DWP v7: plans that delegate, with a record of everything

Deep Work Plan v7 keeps the v6 method: the contract is the authority, the append-only journal is the memory, the scheduler decides what runs next, and a task closes only when recorded evidence satisfies its criteria. v7 adds the ability to delegate, and keeps the same discipline around the result.

A plan that grants agent_delegation can mark a task parallel_safe and hand it to another agent. The delegate’s reply is recorded as data, never as instructions, and stays asserted until the plan’s own gate runner observes the outcome. Only the runner mints observed evidence, so delegation adds reach without lowering the bar for completion.

Four optional addons turn delegation into practical autonomy. Herdr hands a task to an agent in a pane, on any machine. Agentkit puts one ak command over every terminal coding agent, with autonomy by default and an opt-out, and runs bounded tasks headless in a git worktree. Devcontainer gives each repository a reproducible container with no SSH key inside. DeepWorkPlan Vim is a terminal editor with a plan browser and Markdown viewer. Each is pinned by tag to a product with its own repository and works without Deep Work Plan. A repository is fully conformant with none of them, and the registry in .dwp/config.json records which are enabled.

Benchmark and learnings mode records what each plan teaches, so findings can be analysed afterwards. An audit of the whole ecosystem, run as a v7 orchestrator plan with one agent per repository, found no behavioural regression against v6: the pack suite passes 807 of 807 on a clean environment, and instruction load grew 0.1% to 3.9% per flow (4.6% for the whole pack), measured in bytes on both tags rather than estimated as tokens.

v7 is a step change in orchestration and auditability, not yet fully hands-off autonomy. Release 7.2.0 closes the gaps that audit found: benchmark and learnings now measure v7 plans, starting a task before its prerequisites are complete is refused, and the test suite runs hermetically (815 of 815 on a clean environment). Agent outcome non-inferiority has not been measured. Existing plans keep their recorded generation and are never migrated implicitly; new plans use the v7 contract by default.

Installed skill release: 7.5.1, stable since 7.0.0.