dsh-context-mode exposes a self-contained context-mode server as native DeepSeek Harness tools. It registers the full ctx_* catalog at runtime and adds model-facing routing guidance, so dsh-TUI can use sandboxed execution, indexing, FTS5 retrieval, and web fetching without a second MCP client.

The engine is vendored: vendor/context-mode/ carries the source and this package builds its own server.bundle.mjs from it. There is no context-mode npm dependency at runtime, and the eighteen upstream platform adapters (Claude Code, Codex, Cursor, Gemini CLI, …) were removed — DSH is the only supported host.

Two DSH-specific changes sit on top of the vendored engine:

CJK search. Upstream indexes with porter unicode61 and trigram, so Chinese queries largely fail (a run of Han characters becomes one token, and the trigram table cannot match two-character words). The adapter segments CJK text on both sides of the index, which took a five-query Chinese benchmark from 2/5 to 5/5.
Output containment. Results above 40 KB are truncated to head + tail, the full text is spilled to disk, and the model is told how to read it back.

See LICENSING.md: the adapter is MIT, the vendored engine is Elastic License 2.0.

Install

dsh plugin --profile dsh-tui add dsh-context-mode

Restart the profile after installation. Removing the package removes its profile row as well:

dsh plugin --profile dsh-tui remove dsh-context-mode

The npm package ships a prebuilt vendor/context-mode/server.bundle.mjs, so no build step runs at install time. The bundled skills/context-mode/SKILL.md is registered with DSH's skill registry when the profile provides the skills service, so it should appear in /skills after a restart.

Configuration

The default patch starts the bridge with:

enabled: true
the installed context-mode/server.bundle.mjs
the current working directory as the project directory
~/.dsh/context-mode as the isolated database root
a 60-second initialize and tool-catalog timeout

Override the row's config in a profile patch when needed:

- id: dsh-context-mode
  name: dsh-context-mode
  inject: [tools, systemPrompt]
  config:
    projectDir: /absolute/path/to/worktree
    storageDir: /absolute/path/to/context-mode-data
    handshakeTimeoutMs: 120000

serverPath is available for local development and pinned deployments. It may be an absolute path or a path relative to the profile process directory.

Runtime behavior

The plugin starts one long-lived child process per DSH composition. It performs MCP initialize and tools/list, registers each returned tool through the DSH tool registry, and forwards every call over MCP stdio. The child is terminated when the Cordis plugin is disposed.

Context-mode analysis, indexing, search, and diagnostics tools are marked concurrency-safe so independent model tool calls may overlap. ctx_insight, ctx_purge, and ctx_upgrade remain exclusive because they open external UI or mutate installation and stored data. ctx_batch_execute additionally parallelizes its own command batch through its concurrency parameter.

All eleven ctx_* capabilities are exposed on DSH — ctx_execute, ctx_execute_file, ctx_batch_execute, ctx_fetch_and_index, ctx_index, ctx_search, ctx_stats, ctx_doctor, ctx_upgrade, ctx_purge, and ctx_insight. Each registered tool carries DSH-specific routing guidance in its description, and the plugin injects a routing section plus a bundled context-mode skill so the model reaches for these tools by default.

The adapter targets DSH only. Cross-platform code you may see inside the context-mode dependency belongs to that package's prebuilt MCP server, which this adapter does not modify; it ships no Claude Code, Codex, or other host integration of its own. CONTEXT_MODE_PLATFORM=pi is set solely because upstream only accepts its own platform ids, and pi is its neutral MCP-only id — every store stays DSH-owned: CONTEXT_MODE_DIR isolates DSH data and CONTEXT_MODE_PROJECT_DIR pins project hashing to the configured workspace.

Compaction archiving

Compaction replaces the live conversation with a generated summary and prunes the events behind it, so anything the summary omits leaves the model's reach. The plugin listens for compaction/start and files the transcript into the knowledge base first, where ctx_search can still reach it afterwards.

