Files
deepseek-harness/examples/acp-agent
creatixchu 340c9ba576 feat(web): split context injection into producer-declared forms
Injected context reached the transcript as one anonymous shape whose
expanded body was the whole message serialized as inline JSON, so the
escaping collapsed the only readable part — the model-facing prose —
into a single line.

`MessageSource` gains an optional producer-declared `form`, a small
semantic vocabulary of information shapes independent of `kind`: kind
says who produced the context, form says what shape it is, so several
producers may share one presentation. Two values ship.

`instructions` (workspace-context) lists the reconciled files above
their text and keeps the `<system-reminder>` framing verbatim, because
the framing is part of what the model read. `catalog` moves dsh-tool-skill
off the shared plugin kind onto a `skill-catalog` source carrying the
published name/description entries, and the body lists those instead of
re-parsing `<available_skills>` out of the prose. Catalog identity moves
with it: the republish digest now covers the durable entries, deleting
the text-slicing that recovered them from a logged message.

Everything else renders the opaque body — the model-facing text with its
real line breaks, then the remaining provenance as fields. That is the
documented default, not a leftover: a resumed, forked, or foreign log
must render whether or not its producer is mounted here, which is why
the classification lives in the durable source rather than a client-side
table keyed by producer.
2026-08-05 12:15:03 +08:00
..

acp-agent example

English | 中文

Automation-oriented Agent Client Protocol server over JSON-RPC stdio. It is intended for parent agents, subagent providers, and other programmatic clients, not as the product UI.

pnpm run demo:acp             # needs DEEPSEEK_API_KEY (repo-root .env or env)
pnpm run demo:code-mode       # same protocol with the Code Mode tool transport

The leaf loads the ACP app, DeepSeek adapter, sandboxed bash and filesystem stacks, one-shot approval policy, compaction, subagents, workflows, hooks, a derived session-query index, and repeat guard. The app creates one fresh agent per session/new, persists sessions to JSONL, and keeps stdout protocol-pure. session-query.cordis.yml explicitly opts into the workspace-authorized query tools and generic timeout/spill policies for their dedicated snapshot; fs.cordis.yml adds spill storage for filesystem scenarios, code-mode.cordis.yml adds run_code and its generated TypeScript SDK, and web.cordis.yml adds the web seam, the local fetch provider, web_fetch, and a loopback HTML fixture server for the web-fetch snapshot.

Protocol channel

Stdout carries only newline-delimited ACP JSON-RPC. @deepseek-ai/dsh-acp-demo installs no stdout logger; leaf additions must use stderr for diagnostics.

The automation contract — supported methods, baseline prompt content, committed-text output, and the intentionally absent UI surfaces — lives in @deepseek-ai/dsh-acp.

Session workspaces and permissions

Each session/new supplies an absolute cwd. Sandboxed bash and filesystem mutations resolve workspace-write against that session cwd, so concurrent sessions can use separate project roots; platform temporary roots remain shared writable scratch space (sandbox contract). DSH_PERMISSION_MODE selects workspace-write or danger-full-access for deployment and tests.

Under workspace-write, a model retry requesting wider sandbox access triggers session/request_permission with allow_once and reject_once. The client decides programmatically; dismissal or an unavailable answer fails closed. The selected outcome applies only to that retry and is recorded through the normal tool-result/audit path. The server never exposes a permission picker or persists client policy.

Snapshot tests

This example owns the ACP snapshot suite. It boots the real automation server, replays committed model streams through dsh-llm-replay, and compares both normalized protocol output and re-persisted session logs. Recording uses the real model; refresh reuses committed replay input. Overrides cover throw/hang behavior, and optional workspace/ fixtures seed world-state checks.

Most scenarios pin backend behavior rather than ACP-specific behavior; the automation-only ACP decision owns why that coverage remains transport-coupled.