...) / MessageId(peerId:...) constructions in layout-measurement code are a recurring trap for de-Postbox work. Common pattern in this codebase: construct a dummy component instance purely to call .update(...) and read back the returned size. The dummy values are not used meaningfully but naming the types pins import Postbox. When de-Postboxing, grep for PeerId(namespace:/MessageId(peerId: with all-zero args and replace with any convenient real value in scope (context.account.peerId is almost always available).

Net: 1 file changed, +22 / -29 lines (−7 simplification — new switch-over-tuple Equatable is both terser and more idiomatic).

Plan / record: docs/superpowers/plans/2026-04-20-postbox-to-telegramengine-wave-10.md.

Wave 11 outcome (2026-04-20)

Revisits ActionSheetPeerItem — one of the six wave-1 abandonments. The wave-1 blocker was that the public init took postbox: Postbox + network: Network explicitly, forcing the module to import Postbox, and the sole external caller (ShareController, out-of-wave at the time) couldn't be edited. This wave resolves the blocker without any rule-2 violation by routing the pair through AccountStateManager.

Init-surface collapse. ActionSheetPeerItem.init(accountPeerId:postbox:network:contentSettings:peer:…) → .init(accountPeerId:stateManager:contentSettings:peer:…). AccountStateManager is a TelegramCore public class whose public API surface includes postbox: Postbox and network: Network fields; passing the manager as a single handle lets the module hold on to the two values without ever naming Postbox in its own source. The setItem call site becomes self.avatarNode.setPeer(…, postbox: item.stateManager.postbox, network: item.stateManager.network, …) — Swift's type inference resolves Postbox through transitive module visibility (TelegramCore → AvatarNode), no import Postbox needed in the consumer.

Convenience init unchanged in shape. The (context: AccountContext, …) convenience delegates to (accountPeerId:stateManager:contentSettings:…); the two callable forms stay aligned.

Caller (ShareController.swift:1146). Dropped postbox: info.account.stateManager.postbox, network: info.account.stateManager.network → single stateManager: info.account.stateManager. ShareControllerAccountContext (the per-switchable-account protocol) already exposes stateManager: AccountStateManager, so this is a collapse, not a signature divergence. ShareController continues to import Postbox for its own unrelated reasons; no change to its dependency profile.

Reusable pattern. For any wave-1-style module that was abandoned because a public init takes postbox: Postbox, network: Network with avatar-rendering downstream: collapse to stateManager: AccountStateManager (TelegramCore type) and unpack inside the setItem/setPeer body. The pattern applies broadly — most wave-1 abandonments used this param-pair for avatar setup. Candidates to try next: ChatListSearchRecentPeersNode, HorizontalPeerItem, SelectablePeerNode, ItemListPeerItem, ItemListAvatarAndNameInfoItem, ItemListStickerPackItem (verify each by grep first — some may use postbox for non-avatar reasons).

Net: 3 files changed, +8 / -15 lines. Build green (5854 actions, ~6min).

Plan / record: (no plan doc this wave — single-module, low-complexity).

Wave 12 outcome (2026-04-20)

Applies the wave-11 stateManager: AccountStateManager collapse pattern to HorizontalPeerItem — another wave-1-era candidate whose public init leaked postbox: Postbox, network: Network. Additionally ripples the collapse one layer up into ChatListSearchRecentPeersNode's public init so the HorizontalPeerItem call site has stateManager: in scope.

HorizontalPeerItem fully Postbox-free. init(postbox: Postbox, network: Network, …) + matching stored fields → init(stateManager: AccountStateManager, …) + let stateManager. SelectablePeerNode.setup call site routes via item.stateManager.postbox / .network. Module drops import Postbox and //submodules/Postbox:Postbox dep.

ChatListSearchRecentPeersNode public surface migrated, module still imports Postbox. Public init(accountPeerId:postbox:network:…) → init(accountPeerId:stateManager:…). Two private helpers (item(…) on ChatListSearchRecentPeersEntry and preparedRecentPeersTransition(…)) get the same collapse for forwarding. Internal uses of raw postbox (_internal_recentPeers, postbox.peerView, postbox.combinedView, _internal_managedUpdatedRecentPeers) rewritten to stateManager.postbox / stateManager.network — the module stays on import Postbox because of PostboxViewKey / UnreadMessageCountsItem / UnreadMessageCountsView usage inside the peerViews-to-unread-counts pipeline. That pipeline could be rewritten against EngineDataMap + TelegramEngine.EngineData.Item.Peer.Notifications.* in a future wave, but the public surface simplification is valuable standalone.

Two external caller sites migrated:

• ShareController/Sources/ShareControllerRecentPeersGridItem.swift:66-67 — postbox: context.stateManager.postbox, network: context.stateManager.network → stateManager: context.stateManager (ShareControllerAccountContext protocol already exposes stateManager).
• ChatListUI/Sources/ChatListRecentPeersListItem.swift:125-126 — postbox: item.context.account.postbox, network: item.context.account.network → stateManager: item.context.account.stateManager.
• SettingsUI/Sources/DeleteAccountPeersItem.swift:51-52 (call site for HorizontalPeerItem) — postbox: context.account.postbox, network: context.account.network → stateManager: context.account.stateManager.

Lesson reinforcement. The wave-11 collapse pattern is very cheap to ripple through intermediate owners. Whenever a consumer module takes (postbox:Postbox, network:Network) purely to forward them to another call downstream, collapse to stateManager: AccountStateManager — no propagation fan-out required for the raw pair because the stateManager is a single handle. Even when the intermediate owner itself uses raw postbox.peerView internally (like this wave's ChatListSearchRecentPeersNode), the public surface still gets the collapse at zero cost.

Net: 6 files changed, +26 / -36 lines. Build verified green (incremental, 136 actions).

Plan / record: (no plan doc this wave — pattern-application, low-complexity).

Wave 13 outcome (2026-04-20)

Targeted AttachmentTextInputPanelNode at the user's request. On inspection, the module was already Postbox-free at the source level (swept in wave 6) — its two .swift files compile fine without import Postbox. Two leftover items were fixed:

1. Dead //submodules/Postbox:Postbox BUILD dep — wave 6 swept ^import Postbox$ lines from source but never touched BUILD files. AttachmentTextInputPanelNode/BUILD (and, it turns out, 97 other modules' BUILDs — see wave 14) still listed the dep despite no source file needing it. Removed.
2. Two raw peerId?.namespace == Namespaces.Peer.SecretChat checks (lines 436, 2102) migrated to use the existing PeerId.isSecretChat extension at submodules/TelegramCore/Sources/Utils/PeerUtils.swift:615. (First-pass attempt introduced a duplicate isSecretChat extension and failed with "invalid redeclaration" — note for future waves: always grep TelegramCore for an existing helper before adding.)

No new TelegramEngine methods/types introduced. The refactor was smaller than anticipated; the module's migration debt had already been paid down by wave 6's source-level sweep. The BUILD-dep leftover and the namespace-equality sites were the only remaining items. Both are quality-of-life cleanups rather than structural migration.

Observation that drove wave 14. Wave 6's methodology-note in the "Unused-import sweeps" guidance only measured Postbox-freeness by ^import Postbox$ lines in sources. After touching AttachmentTextInputPanelNode/BUILD in this wave, I noticed many other wave-6-swept modules still carry dead BUILD deps, ~= the wave-6 survivor count. That's the whole of wave 14.

Net: 2 files changed, +2 / -3 lines.

Plan / record: (no plan doc this wave — discovery pass).

Wave 14 outcome (2026-04-20)

Build-dep sweep analogous to wave 6's source-import sweep: drop //submodules/Postbox:Postbox (and //submodules/Postbox) from every BUILD whose source files no longer import Postbox.

Methodology.

1. For each submodules/*/BUILD referencing submodules/Postbox, check whether any .swift file in the module's Sources/ tree has ^import Postbox$.
2. If none do, speculatively drop the Postbox dep line from the BUILD via sed -i '' -e '/^[[:space:]]*"\/\/submodules\/Postbox\(:Postbox\)\{0,1\}",[[:space:]]*$/d'.
3. Full Make.py build --continueOnError.
4. Restore any BUILD that now fails to compile (none did).
5. Commit surviving drops.

Result. 98 candidate BUILDs identified. Zero iterations needed — first-pass build came up green (80 incremental actions, no restores). Net: 98 BUILD files, −98 lines (each lost exactly its //submodules/Postbox dep line).

Why zero iterations. Bazel Swift rules require source-level import for symbol resolution. If a module compiled after wave 6's import Postbox sweep, then none of its source files are physically referencing Postbox symbols. The BUILD-level dep was always redundant — it was carried for historical reasons (code likely once imported Postbox but was migrated off) but had no effect on either compilation or the actual dependency graph (Postbox is still transitively pulled in by TelegramCore, which every module depends on). Dropping it is a metadata cleanup with no semantic effect.

Lesson / reusable pattern.

• After every source-level import Postbox sweep (wave-6 shape), run a matching BUILD-dep sweep immediately. Same candidate set, near-zero execution risk, same commit.
• Script for identifying candidates:

  find submodules -name "BUILD" -type f | while read build; do
    dir=$(dirname "$build")
    if grep -q "submodules/Postbox" "$build" 2>/dev/null && [ -d "$dir/Sources" ]; then
      if ! grep -rq "^import Postbox$" "$dir/Sources" 2>/dev/null; then
        echo "$dir"
      fi
    fi
  done
  

• After waves 13+14, 194 modules still list //submodules/Postbox in BUILD — all of them have source files still importing Postbox.

Net (wave 14 alone): 98 files changed, 0 insertions / 98 deletions.

Plan / record: (no plan doc this wave — mechanical sweep).

Wave 15 outcome (2026-04-20)

Applies the wave-11/12 stateManager: AccountStateManager collapse pattern to SelectablePeerNode — another wave-1-era candidate listed in the post-wave-14 shortlist. Module becomes fully Postbox-free (source + BUILD).

SelectablePeerNode fully Postbox-free. Two public setup methods migrated:

• setup(accountPeerId: EnginePeer.Id, postbox: Postbox, network: Network, …) → setup(accountPeerId:, stateManager: AccountStateManager, …).
• setupStoryRepost(accountPeerId: EnginePeer.Id, postbox: Postbox, network: Network, …) → setupStoryRepost(accountPeerId:, stateManager: AccountStateManager, …).

Internal forwards rewired: AvatarNode.setPeer(…, postbox: stateManager.postbox, network: stateManager.network, …) and EmojiStatusComponent(postbox: stateManager.postbox, …). Neither site names Postbox in the consumer — Swift infers through transitive module visibility.

Namespaces.Peer.SecretChat fixup (×3). Replaced peer.peerId.namespace == Namespaces.Peer.SecretChat checks with peer.peerId.isSecretChat at three sites, matching the wave-13 pattern (PeerId.isSecretChat at TelegramCore/Sources/Utils/PeerUtils.swift:611). The third site (updateSelection in the not-selected branch) additionally needed an ?? false fallback — previous expression was self.peer?.peerId.namespace == Namespaces.Peer.SecretChat (optional-equals-non-optional produces Bool), new expression is (self.peer?.peerId.isSecretChat ?? false).

