Disable Anthropic identity compression on proxy path

[pythondatascrape]

Summary

Anthropic requests proxied through Engram became unreliable when the proxy rewrote Claude's system prompt for identity compression. Direct Claude requests worked, but proxied requests intermittently received upstream 429 rate-limit responses.

Type

Bug

Affected Area

Anthropic proxy path ()

Reproduction Steps

1. Configure Claude Code to route Anthropic traffic through Engram on .
2. Start Engram and issue a normal Claude request with the default large Claude system prompt.
3. Observe upstream 429 responses when Anthropic identity/system compression is enabled.
4. Disable Anthropic system/identity compression and retry the same request.

Expected vs Actual

• Expected: Claude requests proxied through Engram should remain reliable while preserving safe compression.
• Actual: Rewriting the Anthropic prompt for identity compression triggered upstream throttling, while keeping the system prompt verbatim restored stability.

Investigation Findings

• Direct Claude requests succeeded while proxied requests failed only when Anthropic system/identity compression was enabled.
• Disabling Anthropic identity/system compression restored reliable requests while window/context compression continued to work.
• The Anthropic preflight also added unnecessary upstream traffic and was removed as part of the fix.
• Relevant files: , ,

Proposed Resolution / Guidance

Keep the Anthropic prompt verbatim and disable identity compression on the Anthropic proxy path. Continue applying conversation window compression and improved upstream error logging, and leave OpenAI-side identity compression unchanged.

Core Application Impact

P1 — High
This significantly degraded the core Claude-through-Engram workflow, but a direct-Claude workaround existed.

Related

• Branch: