@oneciel-ai/ciel-runtime 0.2.37 → 0.2.38

package/docs/journal/2026/09/02/diagnostics/remote/mia/socket-tui-visibility.okf

@@ -0,0 +1,47 @@
+task:
+  id: mia-socket-tui-visibility-20260902
+  request: Determine whether Mia hides socket-delivered AI Net messages because of an old runtime version.
+  status: diagnosed
+
+evidence:
+  remote:
+    host: mia@100.95.132.58
+    hostname: aap-pool-hera
+    tmux_session: aap-agent-86
+  versions:
+    installed_ciel: 0.2.37-nightly.20260902-033557.60846f9
+    npm_nightly_tag: 0.2.37-nightly.20260902-033557.60846f9
+    active_cli: codex-cli 0.152.1
+    installed_claude_npm_package: 2.1.245
+    installed_claude_native_binary: unavailable
+  active_process:
+    ciel_pid: 220324
+    cli_pid: 220566
+    cli_kind: codex
+    codex_router_flag: supports_websockets=false
+    claude_messaging_socket_active: false
+  observed_ai_net_delivery:
+    runtime_input_id: 1241
+    event_id: 1788324452627-0
+    event_type: net.ai-net.room.mentioned
+    event_time_utc: 2026-09-02T04:47:32.627Z
+    receiver: default
+    transport: sse
+    terminal_evidence: >-
+      tmux capture-pane contains the exact event between
+      '[ciel-runtime untrusted external event receiver=default transport=sse]'
+      and '[ciel-runtime end external event]'.
+  local_claude_control:
+    version: 2.1.258
+    frame: Ciel ClaudeSessionSocketClient user frame without from
+    result: >-
+      The live TUI rendered a Held peer message notice and the exact probe body;
+      socket input was not inherently invisible.
+
+conclusion:
+  version_mismatch_caused_mia_visibility: false
+  details: >-
+    Mia runs the current Ciel nightly and its active session is Codex, so Claude
+    cross-session socket rendering is not involved. AI Net SSE input has been
+    observed in the active Codex pane. A separate AI Net listener owned by
+    kevin-codex launches 'claude -p --continue' out of band and is not Mia's TUI.

package/docs/journal/2026/09/02/diagnostics/runtime/alibaba/token-plan/qwen38-latency.okf

