DSH Tool Discovery

Keep frequent tool inputs complete. Load complex parameter details only when needed.

A plugin for DeepSeek Harness and its TypeScript Programmatic Tool Calling (PTC) runtime. PTC lets an agent call several tools from one program. This plugin keeps useful signatures in the model's tool SDK and places detailed documentation in conversation history on demand.

中文 · Design · Evaluation · Reproduce

How it works

1. Small schemas and tools listed in alwaysInline keep complete parameters in the SDK.
2. Other large schemas retain the tool name, description, top-level argument names, types, required fields, and return type. Nested details are deferred.
3. The model calls tool_help with exact tool names or a configured group. Complete schemas return as tool results, without expanding the static SDK.
4. Execution still uses native tool registration, scope checks and full parameter validation. Documentation lookup grants no permissions.

presentation: selective
deferSchemaBytes: 1024
alwaysInline:
  - mcp__excel__format_range
  - mcp__excel__set_sheet_layout

// The model reads the returned definition in its next step.
console.log(await tools.tool_help({ names: ['mcp__excel__create_chart'] }))

A fixed SDK avoids changing its prefix during documentation discovery. It does not guarantee provider cache hits. The host chooses frequent tools before execution; there is no query-keyword router.

Recorded engineering result

On 80 selected successful executions across four spreadsheet tasks, selective documentation with two frequent tools kept inline used 6.5% less estimated model cost and 3.1% fewer model requests, with essentially unchanged mean elapsed time.

Metric	Full PTC	Selective PTC with frequent tools inline	Change
Selected successful executions	40	40	—
Usage-based estimated cost	¥6.7084	¥6.2718	−6.5%
Model requests	416	403	−3.1%
Mean elapsed time	63.92s	63.83s	−0.1%
Non-cached input tokens	1,385,262	1,217,760	−12.1%

Both configurations use deepseek-flash, high reasoning, a 65,536-token output limit, and 78 real MCP tools. Nine complex tool inputs use deferred details in the candidate. Offline SDK size is 92,682 → 72,587 UTF-8 bytes (−21.7%); this is not a token measurement. Two tasks are adapted from public Toolathlon tasks; two are derived tasks using its public spreadsheet data.

The table takes the first successful execution for each task, strategy and repeat. It measures resources for successful execution, not one-shot success probability or all-attempt spending. All 87 numeric attempt records are retained. These four development tasks do not establish general superiority or an official Toolathlon score. Costs are calculated from usage at frozen rates, not an invoice. See selection method and complete accounting.

Try it

This source release is not published to npm. Use a built DSH checkout at d743267388641bc76f17c45ce8b4c231aed1d32c:

node scripts/bootstrap.mjs /path/to/deepseek-harness
npm test
npm run typecheck
npm run build

Mount lib/tool-docs.mjs as a Cordis plugin using TypeScript PTC. See configuration and setup.

Recompute the recorded numbers without a model, credentials or data downloads:

python3 -m unittest discover -s bench -p 'test_*.py'
python3 bench/report.py --check

Repository contents

Location	Purpose
src/	Scoped documentation lookup, pagination, minimum signatures and SDK rendering
tests/	Lifecycle, scope, signature, schema and frequent-tool checks
bench/	Public task retrieval, real tool servers, deterministic graders and fixed-schedule runner
results/	87 numerical attempt records, 887 request usages and the reproducible first-success selection
patches/	Shared MCP output metadata and spreadsheet inspection support

Upstream workbooks and original evaluators are downloaded at pinned revisions. Credentials, private tool collections, full conversations and generated workbooks are not distributed.

License

Project code is MIT. Upstream code and datasets retain their terms; see third-party notices.