codex-as-api 0.8.2__tar.gz → 0.9.2__tar.gz

{codex_as_api-0.8.2 → codex_as_api-0.9.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: codex-as-api
-Version: 0.8.2
+Version: 0.9.2
 Summary: Use ChatGPT/Codex OAuth as a local OpenAI-compatible API server.
 Project-URL: Homepage, https://github.com/Eunho-J/codex-as-api
 Project-URL: Repository, https://github.com/Eunho-J/codex-as-api
@@ -192,9 +192,25 @@ model_auto_compact_token_limit = 244800
 
 ### Supported Models
 
-Models and capabilities come only from the authenticated Codex `/models` response. The proxy keeps each account/base-URL/client-version snapshot in memory for 300 seconds, invalidates it when a Responses `X-Models-Etag` changes, and never serves a static, stale, or cross-account fallback. An unavailable or malformed refresh returns `503 catalog_unavailable` before streaming headers are sent; an upstream HTTP rejection after the one allowed OAuth refresh preserves its status as `upstream_error`.
+Ordinary models and capabilities come from the authenticated Codex `/models` response. The proxy keeps each account/base-URL/client-version snapshot in memory for 300 seconds and invalidates it when a Responses `X-Models-Etag` changes. An unavailable or malformed refresh returns `503 catalog_unavailable` before streaming headers are sent; an upstream HTTP rejection after the one allowed OAuth refresh preserves its status as `upstream_error`. Model snapshots are never shared across accounts or served as an offline fallback.
 
-Use `GET /v1/models` instead of relying on a documentation list. Every row in the fresh authenticated snapshot is returned, including its exact slug, visibility, `supported_in_api`, reasoning-summary controls, and opaque `comp_hash` compatibility identifier when supplied. A valid empty catalog returns `200` with `data: []`; health and model-default resolution then return `503 catalog_unavailable` rather than selecting a fallback. Requests must use an exact live slug; unknown explicit IDs return `404 model_not_found`.
+Use `GET /v1/models` for the effective catalog. It includes exact slugs, visibility, `supported_in_api`, reasoning-summary controls, and opaque `comp_hash` compatibility identifiers when supplied. Unknown IDs return `404 model_not_found`. Automatic default selection uses only the upstream catalog's originally visible models, not the experimental additions below.
+
+#### Experimental hidden models
+
+These three upstream-hidden models returned real responses on October 24, 2026, and are explicitly included in codex-as-api's effective catalog:
+
+| Model ID | Observed upstream purpose |
+| --- | --- |
+| `gpt-6-chat-thinking` | Server-described paid CCA experiment; hidden and marked unsupported in the upstream API catalog |
+| `gpt-reserve` | Official Codex CLI's Luna Reserve routing for reserve usage |
+| `codex-auto-review` | Official Codex CLI's Guardian automatic approval-review model |
+
+A small authenticated smoke test also verified function-tool calling plus tool-result continuation for all three models, and executed a simple generated JavaScript function against four inputs per model. This is a basic functionality check, not a coding-quality benchmark.
+
+**These are hidden models. Upstream support can stop at any time, and availability may differ by account. If a model stops responding, [open an issue](https://github.com/Eunho-J/codex-as-api/issues) with its model ID and a sanitized error; we will remove the discontinued model from this catalog.** Do not include access tokens or other credentials in the issue.
+
+The proxy exposes these verified exceptions as selectable models and prefers their current live capability metadata. Bundled capability records keep only these three IDs available when a successful live catalog omits them; they do not bypass OAuth, upstream rejection, or a failed/malformed catalog fetch. A successful empty upstream catalog can therefore list these experimental entries, while health/default selection still fails until an ordinary upstream default is available. The proxy never selects a hidden experimental model as its automatic default.
 
 To use a different port:
 
@@ -284,11 +300,9 @@ it; continuation uses the complete provider Response items saved behind
 
 ### `POST /v1/messages`
 
-Claude Code compatibility endpoint using the Anthropic Messages shape. It supports streaming (`stream: true`) and non-streaming only for requests carrying one valid `x-claude-code-session-id` header. A live Codex model ID routes directly. A recognized `claude-*` facade ID requires an explicit `CODEX_AS_API_MODEL` or Codex config model that is present in the same fresh catalog; it is not silently routed to a default. Other unknown IDs fail. The response preserves the client-supplied model name.
+Anthropic Messages compatibility endpoint for any client, including Claude Code and Scheherazade. It supports streaming (`stream: true`) and non-streaming without a client-identification header. A live Codex model ID routes directly. A `claude-*` facade ID uses the configured Codex backend model, or the authenticated catalog default when no backend model is configured. Other unknown IDs fail. The response preserves the client-supplied model name.
 
-This compatibility endpoint is scoped to Claude Code requests. The proxy accepts
-and strips the non-forwarded `max_tokens` and `cache_control` compatibility
-fields only when one valid `x-claude-code-session-id` header is present.
+`max_tokens` is a required positive compatibility hint: the private Codex OAuth transport does not expose a matching output-token cap. Valid ephemeral `cache_control` hints are accepted and stripped for every client; their TTL and breakpoint semantics are not forwarded. An optional `x-claude-code-session-id` provides stable hashed cache affinity, not permission to use the endpoint. `metadata.user_id` is accepted as client metadata and is not forwarded.
 
 ```bash
 curl http://localhost:18080/v1/messages \
@@ -325,7 +339,11 @@ curl -N http://localhost:18080/v1/messages \
   }'
 ```
 
