Backend: thread on_fallback callback through analysis chain
(gemini_service → agents → analysis_service → handlers). The handler
sends a 'model_fallback' WebSocket message exactly once per analysis
when the primary model is unavailable.
Frontend: handle 'model_fallback' WS message and show a dismissible
yellow toast at the bottom of the screen with an 8-second auto-dismiss.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
google-genai SDK expects http_options 'timeout' in milliseconds.
Passing 45 (seconds) was interpreted as 45ms → ~1s deadline,
which Google API rejected with 400 INVALID_ARGUMENT
'Manually set deadline 1s is too short. Minimum allowed deadline is 10s.'
Primary: 45_000ms (45s), Fallback: 150_000ms (150s)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
asyncio.wait_for cannot reliably cancel SDK-internal HTTP connections.
Replace with two genai.Client instances — one per model — each configured
with http_options={'timeout': N} so the TCP connection is actually torn
down when the deadline is reached.
Primary model: 45s, Fallback model: 150s
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Log analysis showed fallback model responses up to 154s under parallel
load. 60s was too aggressive and would cause false timeouts.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Primary model (gemini-3.1-pro-preview): 45s timeout
Fallback model (gemini-3-flash-preview): 60s timeout
Without timeouts, the fallback model under high load would wait
indefinitely, causing analysis to hang for 10+ minutes per file.
asyncio.TimeoutError from the primary model is now handled the same
as other exceptions (falls through to fallback).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
gemini_service.py: if the primary model (gemini-3.1-pro-preview) is
unavailable or returns a permission error, all three call sites now
automatically retry with gemini-3-flash-preview before propagating failure.
cloudrun.yaml: new Cloud Run service definition that ensures stable
WebSocket operation — 10-minute request timeout (vs 60s default),
2 vCPU / 4Gi RAM for PDF rasterisation, min 1 warm instance to prevent
cold-start disconnects, and GEMINI_API_KEY sourced from Secret Manager
so the service can actually reach the Gemini API.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replace single-line bullet format with a structured two-part format
(**Issue:** / **Recommendation:**) in all specialist and lead agent
prompts. Update Gemini response schema description to match. Update
frontend formatFeedbackText and formatFeedbackTextForPDF to parse
**bold** markdown and preserve line breaks within multi-line bullets.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When a subsequent revision of a proof is uploaded, the analysis now takes
place in context of the previous version's results. The system identifies:
- Resolved issues: fixed in the new revision
- Outstanding issues: still present from previous version
- New issues: introduced in the new revision
Key changes:
- Add resolvedIssues, outstandingIssues, newIssues fields to SubReview
- Add PreviousReviewContext model for passing previous review data
- Update all specialist agents to accept previous_review context
- Extend GeminiService with include_revision_fields parameter
- Add get_latest_version_review() repository method
- Update LeadAgent to synthesize cross-version context in summary
- Fetch previous analysis in WebSocket handler for revisions
First version analysis continues to work exactly as before with revision
fields set to null.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Changed the AI model used for proof analysis from gemini-2.5-flash
to gemini-3-pro-preview for improved analysis capabilities.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add explicit formatting instructions to agent prompts requesting bullet-point
output instead of verbose paragraphs. Update JSON schema descriptions for
feedback and summary fields to enforce concise, outline-style format.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
PDFs are now converted to PNG images at 200 DPI before being sent to
Gemini for analysis. This fixes the unreliable iframe-based PDF preview
and ensures all pages are properly analyzed.
- Add PyMuPDF dependency for PDF rasterization
- Create pdf_service.py with rasterize() and get_page_count()
- Update agent interfaces to accept list of images for multi-page support
- Add analyze_with_images() to Gemini service for multi-image analysis
- Return rasterized PDF pages via WebSocket for frontend display
- Add page navigation UI for multi-page PDFs in preview components
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
