TAPES (Transition Ai Patches Enforcement Systems)

Most AI coding agents are just LLMs with write access. They hallucinate symbols, overwrite protected files, and flood your codebase with untested patches. TAPES fixes that. TAPES (Transition Ai Patches Enforcement Systems) wraps IBM BobShell — Granite-13b-chat-v2 via watsonx.ai — in a five-stage governed execution pipeline that transforms reckless code generation into controlled, auditable engineering. The pipeline starts before Bob sees a single line of code. An AST-powered Intent Router computes a centrality-weighted token budget and extracts scope directives from the user's intent. A frozen Contract locks in the mutation boundary, constraints, and success criteria. Bob only ever sees hollow skeletons of irrelevant files and the full source of the exact target symbol. That single design decision cuts token consumption by 81%. Bob's patches hit a five-gate Bouncer before touching the filesystem: syntax validation, path boundary enforcement, AST anchor verification, exact-match uniqueness, and size integrity. Protected files are never hard-blocked — patches targeting them redirect silently to sidecar extension files, preserving intent without violating protection invariants. Surviving patches enter a dual-tier sandbox. Tier 1 runs targeted tests against a fast Python-only copy in milliseconds. Tier 2 clones the full virtual environment for final isolation before atomic commit. Failures are classified as transient, structural, or semantic — each carrying a different retry penalty — and prior failure tracebacks are injected directly into Bob's next prompt. Across 8 adversarial benchmarks — protected file massacres, mutation floods, path traversal attacks, hallucinated symbols — raw BobShell violated protected files 8/8 times and cost $7.29. TAPES delivered 0 violations, $1.34 total, 26 Bouncer intercepts, and 6 legitimate patches committed cleanly. Bob writes the code. TAPES decides if it deserves to exist.