-Codex reasoning summaries are not Anthropic signed-thinking blocks. A non-streaming response exposes real upstream reasoning only through the top-level `codex_reasoning` proxy extension. A stream emits `event: codex_reasoning_delta` with `type: "codex_reasoning_delta"`. The proxy never fabricates `thinking`, `thinking_delta`, or `signature_delta` data.
+With `thinking.display: "summarized"` (the enabled/adaptive default), real Codex reasoning summaries are returned as Anthropic `thinking` blocks and streamed `thinking_delta` content, alongside the existing `codex_reasoning` / `codex_reasoning_delta` extensions. When upstream supplies encrypted reasoning, each reasoning item has its own proxy-specific `signature`: `codex-reasoning-v1:` followed by base64url JSON containing the format version, actual backend model ID, and original reasoning item ID, encrypted content and summary. This is an opaque continuation envelope, not an Anthropic cryptographic signature.
+
+On the following assistant-history request, recognized envelopes restore the original Responses reasoning items before that assistant turn's text/tool calls. Historical signed reasoning is accepted even when the current call uses `thinking: {"type":"disabled"}`, including compaction calls: current output-generation settings do not invalidate prior reasoning input. Ciphertext is preserved unchanged; summary text is never converted to ordinary assistant speech. Missing or foreign signatures retain the legacy ignored-thinking behavior, while malformed recognized envelopes return HTTP 400. Reasoning is retained for the same backend model and across the verified `gpt-6.1-sol` / `gpt-6-astra` / `gpt-6-sol` / `gpt-6-luna` / `gpt-daybreak-blue-latest` combinations (25 authenticated source/target pairs); other cross-model envelopes are discarded while ordinary history remains.
+
+Multiple summary parts share one thinking block per reasoning item. Encrypted-only items use an empty thinking string with a non-empty signature. Streaming sends `signature_delta` before the matching `content_block_stop`. `display: "omitted"` hides the summary but can preserve an encrypted-only signed block for continuity; disabled thinking suppresses thinking output. Items without encrypted content retain unsigned-summary compatibility behavior.
 
 Hosted Anthropic `web_search` declarations are rejected with HTTP 400 on
 Messages, token-counting, and compact requests. OpenAI Responses cannot supply
@@ -345,10 +363,7 @@ the completion is nonterminal or final usage is absent; a stream that has
 already started emits an Anthropic `error` event and never emits a successful
 `message_stop`. The final response includes every required usage field, using
 `null` only where the Anthropic type permits it, plus required container,
-citation, context-management, and direct-caller fields. To preserve Claude
-Code's realtime behavior, `message_start` is sent immediately without guessed
-input-token counts; this narrow stream-start shape is Claude Code compatibility
-behavior, not generic Anthropic API parity.
+citation, context-management, and direct-caller fields. `message_start` is sent immediately with numeric zero counters in `message.usage`; these are provisional, not measured counts. The final `message_delta.usage` replaces them with authoritative cumulative totals, including disjoint cache buckets.
 
 Generated `tool_use` blocks with `caller: {"type":"direct"}` and empty
 assistant content arrays can be replayed through the request adapter. Malformed,
@@ -364,9 +379,9 @@ different jobs:
 | `client_metadata.session_id` | Forwarded to Codex, used as the default `prompt_cache_key`, and required when Codex metadata mode is enabled | Not used or forwarded |
 | `client_metadata.thread_id` | Codex metadata thread identity; defaults to `session_id` for a root request and preserves an explicit child thread | Not used or synthesized |
 | `prompt_cache_key` | Explicit value wins; otherwise a non-empty `client_metadata.session_id` is used; otherwise omitted | Optional proxy extension; explicit value wins, otherwise the Claude Code session header is hashed |
-| `x-claude-code-session-id` | Not used | Required for the non-forwarded `max_tokens` and `cache_control` compatibility fields; its exact value also supplies hashed cache affinity |
+| `x-claude-code-session-id` | Not used | Optional cache-affinity header; its exact value supplies a hashed default key |
 | `previous_response_id` | Resolves a known entry from the 256-chain process-local history and replays full input/output over HTTP; never forwarded | Non-null values return Anthropic-style HTTP 400 because normal Messages is stateless |
-| `cache_control` | Not an OpenAI Chat control | Accepted only with the Claude Code session header, validated, then stripped before Codex transport |
+| `cache_control` | Not an OpenAI Chat control | Accepted for every client, validated, then stripped before Codex transport |
 
 Anthropic `cache_control` is a cache-boundary annotation, not an identifier.
 This proxy accepts only `{"type":"ephemeral"}` with optional `ttl: "5m"` or
@@ -378,15 +393,13 @@ responses do not expose the Chat-only `response_id`. The custom
 `previous_response_id` support.
 
 The Anthropic type permits `cache_control: null`; the proxy treats that exact
-value as omission. A non-null cache-control object is validated in full and is
-accepted only on a Claude Code request.
+value as omission. A non-null cache-control object is validated in full for every client.
 
 ### `POST /v1/messages/count_tokens`
 
-Anthropic-compatible token counting helper. Codex OAuth does not expose a count-only endpoint equivalent to Anthropic's native API, so this route counts text locally and returns context-window metadata for the effective backend model. Normalized model-visible messages, tool calls, tool-result metadata, reasoning, and tool schemas are counted once. Request-envelope fields and the approved Claude Code `max_tokens` compatibility hint are excluded; other unsupported generation controls fail instead of becoming no-ops. Image inputs use a separate fixed estimate so inline base64 data is not counted as text.
+Anthropic-compatible token counting helper. Codex OAuth does not expose a count-only endpoint equivalent to Anthropic's native API, so this route counts text locally and returns context-window metadata for the effective backend model. Normalized model-visible messages, tool calls, tool-result metadata, reasoning, and tool schemas are counted once. Request-envelope fields and the `max_tokens` compatibility hint are excluded; other unsupported generation controls fail instead of becoming no-ops. Image inputs use a separate fixed estimate so inline base64 data is not counted as text.
 
-`max_tokens` and `cache_control` remain Claude Code-only compatibility hints on
-this route and require `x-claude-code-session-id` when present.
+`max_tokens` and `cache_control` are optional client-neutral compatibility hints on this route; no Claude Code header is required.
 
 Text estimation uses a bundled port of official `tiktoken` `o200k_base` `encode_ordinary`: the same Unicode pre-tokenization regex, byte-pair merge algorithm, and merge-rank data are implemented in Python, TypeScript, and Rust. It is a local estimator, not model metadata and not a substitute for a provider count API. The project does not depend on a `tiktoken` package or download encoding data at runtime. The last upstream synchronization check was **2026-07-14**, against [`tiktoken` 0.13.0 at `08a5f3b`](https://github.com/openai/tiktoken/tree/08a5f3b2c987ada4fc5aa1f16c643c203fa8acaa); the bundled rank file SHA-256 is `446a9538cb6c348e3516120d7c08b09f57c36495e2acfffe59a5bf8b0cfb1a2d`.
 
