Self-Maintaining Stripe API

Inspired by Y Combinator's Request for Startups theme on self-sustaining APIs — the idea that API providers shouldn't just announce breaking changes, they should apply the fixes. Built a 5-agent pipeline (Planning, Retriever, Synthesizer, Critic/Verifier, Orchestrator) that watches Stripe's API changelog, finds every affected call site in a codebase, drafts a fix, verifies it by running the codebase's real test suite in a sandbox, and opens a human-reviewed pull request.

Technical highlights

• 01

  Architected a 5-agent pipeline (Planning, Retriever, Synthesizer, Critic/Verifier, Orchestrator) that autonomously detects Stripe API changes and generates fix PRs, integrating the Claude API for structured LLM reasoning (forced tool-use calls for classification, code synthesis, and verification), ChromaDB for semantic code retrieval, SQLite for relational state, Python AST analysis for call-site extraction behind a pluggable language-adapter interface, and Docker + GitHub CLI for sandboxed test verification and automated PR delivery.

• 02

  Diagnosed and resolved two failure classes that crashed the pipeline on rerun: non-idempotent re-indexing that silently duplicated call-site records (fixed by clearing state before each index pass) and a fragile LLM-reproduced-text patching strategy that broke on whitespace drift (redesigned to anchor every patch on byte-exact AST source spans via ast.get_source_segment, eliminating the failure mode instead of tolerating it).

• 03

  Validated end-to-end against a synthetic repo modeling a real historical Stripe deprecation (the 2020 Plans-to-Prices migration): reached 100% precision on affected-usage detection after calibrating a semantic-search distance threshold that removed a false-positive match, and verified fixes via real sandboxed pytest execution, correctly distinguishing 2 stale-test failures from actual regressions before opening a fully automated PR.