[Meta] v0.2 provider/generation failure diagnostics and recovery sweep

[Sun-sunshine06]

During the 2026-05-08 open issue triage, many GENERATION_FAILED reports clustered into a few provider/gateway failure modes rather than independent product bugs. This issue tracks the remaining v0.2 provider/generation diagnostics and recovery work in one place so duplicate reports can point here.

Current status

PR #315 has merged into main as 768dc8b. It covers the concrete diagnostics sweep from this triage:

• GENERATION_TIMEOUT now has a dedicated generation-timeout diagnostic with localized copy.
• Timeout failures offer an action that opens Settings -> Advanced so users can adjust the generation timeout.
• 400/404/422 errors involving a models/ prefixed model id now diagnose as model-id-shape, even when the gateway returns no useful body.
• The existing one-click model id normalization action is offered for those prefixed model id failures.
• CI now allows Electron install scripts under pnpm 10, and the onboarding IPC registration test no longer relies on inter-test state.

Closed as covered by this sweep

• Generation timeout duplicates: [Bug]: GENERATION_FAILED (fp: 69213336) #309, [Bug]: GENERATION_FAILED (fp: 69213336) #308, [Bug]: GENERATION_FAILED (fp: 69213336) #303, [Bug]: GENERATION_FAILED (fp: 69213336) #277.
• models/ prefixed model id reports: [Bug]: GENERATION_FAILED (fp: a1ef45df) #282, Check this out. Couldnt try it[Bug]: GENERATION_FAILED (fp: 78910818) #269.
• Repeated terminated / transport-interrupted duplicates consolidated here: [Bug]: GENERATION_FAILED (fp: c85d010d) #312, [Bug]: GENERATION_FAILED (fp: c85d010d) #310, [Bug]: GENERATION_FAILED (fp: c85d010d) #262, [Bug]: GENERATION_FAILED (fp: c85d010d) #237, [Bug]: GENERATION_FAILED (fp: c85d010d) #235.

Still useful follow-up

• Generations silently cancelled at 10 min against local LM Studio / Ollama endpoints, regardless of generationTimeoutSec #199: deeper provider/request-timeout plumbing for local LM Studio / Ollama endpoints. fix: improve generation failure diagnostics #315 improves the recovery UX, but this remains the underlying timeout behavior to revisit if pi-ai exposes a suitable per-request timeout option.
• bug: 第三方中转网关处理 OpenAI Responses API SSE 时，遇到 'response' 事件名会提前截断 #167: third-party gateway / Responses API SSE cutoff behavior that still needs gateway-specific repro notes.
• [Bug]: GENERATION_FAILED (fp: 018597eb) #265: DeepInfra/custom provider 422 behavior where Test Connection succeeds but generation fails.
• using deepseek-v4-pro design data dashborad happen Error "invoking remote method 'codesign:v1:generate': CodesignError: Connection error.[Bug]: GENERATION_FAILED (fp: e443b98c)" #264: long-running custom DeepSeek provider connection error.
• Provider roadmap issues that remain useful: [Feature]: 增强 Provider 错误归一化与恢复提示 / Improve provider error normalization and recovery actions #212, [Feature]: 升级连接诊断为 capability-aware / Improve connection diagnostics with capability-aware detection #213, [Feature]: 加固 Provider / Gateway 兼容性与外部 API 接入 / Harden provider-gateway compatibility and external API integration #214, [Feature]: 对齐“测试连接”和“实际调用”链路 / Ensure connection-test and actual-invocation parity #216, [Feature]: 增加 LiteLLM Gateway 一等预设 / Add a first-class LiteLLM gateway preset #208, [Feature]: 跟进 OpenCode / Codex / CCS 风格配置导入路线 / Follow up OpenCode-Codex-CCS style import roadmap #209, [Feature]: 正式建模模型发现模式 / Formalize provider model discovery modes #210, [Feature]: 统一外部 Provider 导入为共享 Adapter 框架 / Unify external provider imports under a shared adapter framework #211.

Kept separate

• [Bug]: GENERATION_FAILED (fp: a3cd0486) #261 is a reference URL fetch/validation path, not a provider invocation failure. Keep it as its own issue unless it later reproduces as a provider diagnostic problem.

Future follow-up

• If pi-ai exposes per-request timeout configuration for all relevant providers, plumb generationTimeoutSec into the underlying provider request timeout instead of only surfacing the recovery action.
• Add adapter-specific repro notes for third-party gateways that still terminate Responses API SSE streams early.
• Keep connection-test versus actual invocation parity work linked to [Feature]: 对齐“测试连接”和“实际调用”链路 / Ensure connection-test and actual-invocation parity #216.

[github-actions]

Thanks for putting together this comprehensive v0.2 provider/generation diagnostics and recovery sweep, @Sun-sunshine06. It's very helpful to have all the failure modes and follow-ups clustered in one place.

From the codebase (changesets and existing code), I can confirm that several of the items you've listed are already implemented:

• GENERATION_TIMEOUT diagnosis with localized copy and Settings → Advanced action: verified in .changeset/fix-generation-timeout-reason.md (also in the applyError handling in packages/core and renderer error surfaces).
• 400/404/422 models/-prefixed model ID diagnosis (model-id-shape classification): referenced in .changeset/fix-reasoning-flag-openai-compat.md and the provider error classification in packages/providers.
• Diagnostic report loop (error events store + Report bug flow): .changeset/diagnostics-report-loop.md documents the full report loop with i18n error codes.
• Retry for transient provider errors (5xx, 429, network): .changeset/agent-first-turn-retry.md adds a backoff loop for the first generation turn.

The remaining follow-ups (plumbing generationTimeoutSec into provider request timeouts, adapter-specific repro for SSE cutoff, connection-test vs. invocation parity) are not yet visible in the committed changesets or source code — seems like future work as noted.

If there's anything I can help dig into (specific issue clusters, code paths, or testing), let me know.

---

Open-CoDesign Bot