@oneciel-ai/ciel-runtime 0.2.37 → 0.2.38

package/docs/journal/2026/09/03/diagnostics/remote/kevin/codex-gpt-early-turn-completion.okf

@@ -0,0 +1,162 @@
+okf_version: "1.0"
+knowledge:
+  task:
+    id: "kevin-codex-gpt-early-turn-completion-20260903"
+    date: "2026-09-03"
+    status: "implemented_and_live_kimi_verified"
+    objective: >-
+      Diagnose repeated Codex turn completion while requested work remains and
+      replace language-specific recovery with a protocol-level completion gate.
+
+  live_evidence:
+    kevin:
+      host: "kevin-codex@100.95.132.58"
+      hostname: "aap-pool-hera"
+      codex_thread: "01a03716-9bb1-77e2-8771-4329c3b0338f"
+      router_port: 8804
+      active_model_display: "gpt-5.6-terra medium fast"
+      transport: "/backend-api/codex"
+      observed_turn:
+        turn_id: "01a065ec-2ea7-7db0-8d56-e0329a0cfb0c"
+        final_at_utc: "2026-09-03T06:19:58.496Z"
+        completed_at_utc: "2026-09-03T06:19:58.608Z"
+        result: >-
+          The assistant output named another audit action, then the native
+          Responses stream emitted task_complete. Ciel and Codex processes
+          remained alive, so the horizontal line was a completed model turn,
+          not a crash.
+      goal_evidence:
+        at_utc: "2026-09-02T18:53:20.946Z"
+        call: 'tools.update_goal({status:"blocked"})'
+        later_state: "Goal stalled (/goal resume)"
+    cielarvis:
+      workspace: "C:\\Users\\djlov\\cielarvis"
+      provider: "ollama-cloud"
+      model: "kimi-k3"
+      codex_thread: "01a01dff-f0f0-7ef3-8cde-1d9916510f35"
+      missed_completed_turns_utc:
+        - "2026-09-03T06:13:24Z"
+        - "2026-09-03T06:14:23Z"
+        - "2026-09-03T06:18:12Z"
+        - "2026-09-03T06:32:36Z"
+      result: >-
+        Each response contained reasoning and visible status text but no tool
+        item, after which Codex recorded task_complete.
+
+  patch_history_findings:
+    confirmed:
+      - >-
+        Native GPT requests never entered recover_preamble_only_turn; the
+        backend adapter relayed their SSE after reading only a small error
+        preamble. Therefore Kimi recovery commits cannot be the direct cause of
+        Kevin's native GPT completed turn.
+      - >-
+        Commit 57fb375 added a provider-specific bounded Kimi continuation loop,
+        517d07a raised it to six attempts, and f9062db expanded Korean/English
+        regular-expression matching. Those recent changes were language
+        dependent and still missed observed wording.
+      - >-
+        The exact frequency increase cannot be attributed to a specific commit
+        from the collected evidence. The native gap predates the recent Kimi
+        commits, while the Kimi implementation itself did change recently.
+
+  rejected_implementation:
+    description: >-
+      An initial local change expanded Korean action words and applied the
+      expression to additional providers.
+    disposition: >-
+      Deleted before commit at the user's direction. No natural-language word,
+      phrase, language, task domain, or action-name matcher remains in the new
+      completion decision.
+    token_handshake: >-
+      The first protocol revision used a fixed text confirmation token. An
+      installed-runtime probe showed that a validator could repeat that token
+      inside its reasoning summary. The text token was therefore removed and
+      replaced by a typed runtime-only tool call before final deployment.
+
+  implementation:
+    translated_provider_gate:
+      trigger: >-
+        The request supplies tools, and the returned message structurally
+        contains reasoning plus visible text but no tool_use item.
+      handshake: >-
+        A private follow-up requires either the runtime-owned
+        ciel_runtime_confirm_completion tool when all requested actions are
+        complete, or the next real work tool call when work remains.
+      delivery: >-
+        Confirmation retains the original answer and hides the control tool;
+        a real tool call replaces/continues the candidate completed response.
+    native_codex_gate:
+      trigger: >-
+        A completed OpenAI Responses SSE has tools in the request, a reasoning
+        output item, visible assistant text, and no protocol action item.
+      buffering: >-
+        The response is spooled to disk after 1 MiB rather than retained in an
+        unbounded byte array. SSE events larger than 8 MiB disable validation
+        and preserve the original response.
+      state_replay: >-
+        Stored responses use previous_response_id; stateless responses replay
+        the returned output items, including encrypted reasoning. This follows
+        the official Responses state fields.
+      fallback: >-
+        Any parse, context, capacity, transport, or validator-protocol failure
+        relays the untouched original response.
+    native_prompt:
+      result: >-
+        The stable Codex compatibility instruction is now also present on the
+        native route and includes persistent-monitoring goal behavior.
+
+  verification:
+    focused_tests:
+      commands:
+        - "python tests/test_codex_completion_gate.py"
+        - "python tests/test_codex_backend_item_repair.py"
+        - "python tests/test_codex_turn_recovery.py"
+      result: "5 + 21 + 43 tests passed"
+    full_suite:
+      command: "npm test"
+      result:
+        unit: "1366 passed, 45 skipped"
+        router: "1093 passed"
+        channel: "403 passed, 80 skipped"
+        runtime: "276 passed, 12 skipped"
+        exit_code: 0
+    lint:
+      result: "passed"
+    packaging:
+      command: "npm pack --dry-run --json"
+      result: >-
+        Passed; 542 entries and ciel_runtime_support/codex_completion_gate.py
+        included.
+    live_ollama_cloud_kimi:
+      isolated_router_port: 9610
+      initial_response: >-
+        Protocol test acknowledged; standing by for the runtime completion check.
+      router_log: >-
+        codex_turn_retry provider=ollama-cloud reason=completion_check
+        model=ciel-runtime-ollama-cloud-kimi-k3 chars=73
+      delivered_action: 'function_call inspect_workspace {"scope":"all"}'
+      http_status: 200
+      cleanup: "The isolated router was stopped after verification."
+      completion_tool_revision:
+        unfinished_case: >-
+          A 399-character arbitrary response with reasoning and no tool was
+          classified as completion_check and delivered inspect_workspace with
+          scope all.
+        finished_case: >-
+          A 2+2 request logged codex_completion_confirmed attempt=1/6 and
+          returned visible text 4 with zero function calls and no completion
+          tool identifier in the delivered response.
+    native_openai:
+      result: >-
+        Native SSE buffering, private confirmation, action replacement, and
+        unknown future action types are covered by the router integration tests.
+        A live native OpenAI call has not yet been made with this uninstalled
+        source revision.
+
+  official_reference:
+    url: "https://developers.openai.com/api/reference/cli/resources/responses/methods/create"
+    confirmed_contract: >-
+      The Responses API defines previous_response_id for multi-turn state,
+      accepts prior response output items as input, and exposes encrypted
+      reasoning for stateless replay.

