Skip to content

DeepSeek Harness ​

Post-synthesis product case study. Evidence was retrieved 2026-08-20 from DeepSeek Harness release dsh-v0.1.0-rc.8, commit 141eb6fef83422698aef7a981029e843e8161534. Every upstream source link below is pinned to that commit. The TrogonAI callability check used repository commit c8e05872b9a3b156b974d9773b9723f07493cb1d.

ACP status and version ​

DeepSeek Harness has a native ACP agent implementation. The @deepseek-ai/dsh-acp package describes itself as an automation-only JSON-RPC stdio server and is version 0.1.0-rc.8; it pins @agentclientprotocol/sdk 0.25.1 (packages/acp/acp/package.json:2-4, packages/acp/acp/package.json:34-36). That SDK defines PROTOCOL_VERSION as wire version 1 (typescript-sdk/src/schema/index.ts:310), and the server returns that constant during initialize. The wire-reported agent identity is independently hard-coded as deepseek-harness-acp version 0.0.1, so it must not be mistaken for the package release version (packages/acp/acp/src/index.ts:290-301).

This is deliberately not an editor integration. Its supported surface is for programmatic clients, while navigation, transcript replay, commands, modes, elicitation, reasoning, plans, titles, and tool presentation remain outside the ACP package (packages/acp/acp/README.md:5-7).

Capabilities and session lifecycle ​

SurfaceBehavior at the pinned release
TransportNewline-delimited JSON-RPC over stdin/stdout. Stdout is reserved for protocol frames.
PromptsText and resource-link text are supported. Inline raster images are advertised only when a durable attachment store and the exact provider/model route support them. Audio and embedded context are false.
Agent capabilitiesNo load-session, editor, terminal, filesystem, or MCP capability is advertised.
AuthenticationauthMethods is empty and authenticate is a no-op. Provider credentials are a process-launch concern, not an ACP authentication exchange.
Session creationsession/new creates a fresh harness agent and requires an absolute cwd. Non-empty additionalDirectories or mcpServers are rejected.
Prompt concurrencyOne prompt may be in flight per session. The response waits for admission, whole-agent idle, and ordered committed output delivery.
OutputOnly committed assistant text and images become agent_message_chunk updates. Reasoning, tool activity, plans, and live deltas remain in the harness session log.
Cancellationsession/cancel aborts prompt admission or cancels the addressed agent after admission. Unknown session ids are no-ops.

The capability and lifecycle rows follow the package's method-by-method contract (packages/acp/acp/README.md:20-34). One connection may own several independent sessions, but the implementation is fresh-session only: load, list, resume, delete, fork, and per-session close are not implemented. Disconnect or plugin disposal cancels and drains all agents owned by that connection before releasing them (packages/acp/acp/README.md:36-40, packages/acp/acp/README.md:76-81).

Permissions and trust boundary ​

DeepSeek Harness self-serves its tools inside the child process instead of requesting client filesystem or terminal services. When a bridge-owned tool approval has a tool-call id, the ACP agent sends one allow_once and one reject_once option. A cancelled response fails closed, and no durable grant is inferred (packages/acp/acp/src/index.ts:268-285). The runnable example selects workspace-write or danger-full-access through DSH_PERMISSION_MODE; under workspace-write, the client decides each wider retry and the server does not expose a picker or persist client policy (examples/acp-agent/README.md:20-24).

This boundary is narrower than the common editor-host pattern. A compatible client may advertise no optional capabilities because the child owns its filesystem, terminal, model route, and tools. The host still owns the final permission answer and the OS process lifetime.

