AppPulse review pipeline now uses AI code analysis

AppPulse classifies daily app-store reviews with an LLM and runs an AI-driven code analysis that returns a validated AnalysisBrief Pydantic model for each bug [Dev.to]. The initial implementation used a PydanticAI agent with four read-only tools, costing roughly 40-50 tool calls per bug. By generating a repo map upfront and feeding it into the system prompt, the agent cut tool calls by about 60%, dropping the average to 8-12 calls without measurable quality loss [Dev.to]. The pipeline was refactored into a strategy pattern, allowing the analysis backend to be swapped at runtime. The built-in PydanticAI backend stays lightweight, while external agents - Grok Build, Claude Code, and OpenAI Codex - run headless CLI commands, explore the codebase with full grep access, and emit raw markdown. A cheap gpt-4o-mini pass then extracts the markdown into the same AnalysisBrief schema, keeping downstream code unchanged [Dev.to]. Cache-based repo-map fingerprinting makes map regeneration 100-500 ms for a 500-file project, with cache hits at ~0 ms. The author reports a 5x speedup in debugging time, with a batch of five critical bugs now yielding actionable code locations in under two minutes per bug [Dev.to]. The structured briefs enable automation, as CI pipelines can ingest the affected_files and proposed_changes sections to auto-generate PRs or test scaffolding. The backend-agnostic design guards against provider lock-in, allowing for easy swapping between AI vendors with only a config change.