BYO agents Phase 2: `bbox agents add` / `agents init` scaffolding

[JAORMX]

Follow-up to #191 (epic). Phase 1 landed declarative custom agents + bbox agents list|inspect|doctor in #200. This issue adds the remaining Phase 2 CLI for authoring a custom agent without hand-editing YAML.

Scope

• bbox agents add NAME --image IMAGE --command CMD [--env PATTERN ...] [--memory 4g] [--cpus 4] [--egress-profile standard] [--mcp] [--mcp-authz-profile safe-tools] — appends a validated custom agent to the global config file (~/.config/broodbox/config.yaml), then runs the existing ValidateCustomAgent checks and prints the result. Refuses to overwrite a built-in or an existing custom agent unless --force.
• bbox agents init [NAME] — writes a commented starter agents: stanza (or a standalone manifest, see #(Phase 4)) the user can fill in. Should reuse the documentation already in internal/infra/config/writer.go's default template.

Notes / constraints

• Pure mapping (config.AgentFromOverride) and validation (config.ValidateCustomAgent) already exist from Add declarative bring-your-own agent support (#191 Phase 1) #200 — reuse them; the new code is just CLI + a config-file mutation in internal/infra/config.
• Writing config is I/O → lives in internal/infra/config + cmd/bbox; keep the domain pure.
• Must preserve existing comments/formatting on rewrite as much as practical (or document that it normalizes).
• Custom agents remain global-only; agents add must never target a workspace .broodbox.yaml.

Acceptance criteria

bbox agents add aider --image ghcr.io/acme/aider-bbox:latest --command aider --env OPENAI_API_KEY --mcp
bbox agents doctor aider   # PASS
bbox aider

with no manual YAML editing.

[HarperZ9]

One small acceptance-criteria suggestion from the agent-observability side: make agents add/init emit or preserve a config mutation receipt distinct from doctor pass/fail.

Since this command takes user input and writes global config, the receipt would be useful if it captured:

• before/after config fingerprint
• agent name, image, command, egress profile, MCP enabled state, and MCP authz profile
• forwarded env patterns by name only, never values
• validator/schema version used for ValidateCustomAgent
• writer behavior: comments preserved, normalized, or intentionally dropped
• explicit non-effects: no workspace .broodbox.yaml mutation, no secret values read/written, no credential path widening

Then bbox agents doctor aider --json can be the re-checkable predicate: parse config, map to domain agent, validate, resolve required env names as present/missing, and produce evidence refs that are safe to paste into an issue.

The main risk in this class of CLI is not just a bad YAML write. It is later debugging where "this agent was added safely" becomes a self-report. A small JSON stdout receipt or local receipt file gives users something they can attach without leaking credentials.

I am building similar receipt shapes in Project Telos for agent handoffs, so this issue jumped out. Context if useful: https://harperz9.github.io/field-guide.html

[JAORMX]

Thanks for bringing this up, it's quite useful @HarperZ9 !