[Bug]: OCG DeepSeek Chat re-hoists chronological reminders after inbound fix #4161 #4439

Client or integration

Claude Code (Claude Desktop Code surface)

Area

Provider adapter

Summary

On OpenCode Go's deepseek-v4.1-flash Chat route, appending a chronological Claude system reminder still rewrites the leading outbound system message. The generic Chat serializer collects text-only developer messages into that leading message, undoing the chronological representation introduced by #4161. This changes the reusable prompt prefix even when the existing conversation is unchanged.

This is the Chat-outbound follow-up explicitly left out of scope in #4161, which closed #4148. Proposed narrow fix: #4438.

Reproduction

1. Use the canonical OpenCode Go destination (https://opencode.ai/zen/go/v1), Chat adapter, and model deepseek-v4.1-flash.

2. Translate a synthetic Claude request through anthropicToResponsesBody → parseRequest → createOpenAIChatAdapter(...).buildRequest(...):

   {
     "model": "deepseek-v4.1-flash",
     "system": "Stable project instructions.",
     "max_tokens": 100,
     "messages": [
       {"role": "user", "content": "Inspect the synthetic project."},
       {"role": "assistant", "content": "First result."},
       {"role": "system", "content": "Synthetic reminder A."}
     ]
   }

3. Append an assistant turn, a user continuation, and another system reminder B; build the second outbound request.

4. Compare the old-length prefix of the second outbound messages array with the first array. Before the fix, reminder B joins the leading system content and the comparison fails. Expected on this compatible route: the prior serialized messages remain unchanged and the new reminder stays in its chronological position, while pending tool calls/results remain adjacent.

The offline regression in #4438 exercises the actual translation/parser/adapter chain with stabilizePromptCache both false and true. It requires no credentials or model requests. The new suite produced 5 failures on the unmodified base and 15 passes with the patch, including destination/model boundaries and pending-tool behavior.

Version

Observed locally on OpenCodex 2.52.0, Bun 1.4.2. PR base: dcd13b4358befaae0fdca845a8219103943faca0; proposed head: e7bfb08b2823647b09d501692f759193a19d2299.

Operating system

Windows 11, build 26200.

Provider and model

OpenCode Go / deepseek-v4.1-flash, High effort. Live observation includes input growing to 409,130 tokens; it does not validate the full advertised 1M window.

Observed cache behavior

Snapshot through 2026-09-12 22:07:56 UTC, about 2h23m after the local chronological patch became active:

• Same long-session cohort: 9.17% before (16 measured later requests), 99.05% after (194).
• Both patched cohorts combined: 99.06% over 202 measured later requests; latest 50: 99.88%; latest 20: 99.92%.
• Four measured requests were below 90%, retained in those totals. One reset coincided with input shrinking from 214,155 to 36,948 tokens; another followed about 33 minutes without a successful request in that cohort and recovered on the next success. Compaction/cache expiry are possible explanations, not established causes.

Ratios are sum(cached input)/sum(input) from reported upstream usage; input already includes cached tokens. The first successful request per phase/session/effort is excluded as a cold-start approximation, and eligible requests require at least 1,024 input tokens. Failed requests and missing/estimated usage are not counted as cache measurements. The whole patch window contains 218 requests: 204 HTTP 200, 2 client cancellations (499), 12 connection failures (502), and no 400/422. Connection failures are disclosed separately; this issue does not propose a 502 fix.

This is an uncontrolled within-session observation, not a general cache guarantee. The live local endpoint predicate differs from the PR's registry-based predicate, whose boundary is tested offline. The opt-in trailing-notice stabilization setting stayed enabled during the live patch window; the new offline tests cover both setting values.

Related work / duplicate check

• [Bug]: Mid-conversation role: system messages in Claude inbound are aggregated into top-level instructions, destroying prompt cache prefix #4148 / fix(claude): keep mid-conversation system messages in the timeline #4161: inbound chronology; fix(claude): keep mid-conversation system messages in the timeline #4161 explicitly identifies Chat developer-message re-hoisting as a separate unresolved compatibility tradeoff.
• fix(claude): stabilize Responses instructions for Muse/Go prompt cache #4052 / fix(claude): opt in to harness prompt-cache stabilization (carry #4052) #4347: opt-in handling of supported trailing harness notices; does not change this Chat serializer behavior. In one matched cohort that setting alone yielded 7.13%, versus 9.03% before and 99.59% after the chronological patch.
• fix(openai-chat): keep system messages first #397: existing generic system-first compatibility behavior. fix(chat): preserve OCG DeepSeek timeline system instructions #4438 preserves it for other destinations/models.
• Open Experimental: preserve parent prompt-cache prefixes for Desktop side chats #4222 and feat: Change Astra reasoning mid-conversation without breaking cache #4225 concern Codex Responses side chats and Astra effort changes, respectively; neither implements this OCG Chat exception.

Searched open/closed issues and PRs for cache, DeepSeek, system, hoist and chronological terminology. No other PR implementing this specific residual fix was found; #4438 already exists and is reused. It remains draft because broader local test:changed validation is unresolved, as detailed in its body.

Checks

• I searched existing issues and documentation.
• I removed secrets, tokens, account details, request credentials, and personal data.