@oneciel-ai/ciel-runtime 0.2.37 → 0.2.38

package/docs/journal/2026/09/01/implementation/claude/session-socket/windows-direct-input.okf

@@ -0,0 +1,112 @@
+okf: 1
+record:
+  title: Claude Code Windows peer-socket compatibility audit
+  date: 2026-09-01
+  supersession:
+    status: superseded-by-host-owned-socket-delivery
+    record: implementation/claude/session-socket/all-input-paths/default-delivery.okf
+    note: >-
+      The later implementation omits the peer from address and preserves Ciel's
+      explicit Web Chat/MCP reply contract, removing the reply-routing condition
+      that blocked socket-first delivery in this audit.
+  scope:
+    runtime: ciel-runtime
+    branch: nightly
+    feature:
+      family: external-input
+      transport: claude-cross-session-socket
+      platforms:
+        - windows-named-pipe
+        - unix-domain-socket
+  evidence:
+    official:
+      changelog: https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md
+      channels: https://code.claude.com/docs/en/channels
+      observations:
+        - Claude Code 2.1.239 added Windows cross-session messaging.
+        - Claude Code 2.1.243 requires a complete input line within 30 seconds.
+        - Claude Code 2.1.257 accepts gateway-supplied model descriptions.
+        - Claude Code 2.1.258 is the npm latest version measured for this task.
+    local_package:
+      version: 2.1.258
+      build_sha: b3cd543a1f6fcdf4d8fabc0f5e5538d2ee7f38e1
+      protocol:
+        peer_protocol: 1
+        first_line:
+          type: auth
+          credential: socket-bound peerToken
+        second_line:
+          type: user
+          message_role: user
+        framing: newline-delimited JSON
+        windows_path_shape: '\\.\pipe\LOCAL\cc-msg-<nonce>'
+    live_windows_probe:
+      claude_version: 2.1.258
+      session_name: ciel-socket-probe
+      node_client_result: CIEL_SOCKET_PROBE_OK
+      python_named_pipe_result: CIEL_PY_PIPE_OK
+      observed_ui_provenance: Another Claude session sent a message
+    semantic_probe:
+      result: rejected-for-web-chat
+      observation: Raw socket user frames are classified as ordinary peer messages.
+      side_effect_observed: The target attempted SendMessage to an unrelated live Claude peer.
+      containment: Probe sessions were stopped; the unrelated recipient held the benign probe for approval.
+      binary_source_evidence:
+        socket_user_handler: Constructs peer origin without hostInjected.
+        host_injected_path: Exists only for locally prepended stdin cross-session envelopes.
+        unknown_fields: Ignored by the socket user handler.
+      conclusion: The published peer socket cannot preserve Web Chat or remote-human reply semantics.
+  implementation:
+    safety:
+      socket_first_web_chat: not-enabled
+      retained_input_path: existing terminal and request-body delivery
+      reason: Direct peer delivery can redirect the answer through SendMessage.
+    catalog:
+      gateway_discovery_enabled: true
+      description_field: provider route description
+    subagent_model:
+      configured_source: provider.subagent_model
+      routed_alias: true
+      force_environment: CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1
+      fallback_without_configured_model: main model without force
+  verification:
+    gateway_catalog_http:
+      endpoint: http://127.0.0.1:6971/v1/models
+      status: 200
+      returned_models: 17
+      description_observed: Routed through Ciel Runtime provider 'zai-start-plan'.
+    claude_2_1_258_isolated_subagent:
+      server: local Anthropic-compatible capture server
+      exit_code: 0
+      expected_result_received: true
+      main_model: ciel-runtime-deepseek-deepseek-v4-pro-1m
+      subagent_model: ciel-runtime-deepseek-deepseek-v4-flash
+      request_model_order:
+        - ciel-runtime-deepseek-deepseek-v4-pro-1m
+        - ciel-runtime-deepseek-deepseek-v4-flash
+        - ciel-runtime-deepseek-deepseek-v4-pro-1m
+        - ciel-runtime-deepseek-deepseek-v4-pro-1m
+    external_provider_probe:
+      claude_version: 2.1.258
+      startup: succeeded
+      request_result: '405 code 3012: request blocked due to unusual activity'
+      interpretation: External account response, not used as implementation evidence.
+    full_repository:
+      compile: pass
+      unit:
+        passed: 1302
+        skipped: 44
+      router:
+        passed: 1061
+      channel:
+        passed: 320
+        skipped: 80
+      runtime:
+        passed: 246
+        skipped: 12
+      total_passed: 2929
+      total_skipped: 136
+    quality:
+      ruff: pass
+      documentation_metadata: pass
+      npm_pack_dry_run: pass

package/docs/journal/2026/09/02/diagnostics/cache/alibaba/qwen38-post-restart-hit-rate.okf

