[bug] WebSocket error is classified as PI_AI_ERROR, bypassing retries

[Beeee0927]

WebSocket error is classified as PI_AI_ERROR, bypassing retries

Summary

A transient OpenAI Codex transport failure with the exact pi-ai message WebSocket error is classified by llm-pi-ai as PI_AI_ERROR instead of TRANSPORT. Since the default DSH retry policy retries TRANSPORT but not PI_AI_ERROR, the agent turn ends immediately without emitting llm/retry.

This is distinct from the already-covered message WebSocket closed unexpectedly.

Observed behavior

Environment:

• DSH: 0.1.0-rc.6
• Provider: openai-codex
• Adapter: @deepseek-ai/dsh-llm-pi-ai

Persisted session events:

{"type":"assistant/chunk","chunk":{"type":"finish","reason":{"kind":"error","failure":{"message":"WebSocket error","code":"PI_AI_ERROR"}}}}
{"type":"turn/end","reason":{"kind":"error","error":{"message":"WebSocket error","code":"PI_AI_ERROR"}}}

The session contained zero llm/retry / llm/retry-started events.

Root cause

packages/llm/llm-pi-ai/src/stream.ts recognizes this WebSocket wording:

WebSocket closed unexpectedly

but not the shorter wording emitted by pi-ai:

WebSocket error

The latter falls through to PI_AI_ERROR. The default provider retry policy includes TRANSPORT, so classification prevents the existing outer retry mechanism from running.

The Codex OAuth plugin is not the transport owner: it only supplies/refreshes credentials. pi-ai's openai-codex-responses provider owns the WebSocket connection.

Expected behavior

WebSocket error should map to TRANSPORT, allowing the bounded outer DSH retry policy to restart the failed step.

Minimal fix

- WebSocket closed unexpectedly
+ WebSocket (?:closed unexpectedly|error)

Add WebSocket error to the existing table-driven transport-wording test in packages/llm/llm-pi-ai/tests/convert.spec.ts.

Validation

Applied to upstream master at 47f943859bef60e4160492346772ded9b24f765a:

vitest packages/llm/llm-pi-ai/tests/convert.spec.ts: 69 passed
oxlint changed source/test: 0 warnings, 0 errors
pre-commit gates: passed

A portable git format-patch is retained as upstream-fix.patch in this directory.

[lyd123qw2008]

补充一个相关的 Responses/Codex Gateway 失败：Our servers are currently overloaded. Please try again later. 最终被 DSH 记为 PI_AI_ERROR，因此没有进入 provider 的有界重试。

我同意应修复漏分的 WebSocket error，但建议不要继续扩大英文短语表作为主路径。Responses API 在 HTTP failure 中提供 status，在 SSE error / response.failed 中提供 machine code（例如 server_error、rate_limit_exceeded）；pi-ai 0.82.1 在构造 AssistantMessage 前把这些字段压成了 errorMessage，导致 DSH 只能猜文字。

更稳的 seam 是保留可选 errorCode/errorStatus，按 code -> status -> 仅 legacy fallback 文本 分类。这样 server_error 能稳定进入 SERVER 重试，不受提示文案变化影响。完整说明见 #1215。