Root cause: LLM receiving full 34k-char JRXML would regenerate from scratch
instead of modifying coordinates in-place, shrinking output to ~3k chars.
Solution (programmatic node control, not prompt engineering):
- New agent/jrxml_windower.py: decompose JRXML into header (never sent to
LLM) + individual bands. Split bands >4000 chars at element boundaries.
Reassemble with element count validation (>10% change = rollback).
- Rewrite refine_layout: per-band windowed LLM processing (~2-4k chars
each). LLM cannot "reimagine" the entire report.
- Rewrite map_fields: 100% programmatic regex $F{field_N} -> real name
replacement. Zero LLM calls, zero content loss.
- _sanitize_field_name: non-ASCII chars escaped to _uXXXX_ format for
valid JRXML identifiers.
- Tests: 48 new unit tests (windower 28 + map_fields 20). All passing.
Full suite 385 tests, zero regressions.
- process_input sets _failure_recovery flag when injecting pending_failure_context
- classify_intent skips LLM classification when flag is set, directly routes to modify_jrxml
- Smart truncation for intent classify: keep head 200 + tail 300 chars instead of head 500
(prevents user's actual message from being truncated away by long injected context)
- This fixes the bug where "retry" or pasted error messages were misclassified as
consult_question or initial_generation after max retry exhaustion
Root cause: layout_schema.regions is a list of region dicts, not a dict.
_log_ocr_layers() was calling .keys() on it, causing agent_error.
Also fixed: ProcessSection now stays visible after streaming ends (error or
completion), so generated content is not lost. Header shows ✓/✕/pulse indicators.
Error handler now refreshes session state for partial JRXML download.
3-phase pipeline to solve LLM prompt overflow from too many OCR elements:
Phase 1 (generate_skeleton): compressed layout schema → skeleton JRXML
Phase 2 (refine_layout): sampled coordinates → pixel-level position tuning
Phase 3 (map_fields): OCR field names → replace $F{field_N} placeholders
Only triggered when layout_schema.total_rows > 0 on initial_generation intent.
Text requests and all other intents are unaffected (zero behavior change).
- OCR: EasyOCR (primary, ch_sim+en) with PaddleOCR fallback for Windows compatibility
- Validation: _check_minimum_content() rejects empty-shell JRXML (no band/textField)
- Retry: MAX_RETRY 3→5, exhaustion records pending_failure_context for next-turn auto-injection
- Finalize: only saves jrxml_versions on pass, preserves last good final_jrxml on fail
- Extract JRXML: improved empty markdown block handling and XML fragment fallback
- UI: real-time node progress via placeholder updates, initial "analyzing" feedback
- UI: use agent_state (full) instead of node_state (partial) for summary card routing
- UI: unknown template_type now gives LLM meaningful image context instead of metadata
- Docs: updated CLAUDE.md and CODE_GUIDE.md to reflect all v3 changes
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Add rag submodule for semantic JRXML chunk retrieval, refactor
retrieve node to use RAGSearcher, and fix missing api_key in
Anthropic SDK client initialization.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
panda