@@ -0,0 +1,107 @@
+task:
+  id: qwen38-post-restart-cache-hit-rate-20260902
+  request: >-
+    Measure the Alibaba Singapore Token Plan qwen3.8-max cache hit rate after
+    restarting Ciel Runtime, compare it with the immediately preceding calls,
+    and verify whether cache hits benefit Token Plan usage.
+  status: measured
+
+runtime:
+  workspace: C:/Users/djlov/cielarvis
+  router: http://127.0.0.1:9479
+  router_pid: 77820
+  router_started_at_utc: 2026-09-02T18:14:11Z
+  provider: alitoken
+  model: qwen3.8-max
+  protocol: openai_responses
+  session_id: 01a01dff-f0f0-7ef3-8cde-1d9916510f35
+
+observations:
+  before_restart_last_14_calls:
+    input_tokens: 1932589
+    cached_input_tokens: 614400
+    uncached_input_tokens: 1318189
+    weighted_cache_hit_percent: 31.790
+  after_restart_first_5_calls:
+    input_tokens: 723907
+    cached_input_tokens: 0
+    uncached_input_tokens: 723907
+    weighted_cache_hit_percent: 0.000
+    calls:
+      - {time_utc: 2026-09-02T18:15:18.060Z, input: 142849, cached: 0}
+      - {time_utc: 2026-09-02T18:15:46.264Z, input: 143198, cached: 0}
+      - {time_utc: 2026-09-02T18:18:09.654Z, input: 145653, cached: 0}
+      - {time_utc: 2026-09-02T18:19:00.252Z, input: 145963, cached: 0}
+      - {time_utc: 2026-09-02T18:19:49.739Z, input: 146244, cached: 0}
+  after_restart_latest_15_calls:
+    measured_through_utc: 2026-09-02T18:35:53.849Z
+    input_tokens: 2281369
+    cached_input_tokens: 291885
+    uncached_input_tokens: 1989484
+    weighted_cache_hit_percent: 12.794
+    zero_cache_calls: 13
+    high_cache_calls: 2
+    high_cache_call_details:
+      - {time_utc: 2026-09-02T18:23:56.109Z, input: 146623, cached: 145647, percent: 99.334}
+      - {time_utc: 2026-09-02T18:26:11.480Z, input: 149785, cached: 146238, percent: 97.632}
+  change_percentage_points_latest_15_vs_before_14: -18.998
+  result: cache_hit_rate_partially_recovered_but_remained_below_pre_restart_window
+
+configuration_and_wire_evidence:
+  workspace_protocol_header:
+    x-dashscope-session-cache: enable
+  projected_wire_header:
+    x-dashscope-session-cache: enable
+  codex_request_shape:
+    sends_full_input_history: true
+    previous_response_id_observed: false
+
+official_contract:
+  responses_session_cache:
+    url: https://www.alibabacloud.com/help/en/model-studio/compatibility-with-openai-responses-api
+    facts:
+      - session cache is enabled by x-dashscope-session-cache=enable
+      - qwen3.8-max uses explicit cache when session cache is enabled
+      - the documented multi-turn flow links requests with previous_response_id
+  context_cache:
+    url: https://www.alibabacloud.com/help/en/model-studio/context-cache
+    facts:
+      - implicit cache is automatic for supported models
+      - cache hits are not guaranteed even with an identical prefix
+      - cached tokens remain a subset of input tokens in OpenAI-compatible usage
+      - qwen3.8-max cache discount is model-specific and must be checked in the console
+  token_plan:
+    url: https://docs.modelstudio.console.alibabacloud.com/en/model-studio/token-plan-personal-faq
+    facts:
+      - Token Plan Credits deduction is lower for cache hits than misses
+      - official usage analysis is the source of truth for Credits consumption
+
+conclusion:
+  confirmed: >-
+    The first five completed requests after restart reported zero cached tokens.
+    By the fifteenth request, two high-cache responses raised the weighted rate
+    to 12.794%, which was still 18.998 percentage points below the 31.792%
+    weighted rate across the immediately preceding fourteen calls.
+  causal_boundary: >-
+    The header is present and the current Codex full-history request does not
+    use previous_response_id. This differs from Alibaba's documented Responses
+    session-cache flow, but Alibaba did not return a cache-miss reason. The
+    header alone is therefore not recorded as the proven cause of the misses.
+  correction: >-
+    The earlier journal statement that the missing session-cache opt-in header
+    was the confirmed persistent cause is contradicted by this post-restart
+    production measurement and must not be treated as established fact.
+  usage_effect: >-
+    A successful cache hit lowers Token Plan Credits deduction, but the API's
+    total input_tokens value still includes cached tokens; cached tokens are
+    reported as a subset rather than removed from the input count.
+  measurement_correction: >-
+    The first-five-call result remains a valid time-bounded observation but is
+    not the final post-restart rate. This journal now includes the later fifteen-
+    call snapshot so the initial 0% result is not misread as persistent.
+
+evidence:
+  router_activity: C:/Users/djlov/AppData/Roaming/ciel-runtime/router-instances/9479-a3a719fc6552/router-activity.json
+  router_log: C:/Users/djlov/AppData/Roaming/ciel-runtime/router-instances/9479-a3a719fc6552/router.log
+  codex_session: C:/Users/djlov/.codex/sessions/2026/08/20/rollout-2026-08-20T02-08-27-01a01dff-f0f0-7ef3-8cde-1d9916510f35.jsonl
+  credentials_recorded: false

