dsh-checklists

dsh-checklists

Session-scoped checklists for DeepSeek Harness. The agent owns the list, the model always sees it, and a read-only sidebar tab mirrors it.

The problem this solves is not "having a list". It is having to dictate the list again every session. Here the checklist is created once, injected into the model context on every turn, and updated as work progresses — no reminders needed.

How it works

1. The agent creates a checklist (checklist_create), optionally from a template.
2. A compact digest of every checklist is injected into the prompt on each assembly, so the model always knows what is pending.
3. The agent marks items as it finishes them (checklist_check).
4. The sidebar tab shows the same tree, read-only, updated every two seconds.

If the agent forgets to mark something, you say so in chat and it fixes it.

Tools

Items are plain strings. Two leading spaces per level nest them:

["Deploy to staging", "  Smoke tests", "  Rollback plan"]

Install

dsh plugin --profile web add dsh-checklists

From a local build or tarball:

npm install
npm run build
npm pack
dsh plugin --profile dev add ./dsh-checklists-0.7.0.tgz
dsh --profile dev --dump-config        # the tool-checklist row should appear

Install into a test profile first. web and headless are distinct profiles, and so is the desktop app's desktop profile.

Configuration

Override the defaults on the plugin's entry in the profile's cordis.patch.yml:

- id: tool-checklist
  name: dsh-checklists
  config:
    maxDigestLines: 12      # lines of the injected digest
    maxItemsPerList: 200    # per-list item cap
    pruneDays: 14           # prune untouched sessions after N days (0 = never)
    nudge: steer            # steer | observe | off — see "Turn-end gate"
    workspaceView: on-demand   # off | on-demand | always — see "Workspace lens"
    workspaceViewLimit: 10     # sibling checklists returned
    # storageRoot: "D:/data/dsh-checklists"
    # templates:
    #   my-flow: ["Step 1", "  Sub-step", "Step 2"]

Data

$DSH_HOME/storages/dsh-checklist/
├─ sessions/<sessionId>.json    one document per session
└─ templates.json               templates shared across sessions

Writes are atomic (temp file + rename) and serialized per session. A corrupt document degrades to an empty list instead of failing the turn.

The package is dsh-checklists, but its storage directory keeps the dsh-checklist prefix. That path is data, not identity: renaming it would orphan every checklist already saved.

Sidebar tab

Open the right sidebar, then the guide (+) and pick Checklist.

Item states remain agent-only: the tab cannot check anything off. The one write path is text editing, through POST /dsh-checklist/item with expectedText. If the item changed since the tab read it, the host answers 409 rather than overwriting the agent's edit.

HTTP routes

Turn-end gate

A checklist only helps while it is true. The gate closes the most common gap: a turn does real work, forgets to mark it, and the list is stale by the time you look at it.

When a turn is about to close, the plugin checks exactly three things:

1. Did the turn run a tool that changes the workspace (bash, edit, write, …)?
2. Did it skip every checklist write?
3. Is anything still marked doing?

If all three hold, the agent gets one short steering message before the turn ends:

Before you finish: N checklist item(s) are still marked "doing". If you completed any of them this turn, update them with checklist_check and say what proves it. If not, leave them as they are — do not mark work you did not do.

It never guesses what was completed — only that the bookkeeping was skipped. At most one nudge per turn, and none when the turn was pure conversation or the checklist already changed.

Inspect recent decisions at GET /dsh-checklist/nudge:

{ "mode": "steer", "entries": [
  { "source": "turn-stopping", "nudge": true, "steered": true,
    "reason": "1 item(s) still doing after workspace changes", "doing": 1 }
] }

nudge: observe records without interrupting; nudge: off disables the gate.

Workspace lens

Checklists stay per session. The lens is a read-only way to see the checklists of other sessions in the same project — available on request, and never part of the prompt.

Clicking an entry expands it into a read-only tree, fetched once from that session's own state route. Nothing there is editable — the lens can look, not touch.

Membership comes from disk: a session belongs to the workspace directory that contains it under $DSH_HOME/sessions/<workspace>/<sessionId>/. The host's workspace.json is only used for a friendly title, because it is not always complete — on the machine this was written on it was missing a workspace.

Nothing here is writable. To change another session's checklist, open that session: two writers on one list would give the list two truths.

Language

The panel follows the harness locale: Spanish when locale starts with es, English for everything else. If the locale service is unavailable, it falls back to English instead of breaking.

Text addressed to the model — tool descriptions, rendered results, the digest — is English and is not translated: it is prompt material, not interface copy.

Design notes

• One writer. The agent writes state; the sidebar writes text. No optimistic state, no two-tab divergence.
• Digest budget. Each checklist gets an equal share of the line budget. Giving the first list the whole budget starved the rest: with two lists the second showed a single line, and a third never appeared.
• Pending first. Truncation never hides unfinished work.
• Client plugins load through window.__ModuleLoader__ and only require('react'). No bundler needed, but a static inject of a service that never resolves parks the plugin silently — apply never runs and nothing appears. Tab types must be registered lazily (ctx.inject([...])).

Verification

npm run typecheck
npm run test      # vitest
npm run smoke     # same checks, in-process, no child processes

npm run smoke exists because some sandboxes deny spawn with EPERM, and then Vitest cannot start its workers. It covers the model, digest, store, public state and both HTTP routes (23 checks) against the compiled lib/.

Known limits

• No reordering and no drag-and-drop.
• No edit history (who marked what, when).
• Text only: there is no "add a checklist item from the panel" affordance.
• Document format is version: 1; any change must migrate.
• The tab polls every 2 s while mounted — deliberate, the route is read-only.

License

MIT