@oneciel-ai/ciel-runtime 0.2.37 → 0.2.38

package/docs/MCP-Channels.md

@@ -57,8 +57,15 @@ how a request enters the active TUI and how the answer should leave it:
   projection, voice/text hint, and request metadata. `input_mode=tty` puts the
   caller's message text directly on the private terminal-input path without the
   Web Chat input envelope.
-- `input_transport=tty` (default) submits the selected input format through the
-  active CLI terminal. `input_transport=router` keeps the full message in the
+- `input_transport=session_socket` is the default preference for every admitted
+  input. An interactive Claude Code session launched by Ciel receives it through
+  its authenticated messaging socket without terminal typing, including while
+  Claude is processing another turn. If the active runtime does not expose a
+  usable session socket, Ciel falls back to `input_transport=tty` while the TUI is
+  idle; during an active turn the durable input remains queued until terminal
+  injection is safe. An explicit `input_transport=tty` always retains terminal
+  delivery.
+- `input_transport=router` keeps the full message in the
   Runtime Input Gateway: while a model turn is already active, its next request
   consumes the message without a console wake; while the CLI is idle, Ciel types
   a correlated wake turn and replaces that entire turn with the full pending
@@ -67,11 +74,44 @@ how a request enters the active TUI and how the answer should leave it:
   short marker-only wake because their terminal turn behavior is different. In
   both cases the model request receives the queued message exactly once. Router
   transport therefore requires the active model session to use Ciel Router.
-  Aliases `llm`, `context`, and `inband` are accepted.
+  Aliases `llm`, `context`, and `inband` are accepted. Socket aliases `socket`,
+  `claude_socket`, and `messaging_socket` select `session_socket`.
 - `response_mode=web_chat` (alias `ai_net`, default) uses the normal correlated Web
   Chat reply contract. `response_mode=tty` leaves the model's ordinary terminal
   output as-is and does not require a Web Chat tool reply. `response_mode=mcp`
   supplies a one-request MCP routing hint from `response_mcp`.
+- `raw_injection=true` makes the model-facing input exactly equal to the
+  `message` string. Ciel does not add a Web Chat envelope, voice/text marker,
+  attachment description, ACK/reply contract, MCP response hint, correlation
+  marker, or batch separator. It is independent of `input_transport` and
+  `response_mode`; callers selecting a non-terminal response destination must
+  arrange that response behavior outside the injected prompt.
+
+Exact-input example:
+
+```json
+{
+  "message": "Run the queued operation exactly once.",
+  "raw_injection": true,
+  "input_transport": "session_socket",
+  "response_mode": "tty"
+}
+```
+
+Every admitted private Runtime Input has an independent lifecycle identifier.
+The message POST response includes `request_id` and a `request` object whose
+initial state is `queued`; HTTP success does not claim that the TUI accepted the
+input. Query one request with `GET /ca/channel/requests/{request_id}` or list
+latest states with `GET /ca/channel/requests?after=0&status=submitted&limit=100`.
+The `/ca/chat/requests/...` aliases are equivalent. State transitions are
+`queued -> submitted -> replied` or `queued|submitted -> failed`. They are also
+published through `/ca/events/stream?category=runtime_input.status` and
+`/ca/events/ws?category=runtime_input.status`.
+
+The lifecycle store is `runtime-input-status.jsonl`; the delivery queue and its
+cursor both use the private `runtime-inputs.jsonl` ID domain. Public
+`chat-messages.jsonl` IDs are never used to calculate the pending Runtime Input
+count.
 
 The MCP hint is declarative; Ciel does not call the named server itself:
 
@@ -98,6 +138,29 @@ remains supported and maps to the old paired behavior. Unknown modes and an MCP
 response without `response_mcp.server` return HTTP 400; they never silently fall
 back to another route.
 