@@ -0,0 +1,108 @@
+task:
+  id: cielarvis-qwen38-token-plan-latency-20260902
+  request: >-
+    Diagnose whether the nearly stalled Qwen3.8-Max run in C:/Users/djlov/cielarvis
+    is caused by Ciel parameter projection or by the Singapore service.
+  status: diagnosed
+  mutations_to_live_session: none
+
+runtime_identity:
+  workspace: C:/Users/djlov/cielarvis
+  workspace_id: a3a719fc6552
+  router: http://127.0.0.1:9469
+  router_pid: 93732
+  codex_pid: 96776
+  provider: alitoken
+  upstream_base_url: https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1
+  model: qwen3.8-max
+  protocol: openai_responses
+  note: >-
+    The configured endpoint is Alibaba Token Plan, not the separately isolated
+    Coding Plan endpoint at https://coding-intl.dashscope.aliyuncs.com/v1.
+
+confirmed_findings:
+  process_health:
+    router_responding: true
+    codex_responding: true
+    router_working_set_mb: 125.3
+    codex_working_set_mb: 264.9
+    qwen_request_errors_after_launch: 0
+    completed_requests_returned_http_200: true
+  effort_projection:
+    workspace_config_effort: xhigh
+    codex_turn_effort: high
+    launch_catalog_default: xhigh
+    launch_catalog_supported: [low, medium, xhigh]
+    live_router_started_at: 2026-09-02T10:55:03-05:00
+    updated_runtime_installed_at: 2026-09-02T11:02:29-05:00
+    live_projection: high_to_xhigh
+    evidence: >-
+      The launch-time implementation normalized Qwen3.8 Responses effort with
+      _normalize_qwen38_effort, whose documented mapping sends high to xhigh.
+      The already-running Python process predates the later local deployment.
+  live_workload:
+    resumed_session_file_mb: 68.06
+    latest_input_tokens: 99803
+    latest_cache_read_tokens: 8192
+    latest_uncached_input_tokens: 91611
+    wire_payload_before_compaction_bytes: 20681971
+    wire_payload_after_compaction_bytes: 8409549
+    note: >-
+      The resumed session repeatedly sends a large context. Cache reuse varies;
+      the latest observed call had only 8,192 cached tokens.
+  observed_upstream_phase:
+    matched_requests: 47
+    average_seconds: 64.7
+    minimum_seconds: 16
+    maximum_seconds: 197
+    examples:
+      - {seconds: 197, input_tokens: 182055, output_tokens: 6587, reasoning_tokens: 5888}
+      - {seconds: 180, input_tokens: 88987, output_tokens: 6658, reasoning_tokens: 6222}
+      - {seconds: 20, input_tokens: 93151, output_tokens: 86, reasoning_tokens: 17}
+      - {seconds: 16, input_tokens: 92404, output_tokens: 76, reasoning_tokens: 7}
+    interpretation: >-
+      Longer completed calls coincide with thousands of reasoning tokens. Even
+      short-output calls retain a 16-20 second floor at roughly 92K input tokens.
+  controlled_same_endpoint_probe:
+    prompt: Reply with exactly OK.
+    max_output_tokens: 32
+    results:
+      none: {seconds: 3.07, status: completed, input_tokens: 212, output_tokens: 2, reasoning_tokens: 0}
+      low: {seconds: 2.30, status: completed, input_tokens: 236, output_tokens: 23, reasoning_tokens: 18}
+      high: {seconds: 2.35, status: completed, input_tokens: 248, output_tokens: 29, reasoning_tokens: 25}
+      xhigh: {seconds: 2.32, status: completed, input_tokens: 248, output_tokens: 24, reasoning_tokens: 21}
+    conclusion: >-
+      The Singapore Token Plan endpoint was responsive for small requests during
+      the incident; the measurements do not support a general endpoint outage.
+
+official_evidence:
+  token_plan_quick_start:
+    url: https://docs.modelstudio.console.alibabacloud.com/en/model-studio/token-plan-personal-quick-start
+    base_url: https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1
+  responses_api:
+    url: https://www.alibabacloud.com/help/en/model-studio/qwen-api-via-openai-responses
+    default_effort: xhigh
+    supported_efforts: [none, minimal, low, medium, high, xhigh, max]
+    documented_effect: Lower effort accelerates responses and reduces inference tokens.
+  chat_api:
+    url: https://www.alibabacloud.com/help/en/model-studio/qwen-api-via-openai-chat-completions
+    high_mapping: xhigh
+    xhigh_thinking_budget: 262144
+
+diagnosis:
+  general_singapore_service_outage_supported: false
+  parameter_contribution_supported: true
+  workload_contribution_supported: true
+  conclusion: >-
+    The observed delay is dominated by the old resumed session's large, often
+    mostly uncached context plus high being projected to xhigh reasoning. Alibaba
+    performs the slow upstream phase, but the endpoint itself is not generally
+    slow for small requests. Current logs do not expose queue time versus model
+    inference time, so that finer split remains unmeasured.
+
+evidence_locations:
+  config: C:/Users/djlov/AppData/Roaming/ciel-runtime/workspaces/a3a719fc6552/config.json
+  router_log: C:/Users/djlov/AppData/Roaming/ciel-runtime/router-instances/9469-a3a719fc6552/router.log
+  router_activity: C:/Users/djlov/AppData/Roaming/ciel-runtime/router-instances/9469-a3a719fc6552/router-activity.json
+  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/runtime/alibaba/token-plan/qwen38-vs-gpt56-latency.okf

