fix: 修复 3 个后端 BUG + 新增向导恢复/超时重试功能#157
Conversation
- Fix to_api_dict() crash in embedding config update (settings.py:175) - Fix OpenAI provider 500 InternalServerError handling for Responses API fallback - Fix chapter generation word count control: use novel target_words_per_chapter - Fix autopilot max_tokens based on target words per chapter - Add 'Setup Wizard' button in workbench StatsTopBar for existing novels - Add retry/refresh button in wizard when generation times out - Auto-trigger generation when wizard resumes at character/location steps
📝 WalkthroughWalkthroughAdds configurable target-word plumbing for chapter generation, updates daemon token budgeting to use the target, exposes target words via the generation API, introduces a frontend setup wizard and retry flows, broadens OpenAI provider fallback/error detection, and adds an embedding-model fetch endpoint. ChangesChapter Generation Target Words
Frontend: Setup Wizard & Onboarding
OpenAI Provider Fallbacks
Embedding Settings API
Sequence Diagram(s)sequenceDiagram
participant Client as Client
participant API as Engine API
participant NovelSvc as NovelService
participant Workflow as AutoNovelGenerationWorkflow
participant Daemon as AutopilotDaemon
participant LLM as LLM Provider
Client->>API: request generate_chapter_stream(novel_id, ...)
API->>NovelSvc: get_novel(novel_id)
NovelSvc-->>API: novel (target_words_per_chapter)
API->>Workflow: generate_chapter_stream(..., target_words)
Workflow->>Daemon: request writing (beat or non-beat)
alt non-beat
Daemon->>Daemon: compute max_tokens = int(tw * 1.5)
end
Daemon->>LLM: stream_generate(prompt, token_budget)
LLM-->>Daemon: streaming tokens / events
Daemon-->>Workflow: stream events
Workflow-->>API: forward stream to Client
(Note: colored rectangles not used; diagram focuses on component interactions.) Estimated code review effort🎯 3 (Moderate) | ⏱️ ~25 minutes Possibly related PRs
Suggested reviewers
Poem
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches🧪 Generate unit tests (beta)
Tip 💬 Introducing Slack Agent: The best way for teams to turn conversations into code.Slack Agent is built on CodeRabbit's deep understanding of your code, so your team can collaborate across the entire SDLC without losing context.
Built for teams:
One agent for your entire SDLC. Right inside Slack. 👉 Get your free trial and get 200 agent minutes per Slack user (a $50 value). Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Review rate limit: 7/8 reviews remaining, refill in 7 minutes and 30 seconds.Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 4
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (4)
frontend/src/components/onboarding/NovelSetupGuide.vue (1)
954-975:⚠️ Potential issue | 🟠 Major | ⚡ Quick winGuard async resume flow against stale open/close cycles.
After
await detectWizardProgress(), stale callbacks can still run and calltriggerStepGeneration(step)even if the wizard was closed (or reopened), causing unintended backend generation.Suggested fix
const currentStep = ref(1) +const openSessionEpoch = ref(0) watch( () => props.show, async (val) => { if (val) { + const epoch = ++openSessionEpoch.value resetWizardStateForOpen() // 检查已有进度，确定从哪一步继续 const step = await detectWizardProgress() + if (!props.show || openSessionEpoch.value !== epoch) return currentStep.value = step triggerStepGeneration(step) } else { + openSessionEpoch.value += 1 stopGenerationOnClose() } } ) onMounted(async () => { if (props.show) { + const epoch = ++openSessionEpoch.value resetWizardStateForOpen() const step = await detectWizardProgress() + if (!props.show || openSessionEpoch.value !== epoch) return currentStep.value = step triggerStepGeneration(step) } })🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@frontend/src/components/onboarding/NovelSetupGuide.vue` around lines 954 - 975, After awaiting detectWizardProgress() ensure the wizard is still open before proceeding: in both the watcher and the onMounted handler, capture the current visibility (or increment a local generation token/nonce) before the await and then check that props.show is still true (or the token matches) after the await; only then set currentStep.value and call triggerStepGeneration(step); also propagate the same guard to any async paths started by resetWizardStateForOpen so stale callbacks cannot call triggerStepGeneration when stopGenerationOnClose has run.infrastructure/ai/providers/openai_provider.py (1)
126-135:⚠️ Potential issue | 🟠 MajorOnly fall back to Chat Completions before the first chunk is emitted; otherwise re-raise the exception.