Share-Extension boundary — stateManager: over engine:. SelectablePeerNode is used by ShareControllerPeerGridItem, whose context is ShareControllerAccountContext. That protocol exposes stateManager: AccountStateManager and engineData: TelegramEngine.EngineData, but no engine: TelegramEngine — and the Share Extension's ShareControllerAccountContextExtension concrete impl has no Account, so constructing a full TelegramEngine (init(account: Account)) is physically unreachable there. This is the documented "rare but genuine" fallback to stateManager: from the user-preference memory (feedback_postbox_refactor_handle.md) — prefer engine: except when crossing the Share-Extension boundary.

Three external call sites migrated:

• HorizontalPeerItem/Sources/HorizontalPeerItem.swift:227 (wave 12's stateManager: field now forwards directly): postbox: item.stateManager.postbox, network: item.stateManager.network → stateManager: item.stateManager.
• ShareController/Sources/ShareControllerPeerGridItem.swift:237 (setup): postbox: context.stateManager.postbox, network: context.stateManager.network → stateManager: context.stateManager.
• ShareController/Sources/ShareControllerPeerGridItem.swift:273 (setupStoryRepost): same.

Convenience init unchanged. setup(context: AccountContext, …) now delegates with stateManager: context.account.stateManager; signature unchanged — JoinLinkPreviewPeerContentNode.swift:147 (the one caller using the convenience init) needed no edit.

Net: 4 files changed, +12 / -17 lines. Build verified green (193 actions, 131s — Telegram.ipa target built successfully).

Plan / record: (no plan doc this wave — pattern-application, low-complexity).

Wave 16 outcome (2026-04-20)

Two-commit wave targeting ItemListPeerItem. Planning-time inventory (project_postbox_wave16_plan.md) only grepped for Postbox/Network tokens and missed two Postbox-defined public-surface types: EngineMessageHistoryThread.Info? (on threadInfo) and PeerStoryStats? (on storyStats). The first-pass "drop import Postbox" attempt failed at build time. Rather than abandon, the wave split into 16a (move EngineMessageHistoryThread to TelegramCore — clean, independently valuable) and 16b (partial engine: collapse on ItemListPeerItem, keeping import Postbox because PeerStoryStats remains Postbox-defined).

Wave 16a — move EngineMessageHistoryThread to TelegramCore. Before: Postbox declared an empty public final class EngineMessageHistoryThread namespace with a nested public final class Item; TelegramCore's ForumChannels.swift added the .Info nested type via public extension EngineMessageHistoryThread { final class Info … }. The outer name's Postbox residency forced every consumer of .Info to import Postbox too. After: promote Postbox's internal MutableMessageHistoryThreadIndexView.Item to a top-level public type MessageHistoryThreadIndexItem; delete the empty EngineMessageHistoryThread class from Postbox; move the class shell into ForumChannels.swift, collapsing the existing extension into a proper class definition (public final class EngineMessageHistoryThread { class Info … }).

MessageHistoryThreadIndexView.items type changes from [EngineMessageHistoryThread.Item] to [MessageHistoryThreadIndexItem]; its init simplifies (no more wrap/unwrap conversion — the old init re-built items element-by-element just to swap the outer wrapper name). The second public extension on EngineMessageHistoryThread (.NotificationException, at ForumChannels.swift:1318) works unchanged — same-module extension after the class moves.

Zero consumer-site changes: the two Postbox-consumer iteration sites (ChatListUI/Sources/Node/ChatListNodeLocation.swift:229, ShareController/Sources/ShareControllerNode.swift:2086) iterate with for item in view.items (no type annotation) and access only fields that exist identically on both types (id, info, index, pinnedIndex, tagSummaryInfo, topMessage, embeddedInterfaceState).

Commit 3bb22d503c. Net: 2 files, +67 / −111 (Postbox file nets −174 lines, TelegramCore file +4).

Wave 16b — ItemListPeerItem.Context engine: collapse. Wave-11 pattern applied to ItemListPeerItem.Context.Custom. Before: Context.Custom.init(accountPeerId:, postbox: Postbox, network: Network, animationCache:, animationRenderer:, isPremiumDisabled:, resolveInlineStickers:) + matching stored fields; Context had computed postbox: Postbox and network: Network that switched over the .account / .custom cases. After: Context.Custom.init(accountPeerId:, engine: TelegramEngine, animationCache:, animationRenderer:, isPremiumDisabled:, resolveInlineStickers:); Context has one computed engine: TelegramEngine that returns context.engine for the .account case and custom.engine for the .custom case. Six internal forwards rewire from item.context.postbox / item.context.network to item.context.engine.account.postbox / item.context.engine.account.network (three EmojiStatusComponent(postbox:…) sites and three AvatarNode.setPeer(…, postbox:…, network:…, …) sites).

Handle choice: engine: (not stateManager:). The sole external .custom(Custom(...)) construction site codebase-wide is PeerInfoSettingsItems.swift:121 — main-app-only, doesn't cross the Share-Extension boundary. peerAccountContext in that loop is typed AccountContext (from the accountsAndPeers: [(AccountContext, EnginePeer, Int32)] field), so .engine: TelegramEngine is directly available. Per the standing guidance from feedback_postbox_refactor_handle.md, prefer engine: except when physically forced to stateManager: by a Share-Extension boundary.

All 37 other ItemListPeerItem(…) construction sites use the .account(context: AccountContext) convenience overload (at L485) and need no change. PeerInfoScreenMemberItem.swift:223 forwards its own context: ItemListPeerItem.Context field straight through (pass-through) — no change.

Module does not become Postbox-free: PeerStoryStats? remains on the storyStats public-surface field. PeerStoryStats is defined in Postbox/Sources/ChatListView.swift:281 and is deeply baked into Postbox view APIs (PeerView.storyStats, PeerStoryStatsView.storyStats, ChatListEntry.storyStats, MessageHistoryView.peerStoryStats, Postbox.getPeerStoryStats(peerId:)). Moving it would require a cross-module wrapper rewrite across Postbox, TelegramCore, and every view consumer — out of scope for wave 16.

Commit a5432e44a8. Net: 2 files, +17 / −30.

Lessons.

• Public-surface inventory must go beyond the collapse-target tokens. Waves 11/12/15's stateManager/engine collapses were clean because their target modules had no other Postbox-defined public types. Wave 16's planning inventory only grepped for Postbox/Network and missed EngineMessageHistoryThread + PeerStoryStats — both symbols whose names happen to not include Postbox. For future wave-11-pattern candidates, planning-time grep should include the full alphabet of Postbox-defined public types: ^public\s+(class|struct|enum|protocol|typealias)\s+\w+ over submodules/Postbox/Sources/ to build an exhaustive type-name allowlist, then grep for any of those names in the candidate module's public surface.
• "Engine"-prefixed types can still be Postbox-defined. EngineMessageHistoryThread has an "Engine" prefix but was declared in Postbox all along; the .Info nested type living in TelegramCore was a code-organization half-measure that still forced import Postbox on consumers. Don't trust naming conventions; grep for the defining module.
• Splitting a failing wave into a cleanup + a partial collapse is often the right move. Wave 16 could have been abandoned entirely when the build failed — instead, the EngineMessageHistoryThread move (which had been a latent cleanup opportunity for the entire history of the .Info extension) was promoted to a standalone commit (16a), and the partial engine: collapse shipped as a second commit (16b). Both are independently valuable; the wave's "module becomes Postbox-free" goal didn't land but other goals did.
• The "promote internal Postbox Item to top-level, drop Postbox wrapper class, move wrapper class to TelegramCore" pattern generalizes. Any Postbox-defined class whose only role is to namespace a TelegramCore extension is a candidate for this move. Future audit target: grep -l "public extension <ClassName>" submodules/TelegramCore/Sources/ where <ClassName> is a Postbox-defined outer type with no semantic content of its own.

Plan / record: project_postbox_wave16_plan.md (updated with outcome).

Wave 17 outcome (2026-04-20)

Applies the wave-11/12/15 stateManager: AccountStateManager collapse pattern to ItemListAvatarAndNameInfoItem — another wave-1-era candidate. Module becomes fully Postbox-free (source + BUILD). Clean one-shot execution (no abandonment, no replan).

ItemListAvatarAndNameInfoItem.ItemContext enum case collapsed. Before: case other(accountPeerId: EnginePeer.Id, postbox: Postbox, network: Network) + matching destructure at L761 + AvatarNode.setPeer(…, postbox: postbox, network: network, …) internal forward. After: case other(accountPeerId: EnginePeer.Id, stateManager: AccountStateManager) + case let .other(accountPeerId, stateManager): destructure + AvatarNode.setPeer(…, postbox: stateManager.postbox, network: stateManager.network, …) forward. The .accountContext(AccountContext) sister case is unchanged.

Share-Extension-boundary handle choice: stateManager:. The sole external .other(...) construction site codebase-wide is DeviceContactInfoController.swift:413, inside a ternary that fires only when arguments.context is not a ShareControllerAppAccountContext — i.e., when running inside the Share Extension. ShareControllerAccountContext (protocol at AccountContext/Sources/ShareController.swift:16) exposes stateManager: AccountStateManager but not engine: TelegramEngine, and constructing a full TelegramEngine is physically unreachable in the Share Extension's ShareControllerAccountContextExtension impl (no Account). Per feedback_postbox_refactor_handle.md and the wave-15 precedent, use stateManager: at Share-Extension boundaries.

Pre-flight inventory was correct. Running the public-Postbox-type inventory grep returned only Postbox itself (the one enum-case payload leak) — no EngineMessageHistoryThread-style surprises. Wave 17 validates the post-wave-16 lesson: when planning-time inventory uses the full Postbox public-types allowlist (not just Postbox/Network tokens), wave-11-shape candidates execute cleanly.

Single external caller migrated:

• PeerInfoUI/Sources/DeviceContactInfoController.swift:413 — postbox: arguments.context.stateManager.postbox, network: arguments.context.stateManager.network → stateManager: arguments.context.stateManager. The enclosing PeerInfoUI module still imports Postbox for its own unrelated reasons; that stays.

The 5 other ItemListAvatarAndNameInfoItem(itemContext:…) construction sites codebase-wide all use .accountContext(arguments.context) and need no change (ChannelBannedMemberController.swift:321, DeviceContactInfoController.swift:415, ChannelAdminController.swift:370, CreateChannelController.swift:197, CreateGroupController.swift:324).

Pattern-consistency note (reinforced). accountPeerId: EnginePeer.Id is kept as a separate enum-case payload even though AccountStateManager also exposes accountPeerId. This matches waves 11/12/15 (ActionSheetPeerItem, ChatListSearchRecentPeersNode, SelectablePeerNode all kept accountPeerId explicit alongside stateManager). Future wave-11-pattern executions should default to this shape unless a specific reason exists to collapse further.

Net: 3 files changed, +4 / -5 lines (ItemListAvatarAndNameItem.swift: +2 / -3, DeviceContactInfoController.swift: +1 / -1, BUILD: −1). Build verified green for target modules (ItemListAvatarAndNameInfoItem, PeerInfoUI both compiled and linked successfully); the one unrelated failing target in the full build (ChatMessageInteractiveMediaNode.swift) is user-uncommitted work-in-progress that predates this wave.

