ljwzhaoxiaoniu/DLR#dsh-tsm-agent

---

description: "dsh bundle for TSM / DLR: MCP semantic gateway to tsm-core-dlr, the scenario-independent paradigm skill, the TSM status overlay, and the agent rules (AGENTS.md)." kind: "package-bundle"

dsh-tsm-agent

English | 中文

Summary

dsh-tsm-agent wires a dsh profile to a running TSM Core Service (tsm-core-dlr) and adds the scenario-independent paradigm skill, a Web status overlay, and the agent rules (AGENTS.md). It contributes two top-level rows — mcp-semantic-core (MCP client for http://127.0.0.1:28795/mcp, streamable HTTP) and dlr-status (overlay) — and one agent preset, preset-dlr (persona, AGENTS.md instruction loader, skills, compaction). The semantic backend is the separate package tsm-core-dlr (declared as a dependency, so installing this bundle installs it too); the scenario content comes from a separate, optional tsm-scenario-* package or a path. L3 SOP is delivered by index retrieval (dlr_search_sop), so no L3 skill ships with the package.

Table of Contents

• Use this package
• Understand the implementation
• Model Experience
• Known Limitations and Deferred Work
• Dev Note

---

Use this package

Install once per profile. The recommended path is the Harness Plugin Manager (Settings → Plugins → install bundle) or the agent-side plugin_manager tool with install_bundle — both perform package installation and bundle selection. The CLI equivalent installs the package only; the bundle must then be selected (Plugin Manager toggle, or by listing it in the profile's dsh.profile.bundles):

dsh plugin --profile web add dsh-tsm-agent                              # from npm (pulls tsm-core-dlr with it)
dsh plugin --profile web add "<repo>/DSH-based Agent Service/dsh-tsm-agent"   # or from a local checkout
# then select/enable the bundle — installing alone does not select it

Dependencies. One hard dependency, two optional companions:

Package	Relation	Note
tsm-core-dlr	Hard dependency — declared in dependencies, so installing this bundle installs it too; the MCP row refuses activation while its service is not running	Ships native deps (~200 MB); the service itself must be started separately (below)
tsm-scenario-birdmini, tsm-scenario-cloudopsbench	Optional	Scenario content the service serves; install whichever fits and set TSM_SCENARIO=<package> (or point it at your own scenario). The bundle itself is scenario-independent
dsh-tsm-eval	Optional	Benchmark eval harness (run / judge / report); not needed for regular Q&A

Start the semantic backend first — the MCP row refuses activation when it cannot connect:

bash "DSH-based Agent Service/scripts/start_backend.sh"    # in a repo checkout: Neo4j + TS MCP server (:28795)
bash "DSH-based Agent Service/dsh_dlr/run_web.sh"

From an npm-only install (no checkout), the same service ships as the tsm CLI — npm i -g tsm-core-dlr && tsm serve --http 28795 (first run also needs tsm fetch-model and a tsm build over an installed scenario).

The MCP endpoint defaults to http://127.0.0.1:28795/mcp and can be overridden with the TSM_MCP_URL environment variable. The MCP row is top-level, so it also serves headless runs (presets are not available under --profile headless); preset-dlr shapes the interactive sessions that select it.

Rows the bundle contributes:

Row	Layer	Purpose
mcp-semantic-core	top-level	MCP client → http://127.0.0.1:28795/mcp (streamable HTTP), toolCallTimeoutMs: 60000, failOnStartupError: true, reconnect up to 3 attempts
dlr-status	top-level	Web overlay: backend health + semantic-asset counts (browser half: lib/client.js, reads the service's /status)
preset-dlr	agent preset	persona (semantic business assistant) · AGENTS.md instruction loader · filesystem skills · skill tool · compaction (group with isolate)

The tool surface behind the gateway (7 tools): dlr_semantic_query · dlr_search_consensus · dlr_search_sop · get_pe_mapping · get_le_attrs · get_full_data_info · execute_sql.

Understand the implementation

• cordis.patch.yml — all three rows above: the two top-level rows (MCP gateway + status overlay) plus preset-dlr (merged from the former presets/dlr.patch.yml to satisfy the plugin store's single-string dsh.bundle.patch check; the isolate realm on the compaction group is required by preset services — "Preset services require isolate realms").
• skills/paradigm/SKILL.md — paradigm cognition (TSM three levels; DLR structure: LE / PE / anchors and joins); scenario-independent, loaded on demand.
• lib/index.js — node half: exposes the packaged skills/ directory as DLR_SKILLS_DIR at boot; lib/client.js — browser half for the overlay.
• Layering: bundle patch → profile cordis.patch.yml → $DSH_HOME/cordis.patch.yml → --patch (highest). Consumer-local policy — disabled plugins, model choice, default preset, directory-picker fixes — lives in --patch layers (dsh_dlr/dsh.patch.yml, dsh_dlr/dsh-web.patch.yml), not in this package.

Model Experience

Tools register as mcp__semantic-core__<tool>:

Tool	Purpose
dlr_semantic_query	L1 recall: which business entities (LE) the question is about
get_pe_mapping / get_le_attrs	second hop: physical view columns, ARCS, database_url
dlr_search_consensus	L2 domain-consensus entries (terms → columns / values)
dlr_search_sop	L3: retrieval of the per-question clause, if one exists
get_full_data_info	drill-down to physical columns outside the modeled view
execute_sql	run read-only SQL against the dataset (subprocess, 20 s hard timeout)

Known Limitations and Deferred Work

• The bundle depends on tsm-core-dlr (installing it pulls the service package and its native dependencies, ~200 MB). The service still has to be started separately (tsm serve --http 28795, or scripts/start_backend.sh in a checkout) — bundle activation fails loudly (failOnStartupError: true) when the MCP endpoint is unreachable.
• dsh is alpha; this package is verified against @deepseek-ai/dsh@0.2.0-rc.2 (profile composition, referenced plugin packages, the client-half contract and headless --json were all checked). The 500-question benchmark run predates that verification and was produced on 0.1.7-alpha.1 — results were not re-run. Re-check patch rows with --dump-config after any upgrade.
• The semantic backend is out of scope: without it, activation fails (failOnStartupError: true).
• Default model and API credentials are profile-level concerns (dsh_dlr/ patches + .env), not part of the bundle.

Dev Note

Patch rows merge by id, and a config block replaces wholesale on override — restate any field you want to keep. Unmatched ids only warn, so verify edits with dsh --profile web --dump-config.

preset-dlr reads its skills directory from DLR_SKILLS_DIR, which this bundle's node half sets at boot (derived from the package location, never a machine-specific path); launchers may override it. Keep the dlr-status row enabled if you rely on preset-dlr's skill loading, or set the variable yourself.

The repository carries the dsh-plugin GitHub topic for the DSH plugin directory (deepseek1024.com / 1024 Store); the directory lists the package via its static validation (dsh.bundle.patch + committed patch files), and the package is published on npm (dsh-tsm-agent@0.2.1). Install straight from npm: dsh plugin --profile <p> add dsh-tsm-agent (the directory page's auto-populated install command depends on an upstream crawler that is currently stalled).