No Context Flooding
FaultLocalizer parses tracebacks, traverses a bidirectional Python AST call graph, and scores lexical overlap with BM25—cutting prompt size by 96.6% at 100.0% Top-1 recall.
Cut traceback context tokens by 96.6%, catch test-cheating AST diffs in 0.101 ms, and verify fixes inside a byte-exact SHA-256 rollback sandbox.
pip install -e ".[dev]" && patchpilot demo
FaultLocalizer parses tracebacks, traverses a bidirectional Python AST call graph, and scores lexical overlap with BM25—cutting prompt size by 96.6% at 100.0% Top-1 recall.
PatchIntegrityGuard audits AST pairs across 8 deterministic rules (RH001–RH008) to block skipped tests, assert True hacks, and eval/shell=True primitives with 0.0% FPR.
TransactionalPatchWorkspace snapshots files before touching disk, runs verification tests, restores exact pre-state bytes on any failure, and signs a SHA-256 RepairReceipt.
Ship 5 deterministic tools, 2 resources, and 2 prompts over stdio, local HTTP/SSE, or our stateless Vercel gateway at patchpilot-eta.vercel.app/mcp.
Pair PatchPilot AST receipts directly with Nexus Aegis Merkle attestation bundles and zero-risk CI merge gates.
Try a candidate diff in the Interactive Studio or inspect CLI & MCP setup.
Autonomous coding agents often make CI pass by deleting failing assertions, decorating tests with @pytest.mark.skip, wrapping errors in except Exception: pass, or editing test helpers. CI reports green even though the bug remains. PatchIntegrityGuard rejects those diffs at AST parse time in 0.101 ms p50.
A bug inside a 1-hop callee often surfaces in a caller frame. FaultLocalizer weights traceback frames alongside bidirectional AST call-graph edges and BM25 error tokens, packing only the minimal line-numbered function slices into your target budget.
No. For safety, api/index.py puts apply_transactional_patch in HOSTED_BLOCKED_TOOLS and rejects workspace_root arguments on public HTTP requests. Run patchpilot serve-stdio locally for snapshot mutations.