Subagent exhausts max-turns fetching context and returns no summary (174K tokens wasted); salvage partial output + inline context #48

Summary

A read-only Subagent can exhaust its max-turns limit fetching context and then terminate with no summary — burning real wall-clock and a large token spend while returning nothing useful to the parent. Observed live: a Spec-adherence reviewer subagent ran 1m 8s · ↑173.9K input tokens · 12 tools and stopped with stop:max-turns → (subagent produced no summary).

Observed (live, Opus panel review)

The parent dispatched one Subagent ("Spec adherence check for PR 1626"). Its result card:

✓ Subagent · Spec adherence check for PR 1626
subagent · 1m 8s · ↑173.9K ↓2.7K · 12 tools · stop:max-turns
agentId: subagent-toolu_bdrk_017b8X9i1L4g6FqQs1h5KrP2
[subagent stopped: reached its max-turns limit]
(subagent produced no summary)

The subagent spent all 12 turns / 174K input tokens fetching the PR diff and never reached the actual review before hitting max-turns. The parent then said, verbatim: "The spec subagent hit its turn limit. Let me re-run it with the diff inlined (it's already in my context) so it doesn't burn turns fetching." — i.e. the model itself identified that the child wasted its budget re-fetching context the parent already held.

Why this is a problem

1. Pure waste. ~1 minute and 174K input tokens produced zero output. At scale (panel of 6 reviewers), that compounds badly.
2. No partial value on max-turns. The terminal render is (subagent produced no summary) — the parent gets nothing actionable, not even "here's what I found in the turns I had." A max-turns stop currently yields an empty deliverable.
3. Avoidable root cause. The child re-fetched a diff the parent already had in context. The harness has no mechanism nudging the parent to inline context for read-only children, and no guidance steering subagents away from expensive fetches when the budget is tight.

Possible directions (discussion — no fix proposed)

• Salvage partial output on max-turns. When a subagent stops on max-turns (or any non-StopEndTurn benign terminal), prompt it for a one-shot "summarize what you found so far" wrap-up turn before returning, so the parent gets something rather than (produced no summary). (Mirrors how the team lead is driven to a real synthesis even when budget-stopped.)
• Encourage context inlining for read-only children. Surface guidance (Subagent tool description / skill prompt) that the parent should pass excerpts it already holds instead of having each child re-fetch — especially for fan-out reviews where N children would each re-fetch the same diff.
• Right-size the default per-subagent turn budget for fetch-then-act tasks, or make the budget-exhaustion message model-facing enough that the parent reliably switches strategy (here it did, but only after the waste).

References

• Companion agency issue Spurious GPT-5.5 refusal: OpenAI adapter drops the Responses API phase field #46 (model announces work then stops) and perf-visibility issue perf MCP has no visibility into subagent telemetry (child Sink/ToolCallRecorder off by design); add role-tagged child metrics #47 (subagent token/turn spend is invisible to the perf MCP — directly relevant: 174K tokens went unmetered).
• Found debugging a live mecatui session on 2026-06-11.