[FEATURE] Make agent summaries visible to agents

### Preflight Checklist

- [x] I have searched [existing requests](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20label%3Aenhancement) and this feature hasn't been requested yet
- [x] This is a single feature request (not multiple features)

### Problem Statement

In the current version of Claude Code, background agents show a brief summary of what they're doing. This has helped me spot when a background agent is off track, so I could point out to the orchestrating agent that it needs to intervene. It would be better if the orchestrating agent could spot that itself!

I asked the orchestrator (Fable 5.1) if setting a scheduled wake for itself every 5 minutes to check the summaries, reading in more detail to intervene if a subagent appears to have gotten off track, would be viable; it informed me that it cannot see the summaries. That's unfortunate! Claude could get as much use out of them as I can.

### Proposed Solution

Please make the tool that emits subagent status also emit its most recent agent_summary, if any. A "stretch goal" would be adding a hook opportunity when the summary changes (along with a nudge to the summarizer to prefer to leave the summary unchanged if appropriate).

I would like to be able to tell the main agent to efficiently check up on its subagents periodically using the existing agent summary, so it will autonomously intervene when appropriate. I saved a bunch of wasted time and tokens by spotting the problem this time, but I think there's a lot of opportunity to let Claude make better decisions here.

Categorically, when showing new information to the user, it is probably useful to find a surface to disclose it to the agent too.

### Alternative Solutions

Currently, I'm just keeping an eye on summaries myself and pointing out if I notice something that seems off. If I really needed the agent to check in autonomously, I'd have to tell it to use _another subagent_ (probably Sonnet or Haiku) to read the entire transcript of the other agent and come up with its own summary, repeating the work already spent on the summary being shown to me.

### Priority

Medium - Would be very helpful

### Feature Category

API and model interactions

### Use Case Example

Example scenario:

1. I am reviving a stale pull request because, months later, other developers on the project are interested again. I have long since forgotten what I was doing, and why; it is a massive merge conflict and the design of the project has changed somewhat since.
2. I explain this to Fable, asking it to bring the pull request up-to-date with the current status of main, specifically requesting it delegate as much as feasible to Opus 5 subagents to conserve quota.
3. Fable gets started and eventually spins off a subagent to do a code review, and idles pending the result of the review.
4. The subagent gets distracted by failures of tests that were already failing upstream.
5. **New behavior:** Five minutes later, Fable takes a look at subagent summaries and discovers that it is "categorizing pytest failures", reminding it that it already knew about the upstream failures and forgot to tell the subagent. It injects a steering message.
6. The subagent gets back on track with its main task.

Without this feature, the subagent spends an extra half-hour and a quarter million tokens repeating work Fable already did, or it is unable to do so because it's worried about messing up the repository context for other parallel agents and starts a detailed investigation of these test failures not related to the change.

Other possibilities: Fable delegates implementation to five agents in parallel and one of them gets massively distracted and starts writing a to-do list app while it was supposed to be writing unit tests for this pinball game, and the summary lets it intervene; Fable delegates to Sonnet, Sonnet stats making a mess, and the summary "Patching std::vector to work around string comparison issue" tips it off that it is going in a _very_ wrong direction; etc.

### Additional Context

It might be useful to give the summarizer an additional detail -- what the subagent is _supposed to_ be doing -- with the ability to ping the agent that launched the subagent if the agent it is summarizing appears to be off in the weeds. That said, I expect the summarizer model is intended to be small, quick, and cheap, and may not have the capability to make good "is off the rails" determinations? might be worth a try anyway.

No repository file or test is named. Start by locating the tool that emits subagent status and the source of the most recent agent_summary, then trace how status is exposed to the user and orchestrating agent. Done means the orchestrating agent can see updated summaries without duplicating transcript analysis, with coverage for summary presence and changes.

索引モデルが issue の本文から書いたものです。