The exception handlers at lines 126–132 can execute after the Responses stream has already yielded content (line 124). When they complete without returning or re-raising, execution continues to the fallback logic (lines 137–144), which creates and yields from an entirely new Chat Completions stream. Callers like
auto_novel_generation_workflow.pyandautopilot_daemon.pyiterate over chunks sequentially and will receive mixed or duplicated content from both streams, corrupting the generated output. Add a flag to track whether any chunks have been yielded, and only proceed with fallback if the stream hasn't started; otherwise, re-raise.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@infrastructure/ai/providers/openai_provider.py` around lines 126 - 135, The Responses stream exception handlers currently fall back to Chat Completions even if the Responses stream has already emitted chunks; add a local flag (e.g., started_emitting = False) in the streaming generator that you set to True immediately before yielding the first chunk, then in the except blocks (the handlers that call self.__class__._fallback_to_chat_cache.add and logger.info) check that flag and only perform the fallback when started_emitting is False; if started_emitting is True, re-raise the caught exception instead of swallowing it so callers do not receive mixed/duplicated content. Ensure the flag is referenced in the same method handling the Responses stream and use the original exception when re-raising.application/workflows/auto_novel_generation_workflow.py (1)
737-766:⚠️ Potential issue | 🟠 Major | ⚡ Quick winForward
target_wordsinto_build_prompt().
build_chapter_prompt()accepts the new parameter, but it never passes it through. That makes the API misleading and leaves callers on the default 2500-word prompt text, which breaks the new chapter-length control in helper-based call sites.🔧 Suggested fix
return self._build_prompt( context, outline, storyline_context=storyline_context, @@ voice_anchors=voice_anchors, chapter_draft_so_far=chapter_draft_so_far, + target_words=target_words, )🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@application/workflows/auto_novel_generation_workflow.py` around lines 737 - 766, The build_chapter_prompt function accepts a target_words parameter but never forwards it to _build_prompt, so pass target_words through when calling self._build_prompt (e.g., add target_words=target_words to the _build_prompt keyword args); also verify the _build_prompt signature (and any downstream Prompt-building helpers) accepts and uses target_words so the chapter-length control is honored by build_chapter_prompt and its callers.application/engine/services/autopilot_daemon.py (1)
699-714:⚠️ Potential issue | 🟠 Major | ⚡ Quick winCompute the chapter target before the non-beat branch.
twis only assigned inside the beat-generation block, but this branch now uses it too. Ifcontext_builderis absent,use_wfcan still be true and this path will raiseUnboundLocalError. Move the target-word lookup out of the beat-only section and reuse it here.🛠️ Suggested fix
async def _handle_writing(self, novel: Novel): """处理写作（节拍级幂等落库）""" if not self._is_still_running(novel): return + tw = getattr(novel, "target_words_per_chapter", None) or 2500 + # 1. 目标控制：达到目标章节数则自动停止（允许用户设置更高的 max_auto_chapters 作为保护上限） target_chapters = novel.target_chapters or 50 @@ # 5. 节拍放大 beats = [] if self.context_builder: - tw = getattr(novel, "target_words_per_chapter", None) or 2500 beats = self.context_builder.magnify_outline_to_beats( chapter_num, outline, target_chapter_words=int(tw) ) @@ if use_wf: prompt = self.chapter_workflow.build_chapter_prompt( bundle["context"], outline, @@ voice_anchors=voice_anchors, ) cfg = GenerationConfig(max_tokens=int(tw * 1.5), temperature=0.85)🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@application/engine/services/autopilot_daemon.py` around lines 699 - 714, The code uses tw (chapter target words) only inside the beat-generation branch but later references it in the non-beat branch (causing UnboundLocalError if context_builder is absent); compute/assign tw (the chapter target lookup) before the branching so both beat and non-beat paths can use it. Specifically, move the logic that derives tw (the target word count) out of the beat-only block so it runs before the if use_wf: / else: branching that calls chapter_workflow.build_chapter_prompt and GenerationConfig(max_tokens=int(tw * 1.5)), ensuring functions/methods like context_builder, build_chapter_prompt, GenerationConfig, and _stream_llm_with_stop_watch still receive the same tw-derived value and preserving the existing _is_still_running(novel) check and logging.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@frontend/src/components/stats/StatsTopBar.vue`:
- Around line 75-79: The clickable div in StatsTopBar.vue (the element with
class "action-trigger" and `@click`="$emit('open-wizard')") is not
keyboard-accessible; add tabindex="0" and keyboard handlers so Enter and Space
also trigger the same action (either via `@keydown.enter` and `@keydown.space` that
call $emit('open-wizard') or by adding an onKeydown method that emits and
prevents default for Space), keep role="button" and aria-label, and ensure the
handler prevents default for Space to avoid page scrolling.
