deepseek-harness/examples/tui-agent

tui-agent

English | 中文

The full-screen interactive coding agent: DeepSeek V4, local bash and filesystem tools, compaction, subagents, workflows and fresh-agent Ralph iteration, plan mode (/plan enters and exit_plan_mode reviews the exit), timeout/spill policy, and JSONL persistence through @deepseek-ai/dsh-tui-demo, loaded from cordis.yml. The sibling headless-agent runs the same capability class as a one-shot pipe-friendly task, and acp-agent serves it over JSON-RPC.

Run it

# repo root .env (gitignored) or exported env:
#   DEEPSEEK_API_KEY=sk-…
#   DEEPSEEK_BASE_URL=https://…   # optional; defaults to the public API
pnpm run demo:tui

Both the demo script and the installable dsh CLI (apps/cli) boot this example's cordis.yml as the shipped default config; dsh additionally applies the personal overlay from ~/.dsh and uses the invoking directory as the workspace.

Type a coding task. The agent works through the read/write/edit filesystem tools for ordinary file operations and bash (+ the generic task_output / task_list / task_kill for background tasks) for shell commands, searches, and test runs, each in a fresh bash -c (the system prompt tells the model to pass workdir instead of cd). Both the fs tools and bash resolve relative paths against the session workspace. It can also delegate with subagent/subagent_fork.

The todo_write task tracker is opt-in and not in the shipped config: add @deepseek-ai/dsh-tool-todo to cordis.yml (or a personal-config overlay under ~/.dsh) to expose it. Once loaded, the model records a whole-list plan to the session log and the TUI renders it.

The TUI renders Markdown history, reasoning, tool-owned terminal/diff/generic cards, token totals, and — when todo_write is loaded — the latest plan. Long tool bodies keep a head/tail preview; Ctrl+O expands or collapses every card. Enter submits or steers while the agent runs, Ctrl+R toggles reasoning, Escape cancels, and /help lists commands. /plan selects plan mode for the next step; /plan <message> also submits the message into that step, while /plan off selects the default mode without model input. /status expands the current session's identity, activity counts, exact token/cache buckets, context use, and timestamps without interrupting a running turn. /model opens a keyboard selector for the current provider catalog; use Up/Down to focus a model, Shift+Tab to cycle its advertised reasoning efforts, and Enter to select, or use /model <model> and /model <provider>/<model> for direct selection. ask_user_question opens a wide bottom-left keyboard panel with batch progress and numbered options.

Resuming a prior session

Each run starts a fresh session by default (its event log lands under ./.sessions/). To continue a previous conversation, pass its id to the installed dsh CLI — the main agent then rehydrates the persisted log instead of starting fresh, so the model sees the earlier turns as history:

dsh --resume <prior-session-id>

/resume opens a searchable keyboard selector with titles, activity, last-turn results, model route, durable goal phase, and live/persisted state. The installed dsh host flushes and disposes the current app, then replaces the process with dsh --resume <id>. The TUI still prints that command on exit and shows it when a custom host cannot hand off. dsh --resume <id> provides the id on the boot context, which cordis.yml reads (resumeSessionId: !!js "typeof resumeSessionId === 'string' ? resumeSessionId : undefined"); with no flag the agent starts a new session. A missing or unreadable id starts no agent and emits agent-loop/config-start-failed: the TUI prints the failure and exits nonzero. The selector has no cross-process session lock, so deployments with concurrent hosts must coordinate session ownership separately.

Code Mode

code-mode.cordis.yml overlays the same tree with the worker-thread runtime and tools: { mode: code }. The model receives one run_code transport plus a generated TypeScript SDK for the visible tools; only program output returns to model context. Use mode: both to expose native calls alongside run_code. See the Code Mode Agent Note for the execution contract.

pnpm run demo:code-mode        # this overlay under the TUI (default UI)
pnpm run demo:code-mode acp    # the acp-agent example's same-shaped overlay

Try a task that spans several tool calls, e.g.:

Count the lines of every *.md file under docs/ and write the three largest to summary.txt.

and watch the transcript: one run_code call, a program looping over tools, and a result the model curated instead of five round-trips of raw tool output.