@@ -0,0 +1,92 @@
+task:
+  id: cielarvis-qwen38-vs-gpt56-latency-20260902
+  request: >-
+    Compare Qwen3.8-Max Singapore Token Plan latency with GPT-5.6-sol in the
+    same C:/Users/djlov/cielarvis Codex session and determine whether the
+    observed performance is abnormally slower despite the 1M context capacity.
+  status: diagnosed
+
+comparison_scope:
+  workspace: C:/Users/djlov/cielarvis
+  codex_session: 01a01dff-f0f0-7ef3-8cde-1d9916510f35
+  date: 2026-09-02
+  comparable_input_range_tokens: [60000, 110000]
+  qwen_sample_count: 47
+  gpt_sample_count: 55
+  credentials_recorded: false
+
+measurements:
+  qwen3_8_max:
+    provider: alitoken
+    median_seconds: 28
+    average_seconds: 56.5
+    p90_seconds: 120
+    average_input_tokens: 81547
+    average_cached_input_tokens: 23271
+    average_uncached_input_tokens: 58276
+    cache_hit_percent: 28.5
+    average_output_tokens: 1530
+    average_reasoning_tokens: 1302
+    timing_source: >-
+      Exact router interval from provider_responses_wire_compact completion to
+      the matching POST /v1/responses HTTP 200 completion.
+  gpt_5_6_sol:
+    provider: codex
+    median_seconds: 3.0
+    average_seconds: 3.7
+    p90_seconds: 6.1
+    average_input_tokens: 83959
+    average_cached_input_tokens: 81310
+    average_uncached_input_tokens: 2649
+    cache_hit_percent: 96.8
+    average_output_tokens: 374
+    average_reasoning_tokens: 158
+    timing_source: >-
+      Inferred request interval from the preceding Codex token-count boundary
+      to the matching POST /backend-api/codex/responses HTTP 200 completion.
+
+ratios:
+  qwen_to_gpt_median_latency: 9.3
+  qwen_to_gpt_average_latency: 15.3
+  qwen_to_gpt_reasoning_tokens: 8.2
+  qwen_to_gpt_output_tokens: 4.1
+  qwen_to_gpt_uncached_input_tokens: 22.0
+
+additional_checks:
+  qwen_small_prompt_same_endpoint:
+    input_tokens_range: [212, 248]
+    response_seconds_range: [2.30, 3.07]
+    result: completed
+  parallel_alitoken_session:
+    workspace: G:/OneCielDMSUI
+    last_activity: 2026-09-02T10:57:24-05:00
+    sustained_overlap_with_comparison_window: false
+  qwen_errors:
+    http_429: 0
+    timeout: 0
+    completed_status: http_200
+
+official_evidence:
+  qwen_responses_api:
+    url: https://www.alibabacloud.com/help/en/model-studio/qwen-api-via-openai-responses
+    fact: >-
+      A 1M context value is an accepted-capacity limit. The documentation says
+      lower reasoning effort accelerates responses and reduces inference tokens.
+  token_plan_faq:
+    url: https://docs.modelstudio.console.alibabacloud.com/en/model-studio/token-plan-personal-faq
+    facts:
+      - Peak-hour queuing can occur for Personal Edition.
+      - Exact TPM, TPS, and RPM thresholds are dynamically adjusted and not public.
+      - Trimming context is the documented performance optimization.
+
+diagnosis:
+  objectively_slower_than_gpt_5_6_sol: true
+  one_million_context_implies_low_latency: false
+  confirmed_contributors:
+    - Qwen reused only 28.5 percent of comparable input versus GPT's 96.8 percent.
+    - Qwen generated 8.2 times more reasoning tokens per comparable request.
+    - The running pre-update router maps the Codex high effort to Qwen xhigh.
+  unresolved_split: >-
+    Current router logs measure the total Alibaba upstream phase but do not
+    expose time-to-first-token, provider queue time, and model inference time
+    separately. Peak-hour queue contribution therefore remains unmeasured.

package/docs/journal/2026/09/02/implementation/anthropic/context/one-million-defaults.okf