In `@frontend/src/composables/useWorkbench.ts`:
- Around line 74-75: The code uses || to set defaults for novelId.value and
targetChapters.value which can override valid falsy API values; update the
assignments in useWorkbench (novelId.value = novelData.id || slug and
targetChapters.value = novelData.target_chapters || 100) to use the nullish
coalescing operator (??) so they become novelId.value = novelData.id ?? slug and
targetChapters.value = novelData.target_chapters ?? 100, preserving legitimate
falsy values like 0 or "" from the DTO.
In `@infrastructure/ai/providers/openai_provider.py`:
- Around line 66-73: The code is incorrectly treating transient server errors as
permanent "Responses API unsupported" signals by adding base_url to
_fallback_to_chat_cache on InternalServerError and "500"/"Internal Server Error"
checks; update the exception handling in both generate() and stream_generate()
so that only protocol-mismatch / client errors (e.g., 400, 404, "Not Found",
"Account invalid", "INVALID_ARGUMENT") cause adding base_url to
self.__class__._fallback_to_chat_cache, and remove InternalServerError and any
5xx/"500" string checks from the cache-path so transient 5xx failures are left
to normal retry logic instead of being permanently cached as unsupported.
In `@interfaces/api/v1/core/settings.py`:
- Line 175: The update endpoint currently returns updated.to_dict() which omits
created_at/updated_at and breaks the existing response contract; change the
response in the update handler to include the timestamp fields (matching
get_embedding_config()) by merging or adding created_at and updated_at from the
updated model into the returned dict (use the same keys/format as
get_embedding_config()) so callers continue to receive the timestamps after
saving.
---
Outside diff comments:
In `@application/engine/services/autopilot_daemon.py`:
- Around line 699-714: The code uses tw (chapter target words) only inside the
beat-generation branch but later references it in the non-beat branch (causing
UnboundLocalError if context_builder is absent); compute/assign tw (the chapter
target lookup) before the branching so both beat and non-beat paths can use it.
Specifically, move the logic that derives tw (the target word count) out of the
beat-only block so it runs before the if use_wf: / else: branching that calls
chapter_workflow.build_chapter_prompt and GenerationConfig(max_tokens=int(tw *
1.5)), ensuring functions/methods like context_builder, build_chapter_prompt,
GenerationConfig, and _stream_llm_with_stop_watch still receive the same
tw-derived value and preserving the existing _is_still_running(novel) check and
logging.
In `@application/workflows/auto_novel_generation_workflow.py`:
- Around line 737-766: The build_chapter_prompt function accepts a target_words
parameter but never forwards it to _build_prompt, so pass target_words through
when calling self._build_prompt (e.g., add target_words=target_words to the
_build_prompt keyword args); also verify the _build_prompt signature (and any
downstream Prompt-building helpers) accepts and uses target_words so the
chapter-length control is honored by build_chapter_prompt and its callers.
In `@frontend/src/components/onboarding/NovelSetupGuide.vue`:
- Around line 954-975: After awaiting detectWizardProgress() ensure the wizard
is still open before proceeding: in both the watcher and the onMounted handler,
capture the current visibility (or increment a local generation token/nonce)
before the await and then check that props.show is still true (or the token
matches) after the await; only then set currentStep.value and call
triggerStepGeneration(step); also propagate the same guard to any async paths
started by resetWizardStateForOpen so stale callbacks cannot call
triggerStepGeneration when stopGenerationOnClose has run.
