deepseek-harness/packages/ui/stdio-agent/README.md

@deepseek-ai/dsh-stdio-agent

The terminal stdio chat app: a Cordis app plugin that composes the default agent spine (@deepseek-ai/dsh-agent-core) with the front-door cluster a terminal chat needs, and a bin that boots a leaf cordis.yml.

It is the readline counterpart to @deepseek-ai/dsh-acp-agent: both consume the same spine, but each bakes in the OPPOSITE front-door cluster.

What it bakes in

A terminal chat always wants the same cluster, so the package owns it rather than trusting each leaf to re-wire it:

Plugin	Why it is here
@cordisjs/plugin-logger-console	the console logger — stdout is just the terminal here, so logging to it is correct (the ACP app must NOT have this)
@deepseek-ai/dsh-agent-core	the spine, pre-creating one agent under the main config label from this app's model, with process.cwd() as the fresh session cwd and carrying its persona
@deepseek-ai/dsh-session-persistence-jsonl	durable JSONL session log under persistenceRoot
@deepseek-ai/dsh-user-interaction	the human question/answer seam used by confirmation tools
@deepseek-ai/dsh-tool-ask-user	the model-facing ask_user_question tool
stdio-chat (in-package module)	the readline UI, holding the app-owned agent object directly and rendering it as main

@cordisjs/plugin-hmr (the dev/demo edit-reload loop) is deliberately a leaf entry, NOT baked in here: it is a Loader-only, subprocess-only dev plugin — its constructor throws without node --expose-internals + a live loader, and the in-process test tier cannot even import it (so a package whose apply statically pulled it in could never carry the per-file coverage gate). Unlike the console logger, a stray hmr is not a stdout-purity footgun, so leaving it at the leaf costs no safety. The demo:echo / demo:repl leaves load it and pass --expose-internals.

The leaf cordis.yml supplies only the swappable backends — an LLM adapter (llm-deepseek for the real model, or the mock mock-llm for a demo) and a bash executor (bash-local) — hmr, plus this app's Config. The whole plugin tree a run loads is therefore: this app's cluster, the spine inside agent-core, hmr, and the two leaf backends.

Config

Key	Default	Routed to
model	(required)	the pre-created agent's model
persona	—	the deployment persona template (may reference {{model}}/{{cwd}}), routed to dsh-system-prompt
toolOrder	—	explicit model-facing tool order (a name list with one '<unlisted-tools>' rest entry; absent — lexicographic; an unregistered name fails each turn at prompt assembly), routed to dsh-system-prompt
tools	{ mode: 'native' }	tool-registry presentation config (native / code / both), routed through dsh-agent-core
skills	owner defaults	registry-cache, local-provider, and model-facing skill-tool config, routed through dsh-agent-core
persistenceRoot	./.sessions	the JSONL backend's root directory
welcome	ready.	the stdin-chat banner
resumeSessionId	—	resume a persisted session id instead of starting fresh (sourced from an env var in the leaf)

Fresh stdio sessions use the process launch directory as session.header.cwd and mint one combined main-session-<uuid> agent/session id, so durable restarts cannot collide. The app passes that exact opaque id to both its config-created agent and UI; an AgentLoop-only reload resumes materialized history under that id, while the UI's main text remains only a display label and never selects another registry root by prefix or insertion order. Readline buffers nonblank startup input for that identity until agent/session-start, so piped stdin cannot outrun asynchronous exact-id restoration or let EOF discard the queued prompt; agent-loop/config-start-failed instead drains and reports buffered input so a missing or corrupt persisted session cannot hang EOF. A resumed run binds both components to the exact resumeSessionId and keeps the cwd stored in the persisted session header.

The bin

dsh-stdio-agent [path-to-cordis.yml] (default ./cordis.yml) loads a gitignored .env from the cwd (DEEPSEEK_API_KEY / DEEPSEEK_BASE_URL), then drives the cordis Loader against the config and awaits the whole plugin tree before returning. Run it under node --expose-internals: the cordis Loader resolves the config's bare plugin specifiers (@deepseek-ai/dsh-*, npm packages) through its internal module loader, which is only active under that flag. The demo:echo / demo:repl scripts invoke it that way.

Example leaf cordis.yml

# A REPL agent demo: hmr + the DeepSeek adapter + local bash, then this app.
- id: hmr
  name: '@cordisjs/plugin-hmr'
  config:
    root: ['.']
- id: llm-deepseek
  name: '@deepseek-ai/dsh-llm-deepseek'
  config:
    apiKey: !!js process.env.DEEPSEEK_API_KEY
    models: [deepseek-v4-flash]
- id: bash
  name: '@deepseek-ai/dsh-bash-local'
  config:
    timeoutMs: 60000
- id: stdio-agent
  name: '@deepseek-ai/dsh-stdio-agent'
  config:
    model: deepseek-v4-flash
    persona: 'You are a coding assistant powered by the {{model}} model.'

Swap llm-deepseek for a mock-llm leaf plugin and you have the echo demo — "swap the backend, keep the app".

Model Experience

Composed terminal agent request

What the model sees: Through dsh-agent-core, the main agent receives the harness identity, configured persona, skill catalog, and visible tools; this app also composes the generated ask_user_question schema. Each readline submission becomes a user message.

Token effect: Child prompt and schema costs repeat per request; user input and tool history grow until compaction. The welcome banner, logger output, and rendered transcript are terminal-only and add zero model tokens.

Human-answer result

What the model sees: Through dsh-tool-ask-user, successful terminal answers use that package's exact compact JSON shape. Interruption becomes exactly Error: ask_user_question was interrupted before the user answered; a closed stdin becomes Error: ask_user_question cannot be answered because stdin is closed.

Token effect: Only a completed or failed tool call adds retained result tokens; prompts printed while waiting are terminal-only.

Known Limitations and Deferred Work

• One pre-created main agent drives the readline UI — there is no multi-session or concurrent-agent surface in this app; a run is one conversation.
• The front-door cluster is fixed in code — the JSONL persistence backend and the ask-user tooling are baked; a different composition is a leaf-level sibling entry or another app package.
• The question tool is not an approval answerer — this app mounts user-interaction and ask_user_question, but not ctx.approval; a tools/pre-execute ask therefore fails closed unless the leaf composes an approval service and terminal answerer.