dsh-plugin-shiori-role

A native DeepSeek Harness plugin for editable Shiori roles, session-scoped identity, role-owned visuals, and isolated durable memory.

中文说明

Features

• Manage roles in Settings -> Plugins -> Roles. The page contains role cards and a create card; each role opens in a modal for editing its name, introduction, System Prompt, avatar, standing illustration, and image library.
• Select a role from the conversation composer. A blank session can switch roles repeatedly; the first System Prompt assembly commits the selection, and a running or resumed session remains immutable.
• Preview the selected role immediately. The plugin applies role-owned theme tokens and uses the standing illustration as a sidebar/workspace visual layer without changing the user's light, dark, or system preference.
• Store images through the Harness attachment service. Plugin KV tables retain only ImageAttachmentRef metadata, never base64 image blobs.
• Keep durable memories isolated by role and expose memorize, recall_memory, and forget_memory in the bound Agent scope.

All backend services, Remote descriptors, client slots, styles, and theme cleanup live in this plugin. No DeepSeek Harness source patch is required.

Configuration

Configured roles are imported once as editable seeds. Later edits and deletions remain durable and deleted seeds are not recreated after restart.

- id: shiori-role
  name: '@deepseek-ai/dsh-plugin-shiori-role'
  config:
    roles:
      - id: maintainer
        name: Shiori Maintainer
        introduction: Maintains the Shiori workspace.
        prompt: You are the Shiori maintainer for this workspace.

For local development, install the repository as a linked profile dependency:

dependencies:
  '@deepseek-ai/dsh-plugin-shiori-role': link:D://Coding//dsh-plugin-shiori-role

Role Binding

ShioriRoleService stores editable roles, assets, workspace defaults, pending blank-session selections, and immutable session bindings in the shiori_role storage domain.

The composer may stage a role before a blank Agent is created. A blank Agent remains selectable until its first System Prompt assembly; the plugin then resolves and persists the pending or workspace role before model execution continues. A session with no selected or workspace-default role bypasses the plugin entirely: it receives no role prompt, Markdown memory context, or role memory tools. Existing version-1 bindings remain immutable once a session has used them, and non-blank sessions never hot-switch.

Deleting a role is a physical deletion: its catalog record, session bindings, assets references, Markdown files, and role-owned SQLite memory2 database are removed. Existing sessions no longer restore that role. Mutable pending and workspace references are reassigned to the earliest remaining role, and every client surface refreshes from the same catalog mutation signal.

When a deleted role owns local content-addressed image attachments, the plugin removes an attachment object only when no remaining role asset references the same sha256: object. Shared objects and non-local attachment backends are left intact; this GC does not scan message or session-history references outside the role catalog.

Assets And Theme

Supported asset purposes are avatar, portrait, gallery, and theme_background. Avatar and portrait uploads are separate. Binary image data is saved and read through ctx.attachments.saveImage/readImage; role asset rows contain attachment references only.

The selected role controls the composer avatar and role name. Its portrait can influence the workspace and sidebar through plugin-owned CSS and ctx.theme.overrideTokens. Plugin disposal removes the injected stylesheet, theme override, classes, CSS variables, and image URLs.

Role Memory

Each role owns an isolated durable memory scope shared by every DSH workspace. Its editable definition is stored at $DSH_HOME/shiori-plugin/role/<role-id>/role.json; the complete readable memory projection lives under $DSH_HOME/shiori-plugin/role/<role-id>/memory/: MEMORY.md holds stable facts, preferences, and explicit remember requests; HISTORY.md is the append-only shared timeline; PENDING.md buffers long-term candidates; RECENT_CONTEXT.md preserves recent topics and ongoing work; and SELF.md maintains the role's self-model. Searchable structured memory remains in the adjacent memory2.db.

The Shiori semantic layer is active without a memory config block. After every completed turn, a bounded auxiliary request through the Agent's current Harness provider/model extracts long-term memories before the turn closes. An optional extraction endpoint overrides that default call. An optional embedding endpoint adds independent vector recall; without it, retrieval degrades to deterministic text search.

- id: shiori-role
  name: '@deepseek-ai/dsh-plugin-shiori-role'
  config:
    roles:
      - id: maintainer
        name: Shiori Maintainer
        introduction: Maintains the Shiori workspace.
        prompt: You are the Shiori maintainer for this workspace.
    memory:
      embedding:
        endpoint: https://api.openai.com/v1
        apiKey: sk-...
        model: text-embedding-3-small
      extraction:
        endpoint: https://api.openai.com/v1
        apiKey: sk-...
        model: gpt-4o-mini

Retrieval mirrors Shiori's memory2: the keyword lane keeps literal hits while the vector lane recalls semantically similar rows independently (cosine threshold 0.35), then Reciprocal Rank Fusion merges both ranked lanes (1/(60+vec_rank) + 0.5/(60+keyword_rank)) with hotness boosting. Post-turn extraction follows Shiori too: agent/turn-stopping feeds the turn's USER/ASSISTANT conversation to a separate bounded model request, which applies Shiori's long-term memory contract (verbatim user anchors, cross-session durability, source direction, no events) and returns profile / preference / procedure memories with emotional_weight and friends. The listener waits for the write before the turn closes; extraction failure is logged and does not fail the completed response. Changed memory context is materialized by Harness as an append-only runtime-context snapshot, preserving the reusable request prefix.

Memory is stored per role in SQLite (<memoryRoot>/shiori-plugin/role/<role-id>/memory/memory2.db, built-in node:sqlite, zero dependencies). Explicit memorize writes reinforce by content hash; on write, semantically similar old rows are retired automatically (preference/profile similarity ≥ 0.90, high-emotion profile 0.92) and procedures sharing a tool_requirement merge. Successful Harness compaction feeds the exact shadowedSeqs window through Shiori's semantic consolidation rules. Role-scoped maintenance runs serially: it appends history_entries and pending_items, snapshots PENDING.md, immediately merges the snapshot into MEMORY.md, updates SELF.md from the same snapshot, generates RECENT_CONTEXT.md, and commits the snapshot. A failed optimizer restores the snapshot. SQLite stores the corresponding events and long-term candidates under stable per-entry source refs, then records the completed compaction source ref so replay is a no-op. SELF.md, MEMORY.md, and the compact portion of RECENT_CONTEXT.md are injected into the role context; PENDING.md is never injected and HISTORY.md remains a maintained timeline.

This version intentionally does not claim full Shiori default_memory parity. It provides deterministic text retrieval, embedding semantic retrieval with hybrid RRF ranking, exact-match reinforcement, semantic supersede/merge, explicit memory tools, automatic post-turn extraction, compaction-driven Markdown consolidation, immediate role-scoped optimization, and prompt injection. HyDE, query rewriting, scheduled optimization, journal projection, and background ingestion are not implemented.

Development

D:\Coding\deepseek-harness\node_modules\.bin\tsc.cmd --noEmit --project tsconfig.json
node --import file:///D:/Coding/deepseek-harness/node_modules/tsx/dist/esm/index.mjs --test test/*.test.ts
D:\Coding\deepseek-harness\node_modules\.bin\tsc.cmd --project tsconfig.json
D:\Coding\deepseek-harness\node_modules\.bin\tsdown.cmd --config tsdown.config.ts