+### Claude Code cross-session socket boundary
+
+Claude Code 2.1.239 and later publishes its cross-session messaging socket on
+Windows as well as macOS and Linux. For an interactive Claude session, Ciel now
+creates an explicit per-launch socket path, passes it with
+`--messaging-socket-path`, and enables `crossSessionInbound=accept` in Ciel's
+generated runtime settings. Windows uses a local named pipe; Unix platforms use
+an AF_UNIX socket.
+
+Each connection authenticates first with the `peerToken` from Claude's
+user-private `~/.claude/sessions/<pid>.<socket-hash>.key` file and then submits a
+newline-delimited `user` frame. Ciel deliberately omits a peer `from` address:
+the existing private Web Chat/MCP response contract remains authoritative, so a
+socket input does not invent a second Claude peer as its reply destination.
+If the socket is not available before delivery during an active turn, the durable
+Runtime Input remains queued. Once a delivery attempt begins, an unconfirmed
+socket or TTY submission becomes `failed` and is not typed again automatically.
+While the TUI is idle, a failed socket attempt may try TTY once; if that attempt
+is not confirmed, the request remains failed.
+
+Set `claude_code.session_socket_input=false` to disable automatic socket setup
+for a launch, or choose `input_transport=tty|router` on an individual input.
+
 Web Chat attachments are uploaded through `POST /ca/channel/files`. The public
 chat transcript contains only download metadata; the private Runtime Input
 projection adds a validated workspace-local path so the active Claude or Codex
@@ -120,11 +183,16 @@ POST /ca/events/receivers/<receiver-id>
 POST /ca/events/webhooks/<receiver-id>
 ```
 
-The configuration POST accepts `enabled`, `transport` (`webhook` or `sse`), `url` for SSE, an optional `event_types` allow-list, and either `webhook_secret` or `authorization`. Secrets are stored in the local encrypted workspace vault and are never returned by the GET response.
+The configuration POST accepts `enabled`, `transport` (`webhook` or `sse`), `url`
+for SSE, an optional `event_types` allow-list, and either `webhook_secret` or
+`authorization`. `input_transport=auto|session_socket|tty|router` controls how
+admitted events reach the active runtime; `auto` prefers the session socket and
+uses the same safe TTY fallback. Secrets are stored in the local
+encrypted workspace vault and are never returned by the GET response.
 
 Webhook bodies use CloudEvents 1.0 structured JSON and Standard Webhooks `webhook-id`, `webhook-timestamp`, and `webhook-signature` headers. SSE requests negotiate the same structured representation with `Accept: text/event-stream, application/cloudevents+json`; each `data` frame must contain one CloudEvent 1.0 structured JSON object. Ciel validates framing, signatures, replay windows, type filters, and duplicate identities, but preserves the admitted event text exactly for the LLM.
 
-Reconnects use the SSE `id` field and `Last-Event-ID` by default. Streams that carry a cursor inside the CloudEvent can set `cursor_json_pointer` to an RFC 6901 pointer such as `/data/stream_id`. If the producer expects its cursor in the reconnect URL instead of `Last-Event-ID`, set the provider-neutral `cursor_query_parameter` to that query parameter's name. Ciel persists the projected cursor per workspace and receiver; no product-specific event schema is built into the runtime.
+Reconnects use the SSE `id` field and `Last-Event-ID` by default. Streams that carry a cursor inside the CloudEvent can set `cursor_json_pointer` to an RFC 6901 pointer such as `/data/stream_id`. If the producer expects its cursor in the reconnect URL instead of `Last-Event-ID`, set the provider-neutral `cursor_query_parameter` to that query parameter's name. Ciel persists the projected cursor per workspace and receiver; no product-specific event schema is built into the runtime. A failed earlier request blocks later requests, preserving input order and preventing two requests from sharing one unresolved TUI draft.
 
 The workspace router is the sole owner of each outbound SSE subscription. The prelaunch settings process only persists receiver configuration, preventing duplicate connections and duplicate deliveries. The terminal bridge also performs a bounded periodic safety rescan of its durable private input queue so a missed filesystem notification cannot strand an admitted event.
 
@@ -140,7 +208,12 @@ When Web Chat or an explicit Ciel integration needs response tools, the CLI may
 POST /ca/mcp
 ```
 