@@ -405,7 +418,7 @@ curl http://localhost:18080/v1/messages/count_tokens \
 
 ### `POST /v1/messages/compact`
 
-Anthropic-compatible alias for remote conversation compaction. Accepts Anthropic Messages-shaped bodies and returns compacted checkpoint content. Live GPT model IDs select the matching backend model, and Claude Code effort and Fast Mode controls use the same mappings as `/v1/messages`. Its required non-forwarded `max_tokens` and any `cache_control` hints require one valid `x-claude-code-session-id` header.
+Anthropic-compatible alias for remote conversation compaction. Accepts Anthropic Messages-shaped bodies and returns compacted checkpoint content. Live GPT model IDs select the matching backend model, and Claude Code effort and Fast Mode controls use the same mappings as `/v1/messages`. Its required non-forwarded `max_tokens` and any valid `cache_control` hints are accepted for every client; the optional Claude Code session header only supplies cache affinity.
 
 ### `POST /v1/images/generations`
 
@@ -545,7 +558,7 @@ Requests can instead use the Responses-shaped effort/context object:
 - Responses Lite uses `all_turns` as the Codex wire default. An explicitly different context is rejected instead of silently overwritten; use `responses_lite: false` when the backend route supports classic Responses and another context is required.
 - Remote compact keeps its existing private Codex `reasoning_effort` field but does not accept public `reasoning.mode` or `reasoning.context`.
 
-Anthropic `thinking` values map as `enabled → high`, `adaptive → medium`, and `disabled → none`. Claude Code's `output_config.effort` takes precedence over adaptive or enabled thinking and supports `low`, `medium`, `high`, `xhigh`, and `max`. Call-level `thinking.disabled` takes precedence over ambient `output_config.effort` for representable client-side auxiliary calls such as WebFetch; hosted Anthropic WebSearch is rejected before transport because its lifecycle cannot be translated losslessly. The supported reasoning context and verbosity extensions are also available on `/v1/messages`, image generation, and inspection requests; Pro, non-null public cache policy/breakpoints, and any supplied `safety_identifier` fail explicitly on the private Codex provider.
+Anthropic `thinking` values map as `enabled → high` and `adaptive → medium`. Disabled thinking selects `none` when available, otherwise the lowest supported catalog effort, and suppresses reasoning output; it does not claim to eliminate a model's internal reasoning. `output_config.effort` takes precedence over adaptive or enabled thinking and supports `low`, `medium`, `high`, `xhigh`, and `max`. Call-level disabled thinking takes precedence over ambient `output_config.effort` for auxiliary calls such as WebFetch. Explicit unsupported effort requests still fail. Hosted Anthropic WebSearch is rejected because its lifecycle cannot be translated losslessly. The supported reasoning context and verbosity extensions are also available on `/v1/messages`, image generation, and inspection requests; Pro, non-null public cache policy/breakpoints, and any supplied `safety_identifier` fail explicitly on the private Codex provider.
 
 Anthropic's `JSONOutputFormat` carries a schema but no name. When Claude Code
 sends that official shape, the proxy follows pinned Codex 0.153.3 and supplies
@@ -587,7 +600,7 @@ unchanged. The proxy does not inherit speed defaults from Codex config.
 
 Official behavior reference: Codex [0.164.0 service-tier resolution](https://github.com/openai/codex/blob/rust-v0.164.0/codex-rs/tui/src/service_tier_resolution.rs)
 and [stable, enabled-by-default ultrafast feature](https://github.com/openai/codex/blob/rust-v0.164.0/codex-rs/features/src/lib.rs).
-The rest of the private HTTP contract remains pinned to 0.153.3.
+The bundled private-protocol reference records the reviewed Codex source revision. Outbound client identity is separate: the proxy resolves the stable `latest` version of npm [`@openai/codex`](https://www.npmjs.com/package/@openai/codex) and uses that same version for the model-catalog `client_version` query and Codex User-Agent. It is not hardcoded to the bundled reference or inferred from a locally installed CLI. Successful version lookups are cached for one hour; a refresh failure retains the last successfully resolved version and retries after a short backoff. A cold lookup failure returns an explicit availability error rather than advertising an invented version.
 
 Authenticated tests on July 10, 2026 verified both official continuation paths with `gpt-5.6-sol`: a direct private Responses WebSocket completed a second delta request using the exact prior `response.id`, and the private HTTP endpoint completed the same continuation when the full prior input/output history was replayed. The proxy implements the HTTP replay strategy. Direct HTTP forwarding of `previous_response_id`, Pro, public cache controls, and `safety_identifier` was rejected and is not used. See OpenAI's [reasoning mode documentation](https://developers.openai.com/api/docs/guides/reasoning#reasoning-mode) and the Codex 0.153.3 [HTTP/WebSocket request structures](https://github.com/openai/codex/blob/b1a547b1f73ce86205d9222ac19cff334b3b7a2e/codex-rs/codex-api/src/common.rs#L275-L350).
 
@@ -819,9 +832,9 @@ ANTHROPIC_BASE_URL=http://127.0.0.1:18080 ANTHROPIC_AUTH_TOKEN=unused \
   claude --model gpt-5.6-luna --effort medium
 ```
 
-Live Codex IDs are sent to the matching backend model. Built-in Claude/Fable/Opus/Sonnet/Haiku IDs are not Codex models, so recognized `claude-*` IDs require the explicitly configured `CODEX_AS_API_MODEL`; the proxy preserves the selected client name in the Anthropic response.
+Live Codex IDs are sent to the matching backend model. Built-in Claude/Fable/Opus/Sonnet/Haiku IDs are not Codex models, so recognized `claude-*` IDs use the configured Codex model or authenticated catalog default; the proxy preserves the selected client name in the Anthropic response.
 
-`ANTHROPIC_BASE_URL` applies to the whole Claude Code process. Therefore the built-in rows remain available in the picker, but they also pass through codex-as-api and use the explicitly configured Codex backend; the rows do not call Anthropic models in that process. To use an actual Anthropic model and GPT in parallel, keep the gateway variables out of global settings, start the GPT process with the inline command above, and start a second process normally:
+`ANTHROPIC_BASE_URL` applies to the whole Claude Code process. Therefore the built-in rows remain available in the picker, but they also pass through codex-as-api and use the configured or catalog-default Codex backend; the rows do not call Anthropic models in that process. To use an actual Anthropic model and GPT in parallel, keep the gateway variables out of global settings, start the GPT process with the inline command above, and start a second process normally:
 
 ```bash
 claude --model sonnet
@@ -831,7 +844,7 @@ If the gateway variables are already in `~/.claude/settings.json`, shell-level `
 
 Claude Code gateway discovery cannot expose raw `gpt-*` IDs because Claude Code filters discovered IDs to `claude*` and `anthropic*`. The official single custom-model option avoids a misleading facade ID. This integration is an Anthropic Messages-compatible bridge; Anthropic explicitly does not support routing Claude Code to non-Claude models through third-party gateways. See the official [model configuration](https://code.claude.com/docs/en/model-config), [gateway connection](https://code.claude.com/docs/en/llm-gateway-connect), and [gateway protocol](https://code.claude.com/docs/en/llm-gateway-protocol) references.
 
