[Bug] /compact aborts with no retry when an external OpenAI-compatible model's stream drops (WebSocket close 1006) — GPT-6 Astra fails, GPT-5.6 Sol succeeds

提交前确认 · Pre-submission checklist

• I searched existing issues and confirmed this isn't a duplicate (checked is:issue compact; #524 covers request-rebuild context trimming, not compaction retry).
• I've read CONTRIBUTING.md.

问题类别 · Category

对话 / Agent 交互 · Agent chat

涉及的 Agent 框架 · Agent framework

ZCode Agent（自研）

严重程度 · Severity

影响体验 · Major (works but degraded)

复现频率 · Reproducibility

偶现 · Sometimes (model slowness is consistent; the hard failure depends on a transport drop)

问题描述 · Description

Running /compact on a long session while using an external OpenAI-compatible model (OpenAI GPT-6 Astra via a local OpenAI-compatible proxy, configured through the official third-party provider path in model settings):

• With GPT-5.6 Sol (same provider, same endpoint), compaction completes normally: ~2m47s on a session with ~266k tokens before compaction, reasoning effort low.
• With GPT-6 Astra, the summary streams much more slowly (~141 chars/s vs ~255 chars/s for Sol on the same history). In one attempt the stream dropped mid-summary with WebSocket close 1006 and the whole compaction aborted. In another attempt it was still streaming after ~3m48s when I cancelled it. The session stays un-compacted.

The same provider/endpoint works for normal turns and for Sol compaction, so this is not a provider-config problem. It looks like compaction performs a single attempt with no retry on transient transport failures, so one dropped connection mid-summary kills the entire compaction.

复现步骤 · Steps to reproduce

1. Configure an external OpenAI-compatible provider (mine: a local OpenCodex proxy in front of OpenAI models) and select GPT-6 Astra.
2. Grow a session to a large context (~266k tokens before compaction).
3. Run /compact.
4. Observe the summary streaming slowly; if the connection drops mid-stream (e.g. WebSocket close 1006), compaction fails and no retry happens.

期望表现 · Expected behavior

• Transient transport failures (429 / 5xx / dropped stream) during compaction should be retried with backoff instead of aborting the whole compaction.
• Ideally, also consider a compaction timeout/retry policy or a configurable compaction model for slow external models.

实际表现 · Actual behavior

Compaction aborts on the first transport drop. With a slow model the summary stream stays open long enough that drops become likely, so /compact effectively cannot complete. Current workaround: switch the model to GPT-5.6 Sol, run /compact, then switch back.

ZCode 版本 · ZCode version

3.11.2 (3.11.2.6792) — the latest release as of 2026-09-09; bundled CLI harness reports 0.16.5.

设备 / 系统 / 浏览器 · Device / OS / Browser

Apple Silicon Mac, macOS, ZCode desktop

截图 / 录屏 / 日志 · Screenshots / Recordings / Logs

Measured on the same ~266k-token session, reasoning effort low in both runs:

Model	First streamed content	Total	Summary chars	Speed
GPT-6 Astra	9.5s	>3m48s (cancelled; an earlier run aborted with WebSocket close 1006)	30,879	~141 chars/s
GPT-5.6 Sol	12.4s	2m47s (success)	39,468	~255 chars/s

Upstream context that looks directly relevant:

• opencode had the same gap and fixed it on 2026-09-07 in anomalyco/opencode#47806: "Provider compaction made exactly one llm.compact attempt, so a transient 429/5xx/transport failure on the trigger call failed the compaction." Compaction calls are now retried with backoff, which is safe because nothing is installed until the provider returns a checkpoint. ZCode's compaction summary prompt matches opencode's, so the machinery looks related and the same fix pattern should apply.
• openai/codex#43038 tracks GPT-6 Astra compaction being 3–5x slower than Sol on the model side, which lengthens the stream window during which a drop can occur.

Happy to share session logs or timings if that helps.

Start by tracing the /compact command into the compaction provider call and its streaming error handling; the issue does not name repository files or tests. Compare the behavior with anomalyco/opencode#47806. Done means transient 429, 5xx, and dropped-stream failures retry with backoff without losing the session's un-compacted state.