dsh-planning-with-files

Native DeepSeek Harness (dsh) plugin for planning-with-files: persistent file-based planning for AI coding agents. The plan lives on disk in task_plan.md, findings.md and progress.md; this plugin keeps it in the model's context on every prompt and can hold a gated turn open until the plan reports complete.

Install

npx skills add OthmanAdi/planning-with-files --skill planning-with-files -g
dsh plugin --profile web add dsh-planning-with-files

The first line installs the skill text into ~/.agents/skills/planning-with-files/, one of the paths dsh reads natively. The second adds this bundle to the web profile; restart dsh web afterwards. The same dsh plugin --profile <name> add dsh-planning-with-files works for the headless, sdk and acp profiles. dsh --profile web --dump-config shows the dsh-planning-with-files layer.

Built against the published @deepseek-ai/dsh 0.1.5-rc.2 and the @deepseek-ai/* 0.1.5-rc.2 libraries it installs.

What the plugin does

Lifecycle point	Behavior
agent/pre-step	Appends the active plan to the step that carries the user's prompt: the framed head of task_plan.md, the normalized tail of progress.md, a pointer to findings.md. Resolves PLAN_ID; with one named plan, .planning/.active_plan or the plan directory itself; two or more named plans without PLAN_ID get the Multiple plans are available notice instead of a plan; then the legacy root file. Steps that carry only tool results, steering or the plugin's own messages get nothing
tools/post-execute	Attaches a progress reminder to successful write and edit calls, and to mutating str_replace_editor calls (create, str_replace, insert) where a profile mounts that tool, while a plan exists
session/event with a successful compaction/end	The next step of that agent (the continuation of an automatically compacted turn, or the prompt after a manual /compact) carries the compaction note (plan pointer, flush instruction, attestation hash) and the framed plan as one message, so the continuation resumes at the current phase
agent/turn-stopping	Completion gate in gated mode: while an in_progress phase remains, the plugin steers the agent into another step with the gate reason. Block cap PWF_GATE_CAP (default 20) and ledger stall detection release the turn; subagent children are never held

Commands: /pwf [--autonomous|--gated] [--template analytics] [plan name] creates the planning files (a name creates .planning/YYYY-MM-DD-<slug>/ and makes it active; --gated or --autonomous writes the v3 markers and attests the plan) and queues the framed plan for the next model request; /pwf-status reports the active plan. Tools: pwf_init, pwf_status, pwf_check with the same contract as the OpenCode plugin, resolving the plan from the calling agent's session workspace.

Environment, read from the dsh process: PLAN_ID=<slug> binds a plan and ends resolution when the slug names none; PWF_PLAN_ROOT=<absolute path> pins the project root and fails closed when it does not resolve; PLANNING_DISABLED=1 silences every listener, command and tool; PWF_GATE_CAP=<n> caps the gate blocks. Autonomous and gated plans inject only when .attestation (slug) or .plan-attestation (root) matches the SHA-256 of task_plan.md. A live plan in a direct child project makes a cwd guess ambiguous and nothing is injected, with a one-line notice.

Row configuration: enabled (default true) and gate (default true, set false to keep injection and drop the completion gate).

Full guide: docs/deepseek-harness.md.

License

MIT