package/docs/journal/2026/09/02/diagnostics/cache/ollama-cloud/kimi-k3-codex-hit-rate.okf

@@ -0,0 +1,91 @@
+task:
+  id: ollama-cloud-kimi-k3-codex-cache-observation-20260902
+  request: >-
+    Inspect the live Ollama Cloud kimi-k3 Codex sessions and determine whether
+    their prompt-cache hit rate is high.
+  observed_at: 2026-09-02T17:08:32-05:00
+  status: measured_client_metric_but_upstream_cache_hit_rate_not_observable
+
+live_sessions:
+  - router_port: 9465
+    router_pid: 101436
+    workspace: G:/onecieldmsui
+    provider: ollama-cloud
+    model: kimi-k3
+    codex_session: 019fcac4-4538-7a53-91d1-565cd21a8c5e
+    rollout: C:/Users/djlov/.codex/sessions/2026/08/03/rollout-2026-08-03T22-14-48-019fcac4-4538-7a53-91d1-565cd21a8c5e.jsonl
+    completed_usage_samples:
+      - {timestamp: 2026-09-02T22:05:56.843Z, input_tokens: 826809, cached_input_tokens: 0, output_tokens: 354}
+      - {timestamp: 2026-09-02T22:06:12.900Z, input_tokens: 831304, cached_input_tokens: 0, output_tokens: 393}
+    aggregate: {samples: 2, input_tokens: 1658113, cached_input_tokens: 0, reported_weighted_hit_percent: 0.000}
+  - router_port: 9479
+    router_pid: 32968
+    workspace: C:/Users/djlov/cielarvis
+    provider: ollama-cloud
+    model: kimi-k3
+    codex_session: 01a01dff-f0f0-7ef3-8cde-1d9916510f35
+    rollout: C:/Users/djlov/.codex/sessions/2026/08/20/rollout-2026-08-20T02-08-27-01a01dff-f0f0-7ef3-8cde-1d9916510f35.jsonl
+    completed_usage_samples:
+      - {timestamp: 2026-09-02T22:06:55.315Z, input_tokens: 607328, cached_input_tokens: 0, output_tokens: 754}
+      - {timestamp: 2026-09-02T22:07:07.182Z, input_tokens: 610473, cached_input_tokens: 0, output_tokens: 507}
+      - {timestamp: 2026-09-02T22:07:23.556Z, input_tokens: 611608, cached_input_tokens: 0, output_tokens: 271}
+    aggregate: {samples: 3, input_tokens: 1829409, cached_input_tokens: 0, reported_weighted_hit_percent: 0.000}
+combined_reported_aggregate:
+  samples: 5
+  input_tokens: 3487522
+  cached_input_tokens: 0
+  reported_weighted_hit_percent: 0.000
+
+wire_evidence:
+  endpoint: https://ollama.com/api/chat
+  port_9465_prompt_estimates: [809165, 809265, 813789, 813876, 815525, 815606]
+  port_9479_prompt_estimates: [575538, 575630, 578584, 579517, 579614]
+  note: >-
+    These are Ciel request-side token estimates, not upstream cache accounting.
+
+source_evidence:
+  response_decoder: ciel_runtime_support/protocols/ollama_chat.py
+  decoder_behavior: >-
+    decode_ollama_chat_response reads prompt_eval_count and eval_count only.
+  usage_encoder: ciel_runtime_support/protocols/ollama_chat.py
+  encoder_behavior: >-
+    encode_anthropic_message emits usage.input_tokens and usage.output_tokens;
+    it does not emit cache_read_input_tokens or cache_creation_input_tokens.
+  codex_projection: ciel_runtime_support/protocols/openai_responses.py
+  projection_behavior: >-
+    Missing Anthropic cache fields become input_tokens_details.cached_tokens=0
+    and cache_write_tokens=0 in the Responses result consumed by Codex.
+  stream_collector: ciel_runtime_support/ollama_stream_collection.py
+  collector_behavior: >-
+    The collector retains prompt_eval_count, eval_count, and total_duration but