Plan / record: (plan doc project_postbox_wave17_plan.md deleted post-commit per the plan's own post-commit housekeeping instructions).

Wave 18 outcome (2026-04-20)

Mixed-shape wave targeting ItemListStickerPackItem. Originally shortlisted (post-wave-17) as "likely wave-11 shape", but plan-writing-time inspection invalidated that hypothesis — the module's public API doesn't take postbox:/network:. Actual shape combined three existing wave patterns plus a narrow typealias addition. Module becomes fully Postbox-free (source + BUILD).

Three narrow typealiases added to TelegramCore. submodules/TelegramCore/Sources/TelegramEngine/Utils/EnginePostboxCoding.swift grew by 3 lines:

• EngineItemCollectionId = ItemCollectionId — needed at public closure-param positions.
• EngineFetchResourceSourceType = FetchResourceSourceType — needed at var updatedFetchSignal type annotation.
• EngineFetchResourceError = FetchResourceError — same.

Per CLAUDE.md rule 1 these narrow-utility typealiases are explicitly allowed (same shape as the existing EngineMemoryBuffer/EnginePostboxDecoder/… batch). Cheat sheet updated.

Wave-4 enum-payload migration on StickerPackThumbnailItem. Public enum case animated(MediaResource, PixelDimensions, Bool, Bool) → animated(EngineMediaResource, PixelDimensions, Bool, Bool). Equatable == simplified: lhsResource.isEqual(to: rhsResource) → lhsResource == rhsResource (uses EngineMediaResource.== which has identical semantics). Two construction sites wrapped via EngineMediaResource(thumbnail.resource) / EngineMediaResource(itemFile.resource). Two destructure-and-forward sites unwrap via resource._asResource() when handing off to chatMessageStickerPackThumbnail(resource: MediaResource) and AnimatedStickerResourceSource(account:, resource: MediaResource, …). One resource.id site (for shortLivedResourceCachePathPrefix) needs the raw MediaResourceId, handled by a local let rawResource = resource._asResource() that serves both the .id read and the AnimatedStickerResourceSource init in the same block.

Wave-3 facade swap. fetchedMediaResource(mediaBox: item.context.account.postbox.mediaBox, userLocation: .other, userContentType: .sticker, reference: resourceReference) → item.context.engine.resources.fetch(reference: resourceReference, userLocation: .other, userContentType: .sticker). Engine facade (TelegramEngine.Resources.fetch) already exists from wave 3; no new TelegramEngine API needed.

External-caller check confirmed zero source edits needed. StickerPackThumbnailItem has no external consumers (UndoUI declares its own nested-private same-named enum). The 6 external ItemListStickerPackItem(setPackIdWithRevealedOptions:) caller sites all pass closures with inferred param types; EngineItemCollectionId being a typealias to ItemCollectionId makes the types interchangeable. The 3 module-field declarations outside the target module that name (ItemCollectionId?, ItemCollectionId?) -> Void explicitly (SettingsUI/Stickers/ArchivedStickerPacksController.swift:27, SettingsUI/Stickers/InstalledStickerPacksController.swift:27, and the init at L32/L42 of those same files) compile unchanged — those modules still import Postbox for their own reasons, and EngineItemCollectionId == ItemCollectionId so no rename is required.

BUILD dep dropped. //submodules/Postbox:Postbox removed from submodules/ItemListStickerPackItem/BUILD.

Pre-existing ChatMessageInteractiveMediaNode.swift WIP still present at build time — no longer failing. The uncommitted change introduces an allowSticker validation around secret-chat sticker playback (~30 lines added in the currentReplaceAnimatedStickerNode block). Per wave-17's note it had failed to compile; on this wave's full build (bazel build Telegram/Telegram, 565 actions, 258s, 0 errors) it compiled and linked without issue. Either the user fixed it between waves 17 and 18, or the bazel dependency graph simply needed a full rebuild. Either way, wave 18's build was clean end-to-end — Telegram.ipa target built successfully, zero errors across the entire project.

Pattern-consistency note. Wave 18 is the third wave (after 3 and 9) where the cheapest path requires adding narrow TelegramCore-side typealiases rather than keeping import Postbox in the consumer. The threshold is: if the consumer needs to NAME a Postbox-defined type (not just use it via inference), and no engine-prefixed alias exists, adding a narrow typealias is preferred over import Postbox. The alternative of refactoring the code to avoid naming the type (e.g., reshaping var foo: Signal<T, E>? to infer from first assignment) is usually unwieldy when the var is conditionally-assigned; typealiases win on readability.

Net: 3 files changed.

• submodules/TelegramCore/Sources/TelegramEngine/Utils/EnginePostboxCoding.swift: +3 / -0.
• submodules/ItemListStickerPackItem/Sources/ItemListStickerPackItem.swift: ~13 lines touched across 9 sites; net +4 / -4.
• submodules/ItemListStickerPackItem/BUILD: 0 / -1.
• CLAUDE.md: +3 cheat-sheet lines + this outcome paragraph.

Plan / record: memory/project_postbox_wave18_plan.md (deleted post-commit per the plan's own housekeeping instructions).

Wave 19 outcome (2026-04-20)

Single-facade expansion. Additive-only — adds TelegramEngine.Resources.shortLivedResourceCachePathPrefix(id: EngineMediaResource.Id) -> String at submodules/TelegramCore/Sources/TelegramEngine/Resources/TelegramEngineResources.swift:456. Body: self.account.postbox.mediaBox.shortLivedResourceCachePathPrefix(MediaResourceId(id.stringRepresentation)).

No consumer migrations this wave. Known consumers (≥25 call sites across ~15 modules: AvatarVideoNode, DrawingUI, SettingsUI/ThemePickerGridItem, PremiumUI/StickersCarouselComponent, ReactionSelectionNode, ReactionContextNode, ChatSendMessageActionUI, ItemListStickerPackItem, ChatThemeScreen, ThemeCarouselItem, PeerInfoBirthdayOverlay, SettingsThemeWallpaperNode, MediaEditorComposerEntity, ChatQrCodeScreen, ChatMessageAnimatedStickerItemNode, ChatMessageItemView, GiftCompositionComponent) migrate in a follow-up wave using the pattern X.context.account.postbox.mediaBox.shortLivedResourceCachePathPrefix(Y.resource.id) → X.context.engine.resources.shortLivedResourceCachePathPrefix(id: EngineMediaResource.Id(Y.resource.id)).

Why not bundle consumer migration in the same wave? Wave-3's original shape did bundle (3 facades + 1 full consumer module in one commit), but the consumer pool for this particular facade is large (~25 sites) and each call site only partially de-Postboxes its module — the caller modules need full inventory before deciding whether to drop import Postbox. Keeping wave 19 narrow (facade-only) lets follow-up waves approach consumer-module migration on a per-module basis without the facade-addition blocking anything.

Net: 1 file changed, +4 / -0.

Plan / record: (no plan doc this wave — single-method addition, target pre-identified in project_postbox_refactor_next_wave.md).

Wave 20 outcome (2026-04-21)

Consumer sweep for the wave-19 shortLivedResourceCachePathPrefix facade. 22 call sites across 16 modules migrated atomically. Pattern (repeated identically at every site): X.context.account.postbox.mediaBox.shortLivedResourceCachePathPrefix(Y.resource.id) → X.context.engine.resources.shortLivedResourceCachePathPrefix(id: EngineMediaResource.Id(Y.resource.id)).

Modules migrated (alphabetical):

• AvatarVideoNode/Sources/AvatarVideoNode.swift (1 site)
• ChatSendMessageActionUI/Sources/ChatSendMessageContextScreen.swift (1 site)
• DrawingUI/Sources/DrawingStickerEntityView.swift (1 site)
• ItemListStickerPackItem/Sources/ItemListStickerPackItem.swift (1 site; simplified from wave-18's let rawResource = resource._asResource(); …shortLivedResourceCachePathPrefix(rawResource.id) + AnimatedStickerResourceSource(…, resource: rawResource, …) to …shortLivedResourceCachePathPrefix(id: resource.id) + AnimatedStickerResourceSource(…, resource: resource._asResource(), …) — drops the intermediate let rawResource)
• PremiumUI/Sources/StickersCarouselComponent.swift (2 sites)
• ReactionSelectionNode/Sources/ReactionContextNode.swift (2 sites)
• ReactionSelectionNode/Sources/ReactionSelectionNode.swift (6 sites — 4 unique expression templates, handled via targeted Edits against the unique argument expression at each call)
• SettingsUI/Sources/ThemePickerGridItem.swift (1 site)
• TelegramUI/Components/Chat/ChatMessageAnimatedStickerItemNode/Sources/ChatMessageAnimatedStickerItemNode.swift (2 sites)
• TelegramUI/Components/Chat/ChatMessageItemView/Sources/ChatMessageItemView.swift (1 site)
• TelegramUI/Components/Chat/ChatQrCodeScreen/Sources/ChatQrCodeScreen.swift (1 site)
• TelegramUI/Components/ChatThemeScreen/Sources/ChatThemeScreen.swift (1 site)
• TelegramUI/Components/Gifts/GiftAnimationComponent/Sources/GiftCompositionComponent.swift (3 sites)
• TelegramUI/Components/PeerInfo/PeerInfoScreen/Sources/PeerInfoBirthdayOverlay.swift (2 sites)
• TelegramUI/Components/Settings/SettingsThemeWallpaperNode/Sources/SettingsThemeWallpaperNode.swift (1 site)
• TelegramUI/Components/Settings/ThemeCarouselItem/Sources/ThemeCarouselItem.swift (1 site)

One site intentionally skipped: TelegramUI/Components/MediaEditor/Sources/MediaEditorComposerEntity.swift:245. That site uses a local postbox: Postbox init-parameter, not context.account.postbox, so the migration would require changing the init's parameter from postbox: to something engine-based and fanning out to its callers. Out of scope — handled by a future module-scoped wave.

No modules became Postbox-free this wave. Each of the 16 migrated modules still has other Postbox usage (raw Postbox types in signatures, fetchedMediaResource(mediaBox:) calls, postbox.transaction, etc.). Consumer-side shortLivedResourceCachePathPrefix closure is just one of several reasons these modules import Postbox. Future wave-shape: module-scoped de-Postbox per-module inventory.

Pattern validation. This is the most mechanical consumer sweep to date — all 22 sites followed identical shape, allowing replace_all=true for sites with duplicate identical call expressions (ReactionSelectionNode hit this at 3 sites for largeListAnimation, 2 for stillAnimation, 1 for listAnimation). First-pass build was clean (35 actions, 0 errors) — no iteration loop. Confirms the wave-19 facade shape is sound.

Build verification. bazel build Telegram/Telegram --keep_going — 2042 action cache hits + 35 new actions, 0 errors, Telegram.ipa up-to-date.

Net: 16 files changed, all edits mechanical (before → after): +22 insertions / -22 deletions at migrated sites, plus 1 deletion in ItemListStickerPackItem (wave-18 let rawResource line dropped). Approximate total: +22 / -23.