@@ -0,0 +1,85 @@
+task:
+  id: anthropic-context-one-million-defaults-20260902
+  request: >-
+    Correct Claude model defaults so routed Claude Code sessions use the 1M
+    context variant only when the Claude-facing model id carries [1m].
+  status: verified
+
+official_evidence:
+  model_overview:
+    url: https://platform.claude.com/docs/en/models/overview
+    observed:
+      - Claude Fable 5.1, Fable 5, Opus 5, and Sonnet 5 list 1M context.
+      - Claude Haiku 4.5 lists 200K context.
+  claude_code_model_configuration:
+    url: https://code.claude.com/docs/en/model-config
+    observed:
+      - Gateway deployments use the [1m] model suffix to select extended context.
+      - The suffix can be appended to aliases or full model names.
+
+local_evidence:
+  before:
+    claude_cli_version: 2.1.258
+    target_workspace: C:/Users/djlov/cielarvis/mind-core-sand-box
+    configured_model: claude-opus-5
+    launched_model: ciel-runtime-anthropic-claude-opus-5
+    displayed_context_tokens: 200000
+    normalization_probe:
+      input: claude-opus-5[1m]
+      output: claude-opus-5
+  after:
+    migrated_model: claude-opus-5[1m]
+    normalized_model: claude-opus-5[1m]
+    upstream_wire_model: claude-opus-5
+    claude_facing_model: ciel-runtime-anthropic-claude-opus-5[1m]
+    max_context_environment: 1048576
+    auto_compact_environment: 1000000
+    claude_tui_header: Opus 5 (1M context) with xhigh effort
+
+implementation:
+  identity:
+    - Preserve [1m] in Anthropic configuration and Claude-facing aliases.
+    - Remove [1m] only at the upstream Anthropic API boundary.
+  defaults:
+    current_model: claude-opus-5[1m]
+    one_million_variants:
+      - claude-fable-5-1[1m]
+      - claude-fable-5[1m]
+      - claude-opus-5[1m]
+      - claude-opus-4-8[1m]
+      - claude-sonnet-5[1m]
+      - claude-sonnet-4-6[1m]
+  migration:
+    marker: anthropic_default_1m_model_ids_20260902
+    behavior: >-
+      Add the 1M variants and migrate current models whose official default is
+      1M. Keep opt-in Sonnet 4.6 and Haiku 4.5 unchanged.
+
+verification:
+  focused:
+    total: 129
+    result: passed
+    suites:
+      - test_provider_contract_matrix.py
+      - test_anthropic_native_output_tokens.py
+      - test_codex_runtime.py
+  full:
+    command: npm test
+    unit: {total: 1358, skipped: 45}
+    router: {total: 1069, skipped: 0}
+    channel: {total: 403, skipped: 80}
+    runtime: {total: 258, skipped: 12}
+    result: passed
+  runtime_probe:
+    request_sent: false
+    result: Claude Code rendered Opus 5 (1M context) in the target workspace.
+  local_deployment:
+    command: .\\install.ps1
+    target: C:/Users/djlov/.local/share/ciel-runtime
+    version: 0.2.37
+    source_installed_hashes_match: true
+    installed_workspace_projection:
+      configured_model: claude-opus-5[1m]
+      claude_facing_model: ciel-runtime-anthropic-claude-opus-5[1m]
+      max_context_environment: 1048576
+      auto_compact_environment: 1000000

package/docs/journal/2026/09/02/implementation/cache/alibaba/qwen38-cache-hit-improvement.okf