package/docs/journal/2026/09/03/implementation/runtimes/meta/muse-code/default-yolo.okf

@@ -0,0 +1,40 @@
+task:
+  title: Default Muse Code launches to the native yolo permission mode
+  date: 2026-09-03
+  branch: nightly
+
+evidence:
+  installed_cli: Muse Code 1.0.2 (1.0.2-R2040.1)
+  command: python ciel_runtime.py muse --help
+  observed_option: --yolo
+  observed_definition: Disable approval and sandboxing and trust this workspace for this run.
+  exec_help: muse exec also exposes --yolo under its Safety options.
+  resume_help: Root options may appear on either side of resume.
+
+implementation:
+  default_scope:
+    - interactive Muse sessions
+    - muse exec
+    - muse resume
+  excluded_scope:
+    - help and version
+    - config, schema, serve, session-message, and other utility commands
+  duplicate_policy: Preserve an explicit --yolo exactly once.
+  adapter_position: After exec or resume subcommand when present; otherwise in root options.
+
+verification:
+  focused_tests: 9 passed
+  ruff: passed
+  documentation_metadata: passed, 26 Markdown files
+  actual_exec:
+    command: python ciel_runtime.py muse exec --provider echo --json --no-session-log --max-model-steps 2 "Reply exactly MUSE_YOLO_DEFAULT_OK"
+    exit_code: 0
+    trust_evidence: 'muse: workspace trust: trusted source=run-flag'
+    terminal_evidence: 'echo: Reply exactly MUSE_YOLO_DEFAULT_OK'
+    billing: No paid model call; Muse deterministic echo provider used.
+  full_suite:
+    unit: 1377 passed, 45 skipped
+    router: 1094 passed
+    channel: 406 passed, 80 skipped
+    runtime: 276 passed, 12 skipped
+    total: 3153 passed, 137 skipped

package/docs/journal/2026/09/03/implementation/runtimes/meta/muse-code/native-subscription-router.okf

