feat(watch): declarative skill-based pipelines for watch #919

Summary

Allow users to define custom watch pipelines as an ordered list of skills in config, instead of writing TypeScript WatchCapabilities. The watch loop reads the pipeline definition and executes each skill as a fresh Copilot session — no drift by design.

Builds on #918 (external WatchCapability loading) as a higher-level abstraction.

Motivation

Users want to enforce workflows like issue > branch > implement > PR > review > merge without writing code. They already know how to write skills (SKILL.md files). If they could declare a pipeline as a list of skill names, the watch loop could execute them in sequence — each step in a fresh LLM session, transitions enforced by TypeScript.

This is the "no-code" version of #918's "low-code" approach. Both coexist.

Proposed Design

Config format

In config.json or squad.config.json:

{
  "pipeline": [
    { "skill": "implement-issue", "type": "copilot" },
    { "skill": "open-pr", "type": "code" },
    { "skill": "review-pr", "type": "copilot", "context": "different" },
    { "skill": "check-approval", "type": "code" },
    { "skill": "merge-if-approved", "type": "code" }
  ]
}

Step types

Skill resolution

Each skill name resolves to .copilot/skills/{name}/SKILL.md (standard path per #430). The SKILL.md content becomes the prompt for the copilot session.

Separation of concerns

• context: "different" means the review step gets a different prompt context than the implementation step — the reviewer doesn't see the implementer's conversation
• Transitions between steps are TypeScript (check PR state, check approval status) — not LLM decisions
• The pipeline runner is itself a WatchCapability (pipeline-runner), so it plugs into the existing phase system

How it differs from #918

Example user experience

1. Write .copilot/skills/implement-issue/SKILL.md — instructions for implementing an issue
2. Write .copilot/skills/review-pr/SKILL.md — instructions for reviewing a PR (different persona)
3. Add "pipeline": [...] to config
4. Run squad watch --pipeline — issues flow through the pipeline automatically

Not in scope

• Visual pipeline editor
• Conditional branching (if/else in pipeline) — keep it linear for v1
• Parallel pipeline steps — sequential only for v1

Context

Follow-up to the conversation in #918. The insight: users already write skills. Let them compose skills into pipelines declaratively, and let the watch loop handle the execution mechanics.

cc @jbenami — this is the declarative version of what you described. Pipeline steps are skills (markdown), transitions are code (TypeScript). No drift because each step is a fresh copilot session.