@oneciel-ai/ciel-runtime 0.2.37 → 0.2.38

package/docs/journal/2026/09/02/research/providers/ollama-cloud/desktop-model-exhaustive-audit.okf

@@ -0,0 +1,159 @@
+task:
+  id: ollama-cloud-desktop-model-exhaustive-audit-20260902
+  request: >-
+    Deploy Ciel Runtime locally and exhaustively inspect every Ollama Cloud
+    model discoverable by Ollama Desktop.
+  observed_at: 2026-09-02T20:02:00-05:00
+  status: verified_and_locally_deployed
+
+scope_and_sources:
+  official_desktop_repository:
+    url: https://github.com/ollama/ollama
+    commit: b79067b0db7417f20108363bc22adb97f35c966a
+    model_discovery:
+      file: app/ui/app/src/api.ts
+      lines: 199-293
+      behavior:
+        - Read local models with ollama.list.
+        - Read cloud models through the Desktop cloud-model endpoint.
+        - Remove only BERT-only families from the selectable chat list.
+        - Read selected-model capabilities with ollama.show.
+    request_projection:
+      file: app/ui/ui.go
+      lines: 893-911,1847-1939
+      behavior:
+        - Use the show response's thinking capability as the default.
+        - Send true for capability-only thinking models.
+        - Send a string only when the Desktop UI explicitly supplies one.
+        - Preserve assistant thinking and tool calls in conversation history.
+  official_cloud_docs: https://docs.ollama.com/cloud
+  official_thinking_docs: https://docs.ollama.com/capabilities/thinking
+  official_model_cards:
+    - https://ollama.com/library/deepseek-v4-flash
+    - https://ollama.com/library/deepseek-v4-pro
+    - https://ollama.com/library/glm-5.2
+    - https://ollama.com/library/glm-5.3
+    - https://ollama.com/library/glm-5.3-flash
+    - https://ollama.com/library/gpt-oss
+    - https://ollama.com/library/kimi-k3
+
+discovery_evidence:
+  endpoint: GET https://ollama.com/api/tags
+  authenticated: true
+  discovered_models: 19
+  installed_ciel_models: 19
+  set_difference: []
+  credentials_recorded: false
+  show_probe:
+    endpoint: POST https://ollama.com/api/show
+    success: 19/19
+    parameters_field_null: 19/19
+    conclusion: >-
+      Ollama Cloud did not publish fixed sampling parameters for any of the 19
+      models through /api/show. Ciel therefore must not invent temperature,
+      top_p, top_k, or num_predict defaults for the wire request.
+
+model_inventory:
+  columns: [model, context_tokens, capabilities, modalities, verified_think_contract, tool_probe_seconds]
+  rows:
+    - [deepseek-v4-flash:0731, 1048576, completion|tools|thinking, text, high|max_default, 1.11]
+    - [deepseek-v4-pro:0813, 1048576, completion|tools|thinking, text, high_default|max, 0.96]
+    - [gemma4:31b, 262144, completion|thinking|tools|vision, text|image, boolean, 0.53]
+    - [glm-5.1, 202752, thinking|completion|tools, text, boolean, 1.40]
+    - [glm-5.2, 1048576, thinking|completion|tools, text, high_default|max, 1.52]
+    - [glm-5.3, 1048576, completion|thinking|tools, text, low|high|max_default, 1.22]
+    - [glm-5.3-flash, 1048576, completion|thinking|tools|vision, text|image, low|high_default|max, 0.76]
+    - [gpt-oss:120b, 131072, completion|tools|thinking, text, low|medium_default|high, 0.56]
+    - [gpt-oss:20b, 131072, completion|tools|thinking, text, low|medium_default|high, 0.99]
+    - [kimi-k2.6, 262144, vision|thinking|completion|tools, text|image, boolean, 2.33]
+    - [kimi-k2.7-code, 262144, vision|thinking|completion|tools, text|image, boolean, 2.67]
+    - [kimi-k3, 1048576, vision|thinking|completion|tools, text|image, low|high|max_default, 1.27]
+    - [minimax-m2.7, 196608, completion|tools|thinking, text, boolean, 1.60]
+    - [minimax-m3, 512000, completion|tools|thinking|vision, text|image, boolean, 3.05]
+    - [mistral-large-3:675b, 262144, completion|tools|vision, text|image, none, 1.18]
+    - [nemotron-3-nano:30b, 262144, completion|tools|thinking, text, boolean, 0.63]
+    - [nemotron-3-super, 262144, completion|thinking|tools, text, boolean, 3.38]
+    - [nemotron-3-ultra, 262144, completion|thinking|tools, text, boolean, 22.85]
+    - [qwen3.5:397b, 262144, completion|thinking|tools|vision, text|image, boolean, 1.65]
+
+live_inference_verification:
+  endpoint: POST https://ollama.com/api/chat
+  request:
+    stream: false
+    task: Call ping exactly once with value ok and no text answer.
