cb3aa48999
fix: Add wrapper function for workflow LLM calls to prevent max_tokens signature mismatch
...
Deploy Development / deploy (push) Successful in 49s
Build Test / pytest-backend (push) Successful in 4s
Build Test / lint-backend (push) Successful in 0s
Build Test / build-frontend (push) Successful in 16s
Problem: workflow_executor calls openrouter_call_func(prompt, model) but
call_openrouter expects (prompt, max_tokens=4096). This caused the model string
'anthropic/claude-sonnet-4' to be passed as max_tokens, resulting in OpenRouter
requesting 64000 tokens and failing with 402 credit errors.
Solution: Added workflow_llm_call() wrapper in workflows.py that matches the
expected (prompt, model) -> str signature and calls call_openrouter correctly.
Fixes: All workflows failing with 402 'insufficient credits' errors
2026-04-12 13:37:31 +02:00
1f8791f4dd
feat: Phase 2 - Normalisierung + Workflow Executor
...
Deploy Development / deploy (push) Successful in 44s
Build Test / lint-backend (push) Successful in 0s
Build Test / build-frontend (push) Successful in 14s
Backend:
- normalization_engine.py (200 Zeilen): Synonym-Mapping, 5 Statuswerte
* normalize_decision_signal(): Kaskade (exact → case → synonym → invalid)
* apply_synonym_mapping(): DB-basierte Synonyme (case-insensitive)
* normalize_all_signals(): Batch-Processing gegen Katalog
* load_question_catalog(): Lädt normalization_rules aus DB
- workflow_executor.py (440 Zeilen): Sequenzielle Workflow-Ausführung
* execute_workflow(): Traversiert DAG in topologischer Reihenfolge
* execute_node(): Führt analysis nodes aus (start/end = no-op)
* aggregate_results(): Kombiniert analysis_core + normalized_signals
* save_execution_state(): Persistiert in workflow_executions
- workflow_models.py: Erweitert um Phase 2 Models
* SignalStatus Enum (valid, normalized, unclear, invalid, not_decidable)
* NormalizedSignal (question_type, raw_value, normalized_value, status)
* NodeExecutionState (node_id, status, analysis_core, normalized_signals)
* ExecutionResult (execution_id, workflow_id, status, node_states, aggregated_result)
- workflow_engine.py: Neue Funktion get_execution_order()
* Flattened topological sort für sequenzielle Execution
* Phase 7: Wird zu levels (parallele Execution)
- prompt_executor.py: execute_workflow_prompt() Implementierung
* Ruft workflow_executor.execute_workflow() auf
* Konvertiert ExecutionResult zu API-Response
- routers/workflows.py (230 Zeilen): Workflow Execution API
* POST /api/workflows/{id}/execute (mit enable_debug)
* GET /api/workflows/executions/{id} (lädt gespeicherten State)
* GET /api/workflows (listet alle aktiven Workflows)
* GET /api/workflows/{id} (lädt einzelnen Workflow mit Graph)
- main.py: Router-Registrierung (workflows.router)
Tests:
- test_phase2_normalization.py (17 Tests): Alle Normalisierungs-Szenarien
* Exact match, case-insensitive, synonym mapping, invalid, whitespace
* Batch-Normalisierung, not_in_catalog, mixed validity
- test_phase2_workflow_executor.py (10 Tests): Executor + Aggregation
* aggregate_results mit verschiedenen Konstellationen
* execute_node für start/end/analysis/unknown
* Integration mit question_augmenter + result_container_parser
Alle 27 Unit-Tests bestanden.
version: 0.9k (backend)
module: workflow 0.3.0
Konzept: .claude/task/Workflow_engine_prompting_engine/anforderungsanalyse_umsetzungsplan.md (Phase 2)
2026-04-03 21:20:23 +02:00