-The current bridge accepts Claude Code's exact no-op thinking cleanup (`clear_thinking_20251015` with `keep: "all"`). Context edits that would change history, task budgets, enabled beta tool fields such as `strict` or `defer_loading`, malformed output formats, and unsupported image source types return HTTP 400 because the Codex OAuth transport has no lossless equivalent. Valid base64 and URL image blocks plus `tool_result.is_error` are preserved during translation.
+The bridge accepts Claude Code's no-op thinking cleanup (`clear_thinking_20251015` with `keep: "all"`), validates and ignores boolean `eager_input_streaming`, and accepts `defer_loading: false`. Context edits that would change history, task budgets, enabled strict/deferred tool execution, malformed output formats, and unsupported image source types return HTTP 400. Valid base64 and URL image blocks plus `tool_result.is_error` are preserved during translation.
 
 ## Architecture
 

{codex_as_api-0.8.2 → codex_as_api-0.9.2}/README.md

@@ -158,9 +158,25 @@ model_auto_compact_token_limit = 244800
 
 ### Supported Models
 
-Models and capabilities come only from the authenticated Codex `/models` response. The proxy keeps each account/base-URL/client-version snapshot in memory for 300 seconds, invalidates it when a Responses `X-Models-Etag` changes, and never serves a static, stale, or cross-account fallback. An unavailable or malformed refresh returns `503 catalog_unavailable` before streaming headers are sent; an upstream HTTP rejection after the one allowed OAuth refresh preserves its status as `upstream_error`.
+Ordinary models and capabilities come from the authenticated Codex `/models` response. The proxy keeps each account/base-URL/client-version snapshot in memory for 300 seconds and invalidates it when a Responses `X-Models-Etag` changes. An unavailable or malformed refresh returns `503 catalog_unavailable` before streaming headers are sent; an upstream HTTP rejection after the one allowed OAuth refresh preserves its status as `upstream_error`. Model snapshots are never shared across accounts or served as an offline fallback.
 
-Use `GET /v1/models` instead of relying on a documentation list. Every row in the fresh authenticated snapshot is returned, including its exact slug, visibility, `supported_in_api`, reasoning-summary controls, and opaque `comp_hash` compatibility identifier when supplied. A valid empty catalog returns `200` with `data: []`; health and model-default resolution then return `503 catalog_unavailable` rather than selecting a fallback. Requests must use an exact live slug; unknown explicit IDs return `404 model_not_found`.
+Use `GET /v1/models` for the effective catalog. It includes exact slugs, visibility, `supported_in_api`, reasoning-summary controls, and opaque `comp_hash` compatibility identifiers when supplied. Unknown IDs return `404 model_not_found`. Automatic default selection uses only the upstream catalog's originally visible models, not the experimental additions below.
+
+#### Experimental hidden models
+
+These three upstream-hidden models returned real responses on October 24, 2026, and are explicitly included in codex-as-api's effective catalog:
+
+| Model ID | Observed upstream purpose |
+| --- | --- |
+| `gpt-6-chat-thinking` | Server-described paid CCA experiment; hidden and marked unsupported in the upstream API catalog |
+| `gpt-reserve` | Official Codex CLI's Luna Reserve routing for reserve usage |
+| `codex-auto-review` | Official Codex CLI's Guardian automatic approval-review model |
+
+A small authenticated smoke test also verified function-tool calling plus tool-result continuation for all three models, and executed a simple generated JavaScript function against four inputs per model. This is a basic functionality check, not a coding-quality benchmark.
+
+**These are hidden models. Upstream support can stop at any time, and availability may differ by account. If a model stops responding, [open an issue](https://github.com/Eunho-J/codex-as-api/issues) with its model ID and a sanitized error; we will remove the discontinued model from this catalog.** Do not include access tokens or other credentials in the issue.
+
+The proxy exposes these verified exceptions as selectable models and prefers their current live capability metadata. Bundled capability records keep only these three IDs available when a successful live catalog omits them; they do not bypass OAuth, upstream rejection, or a failed/malformed catalog fetch. A successful empty upstream catalog can therefore list these experimental entries, while health/default selection still fails until an ordinary upstream default is available. The proxy never selects a hidden experimental model as its automatic default.
 
 To use a different port:
 
@@ -250,11 +266,9 @@ it; continuation uses the complete provider Response items saved behind
 
 ### `POST /v1/messages`
 
-Claude Code compatibility endpoint using the Anthropic Messages shape. It supports streaming (`stream: true`) and non-streaming only for requests carrying one valid `x-claude-code-session-id` header. A live Codex model ID routes directly. A recognized `claude-*` facade ID requires an explicit `CODEX_AS_API_MODEL` or Codex config model that is present in the same fresh catalog; it is not silently routed to a default. Other unknown IDs fail. The response preserves the client-supplied model name.
+Anthropic Messages compatibility endpoint for any client, including Claude Code and Scheherazade. It supports streaming (`stream: true`) and non-streaming without a client-identification header. A live Codex model ID routes directly. A `claude-*` facade ID uses the configured Codex backend model, or the authenticated catalog default when no backend model is configured. Other unknown IDs fail. The response preserves the client-supplied model name.
 