+    thinking: >-
+      false for ordinary models and low for GPT-OSS, solely to keep the health
+      probe small; this probe did not attempt to benchmark reasoning quality.
+  results:
+    http_200: 19/19
+    tool_name_ping: 19/19
+    tool_argument_ok: 19/19
+    minimum_seconds: 0.53
+    maximum_seconds: 22.85
+    maximum_model: nemotron-3-ultra
+  limits:
+    - A successful minimal tool call proves availability and schema execution.
+    - It does not prove sustained-agent quality or normal workload latency.
+
+confirmed_defects:
+  - id: glm-5.3-effort-profile-missing
+    evidence: >-
+      /api/show reports architecture glm_dsa_moe and the official model card
+      documents low, high, and max with max as default. Ciel had no handler for
+      glm_dsa_moe, so its Codex reasoning-level list was empty.
+    correction: >-
+      Add the glm_dsa_moe/glm-5.3 profile; expose low, high, xhigh to Codex and
+      map xhigh to Ollama max, with xhigh as the default.
+  - id: unverified-generic-string-effort
+    evidence: >-
+      Ollama's official generic thinking documentation says most models use a
+      boolean and identifies GPT-OSS as the generic exception requiring
+      low/medium/high strings. Ollama Desktop likewise only exposes the
+      low/medium/high selector for GPT-OSS. Several Ciel model names had been
+      assigned low/medium/high/max without a published model-specific contract.
+    correction: >-
+      Remove the name-based generic string levels. Capability-only thinking
+      models now send boolean true, matching Desktop. Explicit model contracts
+      remain for DeepSeek V4, GLM-5.2, GLM-5.3, GLM-5.3 Flash, GPT-OSS, and Kimi K3.
+
+verification:
+  focused_ollama_suite:
+    command: python -m unittest discover -s tests -p "test_ollama*.py" -v
+    tests: 76
+    result: passed
+  ruff:
+    command: python -m ruff check ciel_runtime_support/ollama_thinking.py tests/test_ollama_provider_options.py
+    result: passed
+  full_suite:
+    command: npm test
+    unit: {tests: 1363, skipped: 45, result: passed}
+    router: {tests: 1077, result: passed}
+    channel: {tests: 403, skipped: 80, result: passed}
+    runtime: {tests: 261, skipped: 12, result: passed}
+    total_tests: 3104
+  installed_policy_probe:
+    glm-5.3:
+      native_levels: [low, high, max]
+      codex_levels: [low, high, xhigh]
+      default_native: max
+      codex_xhigh_wire_value: max
+    minimax-m3:
+      native_contract: boolean
+      codex_xhigh_wire_value: true
+    mistral-large-3:675b:
+      think: false
+      modalities: [text, image]
+
+local_deployment:
+  command: powershell -NoProfile -ExecutionPolicy Bypass -File .\install.ps1
+  result: passed
+  target: C:/Users/djlov/.local/share/ciel-runtime
+  changed_file_sha256: 0B970B89AD6AD2CC5DAAAC680669C8E7860672BA0DD7B641CD457E1D5AA09F63
+  source_installed_sha256_match: true
+  installed_model_command:
+    command: ciel-runtimectl models ollama-cloud
+    models: 19
+    result: passed
+  active_process_boundary: >-
+    The files are locally deployed. Any Ciel router process that imported the
+    previous module before this deployment must be restarted to load this code.

