Multi-provider stream adapters via a per-agent format field

[Beroni]

Description

Make non-Claude agent CLIs first-class citizens of the live experience: a per-agent format: field selecting a stream adapter, instead of the current implicit Claude-only detection.

Motivation

The execution contract is already provider-agnostic — any CLI that reads the prompt on stdin and writes markdown to stdout works, and an agent can override command/args (so codex exec - or gemini -p run today). But everything rich comes from internal/agent/stream.go, which only parses claude -p --output-format stream-json (detected by isStreamJSON(args)). With any other provider the live log is raw stdout with no per-subagent panes, and token usage and quota never show up.

Proposed Solution

• New optional field on the agent block and on each AgentDef: format: claude-stream | codex-json | plain (default: today's behavior — claude-stream when the args ask for stream-json, plain otherwise).
• Turn the current streamParser into an interface producing the existing internal events (log line with agent attribution, usage delta, limits); stream.go becomes the Claude adapter.
• Add a codex-json adapter for codex exec --json (JSONL) as the second implementation proving the interface; a Gemini CLI adapter can follow the same shape.
• Usage from non-Claude adapters feeds the same counters, so per-step and total token reporting (and Report estimated cost in dollars alongside token usage #3's dollar costs) work across providers.

Acceptance Criteria

• format: is accepted per agent and inherited from the agent block
• Existing configs behave identically without the field
• A codex-json agent shows a readable live log and token usage
• README documents a working recipe per supported provider

Related: #3 (dollar cost reporting — the price table should be keyed by provider/model from day one), #12, #13.