The full-screen terminal app bundle: a Cordis plugin that composes [`@deepseek-ai/dsh-agent-spine-demo`](../agent-spine-demo/README.md), persisted same-session goals, the human-command registry and `/goal` producer, JSONL persistence, keyboard-backed user interaction, a pre-created `main` agent, and [`@deepseek-ai/dsh-tui`](../../ui/tui/README.md). A `cordis.yml` mounts it as one entry; the [`dsh`](../../../apps/cli/README.md) CLI is the front door that boots such a config.
Use [`@deepseek-ai/dsh-cli-demo`](../cli-demo/README.md) for pipes, scripts, and other non-interactive runs. This bundle requires a TTY pair and has no line-oriented fallback.
| `@deepseek-ai/dsh-session-checkpoint-policy` | Semantic durability barriers before model requests and top-level tool effects, plus completed-step checkpoints |
Swappable LLM, bash, filesystem, and other capability providers remain in the leaf config. `@cordisjs/plugin-hmr` also remains a leaf-only development entry because it requires Loader internals.
Fresh runs mint a `main-session-<uuid>` session id and pass it to both the TUI and configured agent. Resumed runs bind both components to `resumeSessionId`. The TUI mounts before the spine so it can render a matching config-start failure instead of leaving a blank terminal. The app composes persistence and session query for `/resume`; an embedding host may additionally provide `tuiResumeHost` for in-place process handoff.
This package ships no bin. The [`dsh`](../../../apps/cli/README.md) CLI is the terminal front door: bare `dsh` boots the shipped `examples/tui-agent/cordis.yml` (which mounts this bundle), and `dsh --config <path-to-cordis.yml>` boots an alternate leaf config that mounts it. It loads the optional cwd `.env`, drives the Cordis Loader, and waits for the full plugin tree. The repository installs Loader's optional native helper, so bare package specifiers resolve under plain Node.
Each non-empty non-command editor submission becomes a user message; a submission during a running turn becomes steering. Slash-command input and output remain human-only, while accepted `/goal` mutations append domain-owned model-visible state. The shared spine contributes the configured persona, workspace instructions, skill catalog, goal controls, and visible tool schemas. TUI rendering itself is not model-visible.
User, assistant, and tool history grows under the normal session and compaction rules. Headers, cards, plans, Markdown styling, and keybindings add no tokens.
#### KV Cache effect
Append-only while the composed prompt, schemas, route, and retained history prefix remain stable. Composition changes and compaction can invalidate reuse from the first changed token.
### Human-question answer
#### What the model sees
`ask_user_question` retains the tool call and the compact answer or stable interruption error defined by `dsh-tool-ask-user`. The question overlay is terminal-only.
#### Token effect
Only the completed or failed tool result adds retained tokens.
#### KV Cache effect
Append-only; the answer follows the reusable request prefix.
## Known Limitations and Deferred Work
- **TTY-only** — stdin and stdout must both be terminals; automation uses `dsh-cli-demo`.
- **One configured terminal session** — the transcript and editor bind to one exact session id.
- **The app cluster is fixed** — JSONL persistence and ask-user tooling are baked in; different policy requires another composition.
- **Approval is separate** — this app answers `ctx.userInteraction`, not `ctx.approval`; permission prompts require an approval service and answerer.