OMP Tandem 3.3.0 — bounded reviews and real verification

OMP Tandem 3.3.0

First useful review, fewer manual steps

• First-run guidance in English, Russian and Simplified Chinese: local project diagnosis, then read-only review of the prepared/staged commit.
• tandem_review_run: stable request identity, bounded start/status/reply/cancel, one independent stage plus at most one author comparison, compact progress and complete final answers. Default 600-second total budget; no automatic edits or shell/test execution.
• Proportional planning: a known local fix gets a brief independent risk check; one comparison, then decide, distinguish by experiment or state the unresolved question. No consensus loops.

Correct capture and cancellation boundaries

• Snapshot capture runs in a separately supervised process with owner/reservation/deadline checks and atomic publication. Late capture cannot claim no_changes or start a native phase.
• A publication gate prevents one snapshot writer from blocking another run's cancellation under the controller guard. Contended stops are explicitly pending until persisted; no false terminal acknowledgement or replay.
• New regression reproduces the old cross-run MCP cancellation timeout and passes after correction; rollback during publication retains snapshot identity without publishing cancelled contents.

Evidence instead of benchmark claims

• Real development case: all three paid episodes retained, including missing-context cancellation and timeout; complete available answers, snapshot hashes, coordinator+peer accounting, and separate runtime follow-up.
• Measured episode wall time totals 598.408s; known cost subtotal USD5.6829274. Full cost and human effort are unknown; prior development/repairs are outside those measurements. This is not a comparative benchmark.
• Real OMP compatibility: official checksum-pinned OMP18.1.13 + SDK revision daf07999c2fee9b22edc7bf8fea1fb6272e0df5e, with a deterministic localhost model fixture and no paid credentials. CI runs the binary, not a fake RPC server.
• Separate four-arm benchmark protocol/schema/offline analyzer; no comparative results or superiority claim.

Verification

• Local full suite: 276 tests and 267 subtests passed; Ruff clean; wheel and deterministic source archive built and verified.
• Actual stdio MCP + official OMP smoke passed: two automatic stages, idempotent start, full 25,200/24,300-character answers, staged-only applicability, stale detection and preserved answers, no native project/shell tools exposed to the review model.
• Real-binary local compatibility passed on Darwin arm64/Python3.13.5: host tools, structured completion, continuation, cancellation, think/analyze/work capabilities. Exact JSON proof is included; this does not certify a broad version range or every provider.
• Two dependency deprecation warnings remain from Authlib; no runtime exceptions suppressed to obtain a pass.

Work mode is not an OS sandbox. A completed review is not proof that code is correct. Channels/hooks remain optional; bounded polling remains supported.

Русский: понятный первый запуск, компактное независимое ревью, ограниченное обсуждение, исправленные deadline/отмена и честный разбор трёх попыток. README и подробное руководство обновлены.

中文：更新首次使用流程、紧凑独立审查、有限轮次讨论、捕获期限与取消修复，以及保留三次尝试的真实案例。中文 README 和完整指南已更新。

Final cross-platform acceptance

• CI run 34459190147 passed for release commit 5f1136b5a32cee5314c120a68d9ff7662162ba37 on Linux/Python3.12 and macOS/Python3.13. Both jobs passed 277 tests and 267 subtests, the actual pinned OMP verification, Ruff, wheel build and distribution manifest check.
• CI follow-up corrected a concurrent-close race and foreign-owner validation order, and synchronized deadline regressions independently of interpreter startup speed. Production timeout limits were not increased.