@oneciel-ai/ciel-runtime 0.2.37 → 0.2.38

package/docs/journal/2026/09/02/operations/release/main-merge-local-deploy.okf

@@ -0,0 +1,36 @@
+task:
+  id: main-merge-local-deploy-20260902
+  request:
+    - Merge the Claude session-socket nightly into main.
+    - Deploy the merged source locally.
+    - Return to the nightly branch.
+  status: complete
+
+source_control:
+  previous_main: 6ecdc61
+  merged_nightly: 04fd5f7
+  merge_mode: fast-forward
+  feature_commit: 60846f9
+
+verification:
+  full_test_suite:
+    command: npm test
+    exit_code: 0
+    groups:
+      unit: "1358 tests; OK; 45 skipped"
+      router: "1064 tests; OK"
+      channel: "403 tests; OK; 80 skipped"
+      runtime: "258 tests; OK; 12 skipped"
+  local_deployment:
+    command: .\\install.ps1
+    target: C:\\Users\\djlov\\.local\\share\\ciel-runtime
+    executable: C:\\Users\\djlov\\.local\\bin\\ciel-runtime.ps1
+    reported_version: 0.2.37
+    module_import: ok
+    windows_pipe_prefix_probe: true
+    source_module_sha256: F7774A5A0405A6016A152D78A81D5CFBD4EC75DF3122F1D4A05A921FCE5427E6
+    installed_module_sha256: F7774A5A0405A6016A152D78A81D5CFBD4EC75DF3122F1D4A05A921FCE5427E6
+
+result:
+  main_contains_claude_socket_feature: true
+  local_deployment_matches_source: true

package/docs/journal/2026/09/02/operations/release/nightly/local-cache-deployment.okf

@@ -0,0 +1,54 @@
+task:
+  id: local-nightly-cache-deployment-20260902
+  request:
+    - Deploy the verified source tree to the local Ciel Runtime installation.
+    - Commit the verified provider, context, transport, and cache corrections on nightly.
+    - Push nightly so the npm publication workflows can consume it.
+  recorded_phase: release_candidate_verified_before_push
+
+source_control:
+  branch: nightly
+  starting_head: 8f9d8b9
+  origin_nightly_before: 8f9d8b9
+  fetch_before_deployment: passed
+  force_push_authorized: false
+  intended_push: git push origin nightly
+
+verification_before_deployment:
+  full_test_suite:
+    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
+  live_alibaba_cache_probe:
+    provider: alitoken
+    input_tokens: 53934
+    cached_tokens: 53895
+    uncached_tokens: 39
+    cache_hit_percent: 99.9277
+
+local_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
+  installed_projection_probe:
+    responses_session_cache: enable
+    chat_session_cache: null
+    configured_effort: xhigh
+  source_installed_sha256_matches:
+    ciel_runtime.py: true
+    ciel_runtime_support/providers/alibaba.py: true
+    ciel_runtime_support/provider_request_access.py: true
+    ciel_runtime_support/provider_responses_passthrough.py: true
+
+runtime_boundary:
+  active_router_restarted: false
+  reason: >-
+    Local installation replacement is complete. Existing Python router
+    processes retain already-loaded modules until their next restart; the
+    active user session was not interrupted for release deployment.

package/docs/journal/2026/09/02/operations/release/nightly/ollama-kimi-recovery-model-audit-deployment.okf