In `@infrastructure/ai/providers/openai_provider.py`:
- Around line 126-135: The Responses stream exception handlers currently fall
back to Chat Completions even if the Responses stream has already emitted
chunks; add a local flag (e.g., started_emitting = False) in the streaming
generator that you set to True immediately before yielding the first chunk, then
in the except blocks (the handlers that call
self.__class__._fallback_to_chat_cache.add and logger.info) check that flag and
only perform the fallback when started_emitting is False; if started_emitting is
True, re-raise the caught exception instead of swallowing it so callers do not
receive mixed/duplicated content. Ensure the flag is referenced in the same
method handling the Responses stream and use the original exception when
re-raising.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro Plus
Run ID: 5643d0b5-38c4-4699-b19c-1e7afc966d92
📒 Files selected for processing (9)
application/engine/services/autopilot_daemon.pyapplication/workflows/auto_novel_generation_workflow.pyfrontend/src/components/onboarding/NovelSetupGuide.vuefrontend/src/components/stats/StatsTopBar.vuefrontend/src/composables/useWorkbench.tsfrontend/src/views/Workbench.vueinfrastructure/ai/providers/openai_provider.pyinterfaces/api/v1/core/settings.pyinterfaces/api/v1/engine/generation.py
| <div class="action-trigger" @click="$emit('open-wizard')" role="button" aria-label="打开设置向导" title="新书设置向导"> | ||
| <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" width="18" height="18"> | ||
| <path fill="currentColor" d="M19 3H5c-1.1 0-2 .9-2 2v14c0 1.1.9 2 2 2h14c1.1 0 2-.9 2-2V5c0-1.1-.9-2-2-2zm-5 14H7v-2h7v2zm3-4H7v-2h10v2zm0-4H7V7h10v2z"/> | ||
| </svg> | ||
| </div> |
There was a problem hiding this comment.
Make the new wizard trigger keyboard-accessible.
The new interactive control is mouse-only right now. Keyboard users can’t reliably trigger it.
Suggested fix
- <div class="action-trigger" `@click`="$emit('open-wizard')" role="button" aria-label="打开设置向导" title="新书设置向导">
+ <div
+ class="action-trigger"
+ role="button"
+ tabindex="0"
+ aria-label="打开设置向导"
+ title="新书设置向导"
+ `@click`="$emit('open-wizard')"
+ `@keydown.enter.prevent`="$emit('open-wizard')"
+ `@keydown.space.prevent`="$emit('open-wizard')"
+ >
<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" width="18" height="18">
<path fill="currentColor" d="M19 3H5c-1.1 0-2 .9-2 2v14c0 1.1.9 2 2 2h14c1.1 0 2-.9 2-2V5c0-1.1-.9-2-2-2zm-5 14H7v-2h7v2zm3-4H7v-2h10v2zm0-4H7V7h10v2z"/>
</svg>
</div>+.action-trigger:focus-visible {
+ outline: 2px solid rgba(255, 255, 255, 0.55);
+ outline-offset: 2px;
+}🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@frontend/src/components/stats/StatsTopBar.vue` around lines 75 - 79, The
clickable div in StatsTopBar.vue (the element with class "action-trigger" and
`@click`="$emit('open-wizard')") is not keyboard-accessible; add tabindex="0" and
keyboard handlers so Enter and Space also trigger the same action (either via
`@keydown.enter` and `@keydown.space` that call $emit('open-wizard') or by adding an
onKeydown method that emits and prevents default for Space), keep role="button"
and aria-label, and ensure the handler prevents default for Space to avoid page
scrolling.
| novelId.value = novelData.id || slug | ||
| targetChapters.value = novelData.target_chapters || 100 |
There was a problem hiding this comment.
Use nullish fallback instead of || for DTO defaults.
|| can overwrite valid falsy values from API payloads. ?? is safer here.
Suggested fix
- novelId.value = novelData.id || slug
- targetChapters.value = novelData.target_chapters || 100
+ novelId.value = novelData.id ?? slug
+ targetChapters.value = novelData.target_chapters ?? 100📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
| novelId.value = novelData.id || slug | |
| targetChapters.value = novelData.target_chapters || 100 | |
| novelId.value = novelData.id ?? slug | |
| targetChapters.value = novelData.target_chapters ?? 100 |
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@frontend/src/composables/useWorkbench.ts` around lines 74 - 75, The code uses
|| to set defaults for novelId.value and targetChapters.value which can override
valid falsy API values; update the assignments in useWorkbench (novelId.value =
novelData.id || slug and targetChapters.value = novelData.target_chapters ||
100) to use the nullish coalescing operator (??) so they become novelId.value =
novelData.id ?? slug and targetChapters.value = novelData.target_chapters ??