Storage is layered rather than filtered — every transcript event is archived, and the layers differ only in the source label they carry, so precision is chosen at query time instead of at write time:

Source	Contents
`session/<id>/constraint`	user messages — requirements, decisions, limits
`session/<id>/finding`	tool results and assistant prose stating a concrete value
`session/<id>/narrative`	remaining assistant prose — reasoning, plans

Nothing is dropped at write, so a misclassification costs a query's precision rather than the content itself. Harness-injected blocks (<active_memory>, <current_runtime_context>, <system-reminder>, <resume_snapshot>) are the one exception: they ride on user/message, are per-turn runtime noise rather than transcript, and are discarded before layering. Set precompact: false to turn archiving off entirely.

Archives written before 0.3.2 may contain those injected blocks. A one-shot cleanup script removes them and leaves everything else alone:

node node_modules/dsh-context-mode/scripts/cleanup-injected.mjs --db <path>            # report only
node node_modules/dsh-context-mode/scripts/cleanup-injected.mjs --db <path> --apply    # delete

Checkpoint transcript and archive index

A checkpoint keeps only a summary of the span it replaces, and the shipped checkpoint format has no section for tool output — a long tool result survives only as whatever the summarizing model chose to keep. precompact files that same span into the knowledge base beforehand, but nothing told the model so, and a summary that omits a detail reads as if the detail never existed.

This package also exports a compaction engine that closes that gap. It extends the shipped BasicCompactionEngine and replaces the lossy summary with the conversation itself, plus a pointer to everything it did not carry:

import DshContextModeCompaction from 'dsh-context-mode/compaction'

Each checkpoint it writes carries three parts:

Part	Content
The shipped summary	The eight-section checkpoint, unchanged
`## Conversation Transcript`	user turns kept whole; assistant replies kept head+tail; tool output reduced to one index line each
`## Archive Index`	The `source` labels to search, and how to query them

Every region that was clipped or dropped leaves a pointer naming the source holding the original and the seq it came from, so the checkpoint says which event lost detail rather than merely that an archive exists:

## [assistant 447]
… first 200 chars …
[... 1581 of 1981 chars elided from seq 447; retrieve with ctx_search(source: "session/<id>/narrative") ...]
… last 200 chars …

## [tool ctx_execute seq 3424, 29808 chars → search `session/<id>/finding`]

A prior checkpoint is never transcribed forward: its text is already in the summary, and copying it would grow the transcript on every compaction.

Mount it in place of the shipped backend, inside the isolate group the shipped one requires:

- id: compaction
  name: cordis:group
  group: true
  isolate:
    compaction: true
    toolResultPruner: true
  config:
    - id: compaction-basic
      name: 'dsh-context-mode/compaction'   # was @deepseek-ai/dsh-compaction-basic
      config:
        thresholdRatio: 0.8   # compact at 80% of the window
        retainRatio: 0.1      # keep the newest 10% verbatim

    - id: command-compact
      name: '@deepseek-ai/dsh-command-compact'

Only compactRegion (to capture the span) and summarize (to append) differ: trigger policy, retention, the transaction bracket, token metering, and prefix-cache-aligned replay all stay on the shipped engine. The appendices are added to the returned summary rather than injected into the summarization instruction, so the shipped output contract is untouched and no model has to follow an amended format. Failing to build them returns the superseded summary unchanged, and restoring the shipped name row disables them entirely.

TRANSCRIPT and per-entry clipping are tunable from dsh-context-mode/transcript: USER_CLIP_AT / USER_CLIP_KEEP for long user turns, ASSISTANT_KEEP for assistant head+tail, and MAX_TRANSCRIPT_CHARS for the whole body.

Development

pnpm install
pnpm build

The plugin needs Node ^22.19 || >=24. The upstream package installs better-sqlite3; a native build may be required on platforms without a prebuilt binary.