-This compatibility endpoint is scoped to Claude Code requests. The proxy accepts
-and strips the non-forwarded `max_tokens` and `cache_control` compatibility
-fields only when one valid `x-claude-code-session-id` header is present.
+`max_tokens` is a required positive compatibility hint: the private Codex OAuth transport does not expose a matching output-token cap. Valid ephemeral `cache_control` hints are accepted and stripped for every client; their TTL and breakpoint semantics are not forwarded. An optional `x-claude-code-session-id` provides stable hashed cache affinity, not permission to use the endpoint. `metadata.user_id` is accepted as client metadata and is not forwarded.
 
 ```bash
 curl http://localhost:18080/v1/messages \
@@ -291,7 +305,11 @@ curl -N http://localhost:18080/v1/messages \
   }'
 ```
 
-Codex reasoning summaries are not Anthropic signed-thinking blocks. A non-streaming response exposes real upstream reasoning only through the top-level `codex_reasoning` proxy extension. A stream emits `event: codex_reasoning_delta` with `type: "codex_reasoning_delta"`. The proxy never fabricates `thinking`, `thinking_delta`, or `signature_delta` data.
+With `thinking.display: "summarized"` (the enabled/adaptive default), real Codex reasoning summaries are returned as Anthropic `thinking` blocks and streamed `thinking_delta` content, alongside the existing `codex_reasoning` / `codex_reasoning_delta` extensions. When upstream supplies encrypted reasoning, each reasoning item has its own proxy-specific `signature`: `codex-reasoning-v1:` followed by base64url JSON containing the format version, actual backend model ID, and original reasoning item ID, encrypted content and summary. This is an opaque continuation envelope, not an Anthropic cryptographic signature.
+
+On the following assistant-history request, recognized envelopes restore the original Responses reasoning items before that assistant turn's text/tool calls. Historical signed reasoning is accepted even when the current call uses `thinking: {"type":"disabled"}`, including compaction calls: current output-generation settings do not invalidate prior reasoning input. Ciphertext is preserved unchanged; summary text is never converted to ordinary assistant speech. Missing or foreign signatures retain the legacy ignored-thinking behavior, while malformed recognized envelopes return HTTP 400. Reasoning is retained for the same backend model and across the verified `gpt-6.1-sol` / `gpt-6-astra` / `gpt-6-sol` / `gpt-6-luna` / `gpt-daybreak-blue-latest` combinations (25 authenticated source/target pairs); other cross-model envelopes are discarded while ordinary history remains.
+
+Multiple summary parts share one thinking block per reasoning item. Encrypted-only items use an empty thinking string with a non-empty signature. Streaming sends `signature_delta` before the matching `content_block_stop`. `display: "omitted"` hides the summary but can preserve an encrypted-only signed block for continuity; disabled thinking suppresses thinking output. Items without encrypted content retain unsigned-summary compatibility behavior.
 
 Hosted Anthropic `web_search` declarations are rejected with HTTP 400 on
 Messages, token-counting, and compact requests. OpenAI Responses cannot supply
@@ -311,10 +329,7 @@ the completion is nonterminal or final usage is absent; a stream that has
 already started emits an Anthropic `error` event and never emits a successful
 `message_stop`. The final response includes every required usage field, using
 `null` only where the Anthropic type permits it, plus required container,
-citation, context-management, and direct-caller fields. To preserve Claude
-Code's realtime behavior, `message_start` is sent immediately without guessed
-input-token counts; this narrow stream-start shape is Claude Code compatibility
-behavior, not generic Anthropic API parity.
+citation, context-management, and direct-caller fields. `message_start` is sent immediately with numeric zero counters in `message.usage`; these are provisional, not measured counts. The final `message_delta.usage` replaces them with authoritative cumulative totals, including disjoint cache buckets.
 
 Generated `tool_use` blocks with `caller: {"type":"direct"}` and empty
 assistant content arrays can be replayed through the request adapter. Malformed,
@@ -330,9 +345,9 @@ different jobs:
 | `client_metadata.session_id` | Forwarded to Codex, used as the default `prompt_cache_key`, and required when Codex metadata mode is enabled | Not used or forwarded |
 | `client_metadata.thread_id` | Codex metadata thread identity; defaults to `session_id` for a root request and preserves an explicit child thread | Not used or synthesized |
 | `prompt_cache_key` | Explicit value wins; otherwise a non-empty `client_metadata.session_id` is used; otherwise omitted | Optional proxy extension; explicit value wins, otherwise the Claude Code session header is hashed |
-| `x-claude-code-session-id` | Not used | Required for the non-forwarded `max_tokens` and `cache_control` compatibility fields; its exact value also supplies hashed cache affinity |
+| `x-claude-code-session-id` | Not used | Optional cache-affinity header; its exact value supplies a hashed default key |
 | `previous_response_id` | Resolves a known entry from the 256-chain process-local history and replays full input/output over HTTP; never forwarded | Non-null values return Anthropic-style HTTP 400 because normal Messages is stateless |
-| `cache_control` | Not an OpenAI Chat control | Accepted only with the Claude Code session header, validated, then stripped before Codex transport |
+| `cache_control` | Not an OpenAI Chat control | Accepted for every client, validated, then stripped before Codex transport |
 
 Anthropic `cache_control` is a cache-boundary annotation, not an identifier.
 This proxy accepts only `{"type":"ephemeral"}` with optional `ttl: "5m"` or
@@ -344,15 +359,13 @@ responses do not expose the Chat-only `response_id`. The custom
 `previous_response_id` support.
 
 The Anthropic type permits `cache_control: null`; the proxy treats that exact
-value as omission. A non-null cache-control object is validated in full and is
-accepted only on a Claude Code request.
+value as omission. A non-null cache-control object is validated in full for every client.
 
 ### `POST /v1/messages/count_tokens`
 
-Anthropic-compatible token counting helper. Codex OAuth does not expose a count-only endpoint equivalent to Anthropic's native API, so this route counts text locally and returns context-window metadata for the effective backend model. Normalized model-visible messages, tool calls, tool-result metadata, reasoning, and tool schemas are counted once. Request-envelope fields and the approved Claude Code `max_tokens` compatibility hint are excluded; other unsupported generation controls fail instead of becoming no-ops. Image inputs use a separate fixed estimate so inline base64 data is not counted as text.
+Anthropic-compatible token counting helper. Codex OAuth does not expose a count-only endpoint equivalent to Anthropic's native API, so this route counts text locally and returns context-window metadata for the effective backend model. Normalized model-visible messages, tool calls, tool-result metadata, reasoning, and tool schemas are counted once. Request-envelope fields and the `max_tokens` compatibility hint are excluded; other unsupported generation controls fail instead of becoming no-ops. Image inputs use a separate fixed estimate so inline base64 data is not counted as text.
 