+    currently does not retain prompt_eval_duration from the terminal chunk.
+
+official_contract:
+  chat_api: https://docs.ollama.com/api/chat
+  usage_api: https://docs.ollama.com/api/usage
+  documented_usage_fields:
+    - total_duration
+    - load_duration
+    - prompt_eval_count
+    - prompt_eval_duration
+    - eval_count
+    - eval_duration
+  cache_token_field_documented: false
+  anthropic_compatibility: https://docs.ollama.com/api/anthropic-compatibility
+  anthropic_prompt_caching_documented_as_supported: false
+
+conclusion:
+  confirmed:
+    - The five current Codex usage samples report zero cached input tokens.
+    - Ciel cannot receive a cache-token count from the documented Ollama chat response schema.
+    - Ciel therefore projects zero cached tokens to Codex for this route.
+  not_confirmed:
+    - The Ollama Cloud server actually missed its internal prefix or KV cache.
+    - The actual upstream cache-hit percentage is zero.
+    - Codex has a worse upstream cache hit rate than Claude for the same model, prompt, and endpoint.
+  measurement_limit: >-
+    The current zero-percent Codex display is an observability result, not a
+    validated Ollama Cloud cache-hit measurement. A same-provider Claude sample
+    with a cache field was not present in the active routers for comparison.
+
+credentials_recorded: false

package/docs/journal/2026/09/02/diagnostics/codex/tui/statusline-cache-metrics.okf

@@ -0,0 +1,44 @@
+task:
+  id: codex-statusline-cache-metrics-20260902
+  request: >-
+    Determine whether Codex can display a Claude-like cache hit rate in its TUI
+    status line.
+  status: investigated
+
+local_evidence:
+  installed_version: codex-cli 0.152.1
+  executable: C:/Users/djlov/AppData/Local/Programs/OpenAI/Codex/bin/codex.exe
+  existing_ciel_statusline:
+    source: ciel_runtime_support/statusline_script.py
+    cache_metrics_supported: true
+    integration_target: Claude Code statusLine command settings
+    codex_integration_present: false
+
+official_source_evidence:
+  status_line_enum:
+    url: https://github.com/openai/codex/blob/main/codex-rs/tui/src/bottom_pane/status_line_setup.rs
+    available_token_items:
+      - used-tokens
+      - total-input-tokens
+      - total-output-tokens
+      - context-remaining
+      - context-used
+      - context-window-size
+    cached_tokens_item_present: false
+    cache_hit_rate_item_present: false
+  status_line_renderer:
+    url: https://github.com/openai/codex/blob/main/codex-rs/tui/src/chatwidget/status_surfaces.rs
+    cached_usage_render_branch_present: false
+  external_command_request:
+    url: https://github.com/openai/codex/issues/20043
+    state: closed_as_duplicate_feature_request
+    implemented_contract_confirmed: false
+
+conclusion:
+  current_official_codex_support: false
+  statement: >-
+    Codex 0.152.1 can show cumulative token and context metrics through fixed
+    built-in status-line items, but it cannot show cached_tokens or cache hit
+    percentage, and it cannot attach Ciel's Claude-style external statusLine
+    command. Adding the metric requires an upstream Codex implementation or a
+    maintained Codex fork; no Ciel configuration-only implementation exists.