Process ownership, invocation, and provider authentication ​

The ACP package attaches an AgentSideConnection directly to the harness process's stdin and stdout (packages/acp/acp/src/index.ts:443-448). Therefore the ACP client host, not the agent package, must spawn and supervise that process. The pinned release documents this exact source-checkout invocation:

sh
DEEPSEEK_API_KEY=... pnpm --dir /path/to/deepseek-harness run demo:acp

The script boots the repository's composed ACP example, and its stdout carries only protocol frames (package.json:141-145, examples/acp-agent/README.md:5-16). DEEPSEEK_API_KEY authenticates the harness to the model provider. It does not authenticate the ACP peer. The ACP package exports library entry points and declares no executable, while the tagged documentation gives the composed repository demo as its runnable surface (packages/acp/acp/package.json:13-31, packages/acp/acp/README.md:42-44). A production host would therefore need a packaged profile or an equivalent composition rather than treating the package export itself as a binary.

Relationship to the ACP subagent provider ​

The same repository also implements the client side in @deepseek-ai/dsh-subagent-acp. This is not another server mode. It is an out-of-process subagent provider that spawns one fresh ACP child for each run, then performs initialize, session/new, and session/prompt. It derives the child cwd from the parent session unless explicitly overridden and owns the child's cancellation, stdin closure, termination escalation, and whole-tree exit proof (packages/subagent/subagent-acp/README.md:5-17).

The provider advertises no optional client capabilities, auto-answers permission requests according to its allow or reject policy, collects only committed agent_message_chunk text, and gives the remote child a fresh context with no parent conversation (packages/subagent/subagent-acp/README.md:19-34, packages/subagent/subagent-acp/README.md:64-84). It is the first-party reference client for the automation server, and it proves the intended process boundary. It does not make the TrogonAI repository capable of hosting DeepSeek Harness because that TypeScript provider is not wired into TrogonAI's Rust/NATS runtime.

Channel mapping ​

No messaging or editor channel is mapped through this ACP surface. The package explicitly leaves interactive presentation and human questions to separate Web host and client modules. Its in-repository consumer is machine-to-machine subagent delegation, so this is evidence for ACP as an internal execution boundary rather than ACP as a channel gateway (packages/acp/acp/README.md:5-7).

Callability from TrogonAI today ​

Verdict: wire-compatible target, not callable end to end at the checked commit. Both sides use ACP wire v1, and DeepSeek Harness's required core method subset is representable by TrogonAI's pinned Rust SDK. No static protocol-version blocker is visible. The blocker is process hosting:

  • acp-nats-stdio accepts ACP from a client on its own stdin/stdout and forwards it to a NATS-backed agent. It occupies the agent-facing direction, not the client host direction (rsworkspace/crates/acp/acp-nats-stdio/README.md:1-18, rsworkspace/crates/acp/acp-nats-stdio/src/main.rs:38-41,83-89).
  • The ACP crate family contains no child-process spawn call at the checked commit. The current decision record likewise identifies the missing component as a client host that spawns an ACP CLI, speaks the client role, and bridges the session onto NATS (docs/research/acp/decision-record.md:33-49).
  • Provider credentials must be supplied at process launch, and the exact upstream runnable is the repository demo command above. An eventual host must also accept a self-served agent that advertises no client filesystem or terminal callbacks and must send empty mcpServers and additionalDirectories.

The minimum integration is therefore the planned acp-host process boundary, a deployable DeepSeek Harness ACP composition, an injected DEEPSEEK_API_KEY, an absolute session cwd, and an explicit permission policy. Until that host exists, documenting the command does not make the product callable from TrogonAI.

Design lessons ​

  • Copy the explicit automation profile: a small negotiated surface, committed output only, one in-flight prompt per session, and connection-scoped teardown make lifecycle ownership auditable.
  • Support agents that self-serve filesystem and terminal operations. A host must not require every ACP agent to call client-owned fs or terminal methods.
  • Keep ACP peer authentication distinct from provider credentials. Empty ACP authMethods does not remove the need for controlled credential injection at process launch.
  • Preserve the policy decision in the host. The child can request a one-shot permission, but a headless client still needs a fail-closed rule for choosing or rejecting it.

Source manifest ​

Released under the Apache License 2.0.