-`max_tokens` and `cache_control` remain Claude Code-only compatibility hints on
-this route and require `x-claude-code-session-id` when present.
+`max_tokens` and `cache_control` are optional client-neutral compatibility hints on this route; no Claude Code header is required.
 
 Text estimation uses a bundled port of official `tiktoken` `o200k_base` `encode_ordinary`: the same Unicode pre-tokenization regex, byte-pair merge algorithm, and merge-rank data are implemented in Python, TypeScript, and Rust. It is a local estimator, not model metadata and not a substitute for a provider count API. The project does not depend on a `tiktoken` package or download encoding data at runtime. The last upstream synchronization check was **2026-07-14**, against [`tiktoken` 0.13.0 at `08a5f3b`](https://github.com/openai/tiktoken/tree/08a5f3b2c987ada4fc5aa1f16c643c203fa8acaa); the bundled rank file SHA-256 is `446a9538cb6c348e3516120d7c08b09f57c36495e2acfffe59a5bf8b0cfb1a2d`.
 
@@ -371,7 +384,7 @@ curl http://localhost:18080/v1/messages/count_tokens \
 
 ### `POST /v1/messages/compact`
 
-Anthropic-compatible alias for remote conversation compaction. Accepts Anthropic Messages-shaped bodies and returns compacted checkpoint content. Live GPT model IDs select the matching backend model, and Claude Code effort and Fast Mode controls use the same mappings as `/v1/messages`. Its required non-forwarded `max_tokens` and any `cache_control` hints require one valid `x-claude-code-session-id` header.
+Anthropic-compatible alias for remote conversation compaction. Accepts Anthropic Messages-shaped bodies and returns compacted checkpoint content. Live GPT model IDs select the matching backend model, and Claude Code effort and Fast Mode controls use the same mappings as `/v1/messages`. Its required non-forwarded `max_tokens` and any valid `cache_control` hints are accepted for every client; the optional Claude Code session header only supplies cache affinity.
 
 ### `POST /v1/images/generations`
 
@@ -511,7 +524,7 @@ Requests can instead use the Responses-shaped effort/context object:
 - Responses Lite uses `all_turns` as the Codex wire default. An explicitly different context is rejected instead of silently overwritten; use `responses_lite: false` when the backend route supports classic Responses and another context is required.
 - Remote compact keeps its existing private Codex `reasoning_effort` field but does not accept public `reasoning.mode` or `reasoning.context`.
 
-Anthropic `thinking` values map as `enabled → high`, `adaptive → medium`, and `disabled → none`. Claude Code's `output_config.effort` takes precedence over adaptive or enabled thinking and supports `low`, `medium`, `high`, `xhigh`, and `max`. Call-level `thinking.disabled` takes precedence over ambient `output_config.effort` for representable client-side auxiliary calls such as WebFetch; hosted Anthropic WebSearch is rejected before transport because its lifecycle cannot be translated losslessly. The supported reasoning context and verbosity extensions are also available on `/v1/messages`, image generation, and inspection requests; Pro, non-null public cache policy/breakpoints, and any supplied `safety_identifier` fail explicitly on the private Codex provider.
+Anthropic `thinking` values map as `enabled → high` and `adaptive → medium`. Disabled thinking selects `none` when available, otherwise the lowest supported catalog effort, and suppresses reasoning output; it does not claim to eliminate a model's internal reasoning. `output_config.effort` takes precedence over adaptive or enabled thinking and supports `low`, `medium`, `high`, `xhigh`, and `max`. Call-level disabled thinking takes precedence over ambient `output_config.effort` for auxiliary calls such as WebFetch. Explicit unsupported effort requests still fail. Hosted Anthropic WebSearch is rejected because its lifecycle cannot be translated losslessly. The supported reasoning context and verbosity extensions are also available on `/v1/messages`, image generation, and inspection requests; Pro, non-null public cache policy/breakpoints, and any supplied `safety_identifier` fail explicitly on the private Codex provider.
 
 Anthropic's `JSONOutputFormat` carries a schema but no name. When Claude Code
 sends that official shape, the proxy follows pinned Codex 0.153.3 and supplies
@@ -553,7 +566,7 @@ unchanged. The proxy does not inherit speed defaults from Codex config.
 
 Official behavior reference: Codex [0.164.0 service-tier resolution](https://github.com/openai/codex/blob/rust-v0.164.0/codex-rs/tui/src/service_tier_resolution.rs)
 and [stable, enabled-by-default ultrafast feature](https://github.com/openai/codex/blob/rust-v0.164.0/codex-rs/features/src/lib.rs).
-The rest of the private HTTP contract remains pinned to 0.153.3.
+The bundled private-protocol reference records the reviewed Codex source revision. Outbound client identity is separate: the proxy resolves the stable `latest` version of npm [`@openai/codex`](https://www.npmjs.com/package/@openai/codex) and uses that same version for the model-catalog `client_version` query and Codex User-Agent. It is not hardcoded to the bundled reference or inferred from a locally installed CLI. Successful version lookups are cached for one hour; a refresh failure retains the last successfully resolved version and retries after a short backoff. A cold lookup failure returns an explicit availability error rather than advertising an invented version.
 
 Authenticated tests on July 10, 2026 verified both official continuation paths with `gpt-5.6-sol`: a direct private Responses WebSocket completed a second delta request using the exact prior `response.id`, and the private HTTP endpoint completed the same continuation when the full prior input/output history was replayed. The proxy implements the HTTP replay strategy. Direct HTTP forwarding of `previous_response_id`, Pro, public cache controls, and `safety_identifier` was rejected and is not used. See OpenAI's [reasoning mode documentation](https://developers.openai.com/api/docs/guides/reasoning#reasoning-mode) and the Codex 0.153.3 [HTTP/WebSocket request structures](https://github.com/openai/codex/blob/b1a547b1f73ce86205d9222ac19cff334b3b7a2e/codex-rs/codex-api/src/common.rs#L275-L350).
 