Plan / record: (no plan doc this wave — mechanical sweep).

Wave 21 outcome (2026-04-21)

Combined wave-19+wave-20 shape: facade addition + consumer sweep in a single atomic commit. Adds TelegramEngine.Resources.completedResourcePath(id: EngineMediaResource.Id, pathExtension: String? = nil) -> String? facade; sweeps 29 consumer sites across 14 files.

Facade added at TelegramCore/Sources/TelegramEngine/Resources/TelegramEngineResources.swift:460. Body: self.account.postbox.mediaBox.completedResourcePath(id: MediaResourceId(id.stringRepresentation), pathExtension: pathExtension). Wraps the Postbox MediaBox.completedResourcePath(id: MediaResourceId, pathExtension: String?) overload; consumers that previously called the resource-taking overload (MediaBox.completedResourcePath(_ resource: MediaResource, …)) migrate through the id path (.resource.id is already MediaResourceId).

28 Shape-A consumer sites + 1 Shape-B (already-id-overload) migrated:

• SettingsUI/Sources/Themes/EditThemeController.swift (1 site)
• BrowserUI/Sources/BrowserPdfContent.swift (1 site)
• BrowserUI/Sources/BrowserDocumentContent.swift (1 site)
• GalleryUI/Sources/SecretMediaPreviewController.swift (1 site)
• TelegramUI/Components/MediaEditor/Sources/MediaEditor.swift (1 site, pathExtension: "mp4")
• TelegramUI/Components/Settings/WallpaperGridScreen/Sources/WallpaperUtils.swift (7 sites across 3 functions; 4 unique expression templates, handled via replace_all=true where identical)
• TelegramUI/Components/Settings/ThemeAccentColorScreen/Sources/ThemeAccentColorController.swift (1 site)
• TelegramUI/Components/Settings/WallpaperGalleryScreen/Sources/WallpaperGalleryController.swift (5 sites; 4 used resource expr identically via replace_all=true, 1 used file.file.resource)
• TelegramUI/Components/MediaEditorScreen/Sources/MediaEditorScreen.swift (1 site, pathExtension: nil)
• TelegramUI/Components/Chat/ChatMessageWebpageBubbleContentNode/Sources/ChatMessageWebpageBubbleContentNode.swift (1 site)
• TelegramUI/Components/Chat/ChatMessageMediaBubbleContentNode/Sources/ChatMessageMediaBubbleContentNode.swift (7 sites, all identical telegramFile.resource — handled via replace_all=true)
• TelegramUI/Components/Chat/ChatMessageAttachedContentNode/Sources/ChatMessageAttachedContentNode.swift (1 site)
• TelegramUI/Sources/OpenChatMessage.swift (1 site)
• TelegramUI/Sources/Chat/ChatControllerMediaRecording.swift (1 site)
• TelegramUI/Components/Stories/StoryContainerScreen/Sources/StoryItemImageView.swift (1 site, Shape B — was already using the id: overload; migrated identically to EngineMediaResource.Id(...))

8 sites intentionally skipped (Shape C/D). Listed in the plan — 5 Shape-C sites that access a raw account: Account parameter (no .engine on Account) and 3 Shape-D sites that carry a local postbox: Postbox stored field. Both shapes need module-scoped init-signature rework rather than per-site sweep; defer to future waves.

No modules became Postbox-free. Each consumer has other Postbox usage (signatures, transactions, other mediaBox calls). Matches waves 19/20's expectation.

Build validation. bazel build Telegram/Telegram --keep_going — clean first-pass build (569 processes, 1556 action cache hits + 30 local + 532 worker, 240s, 0 errors, Telegram.ipa up-to-date).

Pattern validation. Wave-shape G (facade addition + consumer sweep in a single commit) works well when the consumer pool is bounded and mechanical. 29 sites in 14 files is comfortably within the threshold. Kept waves 19 and 20 separate because 25+ sites across that many modules was at the edge of reviewability; wave 21's similar fan-out fits because the plan pre-classified every site by shape. When the plan does the classification work upfront, combined waves are cheaper to review and ship.

Net: 14 files changed. TelegramEngineResources.swift: +4 / -0. Consumer files: +29 / -29 (mechanical rewrite at each site). CLAUDE.md: +outcome paragraph.

Plan / record: memory/project_postbox_wave21_plan.md (deleted post-commit per the plan's own housekeeping instructions).

Wave 22 outcome (2026-04-21)

Follows wave 21's pattern: facade addition + consumer sweep in a single atomic commit. Adds TelegramEngine.Resources.storeResourceData(id: EngineMediaResource.Id, data: Data, synchronous: Bool = false) facade; sweeps 46 consumer sites across 17 files.

Facade added at TelegramCore/Sources/TelegramEngine/Resources/TelegramEngineResources.swift:464. Body: self.account.postbox.mediaBox.storeResourceData(MediaResourceId(id.stringRepresentation), data: data, synchronous: synchronous). Wraps Postbox's MediaBox.storeResourceData(_ id: MediaResourceId, data: Data, synchronous: Bool) full-file overload. The range-store overload (MediaBox.storeResourceData(_:range:data:)) is used at a single site inside HLSVideoJSNativeContentNode.swift:302 via a local postbox: Postbox field (Shape D), which is out of scope for this wave; the range overload gets no facade wrapper this round.

46 Shape-A consumer sites migrated:

• ImportStickerPackUI/Sources/ImportStickerPackControllerNode.swift (2)
• DebugSettingsUI/Sources/DebugController.swift (8 — 6 identical gzippedData batched via replace_all=true; logData, allStatsData handled individually)
• BrowserUI/Sources/BrowserWebContent.swift (1)
• TelegramUI/Sources/CreateChannelController.swift (4)
• TelegramUI/Sources/CreateGroupController.swift (4)
• TelegramUI/Sources/Chat/ChatControllerPaste.swift (1)
• TelegramUI/Sources/Chat/ChatControllerOpenDocumentScanner.swift (3)
• TelegramUI/Sources/Chat/ChatControllerMediaRecording.swift (2)
• TelegramUI/Components/LegacyInstantVideoController/Sources/LegacyInstantVideoController.swift (2)
• TelegramUI/Components/PeerInfo/PeerInfoScreen/Sources/PeerInfoScreenAvatarSetup.swift (2)
• TelegramUI/Components/Settings/WallpaperGridScreen/Sources/WallpaperUtils.swift (6 — 3 thumbnailResource, 3 resource; both handled via replace_all=true)
• SettingsUI/Sources/Themes/ThemePreviewController.swift (1)
• SettingsUI/Sources/Themes/EditThemeController.swift (1)
• TelegramUI/Components/Stories/StoryContainerScreen/Sources/StoryItemSetContainerViewSendMessage.swift (3)
• TelegramUI/Components/VideoMessageCameraScreen/Sources/VideoMessageCameraScreen.swift (2)
• TelegramUI/Components/MediaEditorScreen/Sources/CreateLinkOptions.swift (1)
• TelegramUI/Components/MediaEditorScreen/Sources/MediaEditorScreen.swift (3)

Out of scope — not migrated this wave:

• accountManager.mediaBox.storeResourceData(...) sites (Account-manager-scoped, not account-scoped) — 13+ sites across WallpaperGalleryItem, WallpaperGalleryController, ThemeAccentColorController, WallpaperUtils, WebBrowserSettingsController, ThemeUpdateManager, OpenResolvedUrl, and others. These are a different migration path entirely (not a TelegramEngine.Resources.* target) and stay raw.
• account.postbox.mediaBox.storeResourceData(...) (raw Account, no AccountContext) — ~9 sites in LegacyMediaPickerUI, TelegramCallsUI, InAppPurchaseManager, AuthorizationUI, PeerInfoScreenAvatarSetup closures, WallpaperResources. Shape C from wave-21 taxonomy. Needs per-module rework.
• self.postbox.mediaBox.storeResourceData(...) / postbox.mediaBox.storeResourceData(...) inside TelegramCore internals (TransformOutgoingMessageMedia.swift, AccountStateManager.swift, AvailableReactions.swift, SaveSecureIdValue.swift, PeerPhotoUpdater.swift, NotificationSoundList.swift, Stories.swift, Authorization.swift, WebpagePreview.swift). These are Postbox-internal layer by design — keep as-is.
• HLSVideoJSNativeContentNode.swift:302 — uses the range-store overload via local postbox: Postbox field. Out of scope.

No modules became Postbox-free. Matches waves 19/20/21 expectation — each consumer has other Postbox usage.

Build validation. bazel build Telegram/Telegram --keep_going — clean first-pass build (571 processes, 1554 action cache hits + 30 local + 532 worker, 229s, 0 errors, Telegram.ipa up-to-date).

Pattern validation. Wave-shape G (facade + consumer sweep in one commit) scales well up to 46 sites in 17 files when the pattern is mechanical. Heavy replace_all=true usage where call-text is identical across sites (DebugController's 6 gzippedData sites, WallpaperUtils' 6 sites split into 2 batches by first-arg variable, ChatControllerOpenDocumentScanner's identical (resource.id, data: data, synchronous: true) pattern) keeps diff noise to the minimum. 46 sites, mostly done via replace_all + a few individual edits.

Net: 17 consumer files + 1 TelegramCore file + CLAUDE.md. TelegramEngineResources.swift: +4 / -0. Consumer files: +46 / -46 (mechanical rewrite).

Plan / record: (no plan doc this wave — mechanical sweep following wave-21 recipe).

Wave 23 outcome (2026-04-21)

Smallest wave so far: cancelInteractiveResourceFetch facade addition + consumer sweep. Same shape as waves 21/22.

Facade added at TelegramCore/Sources/TelegramEngine/Resources/TelegramEngineResources.swift:468. Body: self.account.postbox.mediaBox.cancelInteractiveResourceFetch(resourceId: MediaResourceId(id.stringRepresentation)). Wraps Postbox's MediaBox.cancelInteractiveResourceFetch(resourceId: MediaResourceId) overload (the _ resource: MediaResource overload delegates to the id version anyway).

5 of 7 Shape-A consumer sites migrated:

• PeerAvatarGalleryUI/Sources/PeerAvatarImageGalleryItem.swift (1)
• GalleryUI/Sources/Items/ChatAnimationGalleryItem.swift (1)
• GalleryUI/Sources/Items/ChatImageGalleryItem.swift (1)
• GalleryUI/Sources/Items/ChatDocumentGalleryItem.swift (1)
• GalleryUI/Sources/Items/ChatExternalFileGalleryItem.swift (1)

2 sites intentionally skipped: ChatMessageInteractiveMediaNode.swift:1474, 1709 — this file has pre-existing uncommitted WIP (the allowSticker validation around secret-chat sticker playback, carried forward since before wave 17). Editing the 2 sites would mix my wave-23 changes with the user's WIP in a single git diff, which git add can't cleanly separate. Deferred until the WIP lands or a narrow follow-up wave intentionally includes both. Note: a future wave that aims to drop those 2 sites should first either (a) wait for the WIP to be committed or (b) use git stash --keep-index + targeted edits + selective staging to split the diff cleanly.

