better-ccflare 3.5.77 → 3.5.79

package/README.md

@@ -377,6 +377,8 @@ claude
 
 ### Codex CLI as a Client
 
+The Codex VS Code extension uses the same provider configuration. See the [VS Code setup and compatibility guide](docs/codex-vscode.md).
+
 better-ccflare supports [Codex CLI](https://github.com/openai/codex) as a client. Codex speaks the OpenAI Responses API; better-ccflare intercepts requests to `/v1/responses` and `/v1/responses/compact` and translates them to Anthropic `POST /v1/messages` internally, routing through your configured account pool.
 
 Configure Codex CLI to point at better-ccflare in `~/.codex/config.toml`:
@@ -399,7 +401,7 @@ Codex CLI requires an API key to start — use `dummy-key` if better-ccflare API
 Known limitations:
 
 - Regular HTTP does not trust an arbitrary caller-supplied `previous_response_id`. On the authenticated, native outbound Codex route, an operator-controlled gateway request can enable gateway-managed continuation: better-ccflare retains only the last upstream response ID and cryptographic input/configuration digests, resumes only after an exact ordered-prefix match, and otherwise sends the full input without a response ID. This bounded state is in memory, so a restart safely cold-starts. WebSocket transport is not implemented.
-- Built-in tool types (`web_search_preview`, `code_interpreter`, `file_search`) are preserved on the native outbound Codex route. Routes translated to Anthropic skip built-in tools and forward only `type: "function"` tools.
+- Built-in tool types (`web_search_preview`, `code_interpreter`, `file_search`) are preserved on the native outbound Codex route. Routes translated to Anthropic skip built-in tools, forward function tools, and bridge freeform custom tools such as Codex's `apply_patch` through a string input schema. Custom tool grammars are described to the model but cannot be enforced by the Anthropic API.
 - `/v1/responses` rejects unknown top-level request fields with `400 invalid_request_error` instead of silently dropping them. The canonical ChatGPT subscription endpoint also rejects an explicitly supplied `max_output_tokens`; omit it and let the served model choose its output limit. Custom OpenAI-compatible endpoints may accept it.
 - Claude OAuth accounts (Claude Pro/Team, `provider=anthropic` with OAuth tokens) are automatically excluded from Codex CLI traffic — Anthropic bans these when used outside Claude CLI. Anthropic API key accounts are fine and will be used normally.
 
@@ -819,8 +821,8 @@ Inspired by [snipeship/ccflare](https://github.com/snipeship/ccflare) - thanks f
 - [@vansh2408](https://github.com/vansh2408) — [#359](https://github.com/tombii/better-ccflare/pull/359)
 - [@aaapl](https://github.com/aaapl) — [#433](https://github.com/tombii/better-ccflare/pull/433), [#435](https://github.com/tombii/better-ccflare/pull/435)
 - [@mjonuschat](https://github.com/mjonuschat) — [#436](https://github.com/tombii/better-ccflare/pull/436)
-- [@anony253](https://github.com/anony253) — [#452](https://github.com/tombii/better-ccflare/pull/452)
-- [@valentinyanakiev](https://github.com/valentinyanakiev) — [#450](https://github.com/tombii/better-ccflare/pull/450), [#454](https://github.com/tombii/better-ccflare/pull/454)
+- [@anony253](https://github.com/anony253) — [#452](https://github.com/tombii/better-ccflare/pull/452), [#460](https://github.com/tombii/better-ccflare/pull/460)
+- [@valentinyanakiev](https://github.com/valentinyanakiev) — [#450](https://github.com/tombii/better-ccflare/pull/450), [#454](https://github.com/tombii/better-ccflare/pull/454), [#459](https://github.com/tombii/better-ccflare/pull/459)
 
 ## Contributing
 
@@ -841,3 +843,42 @@ MIT - See [LICENSE](LICENSE) for details
 <p align="center">
   Built with ❤️ for developers who ship
 </p>
+
+### Claude Messages continuation on Codex (opt-in)
+
+Set `CCFLARE_CODEX_MESSAGES_CONTINUATION=1` to let the Codex provider retain
+server-owned Responses continuation while serving `/v1/messages` clients. It
+requires an authenticated API key, a selected account, and a nonempty prompt
+cache key (normally derived from Claude session metadata on OpenAI endpoints).
+The first request sends the full translated history, implicit GPT cache options
+with a 30-minute TTL, and a developer-prefix breakpoint. Subsequent exact
+history/configuration matches send only new input and the stored response ID.
+Claude cache markers and signatures never become GPT input artifacts.
+
+State is separated by account, model, authenticated caller, session, and client
+protocol. Changed history/tools, expiration, process restart, unknown output,
+failed or ambiguous terminal events, and cancellation cannot create a reusable
+checkpoint. Streaming remains incremental, but checkpoint promotion requires
+clean upstream EOF. Opaque GPT reasoning remains upstream; the proxy retains
+only bounded in-memory response IDs and replay digests. Public responses remain
+Claude-compatible and cache telemetry continues to label this ingress `legacy`.
+
+This switch does not change model selection or fallback policy. It defaults off;
+disabling it restores the prior Messages request format. It does not transfer
+Anthropic caches or guarantee a cache-hit percentage. Custom endpoints must
+support these Responses controls when this bridge is enabled; authenticated
+Messages requests may then derive cache keys from ordinary session metadata.
+Caller-provided native input, identity, and response IDs remain untrusted.
+
+HTTP continuation requires the endpoint to retain the referenced response. If it
+returns `previous_response_not_found` (or a typed invalid previous-response ID),
+the bridge retries the original full history once on the same account and model,
+then suspends continuation for that chain for 30 minutes. GPT cache controls
+remain active. This does not change `store:false` or enable provider-side
+retention. Backend retention and actual cache reuse need endpoint-specific
+validation; local response-ID existence is not proof of backend availability.
+
+Optionally set `CCFLARE_CODEX_MESSAGES_CONTINUATION_MODELS` to a comma-separated
+list of exact resolved model names to limit this feature to validated models.
+Unset means all models when enabled; an empty list enables none. This scope
+controls adapter features, not model selection or fallback routes.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "better-ccflare",
-  "version": "3.5.77",
+  "version": "3.5.79",
   "description": "Load balancer proxy for Claude API with intelligent distribution across multiple OAuth accounts to avoid rate limiting",
   "license": "MIT",
   "repository": {