@@ -0,0 +1,104 @@
+task:
+  id: qwen38-cache-hit-improvement-20260902
+  request: >-
+    Continue from cache research through implementation and verified
+    improvement for Codex using Alibaba Singapore Token Plan alitoken.
+  status: verified
+
+implementation:
+  stateless_cache_policy:
+    source: ciel_runtime_support/provider_responses_passthrough.py
+    behavior: >-
+      When an alitoken Responses request has no previous_response_id, remove
+      x-dashscope-session-cache from the final upstream headers. This leaves
+      qwen3.8-max on Alibaba's automatic implicit prefix cache. Preserve the
+      configured header when a linked previous_response_id is present.
+    provider_default:
+      responses_session_cache_requires_previous_response_id: true
+  prefix_stable_compaction:
+    source: ciel_runtime_support/prompt_compaction.py
+    behavior: >-
+      Round the omitted-history boundary forward to a deterministic 24-item
+      checkpoint, then advance only as needed to avoid orphaning a function
+      output from its call. Appended turns reuse the same summary until the
+      next checkpoint instead of rewriting item zero on every sliding step.
+    provider_default:
+      responses_cache_checkpoint_items: 24
+  scope:
+    enabled_provider: alitoken
+    other_provider_defaults_changed: false
+
+live_verification:
+  method: >-
+    Forked an existing inactive Codex history into a new session and ran it
+    through source-built Ciel Runtime on isolated port 9490. The production
+    router on port 9479 was not stopped or reconfigured.
+  workspace: C:/Users/djlov/ciel-runtime/.tmp/cache-hit-live-verify
+  session_id: 01a0637b-6c5c-70d3-a419-c10ed1854cc2
+  provider: alitoken
+  model: qwen3.8-max
+  protocol: openai_responses
+  before_stateless_policy:
+    header: {x-dashscope-session-cache: enable}
+    calls:
+      - {input_tokens: 115314, cached_input_tokens: 0, hit_percent: 0.000}
+      - {input_tokens: 115565, cached_input_tokens: 0, hit_percent: 0.000}
+    weighted_hit_percent: 0.000
+  after_stateless_policy:
+    header: deferred_missing_previous_response_id
+    calls:
+      - {phase: implicit_warmup, input_tokens: 115604, cached_input_tokens: 1024, hit_percent: 0.886}
+      - {phase: warmed, input_tokens: 115643, cached_input_tokens: 115072, hit_percent: 99.506}
+      - {phase: repeated_warmed, input_tokens: 115682, cached_input_tokens: 115072, hit_percent: 99.473}
+    warmed_calls_weighted_hit_percent: 99.489
+  upstream_evidence:
+    activity_file: C:/Users/djlov/AppData/Roaming/ciel-runtime/router-instances/9490-8b364f9629ac/router-activity.json
+    log_file: C:/Users/djlov/AppData/Roaming/ciel-runtime/router-instances/9490-8b364f9629ac/router.log
+    observed_log_marker: provider_responses_session_cache_deferred provider=alitoken reason=missing_previous_response_id
+
+prefix_verification:
+  input_change: 633_to_636_items
+  budget_tokens: 2240633
+  first_compacted_items: 274
+  second_compacted_items: 277
+  first_wire_bytes: 8220013
+  second_wire_bytes: 8310319
+  common_prefix_bytes: 8220011
+  common_prefix_percent_of_first: 99.999976
+  first_summary_equal: true
+
+automated_verification:
+  full_suite:
+    command: npm test
+    result: passed
+    groups:
+      - {name: unit, tests: 1361, skipped: 45}
+      - {name: router, tests: 1076, skipped: 0}
+      - {name: channel, tests: 403, skipped: 80}
+      - {name: runtime, tests: 258, skipped: 12}
+    total_tests: 3098
+  lint:
+    command: npm run lint
+    result: passed
+  screenshot:
+    path: C:/Users/djlov/ciel-runtime/docs/journal/2026/09/02/implementation/cache/alibaba/qwen38-cache-hit-improvement.png
+    visible_fact: isolated router 9490 reported upstream success for alitoken qwen3.8-max
+
+cleanup:
+  isolated_router_9490_reachable_after_stop: false
+  production_router_9479_reachable_after_cleanup: true
+  production_router_restarted: false
+  production_session_interrupted: false
+
+evidence_boundaries:
+  confirmed:
+    - Two warmed implicit-cache calls reported more than 99.47 percent cached input.
+    - The same isolated session's preceding two forced-session-cache calls reported zero cached input.
+    - Checkpointed compaction preserved 99.999976 percent of the first serialized wire in the representative oversized case.
+  not_claimed:
+    - Alibaba documents that implicit cache hits are not guaranteed, so every future request is not guaranteed to hit.
+    - The live session did not exceed the ten-MiB wire limit; it verifies the header policy but not live upstream behavior after checkpoint compaction.
+    - Server scheduling is not observable, so the A/B sequence is evidence of improvement, not proof that every difference has one exclusive cause.
+    - previous_response_id stateful continuation was not implemented in this change.
+
+credentials_recorded: false

package/docs/journal/2026/09/02/implementation/cache/alibaba/responses-session-cache.okf

