@oneciel-ai/ciel-runtime 0.2.37 → 0.2.38

package/docs/journal/2026/09/04/diagnostics/claude/context/fable-51-status-200k.okf

@@ -0,0 +1,54 @@
+schema: okf/1
+kind: task-journal
+title: Fable 5.1 status line remained at 200K after the context-window fix
+date: 2026-09-04
+workspace: G:\ciel-bridge-service
+
+observation:
+  screenshot:
+    model: Fable 5.1
+    context: 351812/200000 tokens (175.9 percent)
+  process:
+    ciel_pid: 59272
+    claude_pid: 30032
+    claude_model_argument: ciel-runtime-anthropic-claude-fable-5-1
+  config:
+    current_model: claude-fable-5-1
+    route_through_router: true
+    context_override: absent
+  launch_log:
+    time: 2026-09-04T15:23:54
+    context: 1048576
+    auto_compact_window: 1000000
+
+root_cause:
+  confirmed:
+    - Ciel projected the corrected 1M environment values for Fable 5.1.
+    - The Claude-facing custom gateway alias omitted `[1m]`.
+    - Claude Code 2.1.260 consequently supplied a 200000 context-window size to
+      the status line despite Ciel's 1M launch snapshot.
+
+change:
+  - Append `[1m]` to Claude-facing routed Anthropic aliases for every model in
+    the documented 1M-default set, even when the configured upstream ID has no
+    marker.
+  - Continue stripping the local marker at the Anthropic upstream boundary.
+  - Preserve 200K aliases for models outside the documented 1M set.
+
+verification:
+  focused:
+    result: 115 passed; lint passed
+  local_install:
+    path: C:\Users\djlov\.local\share\ciel-runtime
+  launch:
+    claude_code_version: 2.1.261
+    model_alias: ciel-runtime-anthropic-claude-fable-5-1[1m]
+    context: 1048576
+    auto_compact_window: 1000000
+    exit_code: 0
+  interactive_tui:
+    observed_status: ctx est 89/1,048,576 tok (0.0 percent)
+    test_session_exit_code: 0
+  full_suite:
+    result: 3156 passed; 137 skipped; 0 failed
+  status: behavior_verified

package/docs/journal/2026/09/04/diagnostics/runtime/codex/responses/replayed-item-id-validation.okf

@@ -0,0 +1,74 @@
+task:
+  title: Repair invalid replayed Responses item IDs in a resumed Codex session
+  date: 2026-09-04
+  workspace: C:\Users\djlov\cielarvis
+
+observed:
+  runtime_endpoint: http://127.0.0.1:9610/backend-api/codex/responses
+  session_id: 01a01dff-f0f0-7ef3-8cde-1d9916510f35
+  http_status: 400
+  rejected_index: 160
+  rejected_item_type: reasoning
+  rejected_id_shape: 'rs_<token>:rs_<token>'
+  upstream_rule: IDs may contain only letters, numbers, underscores, or dashes
+  persisted_at: C:\Users\djlov\.codex\sessions\2026\08\20\rollout-2026-08-20T02-08-27-01a01dff-f0f0-7ef3-8cde-1d9916510f35.jsonl:13943
+
+cause:
+  boundary: ciel_runtime_support.responses_input_compatibility.repair_replayed_response_items
+  confirmed_behavior: Existing validation checked the type prefix but not the complete character set.
+  reproduction_before_fix: The observed colon-delimited ID was returned unchanged.
+  provenance:
+    ciel_colon_composition_found: false
+    ciel_backend_response_rewrite: false
+    codex_version: 0.153.0
+    codex_release_tag: rust-v0.153.0
+    codex_response_item_behavior: ResponseItemId.from_server preserves the server value without validation.
+    session_behavior: The composite ID was already persisted as a response_item.reasoning before the failing replay.
+    conclusion: The upstream response supplied the composite ID; Ciel's missing character validation allowed it back into a later request.
+  repeat_failure: Every retry replayed the same invalid history item at input index 160.
+
+repair:
+  policy: Validate every non-empty Responses item ID against the upstream character contract.
+  allowed_characters: ASCII letters, digits, underscore, dash
+  invalid_reasoning_with_encrypted_content: Initially removed only the ID; superseded by the cross-workspace finding below.
+  invalid_reasoning_without_encrypted_content: Retain the existing policy of dropping the foreign summary-only item.
+  specificity: No session ID, provider name, or trigger-word special case