@@ -0,0 +1,58 @@
+task:
+  id: ollama-kimi-recovery-model-audit-nightly-deployment-20260902
+  request: >-
+    Deploy all currently verified Ciel Runtime changes locally and publish them
+    through the nightly branch.
+  recorded_at: 2026-09-02T21:45:00-05:00
+  status: ready_for_nightly_push
+
+release_scope:
+  runtime_changes:
+    - Preserve complete assistant content, including thinking, during Kimi K3 continuation recovery.
+    - Apply a bounded three-attempt continuation loop for Ollama Cloud Kimi K3 progress-only turns.
+    - Keep compacted summary metadata byte-stable across small per-request budget changes.
+    - Add the documented GLM-5.3 low/high/max reasoning profile.
+    - Use boolean thinking for Ollama models without a documented string-level contract.
+  tests:
+    - tests/test_codex_turn_recovery.py
+    - tests/test_context_summary_compatibility_api.py
+    - tests/test_ollama_provider_options.py
+  evidence_journals:
+    - docs/journal/2026/09/02/diagnostics/providers/ollama-cloud/kimi-k3-agent-turn-recovery.okf
+    - docs/journal/2026/09/02/diagnostics/providers/ollama-cloud/kimi-k3-parameters.okf
+    - docs/journal/2026/09/02/diagnostics/cache/ollama-cloud/kimi-k3-codex-hit-rate.okf
+    - docs/journal/2026/09/02/research/providers/ollama-cloud/desktop-model-exhaustive-audit.okf
+    - docs/journal/2026/09/02/monitoring/cache/alibaba/cielarvis-qwen38-live-hit-rate.okf
+
+verification_before_push:
+  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
+  ollama_focused_suite: {tests: 76, result: passed}
+  ruff_changed_files: passed
+  ollama_cloud_live_tool_probe: {models: 19, http_200: 19, correct_tool_calls: 19}
+  okf_parse_validation: {files: 21, errors: 0, result: passed}
+  credential_pattern_scan: {matching_files: 0}
+
+local_deployment:
+  command: powershell -NoProfile -ExecutionPolicy Bypass -File .\install.ps1
+  target: C:/Users/djlov/.local/share/ciel-runtime
+  runtime_source_installed_hash_match: true
+  result: passed
+
+nightly_delivery:
+  branch: nightly
+  base_commit: d56c4f7
+  workflow: .github/workflows/npm-publish.yml
+  package_base_version: 0.2.37
+  trigger: push_to_nightly
+  commit: pending
+  github_workflow_run: pending
+  published_version: pending
+  registry_verification: pending
+
+credentials_recorded: false

package/docs/journal/2026/09/02/operations/release/nightly/qwen38-cache-improvement-deployment.okf

@@ -0,0 +1,72 @@
+okf: 1
+task:
+  date: "2026-09-02"
+  request: >-
+    Deploy the verified Alibaba Token Plan cache improvement locally and push
+    the verified source to the nightly branch.
+  workspace: C:/Users/djlov/ciel-runtime
+
+source_control_before_release:
+  branch: nightly
+  origin_divergence: {behind: 0, ahead: 0}
+  force_push_authorized: false
+  nightly_trigger: push_to_nightly
+  publish_workflow: .github/workflows/npm-publish.yml
+  nightly_version_rule: "0.2.37-nightly.<UTC timestamp>.<7-character commit sha>"
+
+verification_before_deployment:
+  full_suite:
+    command: npm test
+    result: passed
+    total_tests: 3098
+    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}
+  lint: {command: npm run lint, result: passed}
+  documentation: {command: npm run check:docs, result: passed}
+  package:
+    command: npm pack --dry-run --json
+    result: passed
+    version: 0.2.37
+    entry_count: 526
+    packed_size_bytes: 1739943
+    unpacked_size_bytes: 6004414
+  live_alibaba_probe:
+    provider: alitoken
+    model: qwen3.8-max
+    warmed_weighted_cache_hit_percent: 99.489
+    isolated_router_port: 9490
+    isolated_router_stopped_after_capture: true
+    evidence: >-
+      docs/journal/2026/09/02/implementation/cache/alibaba/qwen38-cache-hit-improvement.okf
+
+local_deployment:
+  command: powershell -NoProfile -ExecutionPolicy Bypass -File .\\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
+  installed_provider_projection:
+    model: qwen3.8-max
+    responses_session_cache_requires_previous_response_id: true
+    responses_cache_checkpoint_items: 24
+  source_installed_sha256_matches:
+    ciel_runtime.py: true
+    ciel_runtime_support/providers/alibaba.py: true
+    ciel_runtime_support/provider_responses_passthrough.py: true
+    ciel_runtime_support/prompt_compaction.py: true
+  installed_code_tests:
+    - {file: tests/test_alibaba_provider.py, tests: 42, result: passed}
+    - {file: tests/test_responses_prompt_compaction.py, tests: 11, result: passed}
+    - {file: tests/test_meta_provider.py, tests: 23, result: passed}
+
+runtime_boundary:
+  router_9479_before_install: unreachable
+  router_started_or_restarted_by_deployment: false
+  note: >-
+    The local source installation was replaced without launching a router.
+    A future normal runtime launch will load the installed files.
+
+credentials_recorded: false