@@ -785,9 +798,9 @@ ANTHROPIC_BASE_URL=http://127.0.0.1:18080 ANTHROPIC_AUTH_TOKEN=unused \
   claude --model gpt-5.6-luna --effort medium
 ```
 
-Live Codex IDs are sent to the matching backend model. Built-in Claude/Fable/Opus/Sonnet/Haiku IDs are not Codex models, so recognized `claude-*` IDs require the explicitly configured `CODEX_AS_API_MODEL`; the proxy preserves the selected client name in the Anthropic response.
+Live Codex IDs are sent to the matching backend model. Built-in Claude/Fable/Opus/Sonnet/Haiku IDs are not Codex models, so recognized `claude-*` IDs use the configured Codex model or authenticated catalog default; the proxy preserves the selected client name in the Anthropic response.
 
-`ANTHROPIC_BASE_URL` applies to the whole Claude Code process. Therefore the built-in rows remain available in the picker, but they also pass through codex-as-api and use the explicitly configured Codex backend; the rows do not call Anthropic models in that process. To use an actual Anthropic model and GPT in parallel, keep the gateway variables out of global settings, start the GPT process with the inline command above, and start a second process normally:
+`ANTHROPIC_BASE_URL` applies to the whole Claude Code process. Therefore the built-in rows remain available in the picker, but they also pass through codex-as-api and use the configured or catalog-default Codex backend; the rows do not call Anthropic models in that process. To use an actual Anthropic model and GPT in parallel, keep the gateway variables out of global settings, start the GPT process with the inline command above, and start a second process normally:
 
 ```bash
 claude --model sonnet
@@ -797,7 +810,7 @@ If the gateway variables are already in `~/.claude/settings.json`, shell-level `
 
 Claude Code gateway discovery cannot expose raw `gpt-*` IDs because Claude Code filters discovered IDs to `claude*` and `anthropic*`. The official single custom-model option avoids a misleading facade ID. This integration is an Anthropic Messages-compatible bridge; Anthropic explicitly does not support routing Claude Code to non-Claude models through third-party gateways. See the official [model configuration](https://code.claude.com/docs/en/model-config), [gateway connection](https://code.claude.com/docs/en/llm-gateway-connect), and [gateway protocol](https://code.claude.com/docs/en/llm-gateway-protocol) references.
 
-The current bridge accepts Claude Code's exact no-op thinking cleanup (`clear_thinking_20251015` with `keep: "all"`). Context edits that would change history, task budgets, enabled beta tool fields such as `strict` or `defer_loading`, malformed output formats, and unsupported image source types return HTTP 400 because the Codex OAuth transport has no lossless equivalent. Valid base64 and URL image blocks plus `tool_result.is_error` are preserved during translation.
+The bridge accepts Claude Code's no-op thinking cleanup (`clear_thinking_20251015` with `keep: "all"`), validates and ignores boolean `eager_input_streaming`, and accepts `defer_loading: false`. Context edits that would change history, task budgets, enabled strict/deferred tool execution, malformed output formats, and unsupported image source types return HTTP 400. Valid base64 and URL image blocks plus `tool_result.is_error` are preserved during translation.
 
 ## Architecture
 

{codex_as_api-0.8.2 → codex_as_api-0.9.2}/RELEASES.md

@@ -1,5 +1,81 @@
 # Release Notes
 
+## v0.9.2
+
+### TypeScript version lookup deadline
+
+- Keep the Codex version lookup timeout alive until the lookup settles, including
+  fetch implementations without active I/O handles. Clear the timer on completion.
+- Fix Node 22 resolver timeout cancellation that blocked the v0.9.0 and v0.9.1
+  Release workflows. This release includes their reasoning and security updates.
+
+## v0.9.1
+
+### TypeScript dependency security updates
+
+- Update Express to the 5.3 series and smol-toml to the 1.9.1 series.
+- Refresh the lockfile to body-parser 2.4.0, proxy-addr 2.0.8 and qs 6.16.1,
+  addressing the reported request-parser, IP-trust and TOML-parser advisories.
+- Keep Python, TypeScript and Rust package versions aligned. No API or reasoning
+  continuity behavior changes are included in this patch.
+
+## v0.9.0
+
+### Encrypted reasoning continuity
+
+- Carry Codex reasoning items in a versioned proxy-specific `thinking.signature`
+  envelope and restore their opaque encrypted content on the following request.
+  The envelope is transport metadata, not a fabricated Anthropic signature.
+- Preserve one thinking block per reasoning item, including multiple summary
+  parts and encrypted-only items. Streaming signatures precede their block stops;
+  signed reasoning history stays out of ordinary assistant text.
+- Preserve reasoning across the verified GPT-6.1-Sol, GPT-6-Astra, GPT-6-Sol,
+  GPT-6-Luna and Daybreak Blue model combinations; authenticated replay passed
+  all 25 combinations. Daybreak availability remains account/authorization-gated.
+
+### Current client identity and experimental models
+
+- Resolve outbound Codex client identity from npm `@openai/codex`'s stable latest
+  version instead of a hardcoded source-reference version. Use a consistent
+  version for model-catalog lookup and request headers, with bounded lookup,
+  in-memory caching and last-successful-version retry behavior.
+- Explicitly include verified hidden models `gpt-6-chat-thinking`, `gpt-reserve`
+  and `codex-auto-review` in the effective catalog without choosing them as an
+  automatic default. Prefer current live capability metadata.
+- Document their experimental status: upstream support may end at any time;
+  users should report a discontinued model in an issue so it can be removed.
+  Basic authenticated conversation, function-call continuation and generated-code
+  execution checks passed for all three.
+
+## v0.8.3
+
+### Client-neutral Anthropic compatibility (all runtimes)
+
+- Remove Claude Code identity gates around valid required `max_tokens` and
+  ephemeral `cache_control` hints on Messages, counting and compact routes.
+  These are compatibility hints, not native Codex output caps or cache TTLs;
+  optional Claude session cache affinity remains available.
+- Include provisional zero usage in every initial streaming `message_start`,
+  followed by authoritative final usage with the 0.8.2 disjoint cache accounting.
+- Support summarized thinking as real unsigned thinking blocks/deltas, retaining
+  existing Codex reasoning extensions. Omitted and disabled thinking suppress
+  reasoning output; never fabricate an Anthropic signature.
+- Resolve disabled-derived effort to the model's lowest supported level when
+  `none` is unavailable; explicit effort selection remains capability-checked.
+- Validate and strip replayed assistant thinking history for multi-turn clients.
+  Accept standard `metadata.user_id` and boolean tool streaming hints without
+  inventing backend semantics, and use the live catalog default for unconfigured
+  `claude-*` facade routing.
+
+### Official test quality
+
+- Replace regex diagnostic assertions with literal returned-error checks and
+  remove source-presence, filesystem-permission and worker-liveness oracles.
+  Retain actual input/output, persistence-content and cancellation-state cases.
+- Repair stale client-gate expectations and positive controls, including valid
+  null cache hints, live-default Claude facade routing and mixed thinking/tool
+  replay. Validate all three runtimes with their standard regression runners.
+
 ## v0.8.2
 