package/docs/journal/2026/09/02/diagnostics/input/transport/fallback-behavior.okf

@@ -0,0 +1,31 @@
+task:
+  id: input-transport-fallback-behavior-20260902
+  request: Determine whether unsupported session_socket input automatically changes to tty.
+  status: diagnosed
+
+evidence:
+  advertised_allowed:
+    source: ciel_runtime_support/speech_http_controller.py
+    values: [session_socket, tty, router]
+    meaning: globally accepted request values, not active-runtime capabilities
+  omitted_transport:
+    source: ciel_runtime.py:_default_channel_input_transport
+    result: >-
+      session_socket only when the active runtime is Claude and the Ciel-owned
+      Claude socket is configured; tty otherwise.
+  explicit_session_socket:
+    sources:
+      - ciel_runtime_support/chat_http_controller.py:_message_modes
+      - ciel_runtime_support/channel_pending_injection.py:inject_pending_channel_messages
+      - ciel_runtime_support/claude_session_socket.py:ClaudeSessionSocketClient.send
+    result: >-
+      The request is accepted and stamped session_socket. If no socket target is
+      configured or submission fails, delivery is deferred and the cursor is not
+      advanced. No tty fallback occurs.
+  explicit_router:
+    result: no tty fallback when router transport is unavailable
+
+conclusion:
+  automatic_for_omitted_transport: true
+  fallback_for_explicit_session_socket: false
+  capability_specific_allowed_list: false

package/docs/journal/2026/09/02/diagnostics/providers/ollama-cloud/kimi-k3-agent-turn-recovery.okf

