[Feature Request] Mandatory project skill loading gates (.kimi/require-skills)

Feature Request: Project-level mandatory skill loading gates

Problem

When working on projects that define .agents/skills/<name>/SKILL.md, the AI agent can modify source code (src/, tests/) without ever loading the project's skill. This leads to workflow violations (e.
g., skipping TDD, git worktree isolation, code review requirements).

Real-world impact

In our project (Firefly, a Qt6/C++ system monitor), the skill firefly-dev-workflow mandates:

1. Read knowledge base before any code change
2. Use git worktree isolation
3. TDD (Red-Green-Refactor)
4. Two-stage code review
5. Knowledge base update after completion

The AI skipped all of these and directly modified src/view/DetailPanel.cpp, violating the project's agreed-upon workflow. The user had to manually intervene after the fact.

Proposed solution

Add a mandatory gate mechanism for project-level skills. Options:

Option A: Auto-inject skill content
When .agents/skills/ exists in the project root, automatically prepend the skill's SKILL.md content to the system prompt (or treat it as higher-priority instructions).

Option B: .kimi/require-skills config file
Allow projects to define .kimi/require-skills (similar to .pre-commit-config.yaml):

required:
  - .agents/skills/firefly-dev-workflow/SKILL.md
The CLI should block WriteFile/StrReplaceFile operations on src//tests/ until the skill is acknowledged/loaded.
Option C: Tool-call layer gate In the tool dispatcher, if a WriteFile targets src/ or tests/ and the project has unloaded skills, reject with:
▌ "Project skill firefly-dev-workflow not loaded. Load it first before modifying code."
Preferred option
Option B + C combined: .kimi/require-skills as the declarative mechanism, tool-call gate as the enforcement.
Environment
• kimi-cli version: (your version, run kimi --version)
• OS: Linux

Start by inspecting the tool dispatcher and the existing project skill-loading path, then trace how WriteFile and StrReplaceFile handle targets under src/ and tests/. Compare the proposed .kimi/require-skills configuration with the auto-injection and tool-call gate options. Done means the chosen behavior is specified, enforced for required skills, and verified with tests for acknowledged and unacknowledged skills.