package/docs/journal/2026/09/03/diagnostics/providers/meta/muse-spark-contributor-cache-hit.okf

@@ -0,0 +1,148 @@
+task:
+  title: Muse Spark 1.3 Contributor cache-hit diagnosis
+  date: 2026-09-03
+  scope:
+    workspace: G:/OneCielDMSUI
+    router_port: 9605
+    router_pid: 90340
+    provider: meta
+    model: muse-spark-1.3-contributor
+    codex_session: 019fcac4-4538-7a53-91d1-565cd21a8c5e
+
+correction:
+  discarded_instance:
+    port: 9471
+    workspace: C:/Users/djlov/cielarvis/mind-core-sand-box
+    reason: It is a separate active Claude/Anthropic instance and is not evidence for DMSUI.
+  authoritative_instance:
+    port: 9605
+    workspace: G:/OneCielDMSUI
+    runtime: Codex 0.152.1
+
+observations:
+  rollout_measurement:
+    recent_40:
+      interval: 2026-09-03T07:30:25-2026-09-03T08:06:34
+      weighted_cache_hit_percent: 85.454
+      zero_hit_requests: 5
+      below_90_percent_requests: 7
+    recent_100:
+      interval: 2026-09-03T06:53:20-2026-09-03T08:06:34
+      weighted_cache_hit_percent: 89.037
+      zero_hit_requests: 10
+      below_90_percent_requests: 12
+    pattern: Most requests hit 99 percent or more, with intermittent complete or partial misses.
+    mid_task_evidence:
+      task: 52
+      complete_misses:
+        - {time: 2026-09-03T07:56:48, call: 24, input_tokens: 587410, cached_tokens: 0}
+        - {time: 2026-09-03T07:58:42, call: 26, input_tokens: 588396, cached_tokens: 0}
+      partial_miss: {time: 2026-09-03T07:59:28, call: 27, input_tokens: 590307, cached_tokens: 217329}
+      implication: New user-turn startup alone cannot account for every observed miss.
+
+  codex_wire_probe:
+    method: Codex 0.152.1 was connected to an isolated loopback HTTP capture server.
+    requests: 3
+    confirmed:
+      prompt_cache_key_present: true
+      prompt_cache_key_stable_across_requests: true
+      prompt_cache_retention_present: false
+      instructions_fingerprint_stable: true
+      tool_fingerprint:
+        request_1: ee2c774b9ead15c5
+        request_2: 6f65d7083e8fd376
+        request_3: 6f65d7083e8fd376
+      tool_change_scope: Initial asynchronous app namespace discovery changed five tool namespace payloads once; requests 2 and 3 were stable.
+
+  live_meta_probe:
+    endpoint: http://127.0.0.1:9605/v1/responses
+    upstream: https://api.meta.ai/v1
+    request:
+      prompt_tokens: 7674
+      prompt_cache_key: stable probe key
+      prompt_cache_retention: 24h
+    results:
+      first: {http_status: 200, cached_tokens: 0, elapsed_seconds: 1.78}
+      second: {http_status: 200, cached_tokens: 7665, elapsed_seconds: 2.39}
+      second_hit_percent: 99.88
+    confirmed: Meta accepted the Responses retention field and reused the stable prefix.
+
+official_sources:
+  meta_prompt_caching:
+    url: https://github.com/meta-models/meta-model-cookbook/blob/main/01_api_fundamentals/05_prompt_caching.ipynb
+    confirmed_contract:
+      - Cache matching requires an identical leading token sequence.
+      - A stable prompt_cache_key routes matching prefixes together at scale.
+      - prompt_cache_retention 24h keeps bursty traffic warm across idle gaps.
+      - Retention is a hint and not a guarantee.
+
+implementation:
+  meta_adapter:
+    file: ciel_runtime_support/providers/meta.py
+    behavior:
+      - Default prompt_cache_retention is 24h.
+      - It is injected only for Responses requests carrying prompt_cache_key.
