Jarad-z/dsh-goalmesh

English · 简体中文

Goal-driven multi-agent orchestration for DeepSeek Harness.
Turn one objective into a bounded agent graph—and keep every decision, dependency, and result observable.

Why · How it works · Capabilities · Protocol · Development

One goal in. A coordinated mesh out.

DSH GoalMesh is a DeepSeek Harness Plugin that exposes one model-facing tool: goal_mesh. It validates a declarative task DAG, launches bounded specialist agents, propagates typed dependency results, supports nested decomposition, and renders the full execution trajectory in DSH Web.

[!NOTE] GoalMesh is a DSH Plugin, not a Codex Plugin. It runs inside the DeepSeek Harness Profile and follows Cordis Entry/Fiber lifecycle ownership.

Why GoalMesh

• Goal fidelity — local work remains anchored to the same success criteria and constraints.
• Structured concurrency — admission, deadlines, cancellation, failure propagation, and cleanup are bounded by one coordinator.
• Dependency-aware execution — tasks start only after prerequisites settle, with explicit fail, skip, or partial behavior.
• Safe recursive delegation — child agents can open nested meshes through attempt-fenced, child-scoped leases.
• Stable typed results — every child returns a TaskReport; task identity and input order remain deterministic.
• Durable observability — live events and replay fold into the same Web task tree, including trusted child-Session navigation.

How it works

%%{init: {"theme": "base", "themeVariables": {
  "primaryColor": "#eff6ff",
  "primaryTextColor": "#0f2c52",
  "primaryBorderColor": "#60a5fa",
  "lineColor": "#60a5fa",
  "secondaryColor": "#ffffff",
  "tertiaryColor": "#f0f9ff",
  "fontFamily": "Inter, ui-sans-serif, system-ui"
}}}%%
flowchart LR
  G["Global goal<br/>criteria + constraints"] --> T["goal_mesh"]
  T --> C["GoalMesh Coordinator<br/>single-writer owner"]
  C --> D["DAG Scheduler<br/>capacity + policy + deadlines"]
  D --> A1["Agent A<br/>research"]
  D --> A2["Agent B<br/>build"]
  D --> A3["Agent C<br/>verify"]
  A1 --> R["Typed TaskReports"]
  A2 --> R
  A3 --> R
  R --> C
  C --> E["Durable trajectory"]
  E --> U["DSH Web task tree"]
  A2 -. "bounded nested mesh" .-> D

The coordinator sits above root and child agents. Children never receive the mutable run ledger; they operate through revocable leases, so siblings cannot mutate one another's state. Cordis fibers own every tool registration, listener, service, and live resource from mount through disposal.

Capabilities

Package layout

dsh-goalmesh/
├─ packages/
│  ├─ goalmesh-plugin/       # bundle + cordis.patch.yml
│  ├─ tool-goalmesh/         # orchestration runtime
│  └─ client-ui-goalmesh/    # durable trajectory UI
├─ docs/                     # architecture and execution contract
├─ harness-patches/          # minimal public Harness prerequisites
└─ tests/                    # coordinator, nesting, replay, UI, invariants

The split follows DeepSeek Harness ownership boundaries: Host scheduling works in a headless Profile, while the installable bundle composes Host, invariant, and Web entries.

Tool protocol

A root invocation declares one goal and an invocation-local task graph:

{
  "goal": {
    "statement": "Ship a release-ready API migration plan",
    "success_criteria": [
      "Every breaking change has an owner",
      "Rollback and verification steps are explicit"
    ],
    "constraints": [
      "Preserve backward compatibility during rollout"
    ]
  },
  "tasks": [
    {
      "key": "surface",
      "description": "Map the public API surface",
      "objective": "Identify every affected endpoint and consumer",
      "acceptance_criteria": [
        "The inventory is complete and evidence-linked"
      ]
    },
    {
      "key": "rollout",
      "description": "Design the rollout",
      "objective": "Produce staged migration and rollback steps",
      "acceptance_criteria": [
        "Each stage has a measurable gate"
      ],
      "depends_on": ["surface"],
      "dependency_failure": "partial"
    }
  ],
  "failure_mode": "collect_all"
}

Nested invocations omit goal. Ownership comes from the scoped lease captured by the Host tool—never from model-provided IDs.

Development

GoalMesh v0.3 targets DeepSeek Harness 0.1.0-rc.5 at public baseline 47f943859bef60e4160492346772ded9b24f765a, plus the minimal runtime patch in harness-patches/goalmesh-prerequisites.patch.

[!IMPORTANT] Keep deepseek-harness and dsh-goalmesh as sibling directories. Workspace dependencies intentionally resolve against that layout.

mkdir goalmesh-dev && cd goalmesh-dev
git clone https://github.com/deepseek-ai/deepseek-harness.git
git -C deepseek-harness checkout 47f943859bef60e4160492346772ded9b24f765a
git -C deepseek-harness submodule update --init --recursive
git clone https://github.com/Jarad-z/dsh-goalmesh.git
git -C deepseek-harness apply ../dsh-goalmesh/harness-patches/goalmesh-prerequisites.patch
cd dsh-goalmesh

corepack enable
corepack install --global pnpm@11.7.0
pnpm --dir ../deepseek-harness install --frozen-lockfile
pnpm --dir ../deepseek-harness run build:lib
pnpm install --frozen-lockfile
pnpm check

Useful commands

Build artifacts are emitted under packages/tool-goalmesh/lib/ and packages/client-ui-goalmesh/lib/. The ready-to-compose Profile patch lives at packages/goalmesh-plugin/cordis.patch.yml.

Design guarantees

• The root tool returns only after its invocation and owned resources settle.
• Unknown model fields and forged ownership identifiers are rejected.
• The global goal is read-only; children report only against local task goals.
• Coordinator transitions are serialized and checked by runtime invariants.
• Durable UI events are observational; the Web client never becomes a second scheduler.
• Provider removal stops new admission without abandoning already-owned cleanup.

Read the architecture contract for the complete design. Version boundaries and implementation order are recorded in the execution plan.

Contributing

Contributions are welcome. Start with CONTRIBUTING.md and run pnpm check before opening a pull request. Report security issues through GitHub's private vulnerability reporting flow described in SECURITY.md.

Released under the MIT License.
_{Built for explicit goals, bounded agents, and inspectable outcomes.}