[Bug]: Claude Messages to Astra cache plateau/reset with growing history; Codex CLI comparison #3765

Client or integration

Claude Desktop 3P and ocx claude, compared with Codex CLI.

Area

Proxy and routing: Anthropic Messages to native OpenAI Responses.

Summary

With gpt-6-astra, a Claude Desktop main-agent conversation reused essentially the same 27.6k-token prefix across several continuations, despite growing input. Separate synthetic ocx claude sequences show that this route can grow cache reads, but one also went warm → cold within seconds. A Codex CLI control grew its cached prefix on every continuation.

This reports measured symptoms, not a proven OCX root cause or a claim that every continuation must hit cache. Please investigate prefix preservation and effective session/cache identity at the native forward boundary. Existing logs do not distinguish client prefix changes, OCX translation, and upstream cache placement/expiry.

Version

OpenCodex CLI and running proxy: 2.43.0, no version skew at the time of measurement.

Operating system

macOS 26.5.2.

Provider and model

OpenAI / gpt-6-astra, native OpenAI Responses adapter through a local OCX proxy.

Additional environment details

• Desktop embedded Claude Code: 2.1.260, transcript entrypoint claude-desktop-3p.
• Standalone Claude Code: 2.1.258.
• Codex CLI: 0.153.2.
• Local loopback proxy; native OpenAI account-pool route, openai-responses adapter.
• All successful rows below: resolved gpt-6-astra, requested effort high, HTTP 200, one attempt, the same anonymized account label A. This does not prove a common upstream cache machine/cohort.
• Measurements: 2026-09-06, 10:18–10:34 UTC.

Desktop main-agent evidence

OCX's conversation group also contained a child Astra agent and Luna calls. I matched the local Desktop transcript to OCX usage by input/cache counts and timestamps, selected isSidechain=false, and deduplicated assistant content blocks sharing one message ID. Only the six main-agent requests are below:

Total: 139,520 / 266,773 = 52.3% cache reads; excluding the initial cold request, 62.4%. Requests 2–5 remain at exactly 27,648 cached tokens. The first two responses are text; request 3 returns a Skill call; thinking blocks first appear in response 4. Thus dropping prior thinking blocks cannot explain the earliest plateau by itself. No compaction event appears in this main transcript.

The five child-Astra rows and four Luna rows are excluded from this table. Treating the whole log conversation group as the main agent would give a misleading result.

Synthetic CLI comparisons

Each sequence used a fresh session, the same generated user-text growth pattern, four sequential turns with --resume/exec resume, high effort, and the response OK. No tools were called. Earlier user records were retained; each turn appended 120 new records after an initial 400. Each session's OCX conversation ID remained nonempty and equal across its successful requests.

The last sequence's request starts were 10:33:46, 10:33:49, 10:33:54, and 10:33:58 UTC. Its final successful turn lost all reported cache reads about five seconds after a warm request.

These are minimal controls, not identical wire payloads or a client performance ranking: the clients have different system/tool envelopes, Desktop used real tools, and the standalone/embedded Claude versions differ. Two additional startup rows, one per Claude sequence, returned 499 without reported usage; they are excluded rather than counted as zero-cache successes.

Reproduction

The intermittent miss is not established as deterministic. To repeat the minimal comparison:

1. Use an already running local OCX proxy with the same Astra route. Use a disposable working directory and isolated CLAUDE_CONFIG_DIR for Claude.

2. Generate records using:

   records = '\n'.join(
       f'Record {i:04d}: amber cedar river stone meadow lantern willow garden copper island violet orchard.'
       for i in range(start, end)
   )

   Ranges: [0,400), [400,520), [520,640), [640,760). Each turn asks to retain the reference records, reply only OK, and take no external action.

3. Claude invocation: ocx claude --bare -p --output-format json --model gpt-6-astra --effort high --tools '' --strict-mcp-config --mcp-config '{"mcpServers":{}}' --setting-sources '' --max-turns 1, then --resume <returned-session-id> for later turns. Supply authentication appropriate to the local proxy. For this loopback-only sample, an invocation-scoped apiKeyHelper returned OCX's documented opencodex-proxy marker; no upstream credential or global auth setting was changed. The first Claude sequence additionally supplied a fixed --system-prompt asking only for OK and no tools. The second omitted that override.

4. Codex invocation: codex exec --ignore-user-config --skip-git-repo-check --json -m gpt-6-astra -c 'openai_base_url="http://127.0.0.1:10100/v1"' -c 'model_reasoning_effort="high"' -c 'project_doc_max_bytes=0' -c 'features.shell_tool=false' -c 'features.apply_patch_freeform=false' -c 'sandbox_mode="read-only"' -c 'approval_policy="never"' -. Continue the returned ID with codex exec resume, retaining the same overrides. Prompts go through stdin.

5. Correlate successful requests in usage.jsonl. OCX/OpenAI inputTokens includes cached input; Claude's client input_tokens excludes it. Add Claude input + cache-read + cache-creation before comparing. Client totals and OCX totals agreed in these samples.

Source checks and remaining evidence gap

The following installed files were byte-identical to main at commit 06ec553630fa2ee51a96b5cbf694089021249194:

• inbound.ts: metadata-backed cache key generation; also contains the assistant thinking-drop policy.
• claude-messages.ts: native metadata-backed requests synthesize an underscore session_id header from that key.
• outbound.ts: cache counters are mapped into Anthropic usage.
• src/adapters/openai-responses.ts: includes session_id in forward headers and reads upstream cached-token usage.

The populated conversation groups are consistent with metadata-backed identity, but there was no inbound/outbound wire capture. Stable logged identity/account labels do not establish byte-identical prefixes or the effective forwarded headers. Claude inbound debug was disabled. The next useful capture would record only field presence/equality and prefix-change locations across a warm/cold pair, including tools/instructions/history, cache key, underscore/hyphenated session identifiers, and selected account. No private request content is needed publicly.

Related: #3433 tracks Hermes Chat Completions cache misses with missing identity fields; this report adds the Claude Messages path with populated conversation correlation and a Desktop main-agent plateau. #3719 explicitly concerns intentional Anthropic upstream thinking replay; this report's upstream is OpenAI Astra. No duplicate claiming this exact comparison was found.

Checks

• Searched existing issues and current source.
• Removed personal paths, original prompts, session IDs, account identifiers, credentials, and private tool output.
• No proxy/provider configuration change, restart, or local source patch was made for this investigation.