Pattern note on WIP interference. This is the first wave to hit this failure mode — previous waves' mechanical sweeps happened not to touch ChatMessageInteractiveMediaNode.swift. Future sweeps should grep their candidate set against git status's modified-files list before starting, and either (a) defer sites in WIP files, (b) wait for the WIP to commit, or (c) stage selectively via git add --patch-equivalent paths.

Build validation. bazel build Telegram/Telegram --keep_going — clean first-pass build (558 processes, 1567 action cache hits + 19 local + 532 worker, 236s, 0 errors, Telegram.ipa up-to-date).

Net: 5 consumer files + 1 TelegramCore file + CLAUDE.md. TelegramEngineResources.swift: +4 / -0. Consumer files: +5 / -5.

Plan / record: (no plan doc this wave — mechanical sweep).

Wave 24 outcome (2026-04-21)

moveResourceData facade additions + consumer sweep. Same shape as waves 21-23.

Two facades added at TelegramCore/Sources/TelegramEngine/Resources/TelegramEngineResources.swift:

• moveResourceData(id: EngineMediaResource.Id, toTempPath: String) wraps the (MediaResourceId, toTempPath:) overload.
• moveResourceData(from: EngineMediaResource.Id, to: EngineMediaResource.Id, synchronous: Bool = false) wraps the (from: MediaResourceId, to: MediaResourceId, synchronous:) overload.

Postbox's third overload (MediaResourceId, fromTempPath:) has no consumer-side usage; no facade added this wave (YAGNI).

6 Shape-A consumer sites migrated (5 files):

• TelegramUI/Sources/Chat/ChatControllerMediaRecording.swift (1, toTempPath:)
• TelegramUI/Sources/OverlayAudioPlayerController.swift (1, from:to:synchronous:)
• TelegramUI/Components/ComposePollScreen/Sources/ComposePollScreen.swift (2)
• TelegramUI/Components/Chat/ChatMessagePollBubbleContentNode/Sources/ChatMessagePollBubbleContentNode.swift (2)

Build validation. bazel build Telegram/Telegram --keep_going — clean first-pass build (563 processes, 272s, 0 errors).

Net: 5 consumer files + 1 TelegramCore file + CLAUDE.md. TelegramEngineResources.swift: +8 / -0. Consumer files: +6 / -6.

Plan / record: (no plan doc this wave — mechanical sweep).

Wave 25 outcome (2026-04-21)

copyResourceData facade additions + consumer sweep. Same shape as waves 21-24.

Two facades added: copyResourceData(id: EngineMediaResource.Id, fromTempPath: String) and copyResourceData(from: EngineMediaResource.Id, to: EngineMediaResource.Id, synchronous: Bool = false).

4 Shape-A consumer sites migrated (3 files):

• PeerAvatarGalleryUI/Sources/AvatarGalleryController.swift (2, from:to:synchronous:)
• ImportStickerPackUI/Sources/ImportStickerPackControllerNode.swift (1, from:to: — simplified from localResource._asResource().id to localResource.id since operands are EngineMediaResource)
• TelegramUI/Sources/Chat/ChatControllerPaste.swift (1, id:fromTempPath:)

Minor simplification lesson. When a consumer already has an EngineMediaResource-typed local (e.g., from a wave-18-migrated callee), prefer localResource.id over EngineMediaResource.Id(localResource._asResource().id) — the two are semantically equivalent since EngineMediaResource.id is defined as Id(self.resource.id). This halves the verbosity at the call site and removes a redundant unwrap-and-rewrap.

Build validation. Clean first-pass build (563 processes, 242s, 0 errors).

Net: 3 consumer files + 1 TelegramCore file + CLAUDE.md. TelegramEngineResources.swift: +8 / -0.

Plan / record: (no plan doc this wave — mechanical sweep).

Wave 27 outcome (2026-04-22)

preferencesView consumer sweep (wave-9 pattern continuation). No new TelegramCore facades — leverages existing TelegramEngine.EngineData.Item.Configuration.ApplicationSpecificPreference(key:).

Shape. Replace context.account.postbox.preferencesView(keys: [<key>]) — returning Signal<PreferencesView, NoError> — with context.engine.data.subscribe(TelegramEngine.EngineData.Item.Configuration.ApplicationSpecificPreference(key: <key>)) — returning Signal<PreferencesEntry?, NoError>. Downstream, rename <name>.values[<key>]?.get(<Type>.self) → <name>?.get(<Type>.self) at each closure parameter.

30 consumer files, ~40 call sites migrated across ChatListUI, ContactListUI, DebugSettingsUI, GalleryUI, PeersNearbyUI, SettingsUI, TelegramCallsUI, TelegramUI, TelegramUI/Components, WebSearchUI. Full list in git show --stat <wave-27-commit>.

Multi-key sites (PresentationCallManager). 3 sites used preferencesView(keys: [voipConfiguration, appConfiguration]). Migrated via the two-arg engine.data.subscribe(itemA, itemB) |> take(1) overload, which returns Signal<(PreferencesEntry?, PreferencesEntry?), NoError>. Closures that accessed preferences.values[X]?.get(...) rewritten to preferences.0?.get(...) and preferences.1?.get(...).

Direct-postbox-param helper migrated. AccountContext.swift's getAppConfiguration(postbox: Postbox) helper (one internal caller only) was rewritten to getAppConfiguration(engine: TelegramEngine) in the same commit, switching its single call site from getAppConfiguration(postbox: account.postbox) to getAppConfiguration(engine: self.engine).

Annotation update in NotificationExceptionControllerNode.swift. An explicit signal type Signal<(…, PreferencesView, …), NoError> in a mapToSignal return was updated to Signal<(…, PreferencesEntry?, …), NoError>. The file still imports Postbox because PreferencesEntry is (for now) a Postbox-defined type surfaced through TelegramCore's EnginePreferencesEntry typealias — a future wave-6-style import Postbox sweep would clean up such imports where they're now the only Postbox reference.

Deliberately skipped in this wave.

• TelegramPermissionsUI/Sources/PermissionSplitTest.swift:100 — permissionUISplitTest(postbox: Postbox) is a public API whose product value PermissionUISplitTest itself stores postbox: Postbox to satisfy the SplitTest protocol. Proper migration requires a protocol-level refactor (or wholesale rewrite of the SplitTest abstraction) beyond this wave's scope.
• 5 TelegramCore-internal postbox.preferencesView(...) sites (ChatListFiltering × 3, ContentSettings × 1, ManagedGlobalNotificationSettings × 1) — the refactor only migrates consumer modules, not TelegramCore internals.

Build validation. Clean first-pass build (748 processes, 227s, 0 errors). No new facades to test, shape was validated across 30 files on the first attempt.

Lesson — multi-key preferencesView migration. engine.data.subscribe(itemA, itemB) exists and returns a Swift tuple. When a Postbox preferencesView(keys: [K1, K2]) call is inside a combineLatest(...) whose downstream closure accesses .values[K1] and .values[K2], prefer the two-arg subscribe form (vs. two separate subscribes combined externally) — it preserves combineLatest arity exactly. Rewrite .values[K1]?.get(T.self) → .0?.get(T.self), .values[K2]?.get(T.self) → .1?.get(T.self). The closure parameter name stays (e.g., preferences) because the tuple destructure preserves the variable-name semantics at the call site.

Net: 30 consumer files. No TelegramCore changes. CLAUDE.md facade-inventory table unchanged (no new facades).

Plan / record: memory/project_postbox_wave27_plan.md (deleted post-wave).

Wave 26 outcome (2026-04-21)

resourceRangesStatus + removeCachedResources facade additions + consumer sweep. Combines two independent small sweeps into one commit.

Two facades added:

• resourceRangesStatus(resource: EngineMediaResource) -> Signal<RangeSet<Int64>, NoError> wraps the single (MediaResource) -> Signal<RangeSet<Int64>, NoError> overload. Takes EngineMediaResource (not id:) because Postbox's overload only accepts a resource, not an id — consumers pass .resource already. Facade unwraps via _asResource().
• removeCachedResources(ids: [EngineMediaResource.Id], force: Bool = false, notify: Bool = false) -> Signal<Float, NoError> wraps the ([MediaResourceId], force:, notify:) -> Signal<Float, NoError> overload. Maps ids internally.

import RangeSet added to TelegramEngineResources.swift. The RangeSet<Int64> return type caused a name collision with Swift stdlib's RangeSet (iOS 18+ only) until the local RangeSet module is imported. TelegramCore/BUILD already declared the dep at line 23 (//submodules/Utils/RangeSet:RangeSet), so no BUILD change needed.

4 Shape-A consumer sites migrated (3 files):

• PhotoResources/Sources/PhotoResources.swift (1)
• TelegramUI/Components/Stories/StoryContainerScreen/Sources/StoryChatContent.swift (1)
• ChatListUI/Sources/ChatListSearchContainerNode.swift (2)

For ChatListSearchContainerNode.swift:1398, the caller uses a Set<MediaResourceId> local — wave leaves the local as-is and maps at the call site via resourceIds.map { EngineMediaResource.Id($0) }. Migrating the local to Set<EngineMediaResource.Id> is out of scope (module keeps import Postbox for unrelated reasons).

Build validation. Clean build (563 processes, 265s, 0 errors) on the second attempt after adding import RangeSet.

Lesson — Swift-stdlib-vs-third-party-module name collisions. When a facade signature references a type name that exists both in Swift stdlib (potentially availability-restricted) and in a third-party module, the compiler picks the stdlib one by default. Fix: import the third-party module explicitly. In this codebase, RangeSet is provided by submodules/Utils/RangeSet:RangeSet, and TelegramCore already depends on it. Use import RangeSet at the file top.

Net: 3 consumer files + 1 TelegramCore file + CLAUDE.md. TelegramEngineResources.swift: +9 / -0 (including import RangeSet).

Plan / record: (no plan doc this wave — mechanical sweep).

Wave 31 outcome (2026-04-23)

Second build-verified ^import Postbox$ sweep on consumer modules since wave 6 (2026-04-19). Same methodology: speculative-drop + --continueOnError build loop with pattern-based preemptive restores.

Candidate set narrowing. Initial candidate grep grep -rl "^import Postbox$" submodules --include="*.swift" returned 1184 files. 606 of those live in submodules/TelegramCore/Sources/ — TelegramCore legitimately import Postbox; the TelegramCore files were accidentally included and had to be reverted via git checkout -- submodules/TelegramCore/Sources/ before re-seeding the drop. Final consumer candidate set: 578 files. Lesson for future sweep invocations: the candidate-set grep must filter out submodules/TelegramCore/ as well as submodules/Postbox/ / submodules/TelegramApi/. Wave 6's methodology note at step 1 (line 37) already calls this out, but the TelegramCore carve-out is easy to miss because TelegramCore doesn't @_exported import Postbox, so from a pure re-exports perspective it's indistinguishable from a consumer.

9 build iterations to convergence (plus 1 aborted first iteration for the TelegramCore scope error). Per-iteration failure counts: 18 → 2 → 9 → 12 → 1 → 1 → 3 → 1 → 4 → 0. Surfacing pattern was typical of a speculative-drop sweep: errors bubble one dependency-graph layer at a time.

Per-iteration symbol expansion. The wave-6 preemptive-restore symbol list (CLAUDE.md's "Unused-import sweeps" guidance) needed extensions for this sweep:

• Iter 3 surfaced CodableEntry, CachedMediaResourceRepresentation, CachedMediaRepresentationKeepDuration.
• Iter 4 surfaced PostboxViewKey, OrderedItemListView, UnreadMessageCountsItem, ChatListEntrySummaryComponents, PeerStoryStats, ItemCollectionId (note: typealias EngineItemCollectionId exists but raw name still requires import Postbox), and broadened \bMedia\b, \bMessage\b, \bPeer\b.
• Iter 5 surfaced FetchResourceSourceType (same typealias caveat).
• Iter 6 surfaced StoryId.
• Iter 7 surfaced ChatListIndex.
• Iter 8 surfaced PreferencesEntry (typealias caveat), PeerView, RenderedPeer.
• Iter 9 surfaced declareEncodable.
• Iter 10 surfaced ItemCollectionItemIndex, ValueBoxEncryptionParameters, fileSize, plus a restore-script bug (see below).

Restore-script bug: #if canImport(...) blocks. The naive restore inserter picks the last ^import line and appends import Postbox after it. If the last import sits inside an #if canImport(AppCenter) ... #endif preprocessor block, the restored import Postbox lands inside that block and is only active under that configuration. AppDelegate.swift in submodules/TelegramUI/Sources/ hit this (original had import Postbox at line 7; drop + restore put it inside the #if canImport(AppCenter) block at line 51); the build failed in iter10 on cannot find type 'Postbox' in scope errors even though a literal grep ^import Postbox$ matched. Fixed by manually moving the import out of the #if block. Lesson for future restore-script work: insert the restored import Postbox BEFORE the first #if or #endif line, not after the last import line, to avoid preprocessor-scope traps.

Results: 9 source-level surviving drops + 2 duplicate-import dedups. Final diff: 11 files changed, +2 / -13.

Surviving drops:

• submodules/AuthorizationUI/Sources/AuthorizationSequencePhoneEntryController.swift
• submodules/AuthorizationUI/Sources/AuthorizationSequenceSplashController.swift
• submodules/DebugSettingsUI/Sources/DebugAccountsController.swift
• submodules/LegacyDataImport/Sources/LegacyPreferencesImport.swift
• submodules/MediaPlayer/Sources/ChunkMediaPlayerDirectFetchSourceImpl.swift
• submodules/TelegramUI/Components/Stories/StoryContainerScreen/Sources/StoryItemImageView.swift
• submodules/TelegramUI/Sources/ChatLinkPreview.swift
• submodules/TelegramUI/Sources/ChatSearchResultsController.swift
• submodules/TelegramUI/Sources/MediaManager.swift

Duplicate-import dedups (files had two ^import Postbox$ lines; kept exactly one — unrelated-but-latent cleanup surfaced incidentally by the sweep):

• submodules/TelegramUI/Components/ChatControllerInteraction/Sources/ChatControllerInteraction.swift (2 imports → 1)
• submodules/TelegramUI/Sources/ChatHistoryListNode.swift (2 imports → 1)

Spurious-diff cleanup step (new procedure, adopted this wave). After convergence, git diff --numstat showed 564 modified files but only 9 were genuine drops. The other 553 were "1 addition + 1 deletion" — files where the original import Postbox at line X was deleted by the drop and re-inserted at line Y by the restore (different position because restore inserts after "last import line" regardless of original placement). These aren't semantic changes but do produce noisy diffs. Identified via git diff --numstat | awk '$1 == 1 && $2 == 1 {print $3}' and reverted via xargs -I{} git checkout -- {}. Lesson: the wave-6 methodology should add a post-convergence "revert 1-add-1-del spurious diffs" step before committing. Alternative: improve the restore script to insert at the exact original line. Either way, the final diff should be limited to real semantic changes.

No modules became fully Postbox-free this wave. Each of the five containing modules still has other files importing Postbox (TelegramUI: 350 remaining, LegacyDataImport: 4, MediaPlayer: 9, AuthorizationUI: 2, DebugSettingsUI: 1). By this point most trivially-droppable imports have been drained; the remaining Postbox-importing files mostly carry real usage. Re-run cadence lesson: yield per re-run is declining. Wave 6 yielded 183 drops + 189 modules freed; wave 31 yielded 9 drops + 0 modules freed. Consider spacing future sweeps to every 4–6 facade waves rather than 2–3.

Wave 14 BUILD-dep sweep companion: 0 drops. Ran the wave-14-style find submodules -name BUILD | filter-by-no-source-import check: 0 BUILD candidates. The 191 BUILDs still listing //submodules/Postbox all have at least one Sources/*.swift that actually imports Postbox. One outlier (submodules/SpotlightSupport/BUILD) has zero source files but a non-trivial deps = [...] list including //submodules/Postbox; deliberately left alone (stale-BUILD-on-empty-module is a different class of cleanup and carries unknown side effects).

Net: 11 files changed (9 + 2), +2 / -13 lines. Clean first-attempt verification build without --continueOnError (880 actions, 1354 action cache hits, 262s).

Plan / record: (no plan doc this wave — mechanical sweep).

Wave 32 outcome (2026-04-24)

resourceStatus residue sweep. One new facade overload (status(id:resourceSize:)) + 4 migrated sites across 2 consumer files. Commit 289fc908bc.

Facade added in TelegramEngineResources.swift:

• status(id: EngineMediaResource.Id, resourceSize: Int64) -> Signal<EngineMediaResource.FetchStatus, NoError> wraps Postbox's resourceStatus(MediaResourceId, resourceSize:) overload. Body mirrors the existing status(resource:) facade, converting id via MediaResourceId(id.stringRepresentation) and mapping the result via EngineMediaResource.FetchStatus.init.

4 migrated sites (2 files):

• ChatListSearchContainerNode.swift:1059 — new status(id:resourceSize:) overload. Caller supplies EngineMediaResource.Id(downloadResource.id) directly (String initializer; downloadResource.id: String) — no raw MediaResourceId(...) wrap needed. Mirrors the pre-existing EngineMediaResource.Id(downloadResource.id) usage at line 1107.
• ChatMessageInteractiveMediaNode.swift:1769 — existing status(resource:) facade (wave 3).
• ChatMessageInteractiveMediaNode.swift:1799 — same.
• ChatMessageInteractiveMediaNode.swift:1809 — existing resourceRangesStatus(resource:) facade (wave 26).

Local preserved deliberately. let postbox = context.account.postbox at ChatMessageInteractiveMediaNode.swift:1767 stays because line 1793 feeds postbox to HLSVideoContent.minimizedHLSQualityPreloadData(postbox: Postbox, ...) — that is a third-party-function boundary needing raw Postbox. Only the resourceStatus/resourceRangesStatus call sites within that scope migrate.

Case-pattern sharing. MediaResourceStatus (raw Postbox) and EngineMediaResource.FetchStatus (engine wrapper) have identical case names (.Fetching, .Paused, .Local, .Remote). The inner switch status at 1770-1779 keeps its MediaResourceStatus return type annotation — input case matching works for the engine type, constructed MediaResourceStatus return values still compile (MediaResourceStatus is in scope via import Postbox on line 4). This is the wave-29/30 lesson in action: no enum-case edits required.

Inventory scope narrowing from memory's prediction. The memory's wave 32+ candidates section predicted ~12 Shape-B/C sites in the residue sweep. Execution-time re-grep reclassified most of them:

• Coupled to accountManager.mediaBox.resourceStatus siblings (6 sites in 3 files): ThemePreviewControllerNode:271+277, WallpaperGalleryItem:799+805+834+840, SettingsThemeWallpaperNode:284+285. Each pair has an accountManager-sourced fallback whose return type is raw Signal<MediaResourceStatus, NoError>. Migrating only the account.postbox branch breaks the shared sibling type at the mapToSignal/combineLatest merge point. Deferred until accountManager-side has an engine facade.
• Shape-C init-param refactor (3 sites in 3 files): LegacyWebSearchGallery:248 (free function legacyWebSearchItem(account: Account, ...)), NativeVideoContent:455 (init takes postbox: Postbox), VerticalListContextResultsChatInputPanelItem:229 (item stores account: Account). Each needs an init-param change + caller threading — per-module mini-refactor, not wave-shape-G territory.
• approximateSynchronousValue overload: only call site (SettingsThemeWallpaperNode:284) is in the accountManager-coupled bucket above. Adding the facade now would land dead code.

Effective wave scope: 4 sites (the uncoupled subset). Still worth committing as its own wave — closes the resourceStatus arc for every site where migration is currently unblocked.

Build validation. Clean build (558 processes, 236s, 0 errors). No --continueOnError needed — first attempt green.

Lesson — siblings-define-scope in resource-status migrations. When an assignment uses A.resourceStatus(...) in one branch and B.resourceStatus(...) in another (via if/mapToSignal/combineLatest), the branches' return types must match. If A has an engine facade but B does not (e.g., accountManager.mediaBox has no engine wrapper yet), neither branch is migratable in isolation — the whole group must wait. Pre-flight sibling-check for each resourceStatus hit: is the enclosing statusSignal = ... expression a single source or a multi-source merge?

Lesson — Shape-B/C classification requires read, not grep. The memory's wave-32 candidate table classified sites by single-line grep ("account.postbox.mediaBox.resourceStatus"). That pattern matches both the fully-migratable context.account.postbox.mediaBox.X form (Shape-A via AccountContext) AND the (local) account.postbox.mediaBox.X Shape-C form (requires init-param refactor). Distinguishing requires reading 5-10 lines of context to find the account binding: field? local? init param? closure capture? Add this as a mandatory step in the per-site inventory for future residue waves.

Plan / record: (no plan doc this wave — small residue sweep).

---

Wave 33 outcome (2026-04-24)

loadedPeerWithId consumer sweep. 60 sites migrated across 37 consumer files. No new facades, no typealiases. Commit 16d017853a.

Migration pattern (per user's explicit direction):

context.engine.data.get(TelegramEngine.EngineData.Item.Peer.Peer(id: peerId))
|> mapToSignal { peer -> Signal<EnginePeer, NoError> in
    if let peer {
        return .single(peer)
    } else {
        return .never()
    }
}

This replaces context.account.postbox.loadedPeerWithId(peerId) while preserving signature shape. The mapToSignal wrapper is critical: Postbox's loadedPeerWithId returns .never() (signal never emits) when the peer is missing — it does NOT wait for loading. The engine-data equivalent get(Peer.Peer(id:)) returns Signal<EnginePeer?, NoError> (optional snapshot). Unwrapping with .never()-on-nil preserves original semantics exactly, while keeping the outer shape Signal<EnginePeer, NoError> non-optional so callers' closures don't have to cascade new optional handling.

Category distribution (per pre-flight Explore catalog, 60 sites):

Category	Count	Body change
Cat-A (trivial)	22	Only EnginePeer-compatible members; type swap only.
Cat-B (concrete-type cast)	25	peer as? TelegramUser/Group/Channel/SecretChat → if case let .user(user) (etc.).
Cat-C (feeds Peer-typed API)	13	peer._asPeer() at call point (makePeerInfoController, makeChatRecentActionsController, makeChatQrCodeScreen, FoundPeer.init, SendAsPeer.init).

(Cat-B + Cat-C bumped slightly from Explore's catalog after in-edit reclassifications.)

Engine-access variations:

• Most consumer modules use context.engine.data.get(...) on AccountContext.
• ShareSearchContainerNode.swift uses context.engineData.get(...) because ShareControllerAccountContext exposes engineData: TelegramEngine.EngineData but not a full engine.
• CallStatusBarNode.swift (has raw account: Account from switch case) constructs TelegramEngine(account: account) inline.
• PresentationGroupCall.swift uses self.accountContext.engine.data instead of the stored self.account.postbox.

TelegramCore internal sites (36) unchanged. Postbox.swift (2 defs), State/AccountViewTracker.swift, State/FetchChatList.swift, State/SynchronizePeerReadState.swift, Suggestions.swift, and all TelegramCore/Sources/TelegramEngine/ internal _internal_* helpers still call postbox.loadedPeerWithId(...) — they are the Postbox-facing layer.

Pre-flight efficiency. An Explore subagent cataloged all 60 sites by category from a single prompt (one-line-per-site output). That catalog made the sweep straightforward: most files fell into identical patterns, enabling template-substitution Edits. Total context spent on discovery was small compared to doing 60 per-site full reads in the main thread.

Build validation. First-pass clean build (47 actions, 70s) after sweep completion. Earlier pilot (2 sites, 20s) validated pattern before scaling to all 60.

Lessons:

• loadedPeerWithId returns .never() on missing peer, not a pending Signal. Old common misreading: treating it as a "wait-until-loaded" primitive. Actual Postbox source at Postbox.swift:3925: if let peer = self.peerTable.get(id) { return .single(peer) } else { return .never() }. Preserve this by wrapping engine.data.get in mapToSignal with the .never() fallback — don't replace with plain |> compactMap { $0 } (which would drop the signal entirely rather than completing immediately when peer exists).

• "Keep the signatures to help the typechecker" as a migration principle. The user (2026-04-24) explicitly directed: keep call-site outer Signal signatures stable (Signal<EnginePeer, NoError> non-optional), even at the cost of a 6-line inline mapToSignal wrapper at each site. Rationale: 60 sites × optional-cascade body changes > 60 × 6-line wrapper. This is a general principle for sweeps — if the alternative is rewriting every body to handle optionals, prefer the signal-level wrapper to contain the change.

• Pre-flight cataloging via Explore subagent. For sweeps with variable per-site body shapes (unlike facade-migration-with-identical-call-expression sweeps), a dispatch to Explore with a category-classification prompt collapses inventory cost. Explore's output is small (~60 one-line entries); avoids pulling 60 file fragments into the main thread's context. Required for wave shapes where inventory is non-uniform.

• Shape-C peer-fed-to-API pattern needs _asPeer() at call, not facade. Because makePeerInfoController(peer: Peer) / FoundPeer(peer: Peer, ...) / SendAsPeer(peer: Peer, ...) / makeChatQrCodeScreen(peer: Peer, ...) all stay on raw Peer (they're AccountContext-protocol or TelegramCore struct-init APIs whose migration is its own multi-wave effort), the bridge is a single ._asPeer() at the call. Don't try to also migrate those APIs in the sweep — blast radius too large.

• Engine-access varies by containing context. Plain context.engine.data works for ~85% of sites; the remainder need TelegramEngine(account: account) construction or engineData protocol property. Build a per-site context type check into pre-flight for call-site categories where AccountContext isn't guaranteed.

Plan / record: no plan doc this wave — user specified the migration pattern directly; the Explore catalog + commit message captured decisions.

---

Wave 34 outcome (2026-04-24)

FoundPeer.peer: Peer → EnginePeer. Public field-type migration on the struct in submodules/TelegramCore/Sources/TelegramEngine/Peers/SearchPeers.swift. Atomic 12-file commit fdd5b93998. ~135 insertions / ~134 deletions.

Migration shape. The field-type change is necessarily atomic (half-migrated FoundPeer doesn't compile across consumers), so all edits land in one commit. _internal_searchPeers keeps import Postbox (still calls postbox.transaction etc.) and wraps raw peer values with EnginePeer(peer) at the FoundPeer constructor sites. == body changes from lhs.peer.isEqual(rhs.peer) to lhs.peer == rhs.peer.

Final scope (vs planned ~70 semantic edits → actual ~135 line insertions):

• 5 ._asPeer() bridge-drops at FoundPeer constructor sites (e.g., FoundPeer(peer: peer._asPeer(), ...) → FoundPeer(peer: peer, ...))
• 22+ redundant EnginePeer(peer.peer) wrap drops (the field is now EnginePeer; EnginePeer.init(_ peer: Peer) doesn't accept an EnginePeer argument so the wrap fails to compile)
• 30+ Postbox-concrete-type downcasts (peer.peer as? TelegramX / is TelegramX) rewritten to if case let .X(x) = peer.peer enum-pattern form
• ~10 ._asPeer() outflow bridges added where peer.peer flows into APIs that still take raw Peer: ContactListPeer.peer(peer:), canSendMessagesToPeer(_:), EngineRenderedPeer(peer:) legacy paths

Inventory undercounting — pattern. Original Explore inventory pass missed 4 of 12 final consumer files. The grep grep -rln "FoundPeer\b" only catches files that name FoundPeer as a literal type. Files that USE peer.peer access on FoundPeer values without naming the type itself were invisible to that grep. The build verification pass surfaced them:

File	Surfaced by	Edits needed
TelegramCore/Calls/GroupCalls.swift	iter 1	2 internal FoundPeer constructors needed EnginePeer(peer) wraps
ShareController/ShareSearchContainerNode.swift	iter 2	4 errors: 2 C2 downcasts + 2 outflow-bridge needs
ContactListUI/ContactsSearchContainerNode.swift	iter 3	7 errors: nested if !(peer is X) rewrite + multiple downcasts/outflows
PeerInfoUI/ChannelMembersSearchContainerNode.swift	iter 4	6 errors across 2 near-identical loop blocks
ChatListUI/ChatListSearchListPaneNode.swift (extra site)	iter 5	1 missed C2 site at line 3723 (in .globalPeer(foundPeer, …) enum case body, far from the other ChatListUI edits)

5 build iterations total before clean (each iteration: edit → re-build, ~50–60s incremental). First-pass would have needed a much wider pre-flight grep — see lessons.

Lessons:

• Inventory grep must include the access pattern, not just the type name. For a field-type migration, ALL of:

  • <Type>(peer: constructors
  • <x>.peer.<member> reads (verify <x> type is <Type>, not RenderedPeer/SendAsPeer/etc.)
  • <x>.peer as? / <x>.peer is downcasts
  • <api>(<x>.peer) arg passes (where <api> may take the old protocol)

  Use for x in Y binding-tracing to determine if <x> is the migrated type. The wave-34 pre-flight ran the first three but not the fourth (outflow-arg sites), and partially missed the second (because the Explore agent classified by literal FoundPeer token rather than by peer.peer semantics in context).

• if !(peer is A || peer is B) rewrite uses switch case A, B: break / default: .... When the original Postbox code uses a negated disjunction of type-checks, the cleanest enum-pattern equivalent is a switch with combined cases in one arm — not nested if cases. (Used in ChatListSearchListPaneNode:1024 and ContactsSearchContainerNode:502/544.)

• Inner peer shadowing. Many else if let peer = peer.peer as? TelegramChannel Postbox patterns shadow the loop variable. The enum-pattern rewrite renames the inner binding to channel to avoid double-shadowing the EnginePeer outer loop var. Block-internal references to .info etc. then move from peer.info to channel.info.

• Build iteration = inventory completion. When the inventory undercounting becomes apparent (build surfaces 5+ unexpected sites), don't abandon — iterate. Each build is fast (~50s incremental) and each error is actionable (error: cast from EnginePeer to unrelated type X always fails → C2 rewrite; argument type EnginePeer does not conform to expected type Peer → outflow bridge). The inventory grows by file, fix-then-rebuild converges in 5 iterations even when ~30% of sites were missed up front.

• Bridge sites generated by this wave point to next-ring migration targets. The ~10 ._asPeer() outflow bridges land at ContactListPeer.peer(peer:), canSendMessagesToPeer(_:), and EngineRenderedPeer(peer:) (legacy raw-Peer constructor in some paths — e.g., EngineRenderedPeer(peer: foundPeer.peer) doesn't need a bridge in newer EnginePeer-aware paths but does where the local var was already raw-Peer-extracted). These three signatures are the obvious wave-35+ candidates for the next ring of migration.

Plan / record: docs/superpowers/plans/2026-04-24-foundpeer-engine-peer-migration.md. Spec: docs/superpowers/specs/2026-04-24-foundpeer-engine-peer-migration-design.md.

---

Wave 35 outcome (2026-04-24)

SendAsPeer.peer: Peer → EnginePeer. Public field-type migration on the struct in submodules/TelegramCore/Sources/TelegramEngine/Messages/SendAsPeers.swift. Atomic 7-file commit 583c8b1f7c. 22 insertions / 26 deletions.

Migration shape. Same atomic-field-type pattern as wave 34 but scoped to a smaller consumer surface. The _internal_*SendAsAvailablePeers functions keep import Postbox and wrap raw peer values with EnginePeer(peer) at the 4 SendAsPeer constructor sites. Manual == body dropped in favor of synthesized Equatable (EnginePeer: Equatable, Int32? and Bool already Equatable).

Final scope (vs planned ~15 semantic edits → actual 22/26 line diff):

• 3 ._asPeer() bridge-drops at SendAsPeer constructor sites (ChatControllerLoadDisplayNode:772, ChatTextInputPanelComponent:848, StoryItemSetContainerViewSendMessage:249)
• 7 redundant EnginePeer(peer.peer) / EnginePeer($0.peer) / EnginePeer(value.peer) wrap drops across ChatSendAsPeerListContextItem (4 sites), ChatTextInputPanelNode (1), StoryItemSetContainerViewSendMessage (1), StoryItemSetContainerComponent (1)
• 1 peer.peer as? TelegramChannel downcast rewritten to if case let .channel(channel) = peer.peer (ChatSendAsPeerListContextItem:73) with peer.info → channel.info rename in the shadowed scope
• 2 EnginePeer(channel) wraps added where raw TelegramChannel is constructed into SendAsPeer(peer: ...) (ChatControllerLoadDisplayNode:805, 823)
• 1 signal-chain simplification: (sendAsPeer?.peer).flatMap(EnginePeer.init) → sendAsPeer?.peer at StoryItemSetContainerViewSendMessage:4080
• 1 signal-chain simplification: .map({ EnginePeer($0.peer) }) → .map({ $0.peer }) at StoryItemSetContainerViewSendMessage:4081

Inventory undercount = 1 site (vs wave 34's 5). The pre-flight Explore catalog missed StoryItemSetContainerComponent.swift:3069 (currentPeer: EnginePeer(value.peer) → value.peer). The implementer caught it during the edit phase before the build, so no iteration was needed. The wave-34 explicit pattern grep (including .peer as?/is/outflow-args/EnginePeer(.peer)/._asPeer()) dramatically reduced undercounting — 1/7 sites missed (~14%) vs wave 34's 4/12 (~33%).