@@ -0,0 +1,125 @@
+task:
+  id: ollama-cloud-kimi-k3-agent-turn-recovery-20260902
+  request: >-
+    Diagnose Kimi K3 stopping after a progress sentence in routed Codex, compare
+    Ollama Desktop's official implementation, and apply the relevant behavior.
+  observed_at: 2026-09-02T19:05:11-05:00
+  status: verified_and_locally_deployed
+
+live_failure_evidence:
+  workspace: G:/OneCielDMSUI
+  router_instance: 9465-60524a4bf5eb
+  codex_session: 019fcac4-4538-7a53-91d1-565cd21a8c5e
+  sequence:
+    - at: 2026-09-02T18:38:20-05:00
+      observed: two exec_command tool calls completed
+    - at: 2026-09-02T18:39:06-05:00
+      assistant_text: >-
+        번역 계획(plan) 생성 지점과 메일함 설정 디코더, 이벤트 종류,
+        미리보기 조회를 확인합니다.
+      tool_calls: 0
+      codex_event: task_complete
+  router_log:
+    first_recovery_at: 2026-09-02T18:38:59-05:00
+    first_recovery_reason: preamble_only
+    recovery_attempts_before_patch: 1
+    outcome: >-
+      The one recovery response also contained no tool call. The old code then
+      returned the original announcement and Codex completed the task.
+
+official_ollama_desktop_reference:
+  repository: https://github.com/ollama/ollama
+  commit: b79067b0db7417f20108363bc22adb97f35c966a
+  file: app/ui/ui.go
+  relevant_lines:
+    - 952-955: buffer complete assistant tool-call state
+    - 959-993: run an agent pass loop
+    - 1017-1103: detect and execute tool calls
+    - 1200-1275: persist assistant content and thinking
+    - 1287-1293: continue while tools were executed
+  official_tool_loop_documentation: >-
+    https://github.com/ollama/ollama/blob/main/docs/capabilities/tool-calling.mdx
+  confirmed_difference: >-
+    Ollama Desktop retains the assistant response, including thinking and tool
+    calls, before the next pass. Ciel's internal preamble recovery retained only
+    visible text and discarded thinking.
+
+root_cause:
+  confirmed:
+    - >-
+      Ciel bounded preamble recovery to exactly one retry. The live trace shows
+      that one retry was insufficient for this Kimi K3 turn.
+    - >-
+      body_with_continuation_nudge reconstructed the assistant message from text
+      only, dropping its thinking blocks before the recovery request.
+    - >-
+      A substantive no-tool completion returned by recovery was discarded by
+      the old branch because only a tool call could replace the announcement.
+  not_claimed:
+    - >-
+      The trace does not prove why Kimi K3 chose an announcement instead of a
+      tool call. GitHub Copilot's same-model behavior is a user observation; its
+      private request prompt was not available for direct comparison.
+
+implementation:
+  file: ciel_runtime_support/codex_turn_recovery.py
+  changes:
+    - Preserve the complete assistant content block list in continuation history.
+    - Recognize Kimi K3 through ollama and ollama-cloud routes.
+    - Allow up to three bounded no-tool continuation attempts for Kimi K3.
+    - Accept a concrete non-retryable no-tool completion from recovery.
+    - Disable nested gateway retries inside the bounded recovery loop.
+    - Log attempt number, retryability, and returned visible character count.
+
+verification:
+  focused_tests:
+    command: py -3.14 -m unittest tests.test_codex_turn_recovery -v
+    tests: 36
+    result: passed
+  related_regression_tests:
+    tests: 86
+    skipped: 12
+    result: passed
+  full_suite:
+    unit: {tests: 1363, skipped: 45, result: passed}
+    router: {tests: 1076, result: passed}
+    channel: {tests: 403, skipped: 80, result: passed}
+    runtime: {tests: 261, skipped: 12, result: passed}
+    total_tests: 3103
+  ruff:
+    result: passed
+  direct_ollama_agent_loop:
+    model: kimi-k3
+    think: low
+    pass_1: {tool: inspect_workspace, done_reason: stop}
+    pass_2: {tool: inspect_file, done_reason: stop}
+    pass_3: {content: DONE, done_reason: stop}
+    result: passed
+  installed_router_repeated_preamble_probe:
+    port: 9491
+    input_behavior: two progress-only responses before a required tool call
+    router_log:
+      - {attempt: 1, retryable: true, visible_chars: 20}
+      - {attempt: 2, tool: inspect_workspace}
+    http_status: 200
+    responses_status: completed
+    output_types: [message, reasoning, function_call]
+    result: passed
+  isolated_router_after_probe: stopped
+
+local_deployment:
+  command: powershell -NoProfile -ExecutionPolicy Bypass -File .\install.ps1
+  result: passed
+  target: C:/Users/djlov/.local/share/ciel-runtime
+  source_installed_sha256_match: true
+  installed_behavior_probe:
+    attempts: 2
+    tool_recovered: true
+    thinking_preserved: true
+    nested_gateway_retries: 0
+  active_session_boundary: >-
+    Routers 9465 and 9479 were already running before deployment and retain the
+    old imported Python module until those sessions are restarted. They were not
+    interrupted by this verification.
+
+credentials_recorded: false

package/docs/journal/2026/09/02/diagnostics/providers/ollama-cloud/kimi-k3-parameters.okf