-It exposes only Ciel-owned tools such as `send_message`, `send_file`, `compact_session`, and `llm_options`. The endpoint is stateless and has no MCP GET stream, `/ca/mcp/sse`, session registry, replay cursor, or external notification subscription.
+It exposes only Ciel-owned tools such as `submit_input`, `send_message`,
+`send_file`, `compact_session`, and `llm_options`. External Streamable HTTP MCP
+clients can call `submit_input`; its default transport follows the same
+session-socket-first policy and accepts an explicit `session_socket|tty|router`
+override. The endpoint is stateless and has no MCP GET stream, `/ca/mcp/sse`,
+session registry, replay cursor, or external notification subscription.
 
 ## MCP 2026-07-28
 

package/docs/Managed-Tool-Injection.md

@@ -0,0 +1,25 @@
+# Managed tool launch policy
+
+Ciel-owned tool definitions can declare `injection_mode`: `always`, `native`,
+or `non_native`. The default is `always`, retaining channel/runtime integration.
+`WorkspaceMcpLaunchService.prepare(..., native=...)` filters only its
+`injected_servers` argument. User-owned `workspace_mcp.servers` is not filtered.
+The metadata is removed before generating CLI arguments or MCP JSON.
+
+Claude's generated DuckDuckGo and fetch bundle is non-native-only, including
+when the Ciel web-search override is enabled. Anthropic direct and routed
+connections both retain their own web tools. Codex and Codex app-server pass
+their existing native-provider classification to the same projection policy;
+their current launch paths do not auto-generate a DuckDuckGo bundle.
+
+Selection happens for each launch, without deleting shared files or changing
+another running instance. This does not dynamically reconfigure an already
+running CLI after a provider switch. It does not change the CLI's own search
+settings, install a fallback after an error, or remove user-installed tools.
+
+Codex native launches (direct or routed), including app-server, additionally
+override `mcp_servers.duckduckgo.enabled=false` and
+`mcp_servers.web_fetch.enabled=false` after workspace MCP projection. This
+suppresses replacement web tools inherited from Codex's global configuration
+for this child process only. Non-native launches receive neither override.
+The original global/project configuration files remain unchanged.

package/docs/Module-Map.md

@@ -185,6 +185,10 @@ Claude/Codex transcript의 queued-only wake를 찾아 LLM delivery cursor를 안
 
 대기 중인 channel 메시지의 필터링, 중복 제거, wake claim, prompt 생성 및 터미널 주입을 조정하는 Channel Application Service. 상태 판정, prompt 전략, wake 저장소, I/O를 각각 10필드 이하 포트로 분리한다.
 
+### `ciel_runtime_support/runtime_input_status.py`
+
+private Runtime Input의 `queued -> submitted -> replied|failed` 전이를 JSONL로 보존하고 요청별 조회 및 `runtime_input.status` 실시간 이벤트를 제공하는 lifecycle repository다. terminal 상태에서 이전 상태로 되돌아가는 전이는 거부한다.
+
 ### `ciel_runtime_support/channel_tool_context.py`
 
 외부 채널에서 시작된 tool-use의 원문 context를 bounded, thread-safe 저장소에 보관하고 대응하는 tool-result 후속 요청에 한 번만 주입하는 Channel Application Service. 저장소가 lock·eviction·consume-on-read를 소유하며 메시지 projection과 시간·텍스트·로그 효과는 typed ports로 분리한다.