@@ -0,0 +1,97 @@
+task:
+  id: alibaba-responses-session-cache-20260902
+  request: >-
+    Fix the Qwen3.8-Max Responses reasoning-effort projection bug, identify the
+    measured cache miss cause, and increase cache reuse for Alibaba Singapore
+    Token Plan Codex sessions.
+  status: implemented_and_live_verified
+
+confirmed_problem:
+  workspace: C:/Users/djlov/cielarvis
+  provider: alitoken
+  model: qwen3.8-max
+  protocol: openai_responses
+  old_live_router_port: 9469
+  observations:
+    - latest old-session input was approximately 100K tokens with only 8,192 cached tokens
+    - repeated request bodies were compacted below the service's 10 MiB wire limit
+    - exact reconstructed consecutive compacted inputs retained a 7,966,060-byte common prefix
+    - Ciel did not send x-dashscope-session-cache before this change
+  correction: >-
+    The initial hypothesis that deterministic wire compaction changed the first
+    input item on every request was contradicted by exact reconstruction of the
+    latest consecutive inputs. The compressed prefix was stable in that range.
+    The missing session-cache opt-in header is the confirmed persistent cause.
+
+official_contract:
+  responses_api:
+    url: https://www.alibabacloud.com/help/en/model-studio/compatibility-with-openai-responses-api
+    header: x-dashscope-session-cache
+    enabled_value: enable
+    default_value: disable
+  context_cache:
+    url: https://docs.modelstudio.console.alibabacloud.com/en/model-studio/context-cache
+    prefix_requirement_tokens: 1024
+    placement_rule: repeated content first and unique content last
+
+implementation:
+  provider_defaults:
+    providers: [alims-intl, alitoken, alitoken-individual]
+    protocol: openai_responses
+    header: {x-dashscope-session-cache: enable}
+  generic_header_projection:
+    source: ciel_runtime_support/provider_request_access.py
+    safety_exclusions:
+      - credentials and API keys
+      - cookies
+      - host and content-length
+      - hop-by-hop headers
+      - x-ciel-runtime internal headers
+  effort_fix:
+    source: ciel_runtime_support/providers/alibaba.py
+    responses_values_preserved: [none, minimal, low, medium, high, xhigh, max]
+    chat_mapping_unchanged: true
+  evidence_capture:
+    source: ciel_runtime_support/provider_responses_passthrough.py
+    behavior: upstream dump metadata records sanitized final header names and values
+
+verification:
+  focused_unit_tests:
+    result: 83_passed
+    suites:
+      - {file: tests/test_provider_request_access.py, tests: 12}
+      - {file: tests/test_responses_prompt_compaction.py, tests: 8}
+      - {file: tests/test_alibaba_provider.py, tests: 42}
+      - {file: tests/test_meta_provider.py, tests: 21}
+  isolated_live_router:
+    workspace: C:/Users/djlov/cielarvis
+    port: 9479
+    production_router_modified: false
+    stopped_after_test: true
+    stopped_port_state: closed
+  upstream_wire:
+    request_1_body_bytes: 225933
+    request_2_body_bytes: 226262
+    endpoint: https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1/responses
+    session_cache_header_request_1: enable
+    session_cache_header_request_2: enable
+  provider_reported_second_request_usage:
+    input_tokens: 53934
+    cached_tokens: 53895
+    uncached_tokens: 39
+    cache_hit_percent: 99.9277
+  full_project_checks:
+    compileall: passed
+    unit: {tests: 1358, skipped: 45, result: passed}
+    router: {tests: 1074, result: passed}
+    channel: {tests: 403, skipped: 80, result: passed}
+    runtime: {tests: 258, skipped: 12, result: passed}
+    ruff: passed
+    documentation_metadata: passed
+    npm_pack_dry_run: passed
+  credentials_recorded: false
+
+evidence_locations:
+  sanitized_upstream_metadata: C:/Users/djlov/AppData/Roaming/ciel-runtime/verification/qwen-cache-20260902
+  isolated_router_activity: C:/Users/djlov/AppData/Roaming/ciel-runtime/router-instances/9479-a3a719fc6552/router-activity.json
+  old_router_log: C:/Users/djlov/AppData/Roaming/ciel-runtime/router-instances/9469-a3a719fc6552/router.log

package/docs/journal/2026/09/02/implementation/providers/alibaba/qwen38/0902-parameter-alignment.okf

@@ -0,0 +1,97 @@
+task:
+  id: alibaba-qwen38-0902-parameter-alignment-20260902
+  request: >-
+    Verify the updated Qwen3.8-Max version across Singapore Token Plan and
+    other Alibaba providers, then align model parameters with official docs.
+  status: verified
+
+official_evidence:
+  model_page:
+    url: https://docs.modelstudio.console.alibabacloud.com/en/model-studio/qwen3-8-max
+    context_window: 1000000
+    max_input: 991808
+    thinking_max_input: 983616
+    max_output: 131072
+    max_reasoning: 262144
+  responses_api:
+    url: https://www.alibabacloud.com/help/en/model-studio/qwen-api-via-openai-responses
+    efforts: [none, minimal, low, medium, high, xhigh, max]
+    default_effort: xhigh
+    singapore_supports: [xhigh, max]
+    conflict_rule: reasoning.effort takes precedence over enable_thinking
+  chat_api:
+    url: https://www.alibabacloud.com/help/en/model-studio/qwen-api-via-openai-chat-completions
+    native_efforts: [low, medium, xhigh]
+    mappings: {minimal: low, high: xhigh, max: xhigh, none: enable_thinking_false}
+    mutual_exclusion: reasoning_effort and thinking_budget cannot both be sent
+  snapshot:
+    documented_id: qwen3.8-max-0902
+    unsupported_claim_checked: qwen3.8-max-0920 was not found in official docs
+  token_plan:
+    url: https://docs.modelstudio.console.alibabacloud.com/en/model-studio/token-plan-personal-overview
+    supported_id: qwen3.8-max
+    preview_behavior: qwen3.8-max-preview routes to qwen3.8-max
+
+live_evidence:
+  provider: alitoken
+  region: ap-southeast-1
+  credentials_printed: false
+  model_catalog:
+    live_count: 13
+    qwen_ids: [qwen3.8-max, qwen3.8-max-preview]
+  model_id_probes:
+    qwen3.8-max: {http: 200, status: completed}
+    qwen3.8-max-preview: {http: 200, status: completed}
+    qwen3.8-max-0902: {http: 404, error: MODEL_NOT_FOUND}
+    qwen3.8-max-0920: {http: 404, error: MODEL_NOT_FOUND}
+  responses_effort_probes:
+    accepted: [none, minimal, low, medium, high, xhigh, max]
+    http_status: 200
+    none_reasoning_tokens: 0
+    note: >-
+      The 16-token probes for thinking efforts ended incomplete by design but
+      proved that the endpoint accepted each distinct effort value.
+
+implementation:
+  model_catalog:
+    alims_intl:
+      rolling_default: qwen3.8-max
+      selectable_snapshot: qwen3.8-max-0902
+    alitoken:
+      rolling_default: qwen3.8-max
+      snapshot_added: false
+      reason: live Token Plan returned MODEL_NOT_FOUND for snapshot IDs
+    alitoken_individual:
+      rolling_default: qwen3.8-max
+      snapshot_added: false
+    coding_plan:
+      changed: false
+      reason: official Coding Plan list does not include Qwen3.8-Max
+  parameter_projection:
+    responses: preserve all seven official effort levels
+    chat: retain documented three-level mapping
+    conflicting_thinking_budget_removed: true
+    codex_default_effort: xhigh
+    codex_supported_efforts: [none, minimal, low, medium, high, xhigh, max]
+  migration:
+    marker: alibaba_qwen38_0902_catalog_20260902
+    scope: add qwen3.8-max-0902 only to alims-intl catalogs
+
+verification:
+  focused_tests:
+    total: 151
+    result: passed
+  full_tests:
+    command: npm test
+    unit: {total: 1358, skipped: 45}
+    router: {total: 1071, skipped: 0}
+    channel: {total: 403, skipped: 80}
+    runtime: {total: 258, skipped: 12}
+    result: passed
+  lint:
+    command: npm run lint
+    result: passed
+  local_deployment:
+    command: .\\install.ps1
+    module: C:/Users/djlov/.local/share/ciel-runtime/ciel_runtime.py
+    installed_projection_verified: true