dsh-team-toolkit (team-ext)
dreamperday/dsh-team-toolkit/packages/team-ext
DSH 智能体团队的纯观测审计:把每请求缓存用量与提示前缀重写事件落成 JSONL,不加工具、不改提示词。
安装
dsh plugin --profile web add github:dreamperday/dsh-team-toolkit该插件位于仓库的 packages/team-ext 子路径下。
README
dsh-team-toolkit
Two small, additive plugins for DSH Agent Teams (DeepSeek Harness). Neither plugin edits a shipped package, and neither changes the model-visible request unless you explicitly call its tool.
| Package | What it does | Model surface |
|---|---|---|
packages/team-compact | Manual context compaction for the Lead and every teammate: a scoped compact_context tool plus the /compact-agent command. | Adds one tool + one human command |
packages/team-ext | Observation-only audit: records per-request cache usage and every prompt-prefix rewrite (compaction, prune, series reset) to JSONL, with an analysis script. | Adds no tool and no prompt section |
Both are Cordis bundles (dsh.bundle manifest) and install into a DSH profile without touching
resources/app.
Install
Requirements: DSH Desktop with the experimental Agent Teams profile
(@deepseek-ai/dsh-experimental-agent-team-profile), i.e. the Lead gets spawn_teammate and
teammates exist.
From a local clone (recommended while the packages are not in a registry):
git clone https://github.com/DreamPerday/dsh-team-toolkit.git
# in DSH: ask the agent, or call the plugin manager, with the absolute package directory
plugin_manager action=install_bundle target="<clone>\packages\team-compact"
plugin_manager action=install_bundle target="<clone>\packages\team-ext"
Both packages are link-installed, so the clone must stay where it is. Restart DSH Desktop after
the first install if your profile does not hot-reload plugin modules (host plugin code is cached
by the running process; config changes hot-reload, module changes need a restart).
Remove with:
plugin_manager action=remove_bundle target="dsh-team-compact"
plugin_manager action=remove_bundle target="dsh-team-ext"
Why these exist
While running a real multi-agent workload on DSH 0.2.1, every teammate compaction failed with a
400 because auxiliary LLM calls without an explicit reasoning effort fell through to
reasoning.effort: "none" on an OpenAI-compatible reasoning model. The audit plugin exists to make
that class of failure visible (per-request cache fields plus prefix-rewrite events), and the
compact plugin exists because the official compaction seam only compacts the whole compactable
range, so a manual, targetable entry point is useful when an agent's context has grown large.
Patches (not plugins)
patches/team-model-selection — an in-place patch for DSH
Desktop 0.2.1-rc.2 that adds optional provider / model / reasoning_effort to spawn_teammate,
so the Lead can pin each teammate to a different LLM route. It has to edit the installed bundles
(the tool schema and the team service expose no extension point), so it is version-pinned, fails
loudly on anchor mismatch, and must be re-applied after a DSH update.
Compatibility
Written against DSH 0.2.1-rc.2 host services (agents, commands, agentTeams,
tokenMeter, compaction, toolResultPruner). The experimental Agent Teams packages are
explicitly unstable upstream; check the two package READMEs for the exact services they use.
License
MIT