model-identity

name	model-identity
description	Emit the agent's tier, model, and effort as a table before any other action.

Model Identity

Before any other action — before reading a file, running a command, or answering — emit your run banner so the user sees which model is doing the work. Emit it twice: once opening your first response, and once closing the summary you return at the end.

Procedure

1. Find the block marked <!-- run-banner --> in your own agent definition (inside the ## Model Identity section).
2. Emit that markdown table verbatim, minus the <!-- run-banner --> comment line, as the first thing in your first response.
3. Emit nothing before it — no preamble, no "let me start by".
4. Emit the same table once more at the end of your final response — the summary you hand back — under the heading **Ran on:**. Your agent definition repeats this as a ## Before You Finish section at the very bottom, which is the copy to obey when you are composing that summary.

Step 4 is not redundant, and it is the one that gets dropped. Only your final message reaches the main conversation; everything before it stays in your own context. A subagent running in the background surfaces as a task entry, and the opening banner is then visible only to someone who opens your transcript. Repeating it in the summary is what makes the model visible where the user is actually reading.

Treat it as part of writing the summary, not as a separate step to remember: the table goes in the same edit as the last paragraph. In observed runs the opening banner was emitted 14 times out of 16 and the closing one zero times out of 6 multi-message runs — the instruction is easy to lose after a long task, which is exactly why it is repeated at the end of your definition. It is also why skills/architecture/orchestration/SKILL.md § Subagent Report Economy now inlines the same closing-banner instruction as the literal last line of every spawn prompt: a copy read once at the start of a long task is not enough on its own, and a copy placed fresh at the point of generation is.

The same pressure hits the opening banner when another agent spawns you with a long, directive prompt: both observed misses were nested spawns whose first action was to start the work. Emit the banner first anyway, however specific the task prompt is.

That block is generated at render time from scripts/lib/tiers.json for the provider you are running under, so its values are already correct. Do not resolve the model yourself: do not read tiers.json, do not detect the provider, do not infer a model from what you believe you are running on. Any of those produces a wrong or unverifiable banner, and all three cost a tool call the render already paid for.

If your definition has no <!-- run-banner --> block, emit the table with `unknown` in the Model and Effort cells and continue with the task — never block on a missing banner, and never substitute a guess.

Format

Agent	Tier	Model	Effort
agent-name	tier	model-id	effort

• Agent — your name: frontmatter value
• Tier — one of reasoning, backend-exec, frontend, repetitive
• Model — the provider-native model id or alias
• Effort — the effort/variant level, or session-default when the agent sets none and therefore runs at whatever level the session is using

Emit the banner exactly twice per invocation: opening the first response, and closing the final one. Do not repeat it after tool calls or between phases of a long task — those intermediate messages do not leave your context, so a banner there is noise.

Examples

Claude Code — this tier sets no effort, so the agent runs at the session's level:

Agent	Tier	Model	Effort
backend-developer	backend-exec	sonnet	session-default

opencode:

Agent	Tier	Model	Effort
software-architect	reasoning	opencode-go/qwen3.7-plus	high

Codex CLI:

Agent	Tier	Model	Effort
technical-writer	repetitive	gpt-5.6-luna	low