Subagents should know what model they're running

[JonathanMace]

Describe the feature or problem you'd like to solve

The main agent's system prompt includes <model_information> so it can self-identify. Subagents launched via the Task tool with a model parameter don't receive this — making it impossible to validate model routing in experiments.

Proposed solution

Propagate a <model_information> block into the subagent's system prompt when spawned with a model parameter.

Example prompts or workflows

1. "For each available model, create a subagent specifying the model. Ask each to report their model." — 0/18 could self-identify.
2. "Run this task on claude-haiku-4.5 and claude-opus-4.6 in parallel. Label which model produced each result." — impossible without model identity.
3. "Launch a subagent on gpt-5.4 and confirm it's not running Claude." — subagent can only say "ChatGPT" generically.

Additional context

Tested on v1.0.12. Launched 18 explore subagents (all available models), each asked to self-identify:

Requested Model	Self-Report	Match?
claude-sonnet-4.6	"Claude, likely Sonnet 4"	Close
claude-sonnet-4.5	"Claude" (no version)	Family only
claude-sonnet-4	"Claude" (no version)	Family only
claude-haiku-4.5	"Claude 3.5 Sonnet"	Wrong variant
claude-opus-4.6	"Claude 4 / Sonnet"	Wrong (Opus != Sonnet)
claude-opus-4.6-1m	"Claude 4 / Sonnet"	Wrong
claude-opus-4.5	"Claude" (no version)	Family only
gpt-5.4	"ChatGPT, OpenAI"	Family only
gpt-5.3-codex	"OpenAI assistant"	Family only
gpt-5.2-codex	Refused to answer	Refused
gpt-5.2	"Don't have enough info"	Unable
gpt-5.1-codex-max	"ChatGPT"	Family only
gpt-5.1-codex	"ChatGPT"	Family only
gpt-5.1	"OpenAI ChatGPT-style"	Family only
gpt-5.4-mini	"ChatGPT"	Family only
gpt-5.1-codex-mini	Refused	Refused
gpt-5-mini	"AI language model"	Vague
gpt-4.1	"GPT-4-based"	Close

0/18 exact matches. Root cause: model identity is used for routing but never communicated to the subagent.