@@ -0,0 +1,128 @@
+task:
+  id: ollama-cloud-kimi-k3-parameter-verification-20260902
+  request: >-
+    Verify the parameters Ollama publishes for kimi-k3 and compare them with
+    the Ciel Runtime settings and request projection used by the live Codex sessions.
+  observed_at: 2026-09-02T18:41:26-05:00
+  status: verified
+
+official_api_show:
+  request:
+    endpoint: https://ollama.com/api/show
+    model: kimi-k3
+    authenticated: true
+  result: success
+  modified_at: 2026-07-27T10:00:00-05:00
+  details:
+    parent_model: kimi-k3
+    family: kimi-k3
+    parameter_size: 2812000000000
+    quantization_level: MXFP4
+  capabilities: [vision, thinking, completion, tools]
+  model_info:
+    general.architecture: kimi-k3
+    general.parameter_count: 2812000000000
+    kimi-k3.context_length: 1048576
+    kimi-k3.embedding_length: 7168
+  parameters: null
+  meaning: >-
+    The provider did not publish fixed temperature, top_p, top_k, seed,
+    repeat_penalty, num_ctx, or num_predict values through /api/show.
+
+official_model_card:
+  url: https://ollama.com/library/kimi-k3
+  variant: kimi-k3:cloud
+  context_tokens: 1048576
+  advertised_capabilities: [vision, tools, thinking, cloud]
+  pricing_per_million_tokens:
+    input_usd: 3.00
+    cached_input_usd: 0.30
+    output_usd: 15.00
+
+official_request_contract:
+  ollama_chat_api: https://docs.ollama.com/api/chat
+  ollama_think_values: [true, false, low, medium, high, max]
+  kimi_vendor_source: https://github.com/MoonshotAI/Kimi-K3/blob/main/README.md
+  kimi_thinking_always_on: true
+  kimi_reasoning_effort_values: [low, high, max]
+  kimi_default_reasoning_effort: max
+  benchmark_only_sampling_values:
+    temperature: 1.0
+    top_p_single_step: 0.95
+    top_p_agentic: 1.0
+  benchmark_only_note: >-
+    Moonshot reports these sampling values for its published evaluations. The
+    README does not state that Ollama Cloud uses them as hidden defaults.
+
+workspace_configuration:
+  workspaces:
+    - {hash: a3a719fc6552, path: C:/Users/djlov/cielarvis}
+    - {hash: 60524a4bf5eb, path: G:/OneCielDMSUI}
+  provider: ollama-cloud
+  current_model: kimi-k3
+  base_url: https://ollama.com
+  num_ctx: auto
+  num_ctx_min: 32768
+  num_ctx_max: 1048576
+  think: true
+  ollama_options: {}
+  request_timeout_ms: 300000
+  discovered_metadata:
+    model: kimi-k3
+    architecture: kimi-k3
+    capabilities: [vision, thinking, completion, tools]
+    thinking_levels: [low, high, max]
+    thinking_always_on: true
+
+wire_projection:
+  required_fields:
+    model: kimi-k3
+    stream: true
+    messages: present
+    tools: present_when_client_supplies_tools
+  omitted_provider_owned_defaults:
+    - options.num_ctx
+    - options.num_predict
+    - options.temperature
+    - options.top_p
+    - options.top_k
+    - options.seed
+    - options.repeat_penalty
+    - keep_alive
+  omission_reason: >-
+    num_ctx is auto, output tokens and keep_alive were not explicitly selected,
+    and ollama_options is empty. Ciel therefore leaves those values to Ollama.
+  verified_reasoning_mapping:
+    low: {think: low}
+    high: {think: high}
+    xhigh: {think: max}
+  active_codex_reasoning_effort: high
+  active_wire_think: high
+
+connection_evidence:
+  api_show_authenticated: success
+  router_9465:
+    request_started: 2026-09-02T18:33:59-05:00
+    request_estimated_tokens: 816014
+    first_responses_200: 2026-09-02T18:36:12-05:00
+    later_responses_200_through: 2026-09-02T18:37:48-05:00
+  router_9479:
+    request_started: 2026-09-02T18:34:15-05:00
+    request_estimated_tokens: 579783
+    first_responses_200: 2026-09-02T18:35:45-05:00
+    later_responses_200_through: 2026-09-02T18:38:09-05:00
+  upstream_error_in_examined_window: false
+  conclusion: >-
+    The provider connection and model lookup succeeded. The first observed
+    responses took about 133 seconds and 90 seconds respectively, so a client
+    could appear disconnected while waiting, but the logs do not show a
+    connection failure in this interval.
+
+source_evidence:
+  provider_adapter: ciel_runtime_support/providers/ollama.py
+  thinking_policy: ciel_runtime_support/ollama_thinking.py
+  request_builder: ciel_runtime_support/provider_request_builder.py
+  wire_projection: ciel_runtime_support/ollama_wire_projection.py
+  context_policy: ciel_runtime_support/providers/ollama_context.py
+
+credentials_recorded: false