@@ -0,0 +1,86 @@
+task:
+  title: Native Muse Code subscription runtime in Ciel Router control plane
+  date: 2026-09-03
+  branch: nightly
+
+official_evidence:
+  overview: https://dev.meta.ai/docs/muse-code
+  subscriptions: https://dev.meta.ai/docs/muse-code/subscriptions
+  configuration: https://dev.meta.ai/docs/muse-code/configuration
+  session_messaging: https://dev.meta.ai/docs/muse-code/session-messaging
+  confirmed:
+    - Muse Code is a native terminal and CI agent.
+    - Browser sign-in and its onboarding credential are the subscription path.
+    - Additional API keys are pay-as-you-go and take precedence over browser sign-in.
+    - Native installation is documented for macOS and Linux.
+    - The CLI accepts model and reasoning-effort options and loads AGENTS.md.
+
+installed_binary_evidence:
+  platform: Windows 11 with WSL2 Ubuntu 26.04
+  install_source: https://dev.meta.ai/install.sh
+  installed_path: /home/djlov/.local/bin/muse
+  version_output: Muse Code 1.0.2 (1.0.2-R2040.1)
+  interfaces:
+    - muse exec with JSONL output
+    - muse session-message list and send
+    - muse serve MSP host over stdio
+    - exact MSP schema export from the installed binary
+
+implementation:
+  runtime_command: ciel-runtime muse
+  aliases: [muse-code, launch-muse]
+  model_default: muse-spark-1.3
+  effort_mapping: {max: ultra}
+  authentication_guard:
+    removed_child_environment: [META_API_KEY, MODEL_API_KEY]
+    purpose: Preserve Muse browser sign-in subscription eligibility.
+  windows:
+    transport: wsl.exe
+    discovery: command -v muse in the default WSL2 distribution
+    installer_shell: bash
+  router_control_plane:
+    interactive_input: Ciel channel-aware terminal proxy
+    unsupported_claude_socket_fallback: tty
+    transcript_scope: muse
+    tty_delivery:
+      writes_per_message: 1
+      cursor_commit: immediate and durable
+      rationale: Muse has no Claude/Codex-compatible transcript confirmation signal.
+    model_request_path: Direct Muse Code native connection, not Ciel Model API proxy
+  project_context:
+    instruction_file: AGENTS.md
+    remote_instruction_key: muse_url
+
+verification:
+  focused_tests:
+    muse_runtime: 8 passed
+    architecture_contracts: 219 passed, 42 skipped
+    ruff: passed
+    diff_check: passed
+  full_suite:
+    unit: 1376 passed, 45 skipped
+    router: 1094 passed
+    channel: 406 passed, 80 skipped
+    runtime: 276 passed, 12 skipped
+    total: 3152 passed, 137 skipped
+  cli_version:
+    command: python ciel_runtime.py muse --version
+    result: Muse Code 1.0.2 (1.0.2-R2040.1)
+  utility_passthrough:
+    command: python ciel_runtime.py muse session-message --help
+    result: exit 0; list and send commands displayed
+  real_headless_turn:
+    command: python ciel_runtime.py muse exec --provider echo --json --no-session-log --max-model-steps 2 "Reply exactly MUSE_ROUTER_READY"
+    result: exit 0; run.output.delta and run.terminal.completed contained echo MUSE_ROUTER_READY
+    billing: No paid model call; official deterministic echo provider used.
+  router_tty_end_to_end:
+    runtime: Interactive Muse Code 1.0.2 through WSL2 with provider echo
+    endpoint: POST http://127.0.0.1:9611/ca/channel/messages
+    accepted_message_ids: [4, 5]
+    injection_log:
+      - channel_stdin_proxy_injected count=1 message_ids=4 transport=tty commit_cursor=True
+      - channel_stdin_proxy_injected count=1 message_ids=5 transport=tty commit_cursor=True
+    persisted_cursor: '{"last_id":5}'
+    duplicates_after_fix: 0
+  transient_schema_artifacts:
+    removed: true

package/docs/journal/2026/09/03/operations/deployment/local/session-socket-default-fallback.okf

