Supervisor sessions ingest raw child tool outputs; child results should be reports, not transcripts

[baron]

Impact

When a supervisor session polls or waits on child agent sessions, raw tool-output-heavy child content can flow back into the supervisor context. That multiplies token cost across the session tree and can crowd out useful context.

Actual behavior

Child session results surfaced through agent_run wait/poll/steer responses and related snapshots can include, or encourage re-loading of, large raw tool outputs rather than a distilled child report.

Expected behavior

• The default supervisor-visible artifact of a child run is a structured report: status, outcome, concise summary, key findings, changed files/artifacts, and verification.
• Raw transcript/tool detail remains available only through explicit log/transcript operations.
• Snapshot/wait responses have documented size bounds.

Publishable evidence

• Approximate agent_run op counts show heavy supervisor-child traffic: start 986, wait 1791, respond 461.
• Rollout data contained about 247 MB of raw function-call output, which becomes expensive if reintroduced into parent contexts.

Scope / non-goals

In scope:

• default response shaping for agent_run wait/poll/steer and session snapshots,
• size bounds,
• documentation of the report contract.

Out of scope:

• forbidding raw transcript access; full logs should remain reachable through explicit operations.

Likely affected areas

• Sources/RepoPrompt/Infrastructure/MCP/Agent/
• Sources/RepoPrompt/Features/AgentMode/Runtime/
• Tests/RepoPromptTests/MCP/
• Tests/RepoPromptTests/AgentMode/

Acceptance criteria

• Wait/poll/steer responses return bounded report-style payloads by default.
• Raw transcript content is available only through explicit log/transcript operations.
• A tool-heavy child session test asserts supervisor-visible payload size stays under a defined bound.
• Tool/API documentation describes the report contract.

Interlinks

Part of #406. Related: #398 and #405.

---