+      - An explicit client value such as in_memory is preserved.
+  diagnostics:
+    files:
+      - ciel_runtime_support/responses_cache_diagnostics.py
+      - ciel_runtime_support/provider_responses_passthrough.py
+    recorded_fields:
+      - cache_hit_percent
+      - prompt_cache_key_fingerprint
+      - prompt_cache_retention
+      - request_instructions_fingerprint
+      - request_tools_fingerprint
+      - request_input_head_fingerprint
+      - request_input_items
+      - request_tools
+      - request_bytes
+      - request_uses_previous_response_id
+    privacy: Raw prompt text, tool definitions, and cache keys are not recorded.
+
+verification:
+  focused:
+    command: python -m unittest test_meta_provider test_responses_cache_diagnostics test_responses_usage_observer test_statusline
+    result: 55 tests passed
+  full_suite:
+    command: npm test
+    groups:
+      unit: {run: 1368, skipped: 45, failures: 0}
+      router: {run: 1094, skipped: 0, failures: 0}
+      channel: {run: 403, skipped: 80, failures: 0}
+      runtime: {run: 276, skipped: 12, failures: 0}
+    total: {run: 3141, skipped: 137, failures: 0}
+  static:
+    ruff: passed
+    compileall: passed
+    git_diff_check: passed
+  merged_workspace_config_probe:
+    provider: meta
+    model: muse-spark-1.3-contributor
+    merged_default_retention: 24h
+    normalized_request_retention: 24h
+    source_request_mutated: false
+    discarded_check: A raw config-file-only probe returned null because it bypassed the runtime default merge; it is not runtime-equivalent evidence.
+
+limits:
+  confirmed_not_proven:
+    - The missing retention hint is a verified client gap and the live fix path is verified.
+    - It does not by itself prove the cause of complete misses occurring less than one minute apart.
+    - The new fingerprints are required after restart to distinguish request-prefix mutation from provider-side best-effort cache loss for those events.
+
+deployment:
+  local:
+    command: powershell -NoProfile -ExecutionPolicy Bypass -File ./install.ps1
+    target: C:/Users/djlov/.local/share/ciel-runtime
+    result: installed
+    hash_matches:
+      - ciel_runtime_support/providers/meta.py
+      - ciel_runtime_support/provider_responses_passthrough.py
+      - ciel_runtime_support/responses_cache_diagnostics.py
+    executable_version: 0.2.37
+    active_dmsui_router_restarted: false
+    activation: The existing PID 90340 remains on its already-loaded modules; the next DMSUI runtime start loads this deployment.
+  nightly:
+    branch: nightly
+    target: origin/nightly
+    mechanism: A push to nightly triggers the npm nightly publication workflow.

package/docs/journal/2026/09/03/diagnostics/providers/meta/muse-spark-contributor-required-schema.okf

@@ -0,0 +1,127 @@
+task:
+  id: meta-muse-spark-contributor-required-schema-20260903
+  request: >-
+    Diagnose the Muse Spark 1.3 Contributor failure in G:/OneCielDMSUI and make
+    the real Codex request acceptable to Meta's Responses endpoint.
+  workspace: G:/OneCielDMSUI
+  runtime_workspace_id: 60524a4bf5eb
+  router_instance: 9491-60524a4bf5eb
+  codex_session: 019fcac4-4538-7a53-91d1-565cd21a8c5e
+  status: verified
+
+observed_failure:
+  rollout_path: C:/Users/djlov/.codex/sessions/2026/08/03/rollout-2026-08-03T22-14-48-019fcac4-4538-7a53-91d1-565cd21a8c5e.jsonl
+  occurrences:
+    - line: 7811
+      at: 2026-09-03T04:55:29.475Z
+    - line: 7817
+      at: 2026-09-03T04:57:30.700Z
+  response:
+    type: invalid_request_error
+    message: >-
+      'required' is required to be supplied and to be an array including every
+      key in properties. Missing 'limit'.
