Model returns empty turns (server healthy) — degenerate 1-token completions after large tool results / deep context

[Hrt-Htk]

Summary

While the server is completely healthy, the model intermittently returns an empty turn — a single-token, immediate-stop completion with no usable content. The client (interactive pi and the pi subagent) then appears to "stop responding": every input yields an empty reply. This is the original "pi produces no output" complaint, and it is distinct from #13 (a hung/dead worker). Here nothing is hung — the worker answers every request in <1s; it just returns nothing.

Evidence (2026-07-19 ~20:44–20:47)

Server side fully green — worker 62032 alive and idle, every request 200:

task 164  eval time = 0.00 ms / 1 tokens     -> POST … 200 in 442ms   (1386 bytes)
task 168  eval time = 0.00 ms / 1 tokens     -> POST … 200 in 600ms   (1386 bytes)
task 171  eval time = 0.00 ms / 1 tokens     -> POST … 200 in 545ms   (1385 bytes)
task 174  eval time = 0.00 ms / 1 tokens     -> POST … 200 in 596ms   (1385 bytes)

Each of the last four requests generated exactly 1 token from an identical prompt (sim_best = 1.000, n_tokens = 11751). pi's own session (~/.pi/agent/sessions/--H--personal-assistant--/…) records genuinely empty assistant messages:

[toolResult]  (large web-scrape result)
[assistant]   (empty)
[user] "?"        -> [assistant] (empty)
[user] "afsdfsd"  -> [assistant] (empty)
[user] "asdfas"   -> [assistant] (empty)

The empties began right after a large tool result at ~11.7k tokens of context.

Signature (how to recognise it)

• Server healthy, 200 responses, worker alive — no hang/5xx.
• eval time = 0.00 ms / 1 tokens (model emits an immediate stop token).
• pi session shows empty assistant content (no text, no tool call).
• Often follows a large tool result / deep context; identical prompt repeated (sim_best ≈ 1.0).

Relationship to other issues

• Not Proxy hangs indefinitely on a hung model-worker (no first-byte timeout) #13 (hung worker, server-side). The Proxy hangs indefinitely on a hung model-worker (no first-byte timeout) #13 first-byte timeout does not and should not fire here — the server responds.
• Same family as the think-tag / empty-content problem (issue SSEChunkLogger rescue state machine splits reasoning_content on literal </thinking> text #8). Note: the chat trace under-logs short final content (SSEChunkLogger buffer flush skips log_response), but here pi's session confirms the empties are real, not a logging artifact.

Possible directions (not yet investigated)

• Context-state degeneracy after large tool results / near-limit context.
• Custom chat template interaction (--jinja … preserve_thinking).
• Whether the SSEChunkLogger issue-SSEChunkLogger rescue state machine splits reasoning_content on literal </thinking> text #8 rescue can/should catch the "immediate 1-token stop" variant.
• Capture the exact request body (messages) that produced an empty turn from the chat trace to reproduce.

Impact

Recurring; affects both interactive pi and the pi subagent. Manifests as "the agent went silent / stopped responding" even though the backend is fine.