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

@@ -0,0 +1,233 @@
+task:
+  id: qwen38-cache-hit-rate-improvement-research-20260902
+  request: >-
+    Research evidence-backed ways to increase cache reuse for Codex using
+    Alibaba Singapore Token Plan qwen3.8-max through Ciel Runtime.
+  scope: research_only
+  status: completed
+
+environment:
+  runtime_repository: C:/Users/djlov/ciel-runtime
+  observed_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
+  codex_session_id: 01a01dff-f0f0-7ef3-8cde-1d9916510f35
+  production_mutated: false
+
+measured_state:
+  pre_restart_last_14:
+    input_tokens: 1932589
+    cached_input_tokens: 614400
+    weighted_cache_hit_percent: 31.792
+  post_restart_latest_15:
+    measured_through_utc: 2026-09-02T18:35:53.849Z
+    input_tokens: 2281369
+    cached_input_tokens: 291885
+    weighted_cache_hit_percent: 12.794
+    zero_cache_calls: 13
+    high_cache_calls: 2
+  change_percentage_points: -18.998
+  accounting_cross_check: >-
+    Router activity cache_read_tokens and Codex last_token_usage.cached_input_tokens
+    report the same upstream values. The observed misses are not created by a
+    separate Ciel display-only calculation.
+
+wire_evidence:
+  request_headers:
+    configured: {x-dashscope-session-cache: enable}
+    projected_on_openai_responses: {x-dashscope-session-cache: enable}
+  captured_codex_request:
+    body_keys:
+      - client_metadata
+      - include
+      - input
+      - instructions
+      - model
+      - parallel_tool_calls
+      - prompt_cache_key
+      - reasoning
+      - store
+      - stream
+      - tool_choice
+      - tools
+    prompt_cache_key: 01a06337-93d7-7cb0-9fba-895f0daa2c44
+    store: false
+    previous_response_id_present: false
+    full_input_history_replayed: true
+  isolated_prior_probe:
+    condition: two compact stable-prefix requests with session-cache header enabled
+    second_input_tokens: 53934
+    second_cached_input_tokens: 53895
+    second_weighted_cache_hit_percent: 99.928
+    interpretation: >-
+      The header and upstream cache can produce a near-complete hit. This does
+      not prove that the header alone is sufficient for a long changing Codex
+      transcript.
+
+official_contracts:
+  alibaba_responses_session_cache:
+    url: https://www.alibabacloud.com/help/en/model-studio/compatibility-with-openai-responses-api
+    confirmed:
+      - Session cache is opted into with x-dashscope-session-cache=enable.
+      - qwen3.8-max uses explicit cache for Responses session cache.
+      - The documented multi-turn flow sends only new input and links it with previous_response_id.
+      - A response ID remains valid for seven days.
+    not_documented:
+      - Alibaba does not document prompt_cache_key as its Responses cache control.
+      - The page does not document Responses cache_control markers.
+  alibaba_context_cache:
+    url: https://www.alibabacloud.com/help/en/model-studio/context-cache
+    confirmed:
+      - Implicit caching is automatic for supported models but hits are not guaranteed.
+      - Cacheable repeated content must be a prefix and unique content should follow it.
+      - Explicit and implicit caching cannot be used together on one request.
+      - Explicit cache requires at least 1024 tokens, has a five-minute TTL reset on hit, and supports at most four markers.
+  qwen38_model:
+    url: https://docs.modelstudio.console.alibabacloud.com/en/model-studio/qwen3-8-max
+    confirmed:
+      - The model supports a one-million-token context and context caching.
+      - Singapore International input price is USD 2.00 per million tokens.
+      - Implicit cache read is USD 0.25 per million tokens, 12.5 percent of normal input.
+      - Explicit cache creation is USD 2.50 per million tokens, 125 percent of normal input.
+      - Explicit cache read is USD 0.17 per million tokens, 8.5 percent of normal input.
+  token_plan:
+    url: https://docs.modelstudio.console.alibabacloud.com/en/model-studio/token-plan-personal-faq
+    confirmed:
+      - Cache hits deduct fewer Token Plan Credits than cache misses.
+      - Usage Analysis is the official source of truth for actual Credits deduction.
+  openai_responses:
+    url: https://developers.openai.com/api/reference/cli/resources/responses/methods/create
+    confirmed:
+      - previous_response_id is the Responses multi-turn state link.
+      - Instructions from a prior response are not automatically carried into the next request.
+      - prompt_cache_key is documented as an OpenAI cache bucketing key.
+
+local_root_cause_evidence:
+  current_compactor:
+    source: C:/Users/djlov/ciel-runtime/ciel_runtime_support/prompt_compaction.py
+    behavior: >-
+      compact_responses_input_for_budget finds a newest tail that fits and
+      regenerates the first summary item from every omitted item on each call.
+    confirmed_consequence: >-
+      When tail_start changes, the summary input set changes. The first wire item
+      can therefore change even though the conversation only appended new items.
+  live_router_sequence:
+    examples:
+      - {items_before: 633, items_after: 400, omitted: 233}
+      - {items_before: 636, items_after: 403, omitted: 233}
+      - {items_before: 669, items_after: 436, omitted: 233}
+    conclusion: >-
+      During this segment the omitted count and summary input prefix remained
+      fixed while reported cache results still alternated between zero and
+      near-complete hits. Moving compaction boundaries are therefore a proven