+  router_log:
+    path: C:/Users/djlov/AppData/Roaming/ciel-runtime/router-instances/9491-60524a4bf5eb/router.log
+    first_status: POST /v1/responses HTTP/1.1 400
+
+confirmed_cause:
+  - >-
+    Ciel's Meta Responses normalization preserved Codex function schemas without
+    converting optional properties to Meta's required-all-properties contract.
+  - >-
+    The upstream error names the omitted property limit and rejects the request
+    before model inference.
+  - >-
+    A TRACE capture of the actual Codex 0.152.1 request identified the named
+    limit property at tools[12], type tool_search, parameters.limit. The first
+    implementation normalized function tools only and therefore did not cover
+    this hosted tool schema.
+  - >-
+    After the required-all schema fix, the actual Codex request reached Meta's
+    next validator and was rejected because tools[7] was Codex's custom
+    apply_patch tool; Meta returned `custom` tools are not supported.
+  - >-
+    After custom-tool projection, an actual Codex run showed Meta serializing
+    exec_command.max_output_tokens and yield_time_ms as 1000.0 and 10000.0.
+    Codex 0.152.1 rejected those values as floats while deserializing usize.
+
+live_contract_probes:
+  endpoint: https://api.meta.ai/v1/responses via local router 9491
+  model: muse-spark-1.3-contributor
+  all_required_nonnullable:
+    schema: required [query, limit], additionalProperties false, strict true
+    http_status: 200
+    output_arguments: {query: status, limit: 1}
+  all_required_nullable_optional:
+    schema: required [query, limit], limit type [integer, 'null']
+    http_status: 200
+    output_arguments: {query: status, limit: null}
+  conclusion: >-
+    Meta accepts a required-all-properties schema and accepts null as the
+    representation of an originally optional property.
+
+official_documentation:
+  url: https://dev.meta.ai/docs/api-reference/responses/create-response
+  access_result: login_required
+  note: >-
+    The public fetch cannot read the authenticated page. No undocumented field
+    is inferred: the implementation follows the exact upstream validation error
+    and the two successful live API probes above.
+
+implementation:
+  files:
+    - ciel_runtime_support/providers/meta.py
+    - ciel_runtime_support/provider_responses_passthrough.py
+    - ciel_runtime_support/responses_custom_tool_bridge.py
+    - tests/test_meta_provider.py
+  behavior:
+    - Apply to top-level parameters schemas on Meta's OpenAI Responses tools.
+    - Preserve hosted-tool identity and fields while normalizing its parameters schema.
+    - Attach strict true only to function tools, not hosted tools.
+    - Recursively include every object property in required.
+    - Set additionalProperties false and strict true.
+    - Make formerly optional properties nullable so their optional meaning is retained.
+    - Deep-copy tool declarations so the client request is not mutated.
+    - Project Codex custom tools to strict Meta function tools with one raw input string.
+    - Restore function calls and streamed argument events to Codex custom-tool events.
+    - Project replayed custom tool call/output items to Meta function call/output items.
+    - Normalize whole-number JSON floats in Meta tool arguments to integer JSON values.
+    - Remove stale Content-Length whenever a response body is projected.
+
+verification:
+  focused_tests:
+    command: python -m unittest discover -s tests -p test_meta_provider.py -q
+    tests: 39
+    result: passed
+  lint:
+    command: python -m ruff check on changed Python source and tests
+    result: passed
+  actual_codex:
+    version: 0.152.1
+    router_port: 9605
+    workspace: G:/OneCielDMSUI
+    model: muse-spark-1.3-contributor
+    session: 01a065bf-c640-7981-a8b1-f78a02a1f551
+    actions:
+      - apply_patch created .ciel-runtime-muse-probe.txt
+      - exec_command read the created file in 138ms
+      - final response reported MUSE_CUSTOM_TOOL_OK followed by a newline
+    exit_code: 0
+    cleanup: diagnostic file deleted
+  full_suite:
+    command: npm test
+    passed: 3125
+    groups:
+      - {name: unit, passed: 1366, skipped: 45}
+      - {name: router, passed: 1093, skipped: 0}
+      - {name: channel, passed: 403, skipped: 80}
+      - {name: runtime, passed: 263, skipped: 12}
+    result: passed
+  local_deployment:
+    command: powershell -NoProfile -ExecutionPolicy Bypass -File ./install.ps1
+    installed_version: 0.2.37
+    destination: C:/Users/djlov/.local/share/ciel-runtime
+    changed_source_hashes_match: true
+
+credentials_recorded: false

package/docs/journal/2026/09/03/diagnostics/providers/ollama-cloud/kimi-k3-resumed-session-stall.okf

@@ -0,0 +1,160 @@
+task:
+  id: ollama-cloud-kimi-k3-resumed-session-stall-20260903
+  request: >-
+    Diagnose and improve the Cielarvis Codex session that repeatedly stops after
+    a Kimi K3 progress announcement even though it includes the earlier recovery.