100, preserving legitimate falsy values like 0 or "" from the DTO.
| except (openai.NotFoundError, openai.BadRequestError, openai.InternalServerError, RuntimeError) as e: | ||
| logger.info(f"Responses API unsupported for {base_url}, falling back to chat completions: {str(e)}") | ||
| self.__class__._fallback_to_chat_cache.add(base_url) | ||
| except Exception as e: | ||
| # 某些网关在路径错误时可能不抛严格的 404 而是抛出其他错误，如果消息含有明确路径错误也尝试降级 | ||
| if "404" in str(e) or "Not Found" in str(e) or "400" in str(e) or "Account invalid" in str(e) or "INVALID_ARGUMENT" in str(e): | ||
| if "404" in str(e) or "Not Found" in str(e) or "400" in str(e) or "500" in str(e) or "Internal Server Error" in str(e) or "Account invalid" in str(e) or "INVALID_ARGUMENT" in str(e): | ||
| logger.info(f"Gateway returned error for Responses API ({base_url}), falling back: {str(e)}") | ||
| self.__class__._fallback_to_chat_cache.add(base_url) |
There was a problem hiding this comment.
🧩 Analysis chain
🏁 Script executed:
# First, examine the file structure and the specific lines mentioned
wc -l infrastructure/ai/providers/openai_provider.pyRepository: shenminglinyi/PlotPilot
Length of output: 118
🏁 Script executed:
# Read the context around lines 66-73 and 126-135
sed -n '50,90p' infrastructure/ai/providers/openai_provider.pyRepository: shenminglinyi/PlotPilot
Length of output: 2193
🏁 Script executed:
# Get broader context to understand the class and _fallback_to_chat_cache usage
sed -n '1,65p' infrastructure/ai/providers/openai_provider.pyRepository: shenminglinyi/PlotPilot
Length of output: 2185
🏁 Script executed:
# Check how _fallback_to_chat_cache is initialized and used throughout the file
rg '_fallback_to_chat_cache' infrastructure/ai/providers/openai_provider.py -B 2 -A 2Repository: shenminglinyi/PlotPilot
Length of output: 2349
🏁 Script executed:
# Look at stream_generate method
sed -n '110,145p' infrastructure/ai/providers/openai_provider.pyRepository: shenminglinyi/PlotPilot
Length of output: 2183
🏁 Script executed:
# Search for retry logic in the callers mentioned in the scratchpad
rg -l "generate|stream_generate" application/ai/structured_json_pipeline.py application/workflows/auto_novel_generation_workflow.py application/engine/services/autopilot_daemon.py 2>/dev/null | head -5Repository: shenminglinyi/PlotPilot
Length of output: 214
🏁 Script executed:
# Check one of the callers to see how they handle retries
sed -n '1,150p' application/ai/structured_json_pipeline.py | head -80Repository: shenminglinyi/PlotPilot
Length of output: 2234
🏁 Script executed:
# Look for retry decorators or mechanisms
rg -i "retry|attempt|tries" infrastructure/ai/providers/openai_provider.pyRepository: shenminglinyi/PlotPilot
Length of output: 49
🏁 Script executed:
# Check if _fallback_to_chat_cache is ever cleared or reset
rg -i "clear|reset|_fallback_to_chat_cache\s*=" infrastructure/ai/providers/openai_provider.pyRepository: shenminglinyi/PlotPilot
Length of output: 49
🏁 Script executed:
# Examine how generate is called by checking if any retry wrapper is used
grep -n "await.*generate\|\.generate(" application/ai/structured_json_pipeline.py | head -20Repository: shenminglinyi/PlotPilot
Length of output: 189
🏁 Script executed:
# Look at the broader error handling in OpenAIProvider to see if there's recovery
sed -n '1,350p' infrastructure/ai/providers/openai_provider.py | tail -100Repository: shenminglinyi/PlotPilot
Length of output: 3800
🏁 Script executed:
# Check if there's any process-lifetime reset or TTL mechanism for _fallback_to_chat_cache
rg -i "fallback.*cache|cache.*reset|process.*lifetime" infrastructure/ai/providers/ --type pyRepository: shenminglinyi/PlotPilot
Length of output: 958
🏁 Script executed:
# Verify LLM_MAX_TOTAL_ATTEMPTS and retry policy
cat application/ai/llm_retry_policy.py 2>/dev/null || rg "LLM_MAX_TOTAL_ATTEMPTS" application/ai/Repository: shenminglinyi/PlotPilot
Length of output: 171
🏁 Script executed:
# Check if OpenAIProvider is instantiated per-request or singleton
rg "OpenAIProvider\(" application/ infrastructure/ --type py | head -20Repository: shenminglinyi/PlotPilot
Length of output: 228
Don't cache transient 5xx errors as permanent "unsupported API" signals.