@@ -0,0 +1,41 @@
+task:
+  title: Local deployment of session-socket-first Web Chat transport
+  date: 2026-09-03
+  workspace: C:/Users/djlov/ciel-runtime
+
+source_verification:
+  full_suite:
+    unit: {passed: 1368, skipped: 45}
+    router: {passed: 1094, skipped: 0}
+    channel: {passed: 405, skipped: 80}
+    runtime: {passed: 276, skipped: 12}
+    total_passed: 3143
+    total_skipped: 137
+  ruff: passed
+  compileall: passed
+  git_diff_check: passed
+
+deployment:
+  command: .\\install.ps1
+  result: passed
+  target: C:/Users/djlov/.local/share/ciel-runtime
+  launcher: C:/Users/djlov/.local/bin/ciel-runtime.ps1
+  reported_version: 0.2.37
+  path_priority: Local launcher is resolved before the AppData npm launcher.
+
+installed_evidence:
+  source_sha256_matches:
+    ciel_runtime.py: true
+    ciel_runtime_support/web_ui.py: true
+    ciel_runtime_support/channel_pending_injection.py: true
+    ciel_runtime_support/runtime_input_gateway.py: true
+    ciel_runtime_support/chat_http_controller.py: true
+  confirmed_markers:
+    - Built-in Web Chat sends top-level input_transport session_socket.
+    - Runtime Input Gateway default is session_socket.
+    - HTTP controller default is session_socket.
+    - Installed pending injection contains session_socket-to-TTY fallback paths.
+
+runtime_boundary:
+  active_processes_restarted: false
+  reason: Local deployment replaces the installation for future launches; restarting multiple active user sessions was not authorized.

package/docs/journal/2026/09/03/research/deployment/colab/tailscale/credential-storage.okf

@@ -0,0 +1,41 @@
+task:
+  title: Colab Web Chat Tailscale credential storage and installation audit
+  date: 2026-09-03
+  scope: current nightly source and local Windows configuration
+
+confirmed:
+  source_hardcoding:
+    tracked_real_keys: 0
+    test_fixture_matches: 5
+  installation:
+    package_install_embeds_key: false
+    automatic_key_transfer_to_new_machine: false
+  web_chat_flow:
+    entry: Web Chat > Speech Settings > Saved Tailscale recovery key
+    profile_scope: one encrypted credential slot per Colab account profile
+    vault: '%APPDATA%\ciel-runtime\colab-worker-credentials.vault.json'
+    vault_key: '%APPDATA%\ciel-runtime\colab-worker-credentials.vault.key'
+    observed_profile: default
+    observed_fields:
+      - tailscale_auth_key
+      - tailscale_asr_state
+      - tailscale_tts_state
+    secret_values_read_or_recorded: false
+  deployment_flow:
+    parent_to_script: process environment TAILSCALE_AUTHKEY
+    script_behavior: captures the process value and removes it from its own environment
+    colab_transfer: generates a temporary bootstrap containing a Base64-encoded assignment
+    local_cleanup: temporary bootstrap is deleted in a finally block
+    worker_use: bootstrap restores TAILSCALE_AUTHKEY into the Colab process environment
+  direct_script_flow:
+    input: current PowerShell process environment TAILSCALE_AUTHKEY
+    persistence_by_script: false
+    fallback: authorized Colab Secret named TAILSCALE_AUTHKEY
+
+evidence:
+  vault_implementation: ciel_runtime_support/colab_speech_jobs.py:30-148
+  web_action_handoff: ciel_runtime_support/colab_speech_jobs.py:262-305
+  web_form: ciel_runtime_support/web_ui.py:293-300
+  temporary_bootstrap: scripts/deploy_colab_speech.ps1:93-128
+  bootstrap_secret_lookup: scripts/colab/bootstrap_qwen_asr.py:40-51
+  config_root: ciel_runtime_support/runtime_paths.py:48-51,103

package/docs/journal/2026/09/03/research/providers/meta/muse-code/subscription-billing-boundary.okf

