fix: preserve provider failure taxonomy through retries and Desktop presentation

[xuruiray]

Summary

Local end-to-end validation of #1214 found two separate boundary defects in the provider-failure path:

1. Retryable provider failures lose their taxonomy after Vercel AI SDK wraps the final APICallError in AI_RetryError.
2. Desktop receives a safe, structured ErrorEvent.reason, but reparses only event.message, producing generic or unknown UI copy.

These are existing behavioral gaps, not regressions introduced by #1214. The extraction in #1214 remains byte-for-byte behavior-neutral and should stay separate from these fixes.

Reproduction evidence

Reproduced on 2026-07-19 from commit 44baea50 with:

• the real Electron Desktop renderer and IPC bridge
• SessionManager -> AiSdkBackend -> ModelAdapter -> RuntimeEvent -> TurnRecord
• the real ai@7.0.31 OpenAI-compatible adapter
• an isolated loopback provider and isolated Desktop user-data directory
• no real provider credentials

1. Retry wrapper loses 429 / 503 classification

The loopback provider returned one response per request. AI SDK made three attempts for each retryable failure:

Scenario	Provider requests	Expected final reason	Actual final reason / TurnRecord
HTTP 429	3	rate_limit	no reason / unknown
HTTP 503	3	provider_unavailable	no reason / unknown

A direct runtime probe showed:

429 outer class: AI_RetryError
429 lastError class: RateLimit

503 outer class: AI_RetryError
503 lastError class: ProviderUnavailable

The outer error exposes no top-level provider status. AI SDK keeps the final AI_APICallError in its public lastError field, but Maka currently classifies only the outer Error.

2. Desktop discards the structured reason

The runtime correctly emitted and persisted:

reason: context_overflow
message: Context window exceeded
TurnRecord.errorClass: context_overflow

The visible Desktop toast was still the generic Chinese fallback:

对话运行失败，请稍后重试。

The same pure presentation probe currently produces:

Stable reason	Toast	Failed-turn label
context_overflow	generic failure	unknown error
timeout	generic failure	request timeout
provider_billing	generic failure	unknown error
provider_unavailable	generic failure	model service error
auth / rate_limit / network	categorized	categorized

Current failure path

flowchart LR
  subgraph R["Runtime classification gap"]
    R1["HTTP 429 / 503"] --> R2["AI_APICallError<br/>statusCode available"]
    R2 --> R3["AI_RetryError<br/>lastError available"]
    R3 --> R4["Classify outer error"]
    R4 --> R5["unknown"]
    R3 -. "Unwrap official retry error" .-> R6["rate_limit / provider_unavailable"]
  end

  subgraph D["Desktop presentation gap"]
    D1["ErrorEvent<br/>reason + safe message"] --> D2["Drop reason"]
    D2 --> D3["Reparse safe English message"]
    D3 --> D4["generic toast / unknown label"]
    D1 -. "Allowlisted reason mapping" .-> D5["localized actionable copy"]
  end

Loading

Root cause

Runtime

The normalizer reads statusCode, code, provider data, and response bodies from the error it receives. It does not recognize the official AI SDK retry wrapper or inspect lastError.

Both terminal event projection and telemetry use that classification, so the loss propagates into:

• ErrorEvent.reason
• TurnRecord / AgentRun failure taxonomy
• Runtime Event Log and traces
• headless provider-infrastructure accounting
• Desktop presentation

This does not cause an extra retry today: AI SDK has already exhausted its retry policy, and Maka's additional transport retry remains limited to the existing Network path.

Desktop

ErrorEvent.reason is explicitly the stable machine-readable UI/telemetry routing field. sessionEventErrorMessage() ignores it and passes only new Error(event.message) into the generic redaction classifier.

That safe message is already a projection. Parsing it again is lossy: for example, Request timed out does not contain the current timeout keyword, and Provider returned an error no longer contains a 5xx status.

The failed-turn label has a separate allowlist, but it currently omits at least context_overflow and provider_billing.

Related work

• Follow-up to fix(runtime): derive provider error reason and message from one classification #1074 / fix(runtime): unify provider error projection #1075, which unified provider error reason/message projection but did not cover retry wrappers.
• Related to feat(desktop): make UI language selection coherent and complete #1052's Desktop localization work, but this defect affects both Chinese and English and is primarily a structured-routing problem.
• Discovered while validating refactor(runtime): extract provider error classification #1214.

Non-goals

• Changing AI SDK or Maka retry counts/backoff policy.
• Retrying after partial model/tool output.
• Exposing raw provider response bodies.
• Treating output-token-cap errors as input context overflow.
• Expanding refactor(runtime): extract provider error classification #1214 from a refactor into a behavior change.

[xuruiray]

@Astro-Han Could you help take a look at the two implementation PRs when convenient?

• fix(runtime): preserve provider errors through retries #1218 — preserve provider failure classification through retry wrappers
• fix(desktop): present provider failures from event reasons #1219 — present stable provider failure reasons consistently in Desktop

Both are ready for review and CI is green. Thanks!

[Astro-Han]

Both PR reviewed! We only need rebase for the second PR to merge.