InternalServerError and "500" / "Internal Server Error" string checks represent transient server faults, not proof that Responses API is unsupported. Adding these errors to _fallback_to_chat_cache permanently disables Responses API for that base_url for the process lifetime with no recovery mechanism. This bypasses the retryable error detection in callers (e.g., structured_json_pipeline.py), which expect 5xx failures to be retried. Keep the cache limited to protocol-mismatch errors (404, 400, invalid account). The same issue exists in both generate() and stream_generate() methods; for streaming, it also risks splicing partial content from one API with a different API's continuation.
Suggested adjustment
- except (openai.NotFoundError, openai.BadRequestError, openai.InternalServerError, RuntimeError) as e:
+ except (openai.NotFoundError, openai.BadRequestError, RuntimeError) as e:
logger.info(f"Responses API unsupported for {base_url}, falling back to chat completions: {str(e)}")
self.__class__._fallback_to_chat_cache.add(base_url)
except Exception as e:
- if "404" in str(e) or "Not Found" in str(e) or "400" in str(e) or "500" in str(e) or "Internal Server Error" in str(e) or "Account invalid" in str(e) or "INVALID_ARGUMENT" in str(e):
+ if "404" in str(e) or "Not Found" in str(e) or "400" in str(e) or "Account invalid" in str(e) or "INVALID_ARGUMENT" in str(e):
logger.info(f"Gateway returned error for Responses API ({base_url}), falling back: {str(e)}")
self.__class__._fallback_to_chat_cache.add(base_url)Apply the same change to stream_generate() at lines 126–135.
🧰 Tools
🪛 Ruff (0.15.12)
[warning] 67-67: Use explicit conversion flag
Replace with conversion flag
(RUF010)
[warning] 70-70: Comment contains ambiguous ， (FULLWIDTH COMMA). Did you mean , (COMMA)?
(RUF003)
[warning] 72-72: Use explicit conversion flag
Replace with conversion flag
(RUF010)
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@infrastructure/ai/providers/openai_provider.py` around lines 66 - 73, The
code is incorrectly treating transient server errors as permanent "Responses API
unsupported" signals by adding base_url to _fallback_to_chat_cache on
InternalServerError and "500"/"Internal Server Error" checks; update the
exception handling in both generate() and stream_generate() so that only
protocol-mismatch / client errors (e.g., 400, 404, "Not Found", "Account
invalid", "INVALID_ARGUMENT") cause adding base_url to
self.__class__._fallback_to_chat_cache, and remove InternalServerError and any
5xx/"500" string checks from the cache-path so transient 5xx failures are left
to normal retry logic instead of being permanently cached as unsupported.
| model_path=body.model_path, | ||
| ) | ||
| return updated.to_api_dict() | ||
| return updated.to_dict() |
There was a problem hiding this comment.
Preserve the update response contract.
updated.to_dict() drops created_at / updated_at, so this endpoint no longer matches get_embedding_config() or the prior API payload. That’s a breaking change for any caller that reads the timestamps after saving.
Suggested fix
- return updated.to_dict()
+ return updated.to_api_dict()📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
| return updated.to_dict() | |
| return updated.to_api_dict() |
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@interfaces/api/v1/core/settings.py` at line 175, The update endpoint
currently returns updated.to_dict() which omits created_at/updated_at and breaks
the existing response contract; change the response in the update handler to
include the timestamp fields (matching get_embedding_config()) by merging or
adding created_at and updated_at from the updated model into the returned dict
(use the same keys/format as get_embedding_config()) so callers continue to
receive the timestamps after saving.
…fy_outline_to_beats return value to mock_context_builder - Fix fingerprint test to handle keyword args from state_extractor
There was a problem hiding this comment.
