Conversation
- Added new internal API endpoint for documentation tools, allowing actions such as listing available docs, searching, and fetching specific documentation by ID. - Updated environment configuration to support optional internal secret for enhanced security. - Refactored existing search functionality to utilize the new docs tools API instead of the previous MCP server. - Improved error handling and response parsing for documentation-related requests. - Expanded documentation to clarify the relationship between the new tools and existing API functionalities. This update streamlines the documentation access process and enhances the overall developer experience.
- Introduced error capturing for failed HTTP requests in the docs tools API, improving debugging capabilities. - Updated the API response for unsupported methods to include an 'Allow' header, clarifying the expected request type. These changes enhance the robustness of the documentation tools integration and improve developer experience.
- Updated the key name in the capabilities section of the API documentation to follow a consistent naming convention, improving clarity and maintainability.
The .gitmodules was updated in d22593d to point at apps/backend/src/private/implementation, but the gitlink entry (mode 160000) was never added to the tree. This caused `git clone --recurse-submodules` to silently skip the private submodule. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…on for docs tools - Added `STACK_DOCS_INTERNAL_BASE_URL` to backend `.env` and `.env.development` files for AI tool bundle configuration. - Removed references to `STACK_INTERNAL_DOCS_TOOLS_SECRET` from backend and docs environment files and validation logic from the docs tools API route. - Introduced a new `.env` file for the docs app with essential configuration variables.
… into llm-mcp-flow
|
@greptile-ai review |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
| const rawCost = extractCostFromUsage(usage); | ||
| runAsynchronouslyAndWaitUntil(logAiQuery({ | ||
| ...common, | ||
| stepsJson: serializeSteps(steps), |
There was a problem hiding this comment.
BUG: logging-only serialization can break a successful AI response.
Walkthrough: the AI request can finish successfully, then logAiQuerySuccess tries to prepare the log row. Because serializeSteps(steps) is evaluated before the fire-and-forget logging promise is handed to runAsynchronouslyAndWaitUntil, any JSON.stringify failure here throws synchronously. I reproduced this with a tool result containing a circular object: the answer path is otherwise successful, but logging throws circular before the async boundary.
Expected behavior: logging should be isolated from the user-facing success path. Build/serialize the logging payload inside the async logging task, or catch serialization failures and log a safe placeholder.
| const editor = user.display_name ?? user.primary_email ?? user.id; | ||
| const qaId = BigInt(body.qaId); | ||
| await callReducerStrict("update_qa_entry", [token, qaId, body.question, body.answer, editor]); | ||
| await callReducerStrict("set_qa_published", [token, qaId, body.publish]); |
There was a problem hiding this comment.
BUG: this update is not atomic.
Walkthrough: editing text and changing publish state are two separate reducer calls. I reproduced the failure by mocking update_qa_entry to succeed and set_qa_published to fail. The route returns an error, but the question/answer mutation has already been committed.
Expected behavior: use a single reducer for “update text plus publish state” so the whole operation either commits or fails together. Partial review edits are hard for reviewers to reason about because the UI reports failure while the data has changed.
| onSave(row.correlationId, question, answer, publish); | ||
| Promise.resolve(onSave(row.id, question, answer, publish)) | ||
| .catch(err => captureError("knowledge-base-save", err)); | ||
| setEditingId(null); |
There was a problem hiding this comment.
BUG: failed saves look successful to the reviewer.
Walkthrough: the save starts Promise.resolve(onSave(...)).catch(...), then immediately closes the editor with setEditingId(null). I reproduced this from the component source: a rejected save is only sent to captureError, while the reviewer sees the editor disappear as if the save worked.
Expected behavior: await the save/delete action, keep the editor/dialog open on failure, and show an inline error. Capturing the error is useful for us, but it is not enough feedback for the reviewer performing the action.
| { name: "mark_human_reviewed", args: [wrong, "corr", "reviewer"] }, | ||
| { name: "unmark_human_reviewed", args: [wrong, "corr"] }, | ||
| { | ||
| name: "update_human_correction", |
There was a problem hiding this comment.
BUG: this auth test still references a removed reducer.
Walkthrough: the current correction path uses the newer QA reducers, but this test still calls update_human_correction. I reproduced this statically by checking the test source still contains that reducer name. Against the real module this either fails for the wrong reason or no longer tests the current write surface.
Expected behavior: replace this with the current mutating reducer(s), especially the correction/upsert path that actually ships now.
| args: [wrong, "corr", "q", "a", false, "reviewer"], | ||
| }, | ||
| { name: "add_manual_qa", args: [wrong, "q", "a", false, "reviewer"] }, | ||
| { name: "delete_qa_entry", args: [wrong, "corr"] }, |
There was a problem hiding this comment.
BUG: this does not prove delete_qa_entry enforces auth.
Walkthrough: delete_qa_entry expects a numeric/u64 QA id, but the test passes "corr". That can fail while decoding the reducer arguments before the token check is reached. So a failing reducer call here does not necessarily mean the wrong token was rejected.
Expected behavior: pass a valid numeric QA id shape with the wrong token, then assert the failure message is the invalid-token path.
…gative decimal integer
… and maintainability
… updates with publish state
…nd improve async task management
…tasks and improve serialization robustness
…Dashboard and updateDashboard based on context
… it is a valid integer and within bounds
…r improved clarity and maintainability
aadesh18
left a comment
aadesh18
left a comment
There was a problem hiding this comment.
made changes based on comments
| const editor = user.display_name ?? user.primary_email ?? user.id; | ||
| const qaId = BigInt(body.qaId); | ||
| await callReducerStrict("update_qa_entry", [token, qaId, body.question, body.answer, editor]); | ||
| await callReducerStrict("set_qa_published", [token, qaId, body.publish]); |
| const rawCost = extractCostFromUsage(usage); | ||
| runAsynchronouslyAndWaitUntil(logAiQuery({ | ||
| ...common, | ||
| stepsJson: serializeSteps(steps), |
|
|
||
| let chosenIndex: number; | ||
| if (edit.occurrenceIndex != null) { | ||
| if (edit.occurrenceIndex < 0 || edit.occurrenceIndex >= matches.length) { |
This PR adds an AI analytics dashboard to the internal tool and centralises AI query observability. Every call through the /ai/query and /integrations/ai-proxy endpoints is now logged to SpacetimeDB with token counts, cost, latency, and full conversation replay. The internal tool gains a new "Unified AI Endpoint Analytics" tab backed by a new useAiQueryLogs SpacetimeDB subscription, plus reviewer enrollment and unmark-reviewed flows.
How to review this PR:
Summary by CodeRabbit
New Features
Improvements
Infrastructure