OMP Tandem 3.4.0 — shared tasks and autonomous collaboration

OMP Tandem 3.4.0

One shared task for Claude and OMP

• tandem_work provides a durable, versioned task card with goals, constraints, an agreed plan, checklist, assigned implementers and distinct reviewers, dependencies, blockers and evidence. Both participants use the same project-scoped state; Markdown is a rendering, not a concurrently overwritten source of truth.
• Plan agreement, execution, submission and acceptance are separate transitions. Optimistic revisions and stable operation IDs reject conflicting changes and deduplicate exact retries. Acceptance references an immutable output and plan version.
• Independent modules can run concurrently in separate Git worktrees. A final integration step combines accepted prerequisites. The original checkout changes only after an explicit operator apply --expected-head.

Explicitly authorized autonomous continuation

• New operator CLI: python -m omp_tandem.work_daemon with authorize, run, start, status, stop, revoke, reconcile, show and apply.
• A separately authorized controller can outlive the interactive client. It starts dedicated Claude/OMP attempts under project, tool, wall-time, launch-count and reported-cost limits. Models cannot grant themselves autonomous authority through MCP.
• Workers acknowledge startup through the shared task. Cooperative blockers preserve source checkpoints; after an evidenced unblock, work continues from the saved edits. Rejected submissions also remain available for targeted correction.
• Explicit pause is sticky. Lost acknowledgements, crashes or uncertain effects do not cause automatic replay. Recovery requires an operator decision after inspecting old execution and side effects.
• Attached clients receive advisory change events; durable state, not notification delivery, controls readiness. Unknown cost pauses autonomous work instead of being counted as zero.

Safety and correctness

• Managed worker capabilities are bound to their task, role and attempt. Reviewers cannot accept their own implementation; managed acceptance is committed only after successful worker completion and snapshot integrity checks.
• Raw committed bytes are preserved despite Git line-ending/ident attributes. Unexpected file edits, unsafe source entries, stale outputs, dependency conflicts and changed apply targets fail closed with retained evidence.
• Corrected edge cases include retired claim receipts, manual submission validation, reconciliation revisions, removed unresolved steps, concurrent stop updates and incomplete native cost accounting.
• Git worktrees are not an OS sandbox. --allow-tests grants arbitrary shell capability. Reported monetary limits are estimates/soft limits, not invoice-level guarantees. Local execution is unavailable while the machine is asleep/offline.

Verification

• Local full suite: 336 tests and 289 subtests passed. Ruff, wheel build and source archive manifest checks passed; installed-wheel CLI startup was exercised.
• Actual Claude Code 2.1.267 and OMP 18.1.13 were exercised through real stdio MCP/native RPC. Only model HTTP replies were supplied by a deterministic localhost fixture; no paid provider calls were made.
• Real-client scenarios covered shared agreement, two parallel modules, cross-review, final integration and explicit apply; detached continuation after client exit; checkpoint preservation and duplicate unblock; and explicit pause stopping both agents without a second dispatch. These are integration checks, not model-quality or comparative benchmark results.
• README and detailed guides updated in English, Russian and Simplified Chinese.

Русское руководство · English guide · 中文指南

Publishing or updating this plugin does not enable unattended work for a project. Autonomous execution requires its own explicit operator grant.

Final release CI

Run 34595972302 passed for f3954e667fd287095e03e208859fbc62b2899890 on Linux / Python 3.12 and macOS / Python 3.13. Each platform passed 336 tests and 289 subtests, real pinned OMP verification, Ruff, wheel building and distribution manifest verification.

Post-release CI clarification

The later tag-triggered run 34596370446 failed on macOS despite the earlier passing branch run at the same commit. Its review lifecycle test temporarily replaced methods on the running controller with assertion-raising mocks; the background driver could hit that test-only replacement and interrupt the run.

Test-only correction 40238ee removes those internal implementation assertions while retaining observable reply, repeated-reply, cancellation and preserved-answer checks. No runtime code, timeout, skip or CI trigger was changed. Corrected CI 34596992268 passed on Linux and macOS: 336 tests and 289 subtests on each, plus real OMP verification, lint, formatting, wheel and manifest checks.

The published v3.4.0 tag and assets remain unchanged; the test correction is on main. The historical failed run is retained, not rerun to manufacture a green result. README and detailed shared-task guides were included in v3.4.0 in English, Russian and Simplified Chinese.