+  workspace: C:/Users/djlov/cielarvis
+  runtime_workspace_id: a3a719fc6552
+  router_instance: 9479-a3a719fc6552
+  codex_session: 01a01dff-f0f0-7ef3-8cde-1d9916510f35
+  status: verified_in_isolated_router
+
+scope:
+  included:
+    - ollama-cloud/kimi-k3 routed Codex traffic on port 9479
+    - Codex rollout JSONL for the identified Cielarvis session
+    - Ciel Kimi continuation recovery and Ollama tool argument projection
+  excluded:
+    - Claude Opus sessions
+    - unrelated Cielarvis application behavior
+
+observed_failure:
+  local_window: 2026-09-02T23:37:47-05:00 through 2026-09-02T23:38:16-05:00
+  rollout_evidence:
+    path: C:/Users/djlov/.codex/sessions/2026/08/20/rollout-2026-08-20T02-08-27-01a01dff-f0f0-7ef3-8cde-1d9916510f35.jsonl
+    sequence:
+      - line: 6588
+        action: view_image
+      - line: 6590
+        result: tool output returned
+      - line: 6595
+        assistant_text: >-
+          현재 notice 영역을 정확히 보겠습니다 — 채팅 창 하단 상태바 부분을
+          새 캡처로 읽습니다.
+        followup_tool_calls: 0
+      - line: 6597
+        event: task_complete
+  usage_evidence:
+    last_requests:
+      - {input_tokens: 679087, cached_input_tokens: 0}
+      - {input_tokens: 682635, cached_input_tokens: 0}
+      - {input_tokens: 682885, cached_input_tokens: 0}
+      - {input_tokens: 684307, cached_input_tokens: 0}
+      - {input_tokens: 684564, cached_input_tokens: 0}
+    note: >-
+      These values prove request size and reported cache usage. They do not by
+      themselves prove that context size or cache behavior caused the stall.
+
+router_evidence:
+  path: C:/Users/djlov/AppData/Roaming/ciel-runtime/router-instances/9479-a3a719fc6552/router.log
+  loaded_recovery_confirmed:
+    - "codex_turn_retry provider=ollama-cloud reason=preamble_only"
+    - "codex_turn_retry_result provider=ollama-cloud attempt=1/3 retryable=true chars=34"
+    - "codex_turn_retry_result provider=ollama-cloud attempt=2/3 retryable=true chars=17"
+    - "codex_turn_retry_result provider=ollama-cloud attempt=3/3 retryable=true chars=52"
+  conclusion: >-
+    The earlier Kimi patch was loaded and active. Its three-attempt bound was
+    exhausted by three more progress-only replies, after which Codex received
+    the original announcement and recorded task_complete.
+  malformed_tool_path:
+    - at: 2026-09-02T23:37:13-05:00
+      raw_name: exec
+      matched_name: exec_command
+      missing: cmd
+      emitted_key: js
+      disposition: dropped
+    - at: 2026-09-02T23:37:25-05:00
+      raw_name: exec
+      matched_name: exec_command
+      missing: cmd
+      emitted_key: input
+      disposition: dropped
+    source: C:/Users/djlov/AppData/Roaming/ciel-runtime/router-instances/9479-a3a719fc6552/tool-calls.jsonl
+
+confirmed_causes:
+  - >-
+    The fixed three-attempt Kimi continuation limit is insufficient for the
+    captured turn: all three retries were classified retryable and no tool call
+    survived projection.