@@ -0,0 +1,64 @@
+task:
+  title: Muse Spark 1.3 subscription versus API billing boundary
+  date: 2026-09-03
+  scope:
+    provider: Meta Model API
+    model: muse-spark-1.3
+    runtime_workspace_checked: G:/OneCielDMSUI
+
+official_evidence:
+  browser_capture:
+    captured: true
+    decisive_frame: Choose pay-as-you-go or a subscription
+    visible_statement: Subscription only works through Muse Code CLI while signed in; additional API keys are pay-as-you-go.
+  model_api_overview:
+    url: https://dev.meta.ai/docs/overview
+    confirmed:
+      - Direct Meta Model API requests use Bearer MODEL_API_KEY authentication.
+      - Direct Muse Spark API pricing is pay-as-you-go.
+      - OpenAI, Anthropic, and Responses protocol surfaces use the same Model API auth and token cost.
+  muse_code_auth:
+    url: https://dev.meta.ai/docs/muse-code/auth
+    confirmed:
+      - Muse Code supports browser sign-in or an API key.
+      - META_API_KEY and stored API keys take priority over a stored browser session.
+      - Model API key usage is token-based billing.
+  muse_code_subscription:
+    url: https://dev.meta.ai/docs/muse-code/subscriptions
+    confirmed:
+      - A subscription applies only to the Muse Code API key automatically connected during Muse Code CLI onboarding.
+      - The subscription credential is for Muse Code only.
+      - Additional Model API keys are billed pay-as-you-go.
+      - Subscription usage works only through Muse Code CLI while signed in with the Meta Model API account.
+  public_model_card:
+    url: https://developer.meta.com/ai/models/muse-spark/
+    pricing:
+      muse-spark-1.3: {input_per_mtok_usd: 1.25, cached_input_per_mtok_usd: 0.15, output_per_mtok_usd: 4.25}
+      muse-spark-1.3-contributor: {input_per_mtok_usd: 0.10, cached_input_per_mtok_usd: 0.002, output_per_mtok_usd: 0.20}
+
+local_evidence:
+  ciel_meta_adapter:
+    base_url: https://api.meta.ai/v1
+    authentication: Bearer API key
+    required_environment_name: MODEL_API_KEY
+    supported_runtime_path: Ciel Router provider for Claude and Codex compatible clients
+  muse_code_native_runtime:
+    installed_on_path: false
+    supported_by_ciel_launcher: false
+  checked_instance:
+    workspace: G:/OneCielDMSUI
+    provider: meta
+    router: http://127.0.0.1:9605
+    configured_model_at_measurement: muse-spark-1.3-contributor
+
+conclusion:
+  confirmed_cause: Ciel's Meta provider sends a separately created Model API key directly to api.meta.ai, so Meta classifies it as pay-as-you-go. Selecting muse-spark-1.3 versus contributor changes the API price tier but does not attach a Muse Code subscription.
+  unsupported_shortcut: No documented request header, model suffix, or API-key setting converts a direct Model API request into subscription usage.
+  ciel_support_boundary:
+    possible: Add Muse Code as a native CLI runtime and allow Muse Code to own browser sign-in and its subscription credential.
+    prohibited_by_official_contract: Reuse or export the Muse Code-only subscription credential through Ciel Router, Claude, Codex, or direct Model API calls.
+    critical_environment_rule: A Muse Code subscription launch must not inject MODEL_API_KEY or META_API_KEY because an API key overrides browser sign-in and returns usage to pay-as-you-go billing.
+
+repository_change:
+  product_code: none
+  reason: Meta's official subscription contract does not expose a supported router/API mechanism. Implementing an undocumented billing bypass would be incorrect.

