Problem

Subagent completion currently injects too much child-session content back into the parent/main agent input path. For heavy subagent workloads, this can include long final reports, tool output fragments, or large completion payloads. That makes the parent session inherit the subagent's token pressure and can trigger:

• parent context overflow;
• compaction storms/timeouts;
• session write-lock contention;
• subagent completion announce failures;
• duplicated or stale child output leaking into the user-facing channel;
• parent workflows appearing stalled even when the child work has finished and artifacts exist on disk.

This defeats one of the main reasons to use subagents: isolating large work from the main conversation.

Desired behavior

Subagent execution, transcript, tool outputs, and final verbose content should stay isolated from the parent/main agent by default.

When a subagent finishes, the parent should receive only a small completion envelope, for example:

{
  "type": "subagent_completed",
  "status": "completed|failed|blocked",
  "childSessionKey": "agent:main:subagent:...",
  "childSessionId": "...",
  "runId": "...",
  "taskName": "...",
  "sessionLink": "openclaw://sessions/... or UI URL",
  "summary": "optional short bounded summary, e.g. <= 500-1000 chars"
}

The parent/main agent can then choose whether to inspect the child session with an explicit tool call (sessions_history, subagents, or a future dedicated child-result reader). Large child outputs should not be automatically appended to, injected into, or replayed inside the parent prompt.

Proposed default policy

1. No raw child transcript injection into parent input.

   • Do not inject child assistant messages, tool results, logs, or full final reports into the parent model context automatically.

2. Completion event is metadata-only by default.

   • Include status, run/session identifiers, task name, timestamps, and a link/handle to the child session.
   • Include at most a small bounded summary.

3. Large result transfer must be explicit.

   • If the parent needs details, it should explicitly read a file/artifact or fetch child history with a bounded limit.
   • Child tasks should be encouraged to write large outputs to files/artifacts and return only the path/handle.

4. Delivery to user channel should also be bounded.

   • Auto-announced child completion should not leak raw tool output or long child reports into Feishu/Telegram/Discord/etc.
   • A user-facing completion card/message can show status + child session link + short summary.

5. Configurable compatibility mode if needed.

   • Existing behavior could remain opt-in with something like subagents.completion.includeResultPayload=true, but the safe default should be isolation.

Why this matters

Large subagent outputs can be tens of thousands of tokens. If those are injected into the parent, the main session pays the cost and risk:

• every later parent turn drags the child payload forward;
• compaction may run at exactly the wrong time while completion delivery is pending;
• parent write locks and delivery lanes become fragile;
• a long-running research/coding workflow can appear to stop even though child output exists.

A subagent should behave more like an isolated worker with a result handle, not like a giant hidden paste into the supervisor prompt.

Related issues / prior art

This request is related to, but more specific than:

• Bug: Sudden Context Overflow from Large Tool Results (no preflight guard) #10694 — large tool results can cause sudden context overflow;
• Feature: Tool output compression middleware (like claude-context-mode) #30997 — tool output compression middleware, including subagent output compression;
• Subagent completion event sometimes drops result payload and/or token stats (3 repros, intermittent) #75196 — subagent completion payload can be missing/dropped for large outputs;
• [Bug] Subagent announce leaks raw tool output to Telegram and duplicates parent progress reply #79986 — subagent announce leaked raw tool output into user channel;
• [Bug]: Subagent blocked terminals reported as “completed successfully” with no output #80879 — subagent context overflow not surfaced correctly to parent;
• [Bug]: Session write-lock timeouts block subagent delivery lanes #86538 / fix: preserve subagent delivery after lock stalls #86540 — session write-lock and subagent delivery lane failures;
• Multi-agent orchestration is unstable: concurrent agents add/config overwrites, session-lock failures, and detached child work #43367 — umbrella multi-agent orchestration instability.

Acceptance criteria

• Parent prompt/input does not automatically include raw child transcript or large child result payloads.
• Parent receives a bounded completion envelope with session/run identifiers and a child session link/handle.
• User-channel auto-announce is bounded and does not include raw tool output by default.
• Parent can explicitly inspect child results when needed via bounded tools or file/artifact handles.
• Large subagent workflows no longer cause parent context overflow merely because the child completed with a long result.