+  - >-
+    exec name canonicalization occurred before argument validation, but the
+    generic input alias was not converted to exec_command.cmd, so a non-empty
+    model-supplied command was discarded.
+  - >-
+    Each failed retry was appended to the next retry body. The router token
+    estimates increased from 643708 to 643788 to 643865 during that sequence.
+    This proves cumulative request growth; it does not prove model causation.
+
+not_claimed:
+  - The 1M Kimi context declaration is not itself a guarantee of equal behavior at every context length.
+  - The observed 684K-token context is not claimed as the cause of the stop.
+  - Cache misses are not claimed as the cause of the stop.
+  - JavaScript intended for a Node REPL is not assumed to be a valid shell command.
+
+official_reference:
+  ollama_model_card:
+    url: https://ollama.com/library/kimi-k3
+    confirmed: Kimi K3 is marked tools, thinking, cloud, and 1M context.
+  ollama_tool_calling:
+    url: https://docs.ollama.com/capabilities/tool-calling
+    confirmed: >-
+      The documented agent loop preserves assistant tool-call messages and tool
+      results across subsequent model calls.
+  ollama_responses_tool_choice:
+    url: https://github.com/ollama/ollama/blob/main/openai/responses.go
+    confirmed: >-
+      Ollama's Responses conversion states that tool_choice required is not
+      generally supported, so the recovery cannot rely on forcing tool_choice.
+
+implementation:
+  files:
+    - ciel_runtime_support/codex_turn_recovery.py
+    - ciel_runtime_support/pseudo_tool_parser.py
+    - tests/test_codex_turn_recovery.py
+    - tests/test_tool_name_canonicalization.py
+  behavior:
+    - Increase Kimi progress-only recovery from three to six bounded attempts.
+    - Use a stricter continuation instruction after the third failed reply.
+    - Keep the original request as a stable base and replay only the latest stalled assistant response.
+    - Repair non-empty command, script, or input aliases to exec_command.cmd.
+    - Do not reinterpret a js argument as a shell command.
+
+verification:
+  focused_tests:
+    command: py -3.14 -m unittest tests.test_codex_turn_recovery tests.test_tool_name_canonicalization -v
+    tests: 47
+    result: passed
+  live_isolated_router:
+    port: 9599
+    source_fingerprint: 2486fa4d14d56174
+    provider: ollama-cloud
+    model: kimi-k3
+    simple_tool_probe:
+      http_status: 200
+      responses_status: completed
+      output: function_call inspect_workspace {scope: all}
+      elapsed_seconds: 2.7
+      result: passed
+  lint:
+    command: python -m ruff check on changed Python source and tests
+    result: passed
+  full_suite:
+    command: npm test
+    passed: 3125
+    groups:
+      - {name: unit, passed: 1366, skipped: 45}
+      - {name: router, passed: 1093, skipped: 0}
+      - {name: channel, passed: 403, skipped: 80}
+      - {name: runtime, passed: 263, skipped: 12}
+    result: passed
+  limitation: >-
+    A live Kimi response did not deterministically reproduce four consecutive
+    preamble-only replies, so the exact fourth-retry branch is covered by unit
+    tests rather than claimed as a live reproduction.
+  local_deployment:
+    command: powershell -NoProfile -ExecutionPolicy Bypass -File ./install.ps1
+    installed_version: 0.2.37
+    destination: C:/Users/djlov/.local/share/ciel-runtime
+    changed_source_hashes_match: true
+
+credentials_recorded: false

package/docs/journal/2026/09/03/diagnostics/providers/ollama-cloud/kimi-k3-substantive-dangling-action.okf

@@ -0,0 +1,94 @@
+task:
+  id: ollama-cloud-kimi-k3-substantive-dangling-action-20260903
+  request: >-
+    Diagnose why the active Cielarvis Kimi K3 Codex session still terminates
+    before performing the next announced action.
