Issue 2: Context-mill `agents` content type (parallel to skills)

[gewenyu99]

Issue 2: Context-mill agents content type (parallel to skills)

Epic: Task-queue orchestrator runner · Repo: context-mill · · PR: PostHog/context-mill#181
Depends on: none · Gates: #625 onward

Why

The orchestrator separates WHAT from HOW. Skills already exist in context-mill as
the HOW, the procedure. The WHAT, an agent prompt, is a different kind of content:
it carries intention, which model to run on, the goal, how to tell if it
succeeded, which tools it may use, and which skills to load. Rather than shoehorn
agent prompts into the skill pipeline, context-mill grows a new content type for
them, agents, a flavor parallel to skills. Skills stay the procedure. Agents are
the intention.

Scope / deliverable

1. An agents content type in context-mill, modeled on how skills are already
   authored, built, and served. Find the skills implementation first (the skill
   source directory, the build that produces the served artifact, and the menu or
   index the wizard fetches), and mirror it for agents.
2. Agent file format. Markdown with frontmatter. The frontmatter carries the
   artifacts the wizard executor needs: type, model, skills (which skills to
   load), allowedTools, disallowedTools, dependsOn. The body is the prompt:
   goal, plain-text success criteria, handoff instruction. This is the format Issue 6: Agent-prompt and mini-skill format + seed prompt #625
   defines on the wizard side; keep the two in sync.
3. Serving. Agents are fetched the same way skills are, including the
   localhost:8765 local dev path, so the wizard loader ( Issue 6: Agent-prompt and mini-skill format + seed prompt #625) can pull an agent by
   type with the existing fetch code.
4. An agents menu or index parallel to the skill menu, so the wizard can
   discover the available agent types.
5. Author on a clearly-named experiment branch, the same branch name as the
   wizard side, so the two move together while the experiment is in the dark.

Open work to scope on the branch

• Read context-mill's skill build and serve code, then decide whether agents
  is a sibling directory with its own build target, or the skill machinery
  generalized over a content "kind". Prefer the smallest change that keeps agents
  and skills cleanly separated.

Acceptance criteria

• An agent markdown file with frontmatter builds and serves from context-mill,
  locally via localhost:8765.
• The wizard loader ( Issue 6: Agent-prompt and mini-skill format + seed prompt #625) fetches an agent by type and parses its frontmatter
  and body.
• Skills are unaffected. The two content types are separate and both serve.
• The seed prompt and the three thin-slice agents ( Issue 7: Walking skeleton, end-to-end with stub tasks #626, Issue 8: Real task bodies + full 1:1 integration flow #627) live as agents.