package/docs/Muse-Code.md

@@ -0,0 +1,73 @@
+# Muse Code runtime
+
+Ciel Runtime can launch Meta's official Muse Code CLI while preserving Muse
+Code subscription authentication:
+
+```sh
+ciel-runtime muse
+```
+
+This is a native runtime integration. Muse Code sends its model traffic directly
+with the browser-authenticated credential created by its own onboarding. The Ciel
+Router remains the local control plane for Web Chat, external inputs, remote
+instructions, and workspace memory; it does not proxy the subscription model
+request through the pay-as-you-go Model API.
+
+## Authentication and billing boundary
+
+Native Muse launches remove `META_API_KEY` and `MODEL_API_KEY` from the child
+environment. Meta documents that API-key authentication takes precedence over a
+stored browser session and that additional API keys are billed pay-as-you-go.
+Run `ciel-runtime muse`, then use Muse's `/login` command when sign-in is needed.
+
+Use the existing Ciel `meta` provider with Claude or Codex when direct Model API
+pay-as-you-go routing is desired. That is a separate path from Muse Code
+subscription usage.
+
+## Platforms and installation
+
+On macOS and Linux, Ciel discovers `muse` on `PATH`; when it is absent, Ciel runs
+Meta's official installer from `https://dev.meta.ai/install.sh` with Bash.
+
+Meta currently documents native Muse Code installation for macOS and Linux. On
+Windows, Ciel discovers or installs Muse Code inside the default WSL2
+distribution and launches it through `wsl.exe`. The current Windows directory is
+preserved by WSL path translation.
+
+## Model, effort, and passthrough
+
+The native runtime defaults to `muse-spark-1.3`. When the selected Ciel provider
+is `meta`, its configured Muse model and supported reasoning effort are forwarded.
+Ciel `max` maps to Muse's documented `ultra` tier. Explicit Muse flags win:
+
+```sh
+ciel-runtime muse --model muse-spark-1.3 --reasoning-effort high
+ciel-runtime muse exec --json "Inspect this repository"
+ciel-runtime --ca-runtime muse -- --version
+```
+
+Utility subcommands such as `login`, `auth`, `config`, `schema`, `serve`, and
+`session-message` are passed through without injecting model flags.
+
+Interactive sessions, `exec`, and `resume` include Muse's `--yolo` option by
+default, matching Ciel's Codex launch policy. Muse defines this option as
+disabling approval and sandboxing and trusting the workspace for that run.
+An explicitly supplied `--yolo` is retained once rather than duplicated.
+Utility commands do not receive it.
+
+## Router input delivery
+
+Interactive Muse sessions run through Ciel's channel-aware terminal proxy when
+available. Web Chat and external inputs therefore use the standard
+session-socket-first policy and safely fall back to terminal delivery because
+Muse Code does not expose Claude's session socket. Because Muse does not publish
+Claude/Codex-compatible transcript confirmation events, its terminal fallback
+uses one write and commits the durable channel cursor immediately; it does not
+replay the same accepted message while waiting for an unavailable confirmation.
+Headless `muse exec` and
+utility commands run directly and do not start an interactive input proxy.
+
+Muse Code 1.0.2 also exposes `muse serve` for the Muse Session Protocol (MSP) over
+stdio and `muse session-message` for peer sessions. Ciel preserves these commands
+as native passthrough surfaces; it does not claim MSP lifecycle ownership for an
+ordinary interactive TUI launch.

package/docs/Observability.md

@@ -108,6 +108,45 @@ GET /ca/events/stream
 Server-Sent Events로 실시간 이벤트 전송.  
 쿼리 파라미터: `?level=debug&category=router`
 
