QuAIA Files

New features:

• Test Case Design agent (port 8005) replaces the orchestrator's generate → classify → review chain: it generates, reviews and fixes a user story's test cases in a loop, runs a one-time duplicate check on the final set, then saves them to Zephyr, classifies them and publishes the review (generation, review and classification agents now run in-process inside it)
• Structured review findings (owner, action, severity, category, AC reference) with a shared severity rubric and a new whole-set test suite review
• Requirements Review agent runs parallel focused reviews (REQUIREMENTS_REVIEW_FOCUS_AREA_COUNT) merged into one feedback
• Document RAG sync for Confluence and SharePoint with document grounding of requirements reviews, sync locks and outcome reporting
• Multi-provider model factory (Gemini, Claude, Qwen and others) and prompt overrides
• Per-agent thinking levels via <AGENT>_THINKING_LEVEL (none disables thinking) on every provider; configurable Qwen presence penalties and QWEN_CONTEXT_WINDOW
• Context compaction for every A2A agent loop (tiered, zero-LLM) instead of failing when the context window fills
• Routing split into three output tools (AgentSelected, SuitableAgentBusy, NoSuitableAgent) with all fields required
• Deployment rework: deploy/manifest.yaml with redeploy gate, Qwen settings and thinking levels deployable through the pipeline

Improvements:

• Transport retries honour the provider's RetryInfo delay (capped by PROVIDER_RETRY_DELAY_CAP_SECONDS)
• PDF extraction moved from PyMuPDF to pypdfium2 with pixel caps, image-bomb checks and de-duplicated OCR
• Hermetic smoke suite with a scripted LLM mock (no API key, runs in CI on every push and PR) and a separate manual A/B suite against per-workflow baselines
• CALM model extended with credential-scope and job-invocation controls and the new Test Case Design agent node
• Prompts migrated to Markdown templates; new PYTHON_GUIDELINES.md and reworked agent skills
• NOTICE file listing third-party components and licenses

Bug fixes:

• CustomLlmWrapper merges default and provided model settings key by key instead of dropping the defaults
• Qwen model settings (thinking defaults and sampling) fixed
• Requirements review scope parameters over 400 characters rejected instead of silently truncated

Breaking changes:

• Orchestrator 3.0.0: routing output model and the contract of a routing_instruction_template.md prompt override changed
• test-case-generation, test-case-review and test-case-classification services are no longer deployed separately; deploy test-case-design-agent instead
• QWEN_THINKING_ENABLED replaced by per-agent <AGENT>_THINKING_LEVEL
• See Migration Notes / Breaking Changes in the README

Dependencies:

• pydantic-ai 2.x, httpx2, all versions pinned
• Added pydantic-ai-harness 0.36.0; pymupdf replaced by pypdfium2 5.13.0
• PyJWT 2.14.0 -> 2.15.1 (PYSEC-2026-4141 / GHSA-42vr-xj54-vc7v)
• oauthlib pinned to 4.0.0 (GHSA-hj66-6f7g-4r5v, GHSA-xpv3-w29h-x7cv)

Versions:

• Orchestrator 3.0.0, Requirements Review 1.5.0, Test Case Design 1.2.0 (new), Test Case Classification 1.5.0, Incident Creation 1.4.0