package/docs/journal/2026/09/03/research/web-chat/voice/instruction-format.okf

@@ -0,0 +1,135 @@
+task:
+  title: Web Chat voice-message instruction format audit
+  date: 2026-09-03
+  scope:
+    workspace: C:/Users/djlov/ciel-runtime
+    feature: Web Chat live voice input and runtime prompt injection
+
+confirmed_flow:
+  capture:
+    source: ciel_runtime_support/web_ui.py
+    behavior:
+      - Browser microphone PCM is encoded as WAV.
+      - JSON is posted to /v1/audio/transcriptions with audio_base64, filename, content_type, model, and language.
+      - The ASR text is normalized and passed to sendMessage with inputMode voice.
+  channel_request:
+    endpoint: POST /ca/channel/messages
+    body_fields:
+      - channel
+      - sender_id
+      - recipients
+      - delivery
+      - thread_id
+      - kind
+      - message
+      - meta.source
+      - meta.web_chat_session
+      - meta.input_mode
+      - meta.reply_channel
+      - meta.reply_recipient
+      - meta.response_contract
+      - meta.reply_instruction
+      - meta.attachments
+    voice_marker: meta.input_mode = voice
+    message_content: ASR transcript text
+  runtime_prompt:
+    source: ciel_runtime_support/channel_message_prompt.py
+    transcript_label: asr_transcript
+    framing: one physical line
+    trusted_instruction:
+      - Route replies through ciel-runtime-router.send_message.
+      - Send kind ack immediately and exactly one kind reply after work.
+      - response fields are spoken, overview, and optional details.
+      - The browser speaks only response.spoken.
+      - Do not claim completion in the acknowledgement.
+
+channel_semantics:
+  definition: A logical message namespace and reply-routing address; it is not the network transport.
+  web_chat_value: web-chat-<sessionId>
+  session_id_source:
+    priority:
+      - URL query parameter session or s
+      - browser localStorage
+      - newly generated UUID or timestamp-random fallback
+  uses:
+    - Filters Web Chat history reads to the current logical conversation.
+    - Filters the SSE message stream to the current logical conversation.
+    - Tells the agent which channel to target when returning ack and reply messages.
+  related_fields:
+    thread_id: Groups the conversation thread; Web Chat currently uses sessionId.
+    parent_id: Associates an ack or reply with the specific inbound message.
+    recipients: Controls which consumer can see the message; browser reads as recipient web.
+    input_transport: Selects session_socket, router, or tty independently from channel.
+
+delivery_semantics:
+  confirmed:
+    delivery_llm: Causes a structured inbound message to be mirrored into the private model-input queue.
+    delivery_native: Makes the message eligible for native channel notification projection; it does not select the Claude messaging socket.
+    delivery_web: Outbound browser-only ack or reply; it is not mirrored back into model input unless an LLM route is also present.
+  claude_session_socket_selection:
+    controlling_field: input_transport
+    default: session_socket for every admitted input.
+    fallback: If socket delivery is unavailable or fails, use TTY while idle; retain the durable queued input during an active turn and retry until fallback is safe.
+    explicit_override: POST input_transport can select session_socket, tty, or router.
+    conclusion: delivery native alone does not select the socket; input_transport does.
+  verification:
+    chat_input_mode_tests: 13 passed
+    claude_session_socket_focused_tests: 4 passed
+
+correction_2026_09_03:
+  prior_behavior:
+    source_commit: 60846f9
+    admission_default: Active Claude socket selected session_socket; every other runtime selected tty before admission.
+    socket_failure: Durable queue rollback and retry without transport fallback.
+  mismatch: This did not implement the requested session-socket-first default with unsupported-runtime fallback.
+  implementation:
+    - Built-in Web Chat now explicitly sends top-level input_transport session_socket.
+    - HTTP and Runtime Input Gateway defaults now select session_socket.
+    - Missing socket transport and idle socket-send failure fall back to TTY.
+    - Socket-send failure during an active turn does not type into the terminal; the input remains queued.
+    - Explicit router and tty overrides remain supported.
+  verification:
+    focused:
+      web_chat_and_gateway: 25 passed
+      claude_socket_paths: 4 passed
+      speech_http: 23 passed
+      channel_bridge: {passed: 217, skipped: 80}
+      ruff: passed
+      compileall: passed
+    full_suite:
+      unit: {passed: 1368, skipped: 45}
+      router: {passed: 1094, skipped: 0}
+      channel: {passed: 405, skipped: 80}
+      runtime: {passed: 276, skipped: 12}
+      total_passed: 3143
+      total_skipped: 137
+
+security_boundary:
+  observation: The browser-supplied meta.reply_instruction value is not copied into the generated runtime prompt.
+  verification: A generated prompt containing a sentinel CLIENT_SIDE_TEXT in meta.reply_instruction did not contain the sentinel.
+  result: Actual reply instructions are generated server-side from admitted message routing data.
+
+verification:
+  generated_prompt_probe:
+    result:
+      newline_count: 0
+      client_reply_instruction_present: false
+      voice_label_present: true
+      asr_transcript_present: true
+  tests:
+    commands:
+      - PYTHONPATH=<workspace> python tests/test_channel_message_prompt.py
+      - PYTHONPATH=<workspace> python tests/test_web_chat_ui.py
+    result: 21 tests passed
+  note: An initial direct unittest module invocation failed because tests is not a Python package; direct file execution with workspace PYTHONPATH succeeded.
+
+references:
+  - ciel_runtime_support/web_ui.py:1287
+  - ciel_runtime_support/web_ui.py:1455
+  - ciel_runtime_support/web_ui.py:1742
+  - ciel_runtime_support/channel_message_prompt.py:137
+  - ciel_runtime_support/channel_message_prompt.py:203
+  - ciel_runtime_support/channel_message_prompt.py:323
+  - tests/test_channel_message_prompt.py:85
+  - tests/test_channel_message_prompt.py:118
+  - tests/test_web_chat_ui.py:80

