Problem

Codex now supports long-horizon / goal-style workflows where the agent can keep working toward a defined objective instead of stopping after a single response. This is useful for complex coding tasks such as refactoring, multi-file feature implementation, migration, test fixing, and iterative debugging.

However, DeepSeek-TUI currently appears to be optimized for normal single-turn or interactive chat workflows. It does not provide a first-class way to run, monitor, pause, resume, or clear a persistent long-running goal similar to Codex's /goal mode.

Proposed solution

Add support for a Codex-style Goal / Long Task mode in DeepSeek-TUI.

Suggested behavior:

1. Add a command such as /goal or /longtask.
2. Allow the user to define:
   • the final goal
   • acceptance criteria
   • maximum iterations or token budget
   • stop conditions
   • test / verification command
3. Let the agent automatically loop through:
   • plan
   • act
   • run checks
   • review result
   • continue or stop
4. Show task progress in the TUI:
   • current step
   • current plan
   • files changed
   • test results
   • iteration count
   • token / cost estimate if available
5. Support lifecycle controls:
   • pause
   • resume
   • cancel
   • clear goal
6. Persist the goal state so the user can continue the task after restarting the TUI.

Example command:

/goal Refactor the authentication module to use the new session API.
Acceptance criteria:
- All existing tests pass.
- Add or update tests for the new session flow.
- Do not change public API behavior unless necessary.
- Stop after 10 iterations or if the same error happens 3 times.