dsh-plan-keeper

A DeepSeek Harness plugin: plan persistence + workspace memory, presented as a Deliverables (产物) conversation view.

Part of the dsh-plugins collection.

Layout

/.dsh-outputs/          # per-workspace outputs root (session.header.cwd)
├── plans//
│   ├── plan.md                      # latest plan
│   ├── revisions/NN.md              # immutable snapshots
│   └── meta.json
└── memory/
    ├── MEMORY.md                    # rebuild injection index
    ├── meta.json                    # { enabled, budgetBytes, entryCount }
    ├── user/-.md
    ├── feedback/-.md
    ├── project/-.md
    └── reference/-.md

The output root follows the session workspace (session.header.cwd), not the deployment root.

Install

./scripts/install.sh web plugins/dsh-plan-keeper

or, as a bundle-layer package:

dsh plugin --profile web add ./plugins/dsh-plan-keeper

Then refresh the browser page and open the Deliverables (产物) tab in any session header (after Trace (轨迹)).

Note for sessions that ran the dynamic (in-session) version: the dynamic plugin registers the same deliverables view id with a lower shadowing priority, so it wins while it is running. cordis_stop (or undefine) the dynamic plank-1 plugin before installing this package to use the packaged version.

Architecture

• Host half (lib/index.js): plan capture & sync (tools/result, agent/turn-stopping), memory storage & index, systemPrompt.section injection, the memory_write model tool (tools.register), and a local HTTP RPC route POST /api/dsh-plan-keeper (webServer.register, { method, args } JSON).
• Client half (lib/client.js): the Deliverables (产物) view — Memory (记忆) ring and Plan (计划) ring — via same-origin fetch RPC.

Security note

The HTTP route is unauthenticated and bound to the harness web server (same-origin, local development setups). It validates sessionId against [A-Za-z0-9-]+ and only touches files under the session workspace's .dsh-outputs/. Do not expose the harness web port beyond localhost.

Model tool

memory_write (type / topic / content) — the model is instructed to record corrections, preferences, project decisions and external references, and not code details, transient task state, or anything already in AGENTS.md / CLAUDE.md. Entries are plain markdown; the user can edit or delete them in the Deliverables (产物) view at any time.

License

MIT — see LICENSE.