dsh-generation

dsh-generation

English | 中文

A DeepSeek Harness plugin that forks agent presets and runs a task on the next generation.

It is make, not a compiler. Creator mode is already the compiler: it can inspect the runtime, edit files, and author presets. This plugin only records a lineage — copy a known-good preset, let the meta agent edit that copy with ordinary fs / bash, then start a new working session on it.

Compatible with DeepSeek Harness 0.1.2+ (dsh-v0.1.3-alpha.2 and current master). It needs Session.snapshotEvents(), async ctx.agents.create({ sessionId, meta, signal, setup }), Host-global ctx.tools.register, and dsh.bundle.patch. Do not add @deepseek-ai/* to this package — a second copy of those modules crashes tool dispatch.

Install

dsh plugin --profile web add github:goecho/dsh-generation

Then restart dsh --profile web so the host profile reloads. Pin a SHA from the commit history after you have read the source. The package is plain JavaScript, so git installs do not need a prepare build allowlist. Plugin 0.2.0 is the line that talks to snapshotEvents() and passes the parent AbortSignal into agents.create.

Then open a Creator mode (cordis) session — or a copy of that preset whose composition still has a plugin row name: '@deepseek-ai/dsh-tool-cordis'.

dsh plugin add installs into the host profile, so the two tools are registered globally. This plugin then hides them from every non-Creator session (including user copies of standard / minimal / code) and from every agent with origin: subagent, including the workers generation_run starts. Execute still refuses a non-Creator caller.

This repository already has the dsh-plugin, dsh, and deepseek-harness GitHub topics.

Why

DeepSeek Harness splits composition into a Host plane (sandbox, model route, persistence) and an Agent plane (tools, persona, prompt). A session’s preset is locked once a turn has run, so you cannot hot-swap the working agent’s toolset without stranding logged tool calls.

The useful bootstrap is the same as a C compiler: each generation is a new binary. Stage N writes files for stage N+1; you do not patch the process that is currently compiling.

What it is not

• Not a replacement for Creator mode (cordis preset) or cordis_define / cordis_run
• Not a way to recompose the current session, or to give an in-process subagent a different preset
• Not a YAML/JS generator, scorer, or auto-promoter of the default creator
• Not a Host-plane kernel: it does not add services, session event types, or sandbox backends
• It does not depend on @deepseek-ai/dsh-tools / @deepseek-ai/dsh-llm. Tools are registered as raw definitions so a second copy of those packages cannot split the host’s module identity.

Design

Four rules:

1. The kernel stays assembly. No new Host services, no new SessionEvent types, no unlocking agent-preset-locked.
2. A generation is a new process of composition. generation_run uses agents.create + agentPresets.mount(id), never recompose on the meta session and never composeFrom (in-process children inherit the parent preset).
3. This plugin is make. It does not write agent.cordis.yml. Editing uses tools the creator preset already has.
4. A human is stage3. fork and run go through tools/pre-execute ask. Nothing is promoted to the default creator automatically. Lineage is the meta session log (tool/call / tool/result), not a second database.

meta session  (cordis + this plugin)
  → generation_fork     copy a preset to a new id
  → fs / bash           edit the copy (not this plugin)
  → generation_run      new agent, mount that id, run the task, dispose
  → summary             back into the meta log
  → fork again or stop

Working generations should fork from standard, minimal, or code.

Tools

generation_fork

Returns { ok, id, from, purpose, path, compositionPath } where path is the preset directory.

generation_run

Refuses a worker whose composition still has a plugin row name: '@deepseek-ai/dsh-tool-cordis'. Inherits the meta session’s workspace cwd, records origin: subagent and parentSession so logs chain, and passes the parent tool AbortSignal into agents.create so cancelling the meta tool also aborts worker creation. It does not dump the full worker transcript into the meta context. Worker logs are read with snapshotEvents(). If the worker never goes idle, it is cancelled after 15 minutes (stopReason: "timeout"). Domain failures keep { ok: false } but render as ERROR: so the model notices.

Returns { ok, sessionId, presetId, stopReason, toolsUsed, lastAssistantText }.

Trust

Treat a Creator session that can fork and run generations as shell access. Mitigation in v1:

• Human approval on every fork and every run
• Worker preset cannot be cordis, and cannot still carry a @deepseek-ai/dsh-tool-cordis plugin row
• Edits belong under the new preset directory
• Model-written JavaScript is never auto-mounted via cordis_run

Develop

npm test

There are no @deepseek-ai/* runtime dependencies. The tests mock ctx.tools / ctx.agentPresets / ctx.agents. GitHub Actions runs the same npm test on Node 22.

License

MIT