Subagent transcripts are inlined verbatim and uncapped into the parent session export

[kapral18]

Describe the feature or problem you'd like to solve

When a session uses subagents, the parent's exported transcript inlines the complete transcript of every subagent verbatim, with no summarization and no size bound. Each subagent's full tool-call output is embedded into the parent, and a subagent that runs more than once (e.g. a retried phase) has its full transcript embedded again on each run.

Concretely, a single exported session from a multi-lane review flow came out to ~3.2 MB / ~58k lines: roughly 66% of it was inlined subagent transcripts, and a phase that ran three times added another ~30% because each run's full transcript was inlined again. There's currently no way to make the export summarize or externalize subagent transcripts — the parent always embeds the full child log.

Proposed solution

Add a way to keep the parent export bounded when subagents are involved. Any one of these would help:

• Summarize each subagent in the parent export (final message + tool-call count + key results) instead of inlining its full log.
• Externalize each subagent transcript to its own artifact and reference it from the parent, so full detail is preserved but not inlined.
• A flag/setting to choose inline vs. summarized vs. externalized export.

Full detail stays recoverable; the parent transcript stops growing unbounded with each subagent and each retry.

Example prompts or workflows

• A multi-subagent run (e.g. several parallel analysis/review lanes): today the parent inlines every lane's entire transcript; a summary plus per-lane artifacts would keep the parent readable.
• Re-running an expensive sub-step several times in one session: today each run's full transcript is re-inlined, multiplying size; a summarized/externalized export records each run compactly.
• Long sessions in general: a bounded parent transcript stays reviewable, where a fully-inlined one becomes unmanageable.

Additional context

• Affected version: copilot 1.0.65 (Homebrew cask).
• The behavior appears to come from the export renderer that joins each subagent's timeline entries and embeds the full tool-call output for every entry, with no truncation or summarization branch in that path.
• OS: macOS (ARM).

[MRayermannMSFT]

Hi @kapral18, thank you for the detailed write-up, always helpful to have detailed issues with data!

Before I get into things, when you next reply, can you elaborate on what you mean by exported transcript? I'm not 100% certain what exporting process, or transcript thing you are referring to, so at this time I'm not fully sure how to address your points.

In the meantime, I'd love to share how subagents work under the hood, and affect certain things you might see/use in the CLI, just in case those things are what you are referring to.

When the main agent launches a subagent, that subagent runs in its own separate context window. The only thing that flows back into the parent agent's context is the subagent's final result, a single message*. The subagent's full tool-by-tool transcript is never injected into the parent agent's context window. So even when a subagent runs three times, the parent's actual context stays compact: each run just hands back one result message.

So if by chance you're looking at the events.jsonl file, be aware that it is simply a record and observability artifact, not a reflection of what the model is carrying in context. It is an append-only log of everything that happened in the session, including both the outer agent and every subagent's internal steps. It exists so a session can be replayed, resumed, and inspected. It's intentionally a full record, not a representation of what the root agent "sees".

Or if you're looking at the output of /share, then you should simply treat that as a human-readable transcript meant to show off and share what Copilot did during a session. It deliberately inlines the full detail, including subagent activity, so the shared artifact is a complete account of the run, again not a picture of the agent's context window.

If the exported transcript you're referring to is neither of those, please again make sure to clarify what it is you are referring to in your next reply. Thank you!

* We do have a feature in /experimental that allows multiple turns for a subagent, but after each of those turns, the agent who spawned the subagent is still just simply reading a single message.

[kapral18]

Thanks, that distinction helps. I’m referring to the human-readable /share file / /export file session markdown export — not events.jsonl, and not the parent model context.

The context behavior you described makes sense: the parent agent only receives the subagent’s final result message. The issue I’m raising is specifically about the share/export artifact.

One important use case: I use that exported markdown as a handoff/rehydration artifact for follow-up sessions or adjacent tooling. In that workflow, I want the top-level export to explain what happened and preserve links/references to detail, but not inline every subagent’s full internal transcript. When all subagent activity is expanded verbatim, the export becomes too large and noisy to feed back into a later session effectively.

So the request is to keep today’s complete inline export available, but add an option for a bounded top-level export: summarize subagents in the parent export and/or externalize each subagent transcript into separate artifacts referenced from the parent.