+      cache-breaking mechanism, but not a complete explanation for every miss.
+  synthetic_prefix_measurement:
+    representative_input:
+      first_items: 633
+      next_items: 636
+      repeated_item_chars: 30000
+      compaction_budget_tokens: 2240633
+    current_compactor:
+      first_body_bytes: 8837551
+      next_body_bytes: 8837557
+      common_prefix_bytes: 19163
+      common_prefix_percent: 0.2168
+      first_summary_equal: false
+    deterministic_per_item_cap_prototype:
+      cap_chars_per_item: 12000
+      first_body_bytes: 7629029
+      next_body_bytes: 7665098
+      common_prefix_bytes: 7628972
+      common_prefix_percent: 99.9993
+    boundary: >-
+      The prototype only proves that prefix-stable reduction is technically
+      possible. It does not prove acceptable semantic quality and was not
+      applied to production.
+
+ranked_improvement_paths:
+  - rank: 1
+    name: stateful_alibaba_responses_continuation
+    evidence_strength: official_contract_match_plus_local_wire_gap
+    design:
+      - Correlate a local session using provider, model, credential fingerprint, prompt_cache_key, and semantic configuration hashes.
+      - Capture a completed upstream response ID and persist it with hashes of represented input and output items.
+      - On an exact append-only next turn, send only the delta with previous_response_id.
+      - Resend the current instructions because Responses does not carry prior instructions forward.
+      - Fall back to full replay on divergence, incomplete response, concurrent turn, configuration change, or expiry.
+    expected_effect: >-
+      Matches Alibaba's documented session-cache flow and removes repeated
+      multi-megabyte full-history transport from steady-state turns.
+    unverified:
+      - Whether Alibaba accepts continuation when the incoming Codex body has store=false.
+      - Whether Ciel should opt in to upstream state by forcing store=true; that changes privacy and persistence semantics.
+      - Live hit rate, latency, and Credits improvement in the target Token Plan session.
+  - rank: 2
+    name: prefix_stable_wire_compaction
+    evidence_strength: source_inspection_plus_synthetic_reproduction
+    design:
+      - Keep a persistent compaction checkpoint per prompt_cache_key instead of regenerating the entire first summary every append.
+      - Roll the checkpoint only at fixed boundaries and retain append-only stable chunks.
+      - Preserve function-call and output pairs and validate semantic loss before release.
+    expected_effect: >-
+      Keeps the largest reusable prefix byte-identical and benefits Alibaba
+      implicit caching even without server-side response state.
+    unverified:
+      - End-to-end task quality with checkpointed summaries.
+      - Production hit-rate increase under Alibaba scheduling.
+  - rank: 3
+    name: conditional_session_cache_header
+    evidence_strength: official_flow_differs_from_current_wire
+    design: >-
+      Treat x-dashscope-session-cache as a stateful Responses feature: enable it
+      for requests carrying a validated previous_response_id, or retain an
+      explicit user opt-in until the stateful bridge is proven.
+    boundary: >-
+      Current data proves that the header alone is insufficient for consistent
+      hits. It does not prove that the header causes the misses.
+  - rank: 4
+    name: chat_completions_explicit_cache_fallback
+    evidence_strength: existing_adapter_support_plus_official_chat_cache_contract
+    design: >-
+      Ciel already places up to four explicit cache_control markers on Alibaba
+      Chat Completions messages. A Responses-to-Chat route could use them.
+    risk: >-
+      Translation can lose native Responses tool and reasoning semantics, so it
+      is a fallback rather than the preferred qwen3.8-max Codex path.
+
+non_solutions_or_unproven_claims:
+  prompt_cache_key_alone: >-
+    Codex already sends a stable value. OpenAI documents its cache bucketing use,
+    but the Alibaba pages reviewed do not promise that it controls Alibaba cache.
+  header_alone: >-
+    Near-complete hits and complete misses were both observed while the header
+    was present, so no deterministic improvement claim is supported.
+  smaller_context_only: >-
+    Clearing or shrinking history can reduce billed input but is not itself a
+    cache-hit mechanism and discards context.
+
+implementation_boundary:
+  code_changed: false
+  live_router_restarted: false
+  live_session_interrupted: false
+  next_validation_required_before_claiming_improvement:
+    - Prototype previous_response_id continuation against Alibaba using a Ciel-isolated session without violating Token Plan direct-API restrictions.
+    - Compare identical workloads for weighted cached_input_tokens, first-token latency, total latency, and Usage Analysis Credits.
+    - Property-test append-only prefix stability and call/output pair integrity for checkpointed compaction.
+
+follow_up:
+  status: implemented_and_isolated_live_verified
+  journal: C:/Users/djlov/ciel-runtime/docs/journal/2026/09/02/implementation/cache/alibaba/qwen38-cache-hit-improvement.okf
+  note: >-
+    This research-only snapshot preceded the implementation. The follow-up
+    applies stateless implicit-cache routing plus pair-safe prefix checkpoints;
+    consult the implementation journal for measured results and boundaries.
+
+evidence_paths:
+  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/research/cache/codex/provider-scope.okf

