Plugin passthrough leaves the client an empty turn when the model echoes a render tag

The plugin heads every user and assistant text block with a ref annotation (renderTags=text-only), and the model sometimes echoes that annotation line as its own visible content. Where the echo is the entire visible content of a turn, the tag filter deletes it, the emptied chunk is suppressed, and the client receives a turn with no text and no tool call, which the harness aborts:

✘ upstream stream ended before a completion event; this turn may be incomplete

Releasing the echo instead is not acceptable: #463 keeps a turn that is a pure tag echo swallowed, and #674 plus the degenerate-turn warning exist precisely so that imitated render tags never reach the client. So the turn has to be made usable without leaking the annotation.

Expected: when a turn ends cleanly (stop / end_turn / stop_sequence) with zero visible text and no tool call, the pipe should obtain a usable continuation rather than hand the client an empty turn. Constraints that follow from the existing contracts: at most one retry, the retry must respect the wire's framing (one message_start / message_stop on anthropic, retry block indices offset, the openai terminal being the finish_reason chunk rather than [DONE] so the first attempt's terminal does not arrive twice), and if the retry is also empty the original empty turn must still pass through with the existing [degenerate-turn] warning rather than silently vanishing.

Fixed by PR #871.

docs: cite the filed numbers in the plugin empty-turn retry changelog…

The fix is already filed as PR #871 (branch 2026-09-16_870-plugin-empty-turn-retry). Please review there rather than implementing a parallel change.

docs: cite the filed numbers in the plugin empty-turn retry changelog…

fix(plugin): reach the degenerate-turn retry on the Google wire too (r…

Harden the Responses plugin pipe with the degenerate-turn retry (ranx…

🤖 Powered by ework · qwen3.8-27b

Reviewed #871 as requested (rebased head 8d0a089, onto current master) — no parallel change implemented. Verdict: ready for merge (human-side, per repo rules).

Traced the pipe against every constraint in this issue: one-shot retry gate placed after event processing (coalesced content+finish_reason chunks count their text first), clean-terminal whitelist (stop/end_turn/stop_sequence only), OpenAI terminal = finish_reason chunk so the first attempt's [DONE] never doubles, Anthropic single message_start/message_stop with retry blocks index-offset after the client's own, and retry-also-empty falls back to the original empty turn with the existing [degenerate-turn] warning. All hold; truncation (Plugin mode: upstream stream truncated before terminal event → client gets bare cut-off SSE (no completion/error event) #721) survives takeover too.
Verified locally on the rebased head: typecheck clean, build clean, full suite A/B vs master — delta is exactly the 8 new tests (all passing); the single failing test fails byte-identically on clean master (pre-existing, also explains the red CI test jobs on both sides).
Full write-up incl. four non-blocking observations (thinking-only turns also re-trigger by design per [BUG] In some cases, with Codex's auto model + DeepSeek, it will suddenly terminate automatically. #732/workbuddy报错Empty stream: upstream gateway sent only placeholder chunks without any model output (chunks=0, bytes=0) #821; held-tag-tail swallow on takeover is more-correct-per-plugin 模式 openai SSE：混合文本字段 chunk 因单字段纯标签 echo 被整体丢弃（丢失同 chunk 真实输出） #463; abandoned first reader is cosmetic; optional follow-up tests for retry-truncation + coalesced chunk): fix(plugin): retry a turn whose only visible output was a stripped tag echo (#870) #871