Improve ability for sub-agent returned results (summaries) to be read across conversation compactions

[megasilk53]

Describe the feature or problem you'd like to solve

Background sub-agent results are only available in the active conversation context and are permanently lost if conversation compaction occurs before they are used (e.g. integrated into a persistent file or results summary), since read_agent cannot replay results after compaction.

Proposed solution

Sub-agent results should either be re-readable at any time (persisted server-side and re-fetchable via read_agent regardless of compaction state - or possibly saved as a temp file on the local machine with a pointer which can still be accessed post-compaction), or the system should warn the user that unintegrated agent results are at risk before compaction occurs.

Example prompts or workflows

I have a large multi-page schematic PDF called example_schematic.pdf in my current directory. Please do the following:

1. Use pdftoppm -r 150 -png example_schematic.pdf [page] to convert it to per-page PNG images in a folder named images_from_example_schematic_PDF/.
2. For each page image, create a todo and launch background sub-agents (fleet mode) — 10 or more agents simultaneously — each tasked with analyzing one page image and returning a structured summary
   (ICs, pins, signals, passives, DNI parts).
3. Do not integrate any results into a file yet. Simply collect all agent results via read_agent into your conversation context first.
4. After all agents have completed and you have read all results, then attempt to write everything into a description.txt summary file.

The goal is to observe whether sub-agent results read into conversation context early in the session are still accessible and complete by the time the controller agent ("orchestrator") attempts to utilize them to build the combined description.txt — specifically across any automatic conversation compaction that may occur between steps 3 and 4.

[this prompt was authored by CoPilot and intended to force the issue. Issue occurs sporadically even with limits to 2 fleet agents.]
It may help to force an older model with smaller context, such as GPT-4.1.

Additional context

When a background sub-agent completes and its result is read via read_agent, the result exists only in the active conversation context. If a conversation compaction occurs before the result is integrated into a persistent file, the detailed sub-agent output is permanently lost and cannot be re-retrieved — read_agent on an already-completed agent does not replay the result after compaction.

CoPilot: "Sub-agent results come back through the API — they are injected directly into my conversation context when I call read_agent. They are not written to a file automatically."

What this means for compaction:

• "If I call read_agent and the result lands in my conversation context, then a compaction happens, the compaction summarizer may or may not preserve the full detail of the agent's response (it tends to
  summarize/truncate)."
• "Once compacted, that detailed sub-agent output is gone — I cannot re-read it from a completed agent. read_agent on an already-read, completed agent may return nothing new."