dsh-anchored-standard-1024

中文说明

An experimental DeepSeek Harness agent preset that bootstraps the first model request with the Minimal-aligned system prompt, a two-tool schema (bash + read), a 1024-token output cap, and no auto-injected workspace or skill context — then exposes the complete Standard tool catalog after the first durable tool call or reply.

This is a community project. It is not an official DeepSeek preset and is not affiliated with or endorsed by DeepSeek.

Why

DeepSeek V4 Pro conditions strongly on the API-visible tool catalog and on the first request's shape. A full 25-tool Standard opening and a large output budget often push the model into a "survey everything first" trajectory; the Minimal opening anchors a "pick up the tool and start" trajectory but keeps only two tools forever.

This preset separates initial trajectory selection from later tool use:

1. Keep the Minimal-complete system prompt.
2. First request = bash + read only, maxTokens = 1024, and the auto-injected skill catalog / workspace-instruction digest is stripped.
3. After the first durable signal (tool/call or assistant/message), the full Standard catalog and the normal output budget return and never re-narrow.

Key differences from the upstream xiaobright/dsh-anchored-standard:

Mechanism

request #1                        request #2+ (after promotion)
─────────────────────────────     ─────────────────────────────
system: Minimal-complete          system: unchanged (complete)
tools: [bash, read]               tools: full Standard catalog
maxTokens: 1024                   maxTokens: previous header value
skill catalog: stripped           skill catalog: injected normally
AGENTS.md digest: stripped        AGENTS.md digest: injected normally

Phase is derived from durable session events, so resume/reload preserves it. All three filters (catalog, budget, context) fail open: any filter error degrades to the full catalog instead of breaking the session.

Evidence (n = 1, honest disclosure)

Local experiment, 2026-08-16, deepseek-v4-pro, identical real task ("invest architecture root-cause research + remediation plan") run in two fresh sessions — Standard preset vs this preset — then scored blind (anonymized full-text review, reviewer did not know which document came from which preset):

Read this honestly: the anchored run scored higher and found three more real mechanism defects, but it also used more tool calls. This is a single sample on a single task type — evidence that the preset works as designed, not proof that it is universally better. Try it on your own workload and judge for yourself.

Install

Clone this repository, then copy the preset/ directory into the user preset root under the id anchored-standard-1024:

git clone https://github.com/sishuai33-commits/dsh-anchored-standard-1024.git
mkdir -p ~/.dsh/.agent-presets
cp -R dsh-anchored-standard-1024/preset ~/.dsh/.agent-presets/anchored-standard-1024

Restart DSH, open a new session, and select Anchored Standard (1024) in the session preset picker. You can verify the first request in the session log:

# inside a session dir: <session>.jsonl.zstd
# decompress and inspect the first request/header:
#   config.maxTokens == 1024 and tools == ["bash", "read"]

Uninstall:

rm -rf ~/.dsh/.agent-presets/anchored-standard-1024

Verify (three layers, most reliable first)

1. Unit tests — 14 deterministic tests for the three bootstrap filters, no model endpoint, runs in CI on every push:

   npm test

2. Official composition validation — stages the preset in a temp root and runs the harness's own discoverPresets validation (same code path the preset picker uses; catches broken rows / missing files). Requires a local DeepSeek Harness install:

   npm run verify   # node verify/check-preset.mjs
   npm run check    # syntax + unit tests + composition validation

3. Real-endpoint smoke (manual, optional) — install the preset, open a new session, then inspect session.jsonl.zstd: the first request/header must have config.maxTokens == 1024 and tools == ["bash","read"]. CI does not run this layer because it would need model credentials.

Files

preset/
  preset.yml             display metadata
  agent.cordis.yml       composition (persona, tools, skills, planning, ...)
  tool-bootstrap.mjs     the three first-request filters
test/
  tool-bootstrap.test.mjs

License and attribution

MIT. This preset is derived from xiaobright/dsh-anchored-standard (MIT) — see NOTICE. DeepSeek Harness and its presets are © their respective owners; this project is independent and experimental.