package/docs/journal/2026/09/04/diagnostics/claude/context/early-auto-compaction.okf

@@ -0,0 +1,75 @@
+schema: okf/1
+kind: task-journal
+title: Claude routed session early auto-compaction at an apparent 16.6 percent
+date: 2026-09-04
+workspace: G:\ciel-bridge-service
+
+observation:
+  screenshot:
+    claude_native_notice: Compacting conversation
+    claude_native_remaining: 1 percent until auto-compact
+    ciel_status_context: 165666/1000000 tokens (16.6 percent)
+  process:
+    claude_code_version: 2.1.260
+    configured_provider: anthropic
+    configured_model: claude-opus-5
+    routed: true
+  launch_log:
+    model_alias: ciel-runtime-anthropic-claude-opus-5
+    projected_context: 200000
+    projected_auto_compact_window: 200000
+
+root_cause:
+  confirmed:
+    - Ciel's suffix-only compatibility rule classified every routed Anthropic
+      model without `[1m]` as 200K.
+    - The rule overrode the model metadata capacity of 1048576 and projected
+      CLAUDE_CODE_MAX_CONTEXT_TOKENS=200000 plus
+      CLAUDE_CODE_AUTO_COMPACT_WINDOW=200000.
+    - Claude Code's status input still described Opus 5 as a 1M model, so the
+      Ciel status line divided usage by 1000000 while Claude compacted against
+      the Ciel-imposed 200000 window.
+  official_reference:
+    - https://platform.claude.com/docs/en/build-with-claude/context-windows
+    - The current Anthropic table states that current Fable/Mythos, Opus 4.6+
+      and Opus 5, and Sonnet 4.6+ and Sonnet 5 use 1M by default without a beta
+      header.
+
+change:
+  policy:
+    - Apply the 200K routed fallback only when the selected model is not in the
+      documented 1M-default model set.
+    - Expand that set to include Mythos 5/preview, Opus 4.6, and Sonnet 4.6.
+    - Preserve `[1m]` as a backwards-compatible local marker, not a requirement
+      for current 1M-default Anthropic models.
+  regression_tests:
+    - All documented current 1M-default IDs project a 1000000 auto-compact
+      window and 1048576 maximum context without the marker.
+    - Sonnet 4.5 and Haiku 4.5 remain on the 200000 path.
+
+verification:
+  focused:
+    command: python tests/test_anthropic_native_output_tokens.py && python tests/test_claude_native_provider.py
+    result: 115 passed
+  full_suite:
+    command: npm test
+    result:
+      passed: 3156
+      skipped: 137
+      failed: 0
+  local_install:
+    path: C:\Users\djlov\.local\share\ciel-runtime
+    source_install_hash_match:
+      claude_environment.py: true
+      anthropic_model_policy.py: true
+  target_config_projection:
+    configured_model: claude-opus-5
+    anthropic_model: ciel-runtime-anthropic-claude-opus-5
+    max_context_tokens: 1048576
+    auto_compact_window: 1000000
+  isolated_live_launch:
+    claude_code_version: 2.1.260
+    exit_code: 0
+    launch_snapshot: provider=anthropic model=claude-opus-5 context=1048576 window=1000000
+    router_lifetime: stopped cleanly after Claude --version
+  status: verified