Feature: Skills with typed parameter schemas (structured Bash alternative)

Problem

There's a gap between the two main ways Claude Code calls external tools:

| Mechanism | Typed input | Streaming | Long-running |
|-----------|:-:|:-:|:-:|
| Bash | No — string args | Yes (stderr) | 15 min |
| MCP tools | Yes — JSON schema | No | Broken (no per-tool timeout, progress notifications silently swallowed, streamable HTTP buggy) |

For long-running operations (5+ min) that spawn external processes — like CI pipelines, multi-model AI dispatch, test suites — there's no way to get both typed inputs and streaming output. MCP can't handle these today (competing proposals SEP-1391 and SEP-975 are still unmerged), and Bash has no schema enforcement.

Observation

Skills are 90% of the way to solving this. They already have:

• Definitions Claude reads before invoking (like MCP schemas)
• Descriptions for when to use them (model-invoked)
• Execution via any tool — Bash, MCP, Read, Write (streaming works)
• No timeout limit

The one missing piece: structured parameter schemas.

Proposal

Allow skills to define typed parameters in their SKILL.md frontmatter:

name: dispatch
description: Spawn work to external CLI model worker
parameters:
  cli:
    type: string
    enum: [opencode, gemini, codex]
    required: true
  prompt:
    type: string
    required: true
  timeout:
    type: integer
    default: 300

When Claude invokes this skill, parameters would be validated against the schema before execution — similar to how MCP tool calls validate input against JSON schemas. The skill body would access parameters as structured values (not just $ARGUMENTS string).

Why this matters

This creates a "typed Bash" primitive that fills the gap:

| Mechanism | Typed input | Streaming | Long-running |
|-----------|:-:|:-:|:-:|
| Bash | No | Yes | Yes |
| MCP tools | Yes | No | No (today) |
| Skills with schemas | Yes | Yes | Yes |

Use cases unlocked:

• Multi-model AI orchestration — dispatch work to external AI CLIs with validated routing config, stream progress back
• CI/CD integration — typed deploy commands with environment enum validation, streaming build logs
• Database operations — validated migration commands with streaming output
• Any long-running shell process that benefits from input validation before a potentially expensive execution

Current workarounds (and why they're insufficient)

1. CLAUDE.md documentation — Claude usually gets args right from docs (~95%), but no enforcement. Mistakes on expensive operations are costly.
2. MCP with background job pattern (start + poll) — works but wastes tokens on polling loops, loses real-time streaming, adds complexity.
3. Two-step MCP validation → Bash execution — MCP validates input, returns Bash command, Claude runs it. Over-engineered for what a schema declaration should solve.

Implementation scope

This seems like a smaller change than fixing MCP streaming/long-running support:

• Add optional parameters field to SKILL.md frontmatter parsing
• Validate parameters before skill execution
• Expose structured parameter values to skill body (replacing/augmenting $ARGUMENTS)
• Everything else (Bash streaming, tool access, model-invocation) already works