dsh-project-ops

In Agent-driven workflows, a recurring source of friction is this: the Agent knows which files changed but lacks a bounded map of the project declared tasks to answer which checks should run, how they depend on each other, and whether results can be verified. dsh-project-ops fills that gap as a declarative execution-autopilot Bundle for DeepSeek Harness. It reads only task declarations from package.json, Makefile, Justfile, and pnpm workspaces; it installs no second shell, job registry, or workflow engine. Approvals, cancellation, and output retention remain under the Harness host.

Six model tools carry the core logic. task_list enumerates bounded declared tasks with their dependency edges. task_plan accepts a set of changed files and a verification goal (test, lint, typecheck, or build), performs topological ordering, and restricts auto-selection to those four safe purposes. task_run offers foreground, background, and auto modes; auto is background-first with a bounded wait window and returns a terminal or running receipt. task_collect polls under Harness JobRegistry caller-ownership enforcement. verification_gate rediscovers manifests, recomputes the plan, and issues a passed, pending, failed, or stale verdict on version-2 receipts, returning only task IDs and reason codes without echoing commands, paths, or approval content. capability_search ranks only the calling Agent visible tool schemas and in-project tasks.

Built for teams already running DeepSeek Harness on monorepos or Make/Just workflows that want the Agent to execute checks autonomously and produce auditable receipts, without introducing a parallel shell job system or an extra workflow engine. Execution policy authority stays with the Harness host.