🧹 Nitpick comments (1)
tests/unit/application/workflows/test_auto_novel_generation_workflow.py (1)
37-37: Don't stubmagnify_outline_to_beatsglobally in this shared fixture.This forces every workflow test in this module down the single-segment path, and there are no dedicated tests elsewhere to cover the beat-generation and target-word plumbing introduced by this PR. Keep the default fixture closer to real behavior and stub the beat path only in tests that intentionally need it, or add a dedicated beat-path test.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@tests/unit/application/workflows/test_auto_novel_generation_workflow.py` at line 37, The shared test fixture currently forces every test down the single-segment path by stubbing builder.magnify_outline_to_beats.return_value = [], so remove that global stub from the shared fixture and let magnify_outline_to_beats run its real behavior by default; for tests that need the single-segment path, set builder.magnify_outline_to_beats.return_value = [] locally inside those test functions (or use a per-test monkeypatch), and add at least one dedicated test that exercises the beat-generation path to validate beat creation and target-word plumbing (asserting behaviors produced by magnify_outline_to_beats and consumed by the workflow).
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Nitpick comments:
In `@tests/unit/application/workflows/test_auto_novel_generation_workflow.py`:
- Line 37: The shared test fixture currently forces every test down the
single-segment path by stubbing builder.magnify_outline_to_beats.return_value =
[], so remove that global stub from the shared fixture and let
magnify_outline_to_beats run its real behavior by default; for tests that need
the single-segment path, set builder.magnify_outline_to_beats.return_value = []
locally inside those test functions (or use a per-test monkeypatch), and add at
least one dedicated test that exercises the beat-generation path to validate
beat creation and target-word plumbing (asserting behaviors produced by
magnify_outline_to_beats and consumed by the workflow).
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro Plus
Run ID: 01f688d7-0229-4e12-86a5-458a6ea137a8
📒 Files selected for processing (1)
tests/unit/application/workflows/test_auto_novel_generation_workflow.py
关于 Backend CI 测试失败的说明经过排查，本次 PR 没有引入任何测试失败。CI 中 证据1. 本 PR 涉及的文件，对应测试全部通过
2. 预存失败的 42 个测试（与 PR 无关）在本地执行与 CI 相同命令（ 预存失败的测试分布： 3. 为什么会触发？主仓库的 Backend CI 触发条件是 PR 中修改了 on:
pull_request:
paths:
- "**.py"近期没有修改 Python 文件的 PR，所以没触发 CI。本次 PR 开启了 CI 才发现主仓库存在预存测试失败。 建议
|
1 participant
概述
本次 PR 包含 3 个后端 BUG 修复 和 2 项前端向导功能增强，共涉及 9 个文件。
一、BUG 修复
BUG-1: 嵌入配置更新接口崩溃
文件:
interfaces/api/v1/core/settings.py问题:
update_embedding_config端点调用updated.to_api_dict()，但EmbeddingConfigModel只有to_dict()方法，导致运行时AttributeError崩溃。修复: 将
to_api_dict()改为to_dict()BUG-2: OpenAI Responses API 500 错误未兜底
文件:
infrastructure/ai/providers/openai_provider.py问题: OpenAI 兼容接口返回 500
InternalServerError时，无法触发 Responses API → Chat Completions 的降级，直接抛出异常。修复: 在
generate()和stream_generate()的异常捕获中新增openai.InternalServerError，并在字符串匹配中增加"500"/"Internal Server Error"的判断。BUG-3: 章节生成字数不受控
文件:
application/workflows/auto_novel_generation_workflow.pyinterfaces/api/v1/engine/generation.pyapplication/engine/services/autopilot_daemon.py问题: 用户在创建小说时设置的"每章目标字数" (
target_words_per_chapter) 未传递给章节生成流程，LLM 始终按硬编码的 3000-4000 字生成。修复:
target_words_per_chapter，传递给 workflowmax_tokens根据目标字数动态计算二、前端功能增强
功能 1: 向导恢复入口
文件:
frontend/src/composables/useWorkbench.tsfrontend/src/views/Workbench.vuefrontend/src/components/stats/StatsTopBar.vue问题: 新书设置向导在刷新页面后消失，无法恢复。用户只能进入工作台手动补全未完成的设定（地点、故事线等）。
新增:
StatsTopBar工具栏添加「设置向导」按钮NovelSetupGuide组件，自动检测已有进度并跳到对应步骤功能 2: 向导超时重试
文件:
frontend/src/components/onboarding/NovelSetupGuide.vue问题: LLM 生成较慢时前端轮询超时，仅显示错误提示，无重试入口。即使后台实际已完成生成，用户也无法获取数据。
新增:
影响范围
所有改动均为增量修复，不影响现有功能。
Summary by CodeRabbit
New Features
Improvements