Skip to content

Latest commit

 

History

History
61 lines (38 loc) · 3.85 KB

File metadata and controls

61 lines (38 loc) · 3.85 KB

dsh-director

Director Mode agent preset for DeepSeek Harness: two-layer delegation — a high-value model steers, low-cost models execute.

简体中文

What is this?

dsh-director is an agent preset for DeepSeek Harness (DSH) that implements Director Mode: it splits one agent session into strictly two layers

  • Director layer (high-value model): the main agent runs on the strong-reasoning model you pick in the session and handles only judgment, planning, requirement clarification, task decomposition, contract locking and acceptance — it never writes code itself.
  • Executor layer (low-cost model): all subagents (spawn / fork / workflow workers) are pinned to a low-cost model (default opencode-go / deepseek-v4-flash, reasoning MAX) via agentOptions and do the concrete work: implementation, batch edits, research, audits, tests.

Why a "two-layer protocol"

The two layers differ in more than model choice — each has its own hard-coded protocol:

Director protocol (top) Executor protocol (bottom)
Model high-value model chosen in the session low-cost model pinned by agentOptions
Duties judge / plan / decompose / accept execute directly / verify output / report
Forbidden writing the subagents' implementation code delegating, asking, or pestering the user
Tool layer full toolset + delegation tools delegation & asking tools stripped

Structural guarantees (what a persona alone cannot enforce, locked at the configuration layer):

  • maxDepth: 1 — only the Director (depth 0) may delegate; calls from depth ≥ 1 are rejected at spawn.
  • toolFilter.deny — removes subagent / subagent_fork / send_message / list_agents / interrupt_agent / workflow / ralph / ask_user_question from the executors' tool layer, structurally preventing sub-subagents and user interruptions from below.
  • Executor persona — an identity that only does the task, verifies its output, and reports.
  • Director persona — the full two-layer discipline: contract first, parallel delegation, self-contained prompts, no polling, personal acceptance, rework on failure.

Install

  1. Copy the two files under preset/ into ~/.dsh/.agent-presets/director/:

    mkdir -p ~/.dsh/.agent-presets/director
    cp preset/agent.cordis.yml preset/preset.yml ~/.dsh/.agent-presets/director/
  2. In the DSH Web UI, create a new session, pick the Director Mode preset, and select a high-value model for the session (e.g. deepseek-v4-pro, qwen3.7-max, GPT-5.x family).

  3. Start delegating: tell the Director your goal — it will present a plan first, delegate to subagents in parallel, and accept the results personally.

To change the executor model: edit the two agentOptions.model rows in preset/agent.cordis.yml (and provider if needed). To change the Director model: use the session model selector — never the preset.

Validation

docs/ ships with the full build & validation record:

  • docs/director-test-task.md — the ducheck validation task with an 8-point acceptance checklist (plan first / parallel delegation / self-contained prompts / model layering / tool sanitization / accept-and-rework / contract-driven output / cost layering). All 8 points passed in real runs.
  • docs/director-preset-notes.md — architecture decisions mapped to DSH mechanisms, lessons learned, model price cheat-sheet.

Cost layering

All executor tokens are billed at flash rates; the Director spends premium tokens only on planning, integration and acceptance. In practice: all coding work is done by low-cost models, and the interrupt-and-rework loop (via send_message) is a designed-in mechanism that keeps low-value models in bounds.

License

MIT