+### WebSocket 스트림
+```
+GET /ca/events/ws
+```
+RFC 6455 WebSocket text frame으로 같은 이벤트 JSON을 실시간 전송한다.
+`after`, `level`, `category` 쿼리 파라미터를 SSE와 동일하게 지원하며,
+원격 연결은 라우터의 기존 외부 접근 토큰 검사를 그대로 거친다.
+
+CLI 툴콜만 구독:
+```
+ws://127.0.0.1:<port>/ca/events/ws?category=tool.call
+```
+
+`tool.call` 이벤트는 라우터 변환 스트림과 Claude/Codex/Muse Code 세션
+transcript의 구조화된 툴 시작 레코드에서 생성된다. Windows에서 실행되는
+Muse WSL 세션도 `wslpath`로 확인한 세션 저장소를 직접 추적한다. `data`에는 `call_id`, `name`,
+`call_type`, `arguments`, `runtime`, `model`이 가능한 범위에서 포함된다.
+동일한 `call_id`가 라우터와 transcript 양쪽에서 관측되면 한 번만 전달한다.
+`tool_call_events.include_arguments=false`로 설정하면 모든 `tool.call` 이벤트의
+인자 값을 제외할 수 있다.
+
+검색 응답의 URL도 같은 `tool.call` 카테고리와 기존 WS/SSE/recent 경로로
+전달한다. 결과는 `data.phase == "result"`로 구분하며 다음 형태다.
+
+```json
+{"call_id":"toolu_...","name":"WebSearch","runtime":"claude","phase":"result","urls":["https://example.org/article"]}
+```
+
+검색어·제목·요약·응답 본문은 이 결과 이벤트에 포함하지 않는다. 기존 호출 시작
+이벤트는 유지되므로 결과 URL만 필요한 소비자는 `phase`를 필터링한다.
+Claude Code는 호출 ID에 연결된 `tool_result.content`의 JSON `Links` 목록,
+Anthropic 서버 도구는 `web_search_tool_result`의 URL을 읽는다. Codex는 검색
+호출에 연결된 function/custom output, 완료된 `web_search_call`의 `action.sources`
+및 `results`, assistant의 구조화된 `url_citation`을 지원한다.
+단, Codex 네이티브 rollout은 검색어/상태만 저장하고 결과 URL을 생략할 수 있다.
+그 경우 이 수집기는 URL을 생성하거나 검색어/페이지 열기 입력을 결과로 대신하지 않는다.
+URL 없는 응답은 이벤트를 발행하지 않는다. 호출 연결과 중복 방지 상태는
+transcript cursor와 함께 저장하며 최근 512개로 제한한다.
+
 ### 최근 이벤트 JSON
 ```
 GET /ca/events/recent

package/docs/Providers.md

@@ -50,7 +50,14 @@ token은 Bearer header로 전송하며, 사용자 지정 `base_url`로 사설 ga
 
 ### Alibaba Model Studio Singapore
 
-`alims-intl`의 기본 모델은 Singapore International scope의 `qwen3.8-max`다.
+`alims-intl`의 기본 모델은 Singapore International scope의 rolling alias
+`qwen3.8-max`다. 고정 스냅샷 `qwen3.8-max-0902`도 모델 목록에서 선택할 수 있다.
+Token Plan은 공식 지원 ID인 rolling alias `qwen3.8-max`를 사용하며, Responses
+API에서는 `none`, `minimal`, `low`, `medium`, `high`, `xhigh`, `max` reasoning
+effort를 그대로 전달한다. Chat Completions의 축약 effort 매핑과 혼용하지 않는다.
+Alibaba Responses session cache의 공식 opt-in 계약에 따라
+`x-dashscope-session-cache: enable`을 Responses 요청에 기본 전송한다. Chat과
+Anthropic Messages 요청에는 이 header를 추가하지 않는다.
 공식 한도에 맞춰 context window는 1,000,000, 최대 입력은 일반 모드 991,808,
 thinking 모드 983,616, 최대 출력은 131,072로 취급한다. Codex에는 thinking 모드의
 안전한 입력 상한인 983,616을 model catalog context로 제공한다.
@@ -137,8 +144,10 @@ Imagine image/video, Voice, STT/TTS 모델 ID는 text LLM 선택 메뉴에 섞
 
 - **Native 모드**: Claude Code가 직접 Anthropic API 호출. Router가 중계하지 않음.
 - **Routed 모드**: Router가 중계하며 `ROUTED_COMPAT_PROMPT`를 시스템 프롬프트에 주입.
+- Routed 모드는 기본적으로 표준 200K 컨텍스트를 사용한다. 1M 사용 크레딧
+  beta를 명시적으로 사용할 때만 모델 ID에 `[1m]`을 붙인다.
 - 공개 모델 ID 목록 (`ANTHROPIC_PUBLIC_MODEL_DEFAULT_IDS`):
-  - `claude-fable-5`, `claude-opus-4-8`, `claude-sonnet-4-6`, `claude-haiku-4-5-20251001`, `claude-haiku-4-5`
+  - `claude-fable-5-1`, `claude-fable-5`, `claude-opus-4-8`, `claude-sonnet-4-6`, `claude-haiku-4-5-20251001`, `claude-haiku-4-5`
 - 제한 접근 모델 (`ANTHROPIC_LIMITED_ACCESS_MODEL_IDS`):
   - `claude-mythos-5`, `claude-mythos-preview`
 
@@ -247,6 +256,24 @@ ZCode 버전의 선택된 Desktop provider는 fallback으로만 읽는다. 두
 
 ---
 
+## Meta Model API / Muse Spark
+