+  workspace: C:/Users/djlov/cielarvis
+  runtime_workspace_id: a3a719fc6552
+  router_instance: 9599-a3a719fc6552
+  codex_session: 01a01dff-f0f0-7ef3-8cde-1d9916510f35
+  status: verified
+
+observed_runtime:
+  health:
+    version: 0.2.37
+    source_fingerprint: 2486fa4d14d56174
+    pid: 105292
+    installed_source: C:/Users/djlov/.local/share/ciel-runtime/ciel_runtime.py
+  installed_recovery_sha256: 1EF46C9E4D94B94EBE595B2F8C55175A8C54CE06DA4C1042CA17922B5D6D8587
+  repository_recovery_sha256: 1EF46C9E4D94B94EBE595B2F8C55175A8C54CE06DA4C1042CA17922B5D6D8587
+  conclusion: The previous Kimi recovery patch was loaded in the active router.
+
+observed_failure:
+  rollout_path: C:/Users/djlov/.codex/sessions/2026/08/20/rollout-2026-08-20T02-08-27-01a01dff-f0f0-7ef3-8cde-1d9916510f35.jsonl
+  turn_id: 01a065d3-2083-7021-898d-ead2b90238d3
+  sequence:
+    - {ordinal: 6615, type: function_call, name: view_image}
+    - {ordinal: 6617, type: function_call_output}
+    - ordinal: 6622
+      type: assistant_message
+      normalized_characters: 305
+      ending: 먼저 네이티브 쪽에 마이크 모듈이 있는지 확인합니다.
+      next_tool_calls: 0
+    - {ordinal: 6624, type: task_complete}
+  router_log:
+    response_started_at: 2026-09-03T00:54:02-05:00
+    response_completed_at: 2026-09-03T00:54:17-05:00
+    http_status: 200
+    recovery_log_present: false
+
+confirmed_cause:
+  - >-
+    The existing Kimi follow-up expression recognizes Korean future endings
+    such as 확인하겠습니다 but not a sequenced present-form action such as
+    먼저 ... 확인합니다.
+  - >-
+    The generic preamble policy rejects this 305-character response because it
+    contains a substantive diagnosis after a completed tool result, so the
+    Kimi-specific detector must independently preserve an explicit final
+    sequenced action.
+
+implementation:
+  files:
+    - ciel_runtime_support/codex_turn_recovery.py
+    - tests/test_codex_turn_recovery.py
+  behavior:
+    - Detect a final Korean sequenced action beginning with 먼저, 이제, 다음으로, 계속, 추가로, or 바로.
+    - Limit the rule to Kimi messages that include reasoning and no tool call.
+    - Apply the same detector to every bounded continuation result.
+
+verification:
+  focused_tests:
+    command: python -m unittest discover -s tests -p test_codex_turn_recovery.py -q
+    passed: 40
+  lint:
+    command: python -m ruff check ciel_runtime_support/codex_turn_recovery.py tests/test_codex_turn_recovery.py
+    result: passed
+  live_isolated_cielarvis_router:
+    port: 9607
+    provider: ollama-cloud
+    model: kimi-k3
+    initial_visible_characters: 438
+    initial_ending: 먼저 현재 작업공간을 확인합니다.
+    router_log: codex_turn_retry provider=ollama-cloud reason=promised_followup chars=438
+    continuation_output: function_call inspect_workspace {scope: all}
+    http_status: 200
+    responses_status: completed
+  full_suite:
+    command: npm test
+    passed: 3127
+    groups:
+      - {name: unit, passed: 1366, skipped: 45}
+      - {name: router, passed: 1093, skipped: 0}
+      - {name: channel, passed: 403, skipped: 80}
+      - {name: runtime, passed: 265, skipped: 12}
+    result: passed
+  local_deployment:
+    command: powershell -NoProfile -ExecutionPolicy Bypass -File ./install.ps1
+    destination: C:/Users/djlov/.local/share/ciel-runtime
+    installed_version: 0.2.37
+    source_sha256: C74699E14121542C214E3CB9CC61D793964826C6B9D103B0E359F23B20D6FF97
+    installed_sha256: C74699E14121542C214E3CB9CC61D793964826C6B9D103B0E359F23B20D6FF97
+    hashes_match: true
+
+credentials_recorded: false