feat: /goal — autonomous mission completion without repeated confirmation

GitHub Issue: /goal feature request

Repository: https://github.com/wintrover/kimi-cli/issues/new
Title: feat: /goal — autonomous mission completion without repeated confirmation
Labels: enhancement

Body

Summary

Add a /goal command that sets a high-level mission objective which the agent follows through to completion without repeatedly asking the user for confirmation or re-clarification.

Motivation

Currently, Kimi Code CLI often pauses mid-task to ask clarifying questions, confirm next steps, or re-annotate assumptions. For complex multi-step tasks (e.g., "implement a full CEGAR engine with Z3 integration"), this creates significant friction and context overhead.

Competing tools already have this capability:

• Codex: Autonomous mode with goal-driven task completion
• Claude Code: Plan-driven autonomous execution with user approval at plan stage only
• Hermes: Goal-oriented agent with deterministic convergence protocols

Proposed Behavior

/goal Implement a complete CEGAR engine with Z3 Fixedpoint/PDR backend,
preservation check fix, and E2E tests. Run all tests and commit when done.

When /goal is active:

1. Agent creates an internal plan and executes it sequentially
2. Agent does NOT re-ask questions it can answer from codebase context
3. Agent does NOT pause between sub-steps for confirmation
4. Agent DOES stop and report if it encounters an unrecoverable blocker
5. Agent reports progress at milestones (not every step)
6. Goal is considered complete when all success criteria are met

Exit Conditions

• Success: All sub-tasks completed, tests pass, changes committed
• Partial: Some sub-tasks failed, report what succeeded and what failed
• Blocked: Unrecoverable error (e.g., missing dependency, conflicting requirement)

UX

> /goal Refactor VulnerabilityHit to pure ADT with strictCaseObjects
[goal] Mission set: Refactor VulnerabilityHit to pure ADT
[goal] Plan: 4 phases, ~25 files
[goal] Phase 1/4: Type definitions... done
[goal] Phase 2/4: Concept + dispatch... done
[goal] Phase 3/4: Consumer updates (22 files)... done
[goal] Phase 4/4: Compile verification... done
[goal] Mission complete. 35/35 tests pass.

References

• Codex autonomous mode: https://openai.com/index/introducing-codex/
• Claude Code plan mode: autonomous execution with user approval at plan stage only
• Axiom EDC protocol: Evidence-Based Deterministic Convergence (similar concept)

The issue names no files, tests, or implementation entry points. Start by locating the CLI command-dispatch path and the existing confirmation or planning behavior. Define the /goal lifecycle and success, partial, and blocked outcomes before implementation; done means the stated autonomous flow is implemented and verified by the project’s relevant tests.