+- `muse-spark-1.3`과 `muse-spark-1.3-contributor`를 기본 1.3 계열로 제공한다.
+- Responses, Chat Completions, Messages를 각각 네이티브 프로토콜로 전달한다.
+- Responses의 `web_search`, `tool_search`, `defer_loading` 서버 도구와 검색
+  결과·URL citation을 보존한다.
+- 이미지, MP4 비디오, WAV/MP3 오디오, PDF/텍스트/JSON 문서 블록을 프로토콜별
+  원형으로 전달한다. Muse Spark 1.3 오디오는 Meta가 제한적 지원으로 명시하므로
+  고품질 오디오 작업에는 1.2 또는 Muse Voice Transcribe를 사용한다.
+- 인라인 파일은 최대 50,000,000 bytes이며, `/v1/files`는 파일당 1 GiB와 팀당
+  100 GiB를 지원한다. 추론용 업로드 purpose는 `user_data`이다.
+- PDF는 텍스트 기준 처음 100페이지와 페이지 이미지 기준 처음 50페이지를 처리한다.
+- Files API 업로드·목록·내용 조회·삭제는 로컬 라우터와 인증된 Remote Bridge에서
+  Meta 자격 증명으로 전달된다. 대용량 multipart 업로드는 메모리에 통째로 적재하지
+  않고 스트리밍된다.
+
+---
+
 ## vLLM / LM Studio
 
 - OpenAI 호환 로컬 서버.

package/docs/Remote-Bridge.md

@@ -70,6 +70,10 @@ Use `http://ROUTER_HOST:ROUTER_PORT/v1` as an OpenAI base URL and
 | Anthropic token count | `POST /v1/messages/count_tokens` |
 | OpenAI model list | `GET /v1/models` |
 | OpenAI model detail | `GET /v1/models/{provider}/{model}` |
+| Provider file upload | `POST /v1/files` |
+| Provider file list | `GET /v1/files` |
+| Provider file metadata/content | `GET /v1/files/{file_id}` and `GET /v1/files/{file_id}/content` |
+| Provider file deletion | `DELETE /v1/files/{file_id}` |
 | Bridge status | `GET /ca/bridge` |
 
 `GET /v1/models` preserves the OpenAI-compatible `object`, `data`, and
@@ -116,6 +120,14 @@ Providers that require upstream streaming reject non-streaming Chat or
 Anthropic Messages requests with HTTP 501; use `/v1/responses` when Ciel must
 collect that stream into a non-streaming compatible response.
 
+The Files routes are currently exposed for Meta Model API. Because a multipart
+file upload has no `model` JSON field, select a non-default Meta route with
+`X-Ciel-Runtime-Provider: meta`. Ciel replaces the bridge credential with the
+Meta credential stored on the Router host and streams the multipart body
+upstream without retaining the complete upload in router memory. Meta remains
+authoritative for the `purpose=user_data`, MIME-type, expiry, per-file, and
+team-storage validation rules.
+
 ## Select a provider, model, and provider credential
 
 The portable route syntax is `provider/model` in the standard `model` field:

package/docs/Router.md

@@ -13,6 +13,7 @@ Claude Code의 Anthropic Messages 요청과 Codex의 OpenAI Responses 요청을
 POST /v1/messages          → 모델 라우팅 (핵심 엔드포인트)
 POST /v1/responses         → Codex / OpenAI Responses 라우팅
 GET  /ca/events/stream     → SSE 이벤트 스트림
+GET  /ca/events/ws         → WebSocket 이벤트 스트림
 GET  /ca/events/recent     → 최근 이벤트 JSON
 GET  /ca/health            → 라우터 상태 확인
 GET  /v1/models            → 사용 가능한 모델 목록

package/docs/journal/2026/09/01/diagnostics/anthropic/router/fable51-usage-credits.okf

@@ -0,0 +1,70 @@
+okf: 1
+record:
+  title: Anthropic routed 1M usage-credit regression and Fable 5.1 support
+  date: 2026-09-01
+  scope:
+    runtime: ciel-runtime
+    branch: nightly
+    claude_code: 2.1.258
+    provider: anthropic
+    mode: routed
+  evidence:
+    user_observation:
+      direct_claude: works
+      routed_claude: 'API Error: Usage credits required for 1M context'
+    official_changelog:
+      url: https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md
+      version: 2.1.257
+      fable_model: claude-fable-5-1
+      context_tokens: 1000000
+      input_price_per_million_tokens_usd: 10
+      output_price_per_million_tokens_usd: 50
+      cache_read_price_per_million_tokens_usd: 0.25
+    local_environment_comparison:
+      direct_anthropic_model: unset
+      routed_anthropic_model_before: ciel-runtime-anthropic-claude-fable-5[1m]
+      routed_context_before: 1048576
+    claude_binary_capture_before:
+      automatic_1m_beta: context-1m-2025-08-07
+      same_alias_without_suffix_has_1m_beta: false
+      conclusion: The automatic suffix selected Claude Code's usage-credit 1M request path.
+  implementation:
+    catalog:
+      added: claude-fable-5-1
+      retained: claude-fable-5
+      model_family: fable
+      context_metadata: 1048576
+      max_output_tokens_metadata: 128000
+      thinking_mode: adaptive
+    routed_context:
+      implicit_selection: standard-200k
+      explicit_one_million_selection: model-id-with-[1m]
+      direct_native_mode_changed: false
+      non_anthropic_providers_changed: false
+  verification:
+    claude_2_1_258_live_capture:
+      exit_code: 0
+      expected_response_received: true
+      environment_model: ciel-runtime-anthropic-claude-fable-5-1
+      request_model: ciel-runtime-anthropic-claude-fable-5-1
+      context_tokens: 200000
+      compact_window_tokens: 200000
+      context_1m_beta_present: false
+    full_repository:
+      compile: pass
+      unit:
+        passed: 1302
+        skipped: 44
+      router:
+        passed: 1064
+      channel:
+        passed: 320
+        skipped: 80
+      runtime:
+        passed: 246
+        skipped: 12
+      total_passed: 2932
+      total_skipped: 136
+    quality:
+      ruff: pass
+      documentation_metadata: pass

package/docs/journal/2026/09/01/implementation/claude/session-socket/all-input-paths/default-delivery.okf

@@ -0,0 +1,133 @@
+okf: 1
+record:
+  title: Claude session socket default delivery for all Ciel input paths
+  date: 2026-09-01
+  scope:
+    branch: nightly
+    runtime: claude-code
+    claude_version_tested: 2.1.258
+    platforms:
+      windows: named-pipe
+      linux: AF_UNIX
+
+  evidence:
+    packaged_claude_protocol:
+      socket_argument: --messaging-socket-path
+      framing: newline-delimited JSON
+      authentication:
+        first_frame: '{"type":"auth","token":"<peerToken>"}'
+        key_location: ~/.claude/sessions/<pid>.<sha256-canonical-socket-path>.key
+      input_frame:
+        type: user
+        message_role: user
+        peer_from_field: omitted
+      windows_canonicalization: ASCII-lowercased local named-pipe path
+    live_windows_web_chat_probe:
+      executable: isolated Claude Code 2.1.258 native Windows binary
+      upstream: local Anthropic-compatible SSE capture server
+      http_route: POST /ca/chat/messages
+      http_status: 200
+      public_selected_transport: session_socket
+      private_gateway_transport: session_socket
+      socket_send: true
+      upstream_request_count_before: 1
+      upstream_request_count_after: 2