 ### Hotfix: disjoint Anthropic cache usage (all runtimes)

{codex_as_api-0.8.2 → codex_as_api-0.9.2}/config/codex-upstream-contract.json

@@ -1,9 +1,9 @@
 {
   "upstream": {
     "repository": "openai/codex",
-    "version": "0.153.3",
-    "tag": "rust-v0.153.3",
-    "commit": "b1a547b1f73ce86205d9222ac19cff334b3b7a2e"
+    "version": "0.164.0",
+    "tag": "rust-v0.164.0",
+    "commit": "6aaac694803004a63c6f393d153c96d22521a6b9"
   },
   "models_request": {
     "path": "/models",

codex_as_api-0.9.2/config/verified-hidden-models.json

@@ -0,0 +1,165 @@
+{
+  "verified_at": "2026-10-23T21:33:21.486Z",
+  "warning": "Experimental hidden model; upstream support may stop at any time.",
+  "models": [
+    {
+      "slug": "gpt-6-chat-thinking",
+      "display_name": "GPT-6 Chat Thinking",
+      "description": "Experimental hidden model; upstream support may stop at any time.",
+      "priority": 2,
+      "visibility": "list",
+      "supported_in_api": true,
+      "default_reasoning_level": "medium",
+      "supported_reasoning_levels": [
+        {
+          "effort": "low",
+          "description": "Fast responses with lighter reasoning"
+        },
+        {
+          "effort": "medium",
+          "description": "Balances speed and reasoning depth for everyday tasks"
+        },
+        {
+          "effort": "high",
+          "description": "Greater reasoning depth for complex problems"
+        },
+        {
+          "effort": "xhigh",
+          "description": "Extra high reasoning depth for complex problems"
+        },
+        {
+          "effort": "max",
+          "description": "Maximum reasoning depth for the hardest problems"
+        }
+      ],
+      "multi_agent_reasoning_effort": null,
+      "supports_reasoning_summary_parameter": false,
+      "default_reasoning_summary": "none",
+      "comp_hash": "3000",
+      "context_window": 272000,
+      "max_context_window": 272000,
+      "auto_compact_token_limit": null,
+      "effective_context_window_percent": 95,
+      "input_modalities": [
+        "text",
+        "image"
+      ],
+      "service_tiers": [],
+      "default_service_tier": null,
+      "use_responses_lite": true,
+      "supports_image_detail_original": true,
+      "support_verbosity": false,
+      "default_verbosity": "medium"
+    },
+    {
+      "slug": "gpt-reserve",
+      "display_name": "GPT-Reserve",
+      "description": "Experimental hidden model; upstream support may stop at any time.",
+      "priority": 4,
+      "visibility": "list",
+      "supported_in_api": true,
+      "default_reasoning_level": "medium",
+      "supported_reasoning_levels": [
+        {
+          "effort": "low",
+          "description": "Fast responses with lighter reasoning"
+        },
+        {
+          "effort": "medium",
+          "description": "Balances speed and reasoning depth for everyday tasks"
+        },
+        {
+          "effort": "high",
+          "description": "Greater reasoning depth for complex problems"
+        },
+        {
+          "effort": "xhigh",
+          "description": "Extra high reasoning depth for complex problems"
+        },
+        {
+          "effort": "max",
+          "description": "Maximum reasoning depth for the hardest problems"
+        }
+      ],
+      "multi_agent_reasoning_effort": null,
+      "supports_reasoning_summary_parameter": true,
+      "default_reasoning_summary": "none",
+      "comp_hash": "3000",
+      "context_window": 272000,
+      "max_context_window": 872000,
+      "auto_compact_token_limit": null,
+      "effective_context_window_percent": 95,
+      "input_modalities": [
+        "text",
+        "image"
+      ],
+      "service_tiers": [
+        {
+          "id": "priority",
+          "name": "Fast",
+          "description": "1.5x speed, increased usage"
+        }
+      ],
+      "default_service_tier": null,
+      "use_responses_lite": true,
+      "supports_image_detail_original": true,
+      "support_verbosity": true,
+      "default_verbosity": "low"
+    },
+    {
+      "slug": "codex-auto-review",
+      "display_name": "Codex Auto Review",
+      "description": "Experimental hidden model; upstream support may stop at any time.",
+      "priority": 43,
+      "visibility": "list",
+      "supported_in_api": true,
+      "default_reasoning_level": "medium",
+      "supported_reasoning_levels": [
+        {
+          "effort": "low",
+          "description": "Fast responses with lighter reasoning"
+        },
+        {
+          "effort": "medium",
+          "description": "Balances speed and reasoning depth for everyday tasks"
+        },
+        {
+          "effort": "high",
+          "description": "Greater reasoning depth for complex problems"
+        },
+        {
+          "effort": "xhigh",
+          "description": "Extra high reasoning depth for complex problems"
+        },
+        {
+          "effort": "max",
+          "description": "Maximum reasoning depth for the hardest problems"
+        }
+      ],
+      "multi_agent_reasoning_effort": null,
+      "supports_reasoning_summary_parameter": true,
+      "default_reasoning_summary": "none",
+      "comp_hash": "3000",
+      "context_window": 272000,
+      "max_context_window": 872000,
+      "auto_compact_token_limit": null,
+      "effective_context_window_percent": 95,
+      "input_modalities": [
+        "text",
+        "image"
+      ],
+      "service_tiers": [
+        {
+          "id": "priority",
+          "name": "Fast",
+          "description": "1.5x speed, increased usage"
+        }
+      ],
+      "default_service_tier": null,
+      "use_responses_lite": true,
+      "supports_image_detail_original": true,
+      "support_verbosity": true,
+      "default_verbosity": "low"
+    }
+  ]
+}