[chriscoveries]

An unrelated agent-state write invalidates the whole-Context if_version token, even when the task being changed is untouched. Callers cannot distinguish that case from another caller changing the same task.

Tasks now expose an unsigned rev: older task files without one read as 0, and new tasks start at 1. Existing-task operations accept if_task_rev. A mismatch rejects the batch with task_conflict, naming the task and its expected and actual revisions. The existing whole-Context guard remains independent; callers may use either guard or both.

Each semantic task change increments its revision. An identical patch retains the revision and derived timestamp. Operations in a batch see preceding in-memory revisions; a later conflict commits none of the batch's task or Context changes. Targetless creates reject if_task_rev, and callers cannot assign rev themselves. Existing task-ID non-reuse protects against a deleted task being mistaken for a later incarnation. Direct YAML editors must follow the same revision protocol; API guards cannot fence writes that bypass the store.

How we built it. Revision checks and advancement share the common Context operation boundary under its existing lock. Advancement compares all tasks after each operation, including indirectly changed tasks; legacy migration participates too. Read-only loading supplies revision zero without rewriting old files. Full/summary Context reads, individual tasks, task pages, the compact index and MCP bootstrap previews retain the revision. The daemon exposes the structured conflict. MCP preserves a guard once when one call expands into update plus move, and the schema and Context/IPC contracts document it.

How we tested it.

• Eight added core, daemon and MCP tests use projections of three recorded task files. They cover unrelated writes, stale same-task writes, dry runs, late-conflict atomic rejection, sequential guards, identical patches, create/restore/delete, malformed and exhausted revisions, legacy migration, wrapper expansion and read surfaces. Private titles, notes and assignees are omitted from the fixture; test mutations are identified as mutations rather than historical records.
• GitHub Actions CI on the fork at exact PR head 3969f7970f742760dc3b255ec58c611078c480ce passed all six jobs, including rust-linux: run 39777285737. The test-only reference run 39763090179, based on upstream main a212cb04 before it advanced to f1d27838, fails exactly the eight added tests while the other 2,715 pass; it is historical baseline evidence, not a test of current main.
• Real-daemon CI probes use source-pinned CI builds, launch each daemon in a fresh private home, and retain raw Unix-socket requests/responses plus source, binary and fixture hashes. Stock 0.4.44 and upstream main each fail six revision checks; their existing whole-Context guard and ordinary update controls pass. The candidate passes all eight, including named conflicts, no stale-guard writes and atomic batch rejection. Every owned daemon is stopped. Setup failures cannot count as a negative reference.
• Non-author source review passes the feature head. The separate evidence branch is not part of this PR.

Tested on: GitHub-hosted Ubuntu 24.04 x86_64 runners, 4 vCPUs, Rust 1.88.0 (daemon probes); fork CI also passes Windows smoke.