+      injected_marker_present_in_second_request: CIEL_ACTUAL_SOCKET_MESSAGE_7F31
+      claude_exit_code: 0
+      debug_turns_observed: 2
+    live_unix_transport_probe:
+      environment: WSL2 Ubuntu 26.04
+      transport: real AF_UNIX SOCK_STREAM server and client
+      authentication_frame_received: true
+      user_frame_received: true
+      unicode_safe_json: true
+      result: pass
+
+  implementation:
+    launch:
+      generated_path:
+        windows: '\\.\pipe\LOCAL\cc-msg-<32-hex>'
+        unix: /tmp/cc-socks-<uid>/ciel-<32-hex>.sock
+      claude_argument: --messaging-socket-path
+      generated_setting: 'crossSessionInbound: accept'
+      lifecycle:
+        configure_before_child: true
+        clear_after_child: true
+      disable_config: claude_code.session_socket_input=false
+    admission:
+      shared_gateway: ciel_runtime_support/runtime_input_gateway.py
+      runtime_default:
+        active_ciel_claude_socket: session_socket
+        otherwise: tty
+      explicit_overrides: [session_socket, tty, router]
+      routes:
+        web_chat_rest:
+          - POST /ca/chat/messages
+          - POST /ca/channel/messages
+        notification_rest:
+          - POST /ca/chat/notify
+          - POST /ca/channel/notify
+        external_events:
+          - POST /ca/events/webhooks/<receiver-id>
+          - configured SSE receiver
+        mcp_streamable:
+          endpoint: POST /ca/mcp
+          tool: submit_input
+        telemetry_notice:
+          source: OTLP durable log notification
+    delivery:
+      durable_queue_retained_until_socket_send: true
+      active_turn_delivery: true
+      active_tool_call_delivery: true
+      missing_key_or_socket:
+        cursor_committed: false
+        delivery_mark_rolled_back: true
+        retry: bounded polling of the durable queue
+      reply_contract: existing Ciel Web Chat or MCP response contract
+    receiver_option:
+      values: [auto, session_socket, tty, router]
+      auto_behavior: active Claude socket else TTY
+
+  verification:
+    focused:
+      windows_protocol_and_hash: pass
+      unix_AF_UNIX_real_socket: pass
+      ciel_launch_argument_and_lifecycle: pass
+      web_chat_rest_default_and_alias: pass
+      active_turn_socket_delivery: pass
+      external_event_socket_delivery: pass
+      mcp_submit_input_socket_delivery: pass
+      external_receiver_transport_option: pass
+      runtime_gateway_all_sources: pass
+    repository:
+      unit:
+        passed: 1358
+        skipped: 45
+      router:
+        passed: 1064
+      channel:
+        passed: 403
+        skipped: 80
+      runtime:
+        passed: 258
+        skipped: 12
+      npm_test:
+        result: pass
+        total_passed: 3083
+        total_skipped: 137
+      ruff: pass
+      architecture_budget:
+        result: pass
+        ciel_runtime_lines: 4936
+        maximum: 4980
+      metadata: pass
+      git_diff_check: pass
+      npm_pack_dry_run:
+        result: pass
+        package_version: 0.2.37
+        packed_bytes: 1685026
+    deployment:
+      commit: ready-after-full-verification
+      nightly_push: ready-after-full-verification
+      github_actions: pending
+      npm_nightly: pending