First-pass clean build. No errors surfaced by the Bazel build at all. 461 total actions, 196.583s elapsed, INFO: Build completed successfully. Contrast with wave 34's 5 build-iterations-to-converge.

Lessons:

• Wave 34's explicit-pattern pre-flight inventory works. For future Peer-typed-API waves, the minimum grep pattern set is: <Type>\b literal token, \.<fieldName>\s+(as\?|is)\s+Telegram, EnginePeer\(\w+\.<fieldName>\), <api>\(<x>\.<fieldName> for known outflow APIs, and \._asPeer\(\) (to catch bridge-drop opportunities). Wave 35 used this full pattern set and hit ~14% undercount vs wave 34's ~33%.

• Smaller target + validated pattern = faster wave. Wave 35 went from spec-commit (72d4384af0) to outcome-commit in a single session with one clean build, versus wave 34's multi-iteration convergence. When the wave is a replay of a just-validated pattern on a smaller surface, expect minimal iteration.

• Inner-peer shadowing rename works. The wave-34 lesson about renaming peer → channel in if case let .channel(channel) = peer.peer applied cleanly. Single instance this wave (ChatSendAsPeerListContextItem:73) — no issues.

• Name collisions remain a scope hazard. Pre-flight identified sendAsPeers: [EnginePeer] (LiveStreamSettingsScreen, ShareWithPeersScreen) and availableSendAsPeers: [EnginePeer] (ChatSendStarsScreen) as name-only collisions — different type, same identifier. Confirmed these stayed untouched and out of scope. Future Peer-typed-API waves should continue the name-collision disambiguation pass.

• Bridge sites generated by this wave — zero new outflow bridges. Unlike wave 34 (which added ~10 ._asPeer() outflow bridges pointing to ContactListPeer.peer(peer:) / canSendMessagesToPeer(_:) / EngineRenderedPeer(peer:) as next-ring targets), wave 35 added no outflow bridges. All consumer-side .peer flows either stayed as .peer.id accesses (PeerId unchanged) or were simplifications of existing EnginePeer(.peer) wraps. Net: no new next-ring targets surfaced from wave 35.

Plan / record: docs/superpowers/plans/2026-04-24-sendaspeer-engine-peer-migration.md. Spec: docs/superpowers/specs/2026-04-24-sendaspeer-engine-peer-migration-design.md.

---

Wave 36 outcome (2026-04-24)

ContactListPeer.peer(peer: Peer, isGlobal:, participantCount:) → peer: EnginePeer. Enum-case payload migration on the public type in submodules/AccountContext/Sources/ContactSelectionController.swift. Atomic 15-file commit 069a060de1. 57 insertions / 59 deletions.

Migration shape. Same atomic-payload-type pattern as wave 34/35 but wider: 15 consumer files vs wave 35's 7, vs wave 34's 12. Beyond the payload change, the cascading ContactListPeer.indexName return type changed from PeerIndexNameRepresentation to EnginePeer.IndexName — an unexpected discovery during plan-writing that dropped 2 additional EnginePeer.IndexName(...) wraps at ContactListNode:517.

Final scope (vs planned 8 files / ~41 semantic edits → actual 15 files / 57/59 diff):

• Definition (1 file): AccountContext/ContactSelectionController.swift — case payload type, indexName return type, == operator body (lhsPeer.isEqual(rhsPeer) → lhsPeer == rhsPeer).
• 20 ._asPeer() outflow bridge-drops across ContactListNode (12), ContactsSearchContainerNode (3), ContactMultiselectionController (2), ContactMultiselectionControllerNode (1), ContactSelectionControllerNode (2). replace_all=true on ._asPeer(), isGlobal: was the unifying substring.
• 20+ EnginePeer(peer) inflow wrap-drops at destructure sites across ContactListNode (4), ContactsController (1), ContactsSearchContainerNode (4), ContactMultiselectionController (4), ContactMultiselectionControllerNode (1), ContactSelectionController (2), PeerSelectionControllerNode (3), SharedAccountContext (2).
• 2 EnginePeer.IndexName(...) wrap-drops at the sort-comparator at ContactListNode:517 (enabled by the cascading return-type change).
• 8 Postbox-concrete cast rewrites to EnginePeer case patterns across ContactListNode:182-186/1968 (4 sites, including the 3-branch user/group/channel cast-chain), CallController:524/542 (the intermediate let peer = EnginePeer(peer) lines became redundant after migration), StoryItemSetContainerViewSendMessage:2041/2074, DeviceContactInfoController:1419, ChatSendAudioMessageContextPreview:89, ChatControllerOpenAttachmentMenu:557/610/1746/1788 (4 identical sites, replace_all on the full line).
• 2 ._asPeer() outflow bridges ADDED at ContactMultiselectionController:386/403 where the destructured peer flows into peerTokenTitle(peer: Peer) (out-of-scope callee; future-wave bridge target).

Inventory undercount = 7 files / ~20 sites (vs wave 35's 1 site). Much higher miss rate than wave 35 — ~46% by file count. Root cause: the pre-flight grep for ContactListPeer destructures used literal \(peer, _, _\) binding; binding names varied in practice (contact, lhsPeer, rhsPeer, contactPeer, id). Files missed:

1. DeviceContactInfoController.swift:1418/1419 — case let .peer(contact, _, _) + contact as? TelegramUser
2. CallController.swift:523/541 — case let .peer(peer, _, _) + redundant let peer = EnginePeer(peer) pattern
3. ChatSendAudioMessageContextPreview.swift:88/89 — case let .peer(contact, _, _) + contact as? TelegramUser
4. PeerSelectionControllerNode.swift:901-903/1590-1592 — 2 destructures with EnginePeer(peer) inflow wraps
5. StoryItemSetContainerViewSendMessage.swift:2040-2041/2073-2074 — 2 contact as? TelegramUser casts
6. ChatControllerOpenAttachmentMenu.swift:556-1787 — 4 contact as? TelegramUser casts
7. SharedAccountContext.swift:3295-3302 — case let .peer(peer, _, _) + 2 EnginePeer(peer) inflow wraps

Six build iterations to converge vs wave 35's single first-pass-clean. Iterations 1-6 surfaced errors in batches of 2-8 errors; each was a mechanical fix (drop wrap, rewrite cast, add ._asPeer() bridge for outflow to out-of-scope peerTokenTitle). Final iteration (#6) clean.

Lessons:

• Pre-flight grep must use \(\w+, _, _\) not \(peer, _, _\) for enum-payload destructures. Swift destructure patterns bind the payload to any legal identifier; the variable name is not semantic. Future Peer-typed-enum-payload waves should use case let \.<caseName>\((\w+), (or similar wildcard binding) and then per-destructure scan the next ~15 lines for <binding> as\?/<binding> is/EnginePeer\(<binding>\) / outflow-arg patterns.

• "No-edit consumer" claims need stricter verification. Wave 36's "verify-only" list included ChatSendAudioMessageContextPreview because the initial inventory found only [ContactListPeer] at collection level. The deeper scan missed a case let .peer(contact, _, _) + contact as? TelegramUser pattern inside the file's update(...) method. For future waves, "no-edit" claims should run the wildcard-binding destructure grep described above, not just a construction-site grep.

• Outflow-to-out-of-scope-API bridges may need addition during the wave. ContactMultiselectionController:386/403 needed ._asPeer() bridges added where none existed pre-migration — the pre-migration code passed raw Peer to peerTokenTitle(peer: Peer) because the destructured peer was raw Peer. Post-migration, the destructured peer is EnginePeer, so a bridge is required. Future waves with same-scope outflow to not-yet-migrated Peer-typed APIs should pre-flight expect to add bridges.

• Cascading computed-property return type migration (here: ContactListPeer.indexName from PeerIndexNameRepresentation to EnginePeer.IndexName) is a legitimate scope expansion when the enum's properties leak Postbox-typed values. Wave 36 caught this during plan-writing, not execution — a successful plan-review win. Future waves should grep the enum's definition file for computed properties returning Postbox-defined types.

• Build-iteration convergence is acceptable when the wave's surface is large and pre-flight undercount is non-trivial. The cost of 6 build iterations (~5-20 minutes each in the Telegram-iOS build) is real but manageable. The alternative — exhaustive pre-flight to achieve first-pass-clean — is more expensive in plan-writing tokens and controller wall time. For waves expected to have >5 file touches, plan should explicitly budget for 3-5 build iterations.

• Ratchet effect confirmed. Wave 36 was predominantly bridge-removal (20 outflow + 20 inflow + 2 IndexName) with only 2 bridge additions. Matches the expected ratchet behavior: earlier waves 33/34/35 added bridges at Peer/EnginePeer boundaries precisely so wave 36 could drop them atomically. The 2 new bridges added (ContactMultiselectionController:386/403 → peerTokenTitle) become next-wave drop candidates once peerTokenTitle(peer: Peer) migrates.

Plan / record: docs/superpowers/plans/2026-04-24-contactlistpeer-engine-peer-migration.md. Spec: docs/superpowers/specs/2026-04-24-contactlistpeer-engine-peer-migration-design.md.

---

Modules currently free of import Postbox (running tally)

Consumer modules that no longer import Postbox, across all waves and standalone commits:

• ChatInterfaceState (wave 1)
• ChatSendMessageActionUI (wave 1)
• ContactListUI (wave 1)
• DrawingUI (wave 1)
• StickerPeekUI (standalone cleanup, 2026-04-17 — import was unused)
• PromptUI (standalone cleanup)
• PresentationDataUtils (standalone cleanup)
• MapResourceToAvatarSizes (wave 2)
• SaveToCameraRoll (wave 3)
• SecureIdVerificationDocumentsContext (wave 5)
• Wave 6 batch: 189 additional modules — see git show 7b2b74e79b --stat for the commit that swept unused import Postbox lines across 183 files in 16 consumer submodules. Not individually enumerated here for brevity.
• StorageUsageScreen (waves 8–10)
• ActionSheetPeerItem (wave 11; revisits wave-1 abandonment)
• HorizontalPeerItem (wave 12; applies wave-11 pattern)
• SelectablePeerNode (wave 15; applies wave-11 pattern; ShareExtension-boundary stateManager fallback)
• ItemListAvatarAndNameInfoItem (wave 17; applies wave-11 pattern; ShareExtension-boundary stateManager fallback)
• ItemListStickerPackItem (wave 18; mixed-shape — 3 narrow TelegramCore typealiases + wave-4 enum-payload migration + wave-3 facade swap)
• AttachmentTextInputPanelNode BUILD cleanup (wave 13; source was already clean from wave 6)
• Wave 14 BUILD-dep sweep: 98 modules' BUILDs cleaned — same modules as the wave-6 batch; this sweep fixed their leftover //submodules/Postbox:Postbox BUILD deps. Candidate list in /tmp/postbox-dep-candidates.txt at commit time; derivable by the script in "Wave 14 outcome".