What each leaf entry demonstrates

This example is a thin leaf cordis.yml: it picks the swappable backends, loads one app package, and adds product tools that are intentionally outside the shared spine. The spine (sessions, system-prompt, tools, agents, invariants, agent-loop) and the front-door cluster (JSONL persistence, the pi-tui channel, the pre-created main agent) live inside the @deepseek-ai/dsh-tui-demo app and the @deepseek-ai/dsh-agent-spine-demo bundle it loads; the leaf wires the backends and model-facing optional tools:

Entry	Demonstrates
hmr (@cordisjs/plugin-hmr)	the dev/demo edit-reload loop — a leaf entry (not baked into the app) because it depends on the Loader's internal module access
llm-deepseek	the default native adapter
bash (dsh-bash-local)	the executor implementation — the swappable half of the bash seam. The model-facing bash schema (tool-bash) and generic task_* controls (tool-tasks) come from dsh-agent-spine-demo, so only the executor is a leaf choice
tui-agent (@deepseek-ai/dsh-tui-demo)	the app bundle: the agent-spine demo + JSONL persistence + the pi-tui channel + a pre-created main agent
subagent, subagent-spawn, subagent-fork	the subagent provider registry plus the two in-process backends: a fresh child and a child seeded with the parent's completed-turn prefix
tool-subagent, tool-subagent-fork	two model-facing dsh-tool-subagent loads, each bound to a different provider and exposed under a distinct tool name (subagent, subagent_fork)
workflow-workerthread, tool-workflow	the worker-thread workflow engine and its model-facing workflow tool, with child calls routed through the spawn backend
plan-mode	the plugin-owned /plan [message] entry and /plan off exit commands, plan-mode prompt policy, tool restrictions, and reviewed exit_plan_mode transition
fs-local, fs-policy, tool-fs	the filesystem stack: the local ctx.fs provider, the read-before-write/edit policy gate (on the fs/* event gate), and the model-facing read/write/edit tools. Relative paths resolve against the session workspace

End-to-end tests (pnpm run test:e2e)

The UI-independent with-key suites assemble the full stack programmatically through tests/harness.ts (no PTY, no Loader):

• tests/full-loop.e2e.ts — the canary: real model runs echo e2e-ok through the real bash tool; asserts tool/call/tool/result session events and the final answer.
• tests/coding-task.e2e.ts — the swebench-style smoke: a temp dir holds add.js (with a - b where a + b belongs) and a failing add.test.js; the agent must fix the bug and verify. The test re-runs node add.test.js ITSELF and inspects the files — agent claims are not trusted.
• tests/resume.e2e.ts — durable continuity across processes: run 1 tells the real model a secret code and persists the turn to a temp JSONL root, then the whole context is disposed; run 2 is a fresh context over the same root that RESUMES the session id and asks the model to recall the code. The recall can only come from the rehydrated log.
• tests/compaction.e2e.ts — the compaction smoke: a real multi-step bash task runs with a deliberately tiny context window so the auto-compaction listener fires MID-SESSION. Verifies the WORLD — a compact/start…end pair landed in the real log, the surface shrank (a replace node shadowed older nodes), and the agent still produced a correct final answer after compaction.
• tests/todo-write.e2e.ts — loads the opt-in todo_write tool, then a real model drives it and the test verifies the resulting todo/write session event.
• tests/code-mode.e2e.ts — the with-key Code Mode proof: a real model, a two-tool task, asserting the wire tool list was exactly [run_code], the tool/code-dispatch events landed under the parent call, and the curated answer came back.

These self-skip without DEEPSEEK_API_KEY. The keyless tests/tui-keyless-smoke.e2e.ts boots the real Loader tree in a PTY (the one sanctioned PTY surface): the base boot + /plan + /exit, a scripted-LLM conversation with a question dialog and tool round-trip, the Code Mode overlay welcome line, and the resume-failure exit path.

Snapshot tests

tests/snapshots/<scenario>/session.jsonl supplies recorded user prompts and model chunks; sibling child logs drive subagents and workflows. The keyless suite executes those scripts through the real loop and tool implementations, then compares readable expected terminal cell/style output. Use pnpm run test:snapshot:refresh for presentation-only changes and pnpm run test:snapshot:record with a DeepSeek key when a recorded model journey changes. The implemented TUI snapshot Agent Note owns the scenario matrix and the split between recorded journeys, transient package snapshots, and PTY coverage.