Recurring 502 "Stream ended unexpectedly before completion" on large sessions — turns lost after retry exhaustion (macOS, v0.1.47) #991

Summary

Turns on large sessions repeatedly fail with:

502 Stream ended unexpectedly before completion (no finish event) — response was truncated

The failure hits mid-stream. The client retries internally (stream_restart up to 3×, api_retry up to 8× with ~10s backoff), but when retries exhaust (~10 min) the entire turn is lost and the app ends it as run_error. Retrying later sometimes succeeds. Small/medium sessions were fine during the same window; the failures concentrate on large sessions (transcript ≈ 27 MB, requests ≈ 350–400K tokens including many base64 image blocks). Even a 2-character prompt ("hi") failed after ~10 min of retries.

Expected Behavior

The turn completes normally. If the upstream stream truncates, automatic retries recover (ideally by resuming the partial response) instead of losing the whole turn.

Actual Behavior

Turns end with "502 Stream ended unexpectedly before completion (no finish event) — response was truncated"; after internal retries exhaust, the app ends the turn with run_error and the response is lost. Failed attempts consume 0 tokens.

Traces (times UTC+8; session A unless noted):

• Oct 5 22:03 — b85bc5fd3a4e49d382dab12462f34ec9
• Oct 5 22:35 — 7c64a91e28e9eee75ba94ecda838fc56
• Oct 5 22:50 — ecb6851860fe19c292187e3c158d7dce
• Oct 5 23:02 — 328b5cd1d47675ea88a464dc4618c035
• Oct 6 00:04 — 5a6eb72090d1b091c6ca6cf36ee3a6a3
• Oct 6 00:17 — 5c8fe322d8b2c029063623048a7da8a9
• Oct 6 02:26 — 85f2080cb801b9fe1be786c664bdac90
• Oct 5 17:48 — 1ae311d885cfc033964bc7c929e2fdc1 (session B)
• Oct 5 17:51 — 0dd705cddacc282fb3c9763be47ae635 (session B)
• Oct 5 18:36 — 6cdcd37f8bd0af0a5eb5a617565f53ef (session B)

Steps to reproduce the issue

1. Open a large session (transcript ≈ 27 MB; recent context includes many image tool results).
2. Send any message (short ones included) with model deepseek/deepseek-v4.1-flash at reasoning effort max.
3. The response stream truncates mid-way; after internal retries exhaust the turn ends with run_error.
4. Same behavior via headless: cmd --session -p "" — retries for minutes, then exits with "The API server encountered an error. Please try again later."

Command Code Version

0.1.47

Operating System

macOS

Terminal/IDE

N/A (desktop app)

Shell

zsh

Session file (optional)

No response

Fix prompt (optional)

Check the gateway logs for the traces above. Investigate stream truncation (no finish event) on very large requests (large context + image blocks), and make failures more recoverable — e.g. resume/continue the truncated stream instead of discarding the whole response, and clearer error semantics when retries exhaust.

Additional context

Environment: macOS (Apple Silicon); Command Code desktop 0.1.47; model deepseek/deepseek-v4.1-flash, reasoning effort max. Reproduced in the desktop app and via headless CLI on the same night.

Notes:

• Workaround that helps: instructing the model to keep responses small / step-by-step — turns then usually complete.
• Failed attempts show 0 tokens consumed; sessions stayed intact and resumable (no data loss).
• Healthy small sessions in the same window suggests this is tied to very large request payloads / long streams.