@klarkxy/dsh-recap

Writes session recaps in the background and gives the agent bounded progress checkpoints: after a long turn, or when you come back later, the session gets a short recap to read, and the model gets a length-capped progress note at key points. The plugin starts enabled and can be switched off under Settings → Plugins; recaps, agent checkpoints, and semantic checkpoints use fixed automatic defaults.

简体中文

Requires Node.js ≥22 and DSH 0.1.7-rc.2. Auxiliary generation calls the host llm service directly. @deepseek-ai/dsh-llm and @deepseek-ai/dsh-session are required peers.

npm install @klarkxy/dsh-recap
dsh plugin --profile web add @klarkxy/dsh-recap

Recaps

Recaps appear on the chat events seat — a quiet lifecycle controller that renders nothing until the session has a stored recap, then shows the recap cards with Generate, Cancel, and Regenerate (regenerating a finished recap asks for a second click).

A recap is written in the background after a long turn ends, or when you come back after the idle interval. The interval defaults to 15 minutes of focused document and user activity: assistant streaming does not count, and neither does time spent while the chat panel is hidden.

Reading status never calls the model. Each watermark gets at most one generation, unless you retry that card or request a manual refresh.

Checkpoints

Checkpoints are injected only at host agent/pre-step, and only on meaningful boundaries, using native createUserMessage. The injected snapshot is bounded — it lists at most 12 items and is truncated to 2000 characters — so a long turn cannot flood the model with context.

Checkpoint items are collected from the session log. With semantic checkpoints on, a model only refines the checkpoint's next step. When Mood is active in the same profile, Recap reads its task contract through ctx.aiMood.getContract(sessionId) for checkpoint context.

Plugin page settings

Settings → Plugins → Task Recap has switches for recap cards, task checkpoints, and semantic checkpoints, the away interval in minutes (1–180), and one model row per feature: Recap card model for recap cards and Semantic checkpoint model for semantic checkpoints. Both are optional. A saved route is used for that feature's own call as an explicit model; an empty selection follows the live session model and then the host default chat model.

The same switches can also be changed through the update RPC. Turning an ability off cancels in-flight generation and stops auto injection; stored recaps remain. Unloading the plugin waits for pending writes and ignores later session events.

Boundaries

• Recaps are display-only and are not model context; checkpoints are the only thing injected into the model.
• Semantic checkpoints only run while task checkpoints are on.
• A saved route affects only these two features, and it selects a model — it is not evidence of connectivity.

Host RPC

Channel /dsh-recap requires the host authorization policy.

• status — settings, storage health, and stored cards and checkpoints (optionally for one sessionId). Never calls the model.
• update — write settings (cardsEnabled, checkpointsEnabled, semanticCheckpointsEnabled, idleReturnMs) with compare-and-swap (expectedRevision).
• cards / checkpoints — the stored recap cards or checkpoints, optionally for one session.
• cancel / retry — cancel a card's in-flight generation, or regenerate that card ({ cardId, sessionId }).
• idle.return — run the idle-return check for a session immediately.
• refresh — generate a recap for a session on demand (manual trigger), independent of the one-generation-per-watermark rule.

Development

From the repository root:

pnpm --filter @klarkxy/dsh-recap typecheck
pnpm exec vitest run packages/dsh-recap/src
pnpm --filter @klarkxy/dsh-recap build

Hosts that bundle this feature usually enable it by default; where the host supports live plugin switching, toggling needs no restart. Installation or removal may require a restart when the host asks for one.