perf(ios): publish refreshed chats independently of model catalogs

[sambitcreate]

A successful iOS chat refresh previously waited for the model catalog, and a catalog failure discarded the fresh transcript result. Publish the transcript independently while retaining structured task ownership, cache admission, live model selection and credential-revocation fences. Overall load completion and progress bootstrap still await the catalog. If revocation arrives after publication, the removal path now clears mounted messages synchronously before cleanup can suspend.

Validation: the initial loader regression fails seven assertions on original source. The follow-up held-purge regression reproduces two failures before redaction; fixed production passed all 213 chat XCTests. Final focused tests passed 2/2 after one documented simulator-launch retry. Fresh-context follow-up review independently passed three targeted race/error cases without retry, with no actionable findings. Release-policy checks passed. Ordinary catalog 503 retains the transcript and selected model.

Evidence: docs/performance/ios-chat-loading.md. No shared wire contract changes; Android already loads chat/catalog independently. No physical-device latency or energy result is claimed.

[very-hermes-bot]

Hermes Review Bot

Confidence: 5

Engine: agy/gemini-3.8-flash-high
Review mode: full
Head: cb53bbe6a8bed4f31299a0be9363671b92ec5008
Generated: 2026-09-28T03:11:28+00:00
Reviews: 1

Summary

Decouples transcript retrieval from model catalog loading during iOS chat initialization, resolving an issue where a successful transcript fetch was blocked by or discarded due to catalog latency or failures. In AidenChatViewModel.load(), refreshModelCatalog(context:) is now spawned as an independent structured child task while the chat transcript GET completes and publishes immediately via acceptRemoteChat(_:context:writeToken:). The catalog failure path swallows non-revocation errors (such as HTTP 503) without overriding visible transcripts, while credential revocation (401) fences pending operations and synchronously clears in-memory chat.messages in handleRemoval() before coordinator cleanup finishes. Maintainers should double-check that downstream progress observation or approval restoration in load() does not rely on newly advertised catalog metadata before catalogRefresh resolves.

Confidence Score: 5/5

Fully traced and verified across iOS ViewModel lifetime rules, actor isolation, cancellation handling, Android parity (AidenChatViewModel.kt), and corresponding XCTests.

📁 Important Files Changed

• ios/AidenOnTheGo/Features/Remote/AidenChatFeature.swift: Splits the previous (chatRequest, catalogRequest) tuple await into an immediate transcript publication step and an independent refreshModelCatalog(context:) method; clears chat.messages synchronously in handleRemoval().
• ios/AidenOnTheGoTests/AidenChatTests.swift: Adds regression tests covering catalog hold/release (including HTTP 503), chat failure with catalog success, invalidation/unpair while catalog is held, published transcript redaction prior to cleanup completion, and catalog revocation fencing held transcripts.
• docs/performance/ios-chat-loading.md: Documents deterministic timing and failure evidence on the iPhone 17 Pro / iOS 27 simulator, measurement boundaries, and revocation redaction follow-up behavior.
• .memory/perf-ios-chat-loading.md: Tracks project context, test run results, and architectural decisions.

Findings

No findings.

Sequence Diagram

sequenceDiagram
    autonumber
    participant V as View (AidenChatDetailView)
    participant VM as ViewModel (AidenChatViewModel)
    participant C as RemoteClient
    participant CO as Coordinator

    V->>VM: load()
    par Fetch Transcript
        VM->>C: chat(id)
    and Fetch Catalog (Child Task)
        VM->>C: modelCatalog()
    end
    C-->>VM: remoteChat (200 OK)
    VM->>VM: acceptRemoteChat()
    VM-->>V: Publish updated chat transcript
    alt Catalog Succeeds
        C-->>VM: remoteCatalog (200 OK)
        VM->>VM: resolveModelSelection()
    else Catalog Transient Failure (e.g. 503)
        C-->>VM: error (503)
        Note over VM: Suppress catalog error; retain existing transcript & selection
    else Credential Revoked (401)
        C-->>VM: error (401 credential_revoked)
        VM->>CO: handleCredentialRevocation()
        VM->>VM: handleRemoval() (chat.messages = [])
        VM-->>V: Synchronously redact messages
    end

Loading

[]

---

Last reviewed commit: cb53bbe6a8be
Reviews (1) · Comment /hermes review to trigger a new review · /hermes review full for full re-review

[pullfrog]

Important

A catalog credential_revoked response can arrive after transcript publication and leave the newly fetched transcript renderable while the installation-data purge runs.

Reviewed changes This review covers the iOS transcript/catalog refresh decoupling, its race tests, and the performance evidence.

• Independent transcript publication. load() now publishes the fetched transcript while the structured catalog request is pending; overall load and progress bootstrap still await catalog completion.
• Optional catalog outcome. Ordinary catalog failures no longer discard the transcript or existing model choices, and successful refresh resolves the live selection.
• Race coverage and evidence. Tests cover held-catalog success/failure, selection preservation, chat failure, removal/unpair, and revocation while the transcript is held. The report states that 212 chat XCTests passed and makes no physical-device latency claim.

  ^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using GPT Luna ｜ 𝕏}

[pullfrog]

✅ No new issues found.

Reviewed changes This review covers the revocation-redaction follow-up since the prior Pullfrog review.

• Redacted removed chat content. handleRemoval() now clears mounted transcript messages synchronously before asynchronous installation cleanup, and the regression holds cleanup to verify redaction before the shell switches to pairing.
• Strengthened catalog-failure coverage. The held-catalog scenario now confirms the fresh transcript remains after an ordinary catalog failure is released.
• Recorded follow-up evidence. The performance notes document the late-revocation ordering, simulator results, and remaining measurement limits.

  ^{｜ View workflow run ｜ Using GPT Luna ｜ 𝕏}