[Bug] Subagent epoch that burns its entire output budget on reasoning (zero tool calls, zero output) settles as completed with an empty closing message

[jackdizhu]

Hello — I searched the discussions before posting; the closest threads I found are #9124 (terminated sub-agent output unrecoverable), #5903 (thinking-only final response discards a workflow run), and #4666 (spawn backend drops reasoningEffort). The case below looks like a distinct sibling on the normal-settlement path of continuable subagents: no termination, no schema — just a completed verdict over a run that never called a tool. If my analysis is off, corrections are very welcome.

Symptom

A background continuable child (subagent({ run_in_background: true })) settled as completed after ~20 minutes with:

• an empty closing message (no Its closing message: block in the notice);
• zero disk artifacts (the child never created its working directories);
• get_subagent_status before settle: "no activity observed".

The dispatch was a self-contained mechanical rewrite brief (rewrite 13 markdown files, script-verified acceptance). The parent could not tell "the child never started working" from "the child worked and failed" — the same ambiguity class as #9124, but reached via a normal settlement.

What the child's session log shows

The child session log (.dsh/sessions/<workspace>/<subagentId>/session.v4.jsonl.zstd, multi-frame zstd) tells the whole story:

• exactly 1 turn / 1 step for its whole lifetime;
• the assistant message's content blocks are ["reasoning"] only — no text, no tool_use;
• usage: inputTokens: 32,853, outputTokens: 32,000 (exactly the output cap), reasoningTokens: 31,955;
• turn/end reason: completed.

So the model spent the entire output budget on reasoning, the provider ended the response cleanly (no truncation error surfaced), and the epoch folded to completed.

Source localization (0.2.0-rc.2)

In @deepseek-ai/dsh-subagent/lib/index.js:

• createActivationObserver initializes captured = { stopReason: "completed" } as the default (~L301);
• epochStopReason (~L343–352): case void 0: case "completed": return droppedUnrun ? "aborted" : "completed" — an epoch whose event suffix contains no closing turn/end still reports completed;
• observeRun attaches lastAssistantMessage only when output.length > 0 — hence a settlement notice without a closing message.

Net effect: a reasoning-only epoch (zero tool calls, zero assistant text) is indistinguishable, from the parent side, from a successful run. The settlement machinery has no signal a parent could use to treat this as anything but success.

Why it matters

The dispatch's value lives entirely on disk. A completed with an empty closing message reads as success, so the orchestrator may proceed (or report success) while the deliverable never existed. Cost side: the child burned its full output budget (~32k tokens) producing nothing observable.

Suggested directions (non-prescriptive)

1. When an epoch ends with zero tool calls and zero non-empty assistant text while output tokens are at/near the cap, consider reporting it as max-tokens (or a dedicated reason) rather than completed.
2. Lighter alternative: append a one-line diagnostic to the settlement notice when output.length === 0 — e.g. "(no assistant text, no tool calls — likely reasoning-only exhaustion)". That alone converts a silent loss into an actionable one, and aligns with [Bug] Terminated sub-agent output is unrecoverable: no closing message, no registry entry, no read-back path #9124's ask on the terminated path.
3. Surface reasoning/output token usage in the notice or in get_subagent_status, so parents can cheaply detect this shape.

Current workaround (confirmed effective)

• pass reasoning_effort: "low" explicitly for mechanical/rewrite briefs (with provider + model supplied paired);
• prepend an execution-discipline preamble ("first action must be a tool call; every turn must advance disk artifacts");
• before trusting completed, run a three-point check: closing message non-empty / heartbeat file ends with done / disk artifacts exist.

With the workaround, the same brief completes in ~2 minutes with all deliverables verified on disk.

Environment

dsh	0.2.0-rc.2 (@deepseek-ai/dsh-subagent 0.2.0-rc.2)
Node	v24.21.0
OS	Windows (desktop profile)
Dispatch	subagent({ run_in_background: true }), continuable
Route	deepseek-official/deepseek-flash, reasoning_effort not passed → default high

Thanks for reading — happy to provide the full session-log excerpt (already sanitized) if useful.