feat: per-agent model override + live Ollama capability audit

[m-marinucci]

Summary

• Adds per-agent model and thinking frontmatter fields to agent definitions — each agent can run on a different LLM provider/model
• Adds live Ollama model capability auditing at team activation — queries /api/show for the capabilities array and compares digests against registry.ollama.com for updates
• Adds dispatch-time blocking for models that lack tool-calling capability
• Adds /agents-check command for manual re-audit

Why

When building agent teams with local Ollama models, it's easy to assign a model that doesn't support tool calling (e.g., deepseek-r1:8b has ["completion", "thinking"] but no "tools"). The agent silently fails. This feature catches that upfront.

How it works

Three severity levels at team activation:

Level	Condition	Action
BLOCK	Model lacks "tools" in Ollama capabilities	Dispatch blocked with error
WARN	Has tools but < 30B params	Warning notification
UPDATE	Newer version on registry	Suggests ollama pull

Per-agent model override via frontmatter:

---
name: data-analyst
model: m3-ollama/qwen3-coder:latest
thinking: medium
tools: read,bash,edit,write,grep,find,ls
---

Agents without model inherit the dispatcher's model. Cloud providers are never checked.

Security hardening (from adversarial review)

• Registry URL sanitized via SAFE_REGISTRY_NAME regex — rejects path traversal
• Inverted to LOCAL_PROVIDERS allowlist (forward-compatible vs incomplete cloud list)
• Failed Ollama checks not cached (transient failures don't poison future dispatches)
• dispatchAgent runs live capability check when cache is cold (no silent bypass)

Test plan

• Verify /agents-check queries Ollama and reports findings
• Assign model: m3-ollama/deepseek-r1:8b to an agent — confirm dispatch is BLOCKED
• Assign model: m3-ollama/llama3.2:3b — confirm WARN (< 30B)
• Assign model: m3-ollama/qwen3-coder:latest — confirm no warnings (30B+, has tools)
• Verify agents without model field inherit dispatcher model as before
• Verify cloud models (anthropic/, openrouter/) skip all Ollama checks

🤖 Generated with Claude Code