@@ -0,0 +1,77 @@
+task:
+  id: codex-cache-provider-scope-20260902
+  request: >-
+    Determine whether the cache behavior found with Alibaba Singapore Token
+    Plan applies equally to every provider used through Codex.
+  status: answered
+
+confirmed_scope:
+  ciel_prefix_breaking_wire_compaction:
+    common_to_all_providers: false
+    call_path:
+      - openai_responses_router._handle_provider_responses_route
+      - ProviderResponsesPassthrough.forward
+      - ProviderResponsesPassthrough._fit_provider_request
+      - compact_responses_input_for_budget
+    activation_conditions:
+      - selected upstream protocol is openai_responses
+      - provider returns a non-null responses_request_max_bytes
+      - encoded request exceeds ninety percent of that limit
+    provider_overrides_found:
+      alitoken: non_null_limit
+      alitoken_individual: null
+      all_other_adapters: inherited_null
+    conclusion: >-
+      The exact summary-regeneration and prefix-destruction mechanism measured
+      in the qwen3.8-max Singapore Token Plan session currently activates only
+      for the alitoken adapter's oversized native Responses requests.
+
+  codex_request_shape:
+    observed:
+      full_input_history_replayed: true
+      store: false
+      prompt_cache_key_present: true
+      previous_response_id_present: false
+    passthrough_behavior: >-
+      ProviderResponsesPassthrough copies and normalizes the incoming body but
+      does not synthesize previous_response_id. Native Responses providers can
+      therefore receive this same stateless full-history shape.
+    conclusion: >-
+      This is a shared exposure for native Responses routes, not proof of a
+      shared cache defect. A provider may cache repeated prefixes without
+      previous_response_id, may offer a different state mechanism, or may not
+      expose prompt caching at all.
+
+route_classes:
+  native_responses:
+    exact_alitoken_compactor_issue: conditional_as_above
+    stateful_continuation_opportunity: provider_contract_dependent
+  translated_openai_chat_or_anthropic:
+    exact_alitoken_compactor_issue: false
+    reason: >-
+      These routes convert the Codex Responses body before provider transport
+      and do not call ProviderResponsesPassthrough._fit_provider_request.
+    cache_behavior: provider_specific
+  native_codex_backend:
+    exact_alitoken_compactor_issue: false
+    reason: >-
+      The codex-routed branch uses the Codex backend adapter rather than the
+      provider Responses passthrough.
+
+causal_boundary:
+  confirmed:
+    - The alitoken wire-size compactor can change the reusable prefix.
+    - The compactor is not globally enabled for every provider.
+    - Native Responses passthrough does not add previous_response_id.
+  not_confirmed:
+    - Other providers have not been measured to show Alibaba's 12.794 percent hit rate.
+    - Absence of previous_response_id does not by itself prove cache misses.
+    - A single cache strategy is not valid for every provider without checking its official contract.
+
+source_evidence:
+  router_dispatch: C:/Users/djlov/ciel-runtime/ciel_runtime_support/openai_responses_router.py
+  native_passthrough: C:/Users/djlov/ciel-runtime/ciel_runtime_support/provider_responses_passthrough.py
+  compactor: C:/Users/djlov/ciel-runtime/ciel_runtime_support/prompt_compaction.py
+  provider_default: C:/Users/djlov/ciel-runtime/ciel_runtime_support/architecture.py
+  alibaba_override: C:/Users/djlov/ciel-runtime/ciel_runtime_support/providers/alibaba.py
+  code_changed: false