DriftZero

DriftZero migrates a codebase across a breaking API change and proves the migration before opening a pull request. The agent does the work; it never grades it. The problem: an upstream API ships v2. Your code still speaks v1, and so do your tests, because they mock the old response. Hand that to a coding agent and "all tests pass" proves nothing: it may have weakened the tests or shaped the code around them. The pipeline, one command (driftzero migrate --repo <github-url> --changelog ... --openapi-v1 ... --openapi-v2 ...): Differ: a deterministic JSON diff of the two OpenAPI specs, plus one Muse Spark 1.3 call (Responses API, strict JSON schema) that annotates it with migration formulas from the changelog. The model cannot add or drop entries. Blast radius: Python's own ast walks the target repo and finds every function, class and test that touches the drifted fields. No model. Repair: Muse Code (muse exec, muse-spark-1.3) edits the code in an isolated git worktree. Main is never touched. Guard: an AST diff of the tests rejects removed tests, fewer asserts, skips, pytest-config changes and stubbing of the target dependency; allowed edits are listed for human review. Verify: a clean Modal sandbox runs the agent's own tests and a set of hold-out contract tests written from the v2 spec that the agent has never seen (not in its prompt, hidden from its worktree). On failure only the test name and a one-line reason go back, up to 3 iterations. PR: pushed to the target repo with hold-out names and results, sandbox id, tokens, cost and links to raw logs. Review: Muse Spark Open (meta-models/muse-spark-open-1026 on Baseten) reads the diff plus the unchanged tests covering changed code and reports what static analysis can't see, each finding with a confidence; optionally triggers one fix iteration. Results (all logs committed in docs/runs/): on the demo run the agent reported 25/25 passing, the hidden hold-out said 7/8, one redacted line went back, iteration two passed 8/8, PR opened, $0.65, about 8 minutes, no human in the loop. Pointed at two public repos I had never seen (JobTracker and wolverine, stuck on OpenAI Python SDK v0) with the official migration guide as input: both verified, $0.40 and $0.39. The first attempt on each failed; those runs are in the README, including the time the agent stubbed out the entire SDK so its tests would import. The guard rejects that now. Limits, stated in the README: Python only, hold-outs written per target, the reviewer is a second opinion rather than a gate, a wrong hold-out stalls the run. Built solo, Oct 30–Nov 4. 171 tests, 3 target repos, 6 PRs opened by the tool.