[MikeTuev]

Summary

• What was broken: fix(sse): keep mid-conversation system turns in place on anthropic-compatible relays #16236 keeps mid-conversation role: "system" turns inside messages[] for plain
  anthropic-compatible-* relays only when the global ANTHROPIC_COMPATIBLE_MID_CONVERSATION_SYSTEM flag is on
  and the client itself sends the mid-conversation-system-2026-04-07 anthropic-beta. Several Claude Code
  builds (cc_entrypoint=cli 2.1.299, claude-desktop-3p 2.1.308 on our deployment) put those turns in the body
  but do not send the beta. With the flag on, their turns are still hoisted into the top-level system.
  system then grows by one <total_tokens> line every turn and the cached history after it never matches:
  one session read 30461 → 30482 → 30503 → 30524 … from cache (exactly +21 tokens per turn) while ~110–120K was
  re-written every turn. A client on the same relay that does send the beta read 93% from cache.
• Root cause: whether role: "system" inside messages[] is accepted is a property of the relay, not of
  the client. Gating on the client's beta silently disables the fix for clients that never negotiate it, and a
  global flag cannot distinguish relays that 400 on the shape from relays that accept it.

Fix: the decision moves to the connection.

• New providerSpecificData.keepMidConversationSystem (strict boolean, default off), shown as a toggle in the
  edit-connection modal for plain anthropic-compatible-* only (not CC-compatible / native claude), en/ru
  strings, validated + normalized like the other boolean knobs.
• resolveClaudeMidConversationSystemPolicy() now reads that opt-in from the connection's
  providerSpecificData (plumbed through normalizeClaudeUpstreamMessages → liftSystemRoleMessages) and no
  longer requires the client beta. The client's own beta is still forwarded when present
  (applyClientAnthropicBeta, unchanged), never invented.
• The global ANTHROPIC_COMPATIBLE_MID_CONVERSATION_SYSTEM flag is removed: definition, i18n description, and
  the FEATURE_FLAGS.md rows / counts (111 → 110 flags, Runtime 53 → 52) in all locales.
• Default behaviour (opt-in off) is byte-for-byte the legacy full hoist, as before.

Migration note: an operator who had the global flag on needs to enable the toggle on the relevant
connections. The flag's DB override row becomes inert.

Related Issues

• Refs fix(sse): keep mid-conversation system turns in place on anthropic-compatible relays #16236

Target Branch

• Base is the active release/v3.8.x branch (the highest open release/v*), not main

Proof Of Fix (required)

• TDD: tests/unit/claude-mid-conversation-system-compatible-passthrough.test.ts drives the real
  handleChatCore with fetch captured. New cases: beta-less client and client without any anthropic-beta
  stay in place with the beta never invented; a beta-less client across successive turns keeps system stable
  and earlier messages a prefix; strict-boolean policy; validation/normalization.
  Production files reverted to the base, test kept:

  ✖ keepMidConversationSystem: validated and normalized as a strict boolean
  ✖ opt-in on, client without the beta: turns stay in place and the beta is never invented
  ✖ opt-in on, client without any anthropic-beta header: in place, no beta header invented
  ✖ opt-in on, beta-less client: successive turns keep system stable and messages a prefix
  ✖ policy: scope and strict per-connection opt-in
  … (11 fail / 5 pass)
  

  After the fix: all pass.
• Real-environment smoke — runner-base image, throwaway container, plain anthropic-compatible-*
  connection pointed at a capture relay, Claude Code-shaped request with mid-conversation system turns and a
  client anthropic-beta without the mid-conversation token:
  • toggle off: upstream got roles [user, assistant, user], system 3 blocks (hoisted);
  • toggle on: upstream got [user, system, assistant, user, system], system 1 block, no beta invented, 200.
• Production evidence (read-only analysis of call logs on our deployment) is summarized above.

Scope

• Only the files needed for the opt-in move are touched (+ flag removal in docs/i18n, the two flag-count tests)
• No test was weakened or deleted; the "flag on + no beta → hoisted" cases were inverted to the new contract

Validation

• Change type: routing (+ small UI/i18n, docs)
• 44 neighbouring unit files (mid-conversation, claude passthrough, system role, anthropic headers/beta,
  feature-flags incl. doc-sync, provider-specific-data, executor-default, cc-compatible, chatcore translation):
  533/533
• npm run typecheck:core, eslint on touched files
• npm run check:docs-all: 81 STRICT drifts, identical on the base (translated README migration counts 210 vs 211); none from this PR

Reviewer Notes

• Removing a feature flag changes the public flag count; I updated every count I found (docs + the two tests).
• If a relay rejects role: "system" in messages[] it answers 400 and the operator turns the toggle off for that connection only.
• Touches the same edit-connection modal / validation files as fix(sse): stop prefill 400s from locking models; opt-in user-last for anthropic-compatible #18332 (both add one boolean next to disableCooling); whichever lands second needs a trivial merge.

[freefrank]

Same here on 3.8.53, anthropic-compatible connection to CLIProxyAPI 8.0.13 (Claude OAuth).

• Claude Code 2.1.312 sends the <total_tokens> system turns without mid-conversation-system-2026-04-07. It also drops that token from ANTHROPIC_BETAS (debug log: "is generally available; the header is not sent"), so there's no client-side way to opt in.
• With only the client-beta check removed from resolveClaudeMidConversationSystemPolicy, cache reads went from stuck at ~40k to 93–99% from the second turn.
• CLIProxyAPI accepts the in-place system turns without the beta: 200, turn 2 reads all of turn 1.

[coderabbitai]

Caution

Review failed

An error occurred during the review process. Please try again later.

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts