OpenAI 兼容网关断流时 finish_reason: network_error 被误判为 PI_AI_ERROR 且默认不重试

[Billmvp73]

OpenAI 兼容网关中途断流时，finish_reason: network_error 被误判为 PI_AI_ERROR，默认重试策略不重试。

复现、预期与验收

• 复现步骤：

  1. 用 @deepseek-ai/dsh-llm-pi-ai 对接 OpenAI 兼容网关（如 OpenCode Go 的 ox-alpha-free 免费路由）；
  2. 触发网关中途断流（免费/stealth 路由在持续 agent 流式下易掉线），网关在 SSE 里返回 finish_reason: "network_error"；
  3. 会话报错 本轮运行失败 Provider finish_reason: network_error / PI_AI_ERROR。

• 实际结果：
  packages/llm/llm-pi-ai/src/stream.ts::classifyPiAiError 的正则 /\b(?:network|connection|socket|fetch)\b/ 匹配不到 network_error（_ 是单词字符，k 与 _ 之间无词边界），落到 stream.ts:64 返回 PI_AI_ERROR；而默认重试码 packages/llm/llm/src/retry-policy.ts:18-24（EMPTY_RESPONSE/RATE_LIMIT/SERVER/TIMEOUT/TRANSPORT）不含 PI_AI_ERROR，故不会自动重试。已实测 classifyPiAiError('Provider finish_reason: network_error') 返回 PI_AI_ERROR。

• 预期结果：
  将非标准的 provider finish_reason: network_* / connection_* / socket_* 归类为 TRANSPORT，使其进入默认重试策略自动重试。

• 环境：

  • 仓库 packages/llm/llm-pi-ai，@earendil-works/pi-ai@^0.82.1
  • 路由 opencode-go-free（OpenCode Go），模型 ox-alpha-free，https://opencode.ai/zen/go/v1/chat/completions

• 验收条件：

  • classifyPiAiError('Provider finish_reason: network_error') === 'TRANSPORT'（含 connection_error / socket_error）；
  • content_filter 仍为 PI_AI_ERROR（不发生误伤）；
  • 在 packages/llm/llm-pi-ai/tests/convert.spec.ts 的 transport it.each（约 827 行）增补对应用例；
  • 默认重试策略对该错误自动重试。

[xryanv]

We verified and published a standalone fix for this error classification issue:

Implementation

• Branch: fix/pi-ai-transient-error-classification
• Commit: 93a863b154cb7b9ff6c72e2764b8be568aa69bdf
• Changes: Updated classifyPiAiError (packages/llm/llm-pi-ai/src/stream.ts) to normalize network_error, stream_read_error, connection_reset, socket_error, fetch_error to TRANSPORT, and server_error to SERVER.
• Permanent Error Non-Looping: Verified that genuine non-retryable errors (401, 403, 400, UNKNOWN_MODEL, CONTEXT_WINDOW_EXCEEDED) terminate immediately on attempt 1 without consuming retry budgets.
• Tests: 354/354 passed across packages/llm/llm-pi-ai and packages/llm/llm-retry with end-to-end mock stream recovery tests.

(Cross-reference: also addresses related reports in #3158 and #3407).

[AnhLead]

Complementary gap for the same classifyPiAiError / transport family — different error surface than network_error.

When an openai-completions stream is truncated mid-SSE-frame, the OpenAI SDK can throw:

Unterminated string in JSON at position N

That still maps to PI_AI_ERROR today (including on current master), so default retry does not run. Same transport-truncation class as the 2026-07-22 note, but not covered by network_error / stream_read_error patterns.

Proposed patch (PRs disabled here — compare link):

master...AnhLead:fix/pi-ai-truncated-json-transport

• Maps unterminated string in json and generic stream was interrupted wording → TRANSPORT
• Tests in packages/llm/llm-pi-ai/tests/convert.spec.ts

Happy to rebase onto fix/pi-ai-transient-error-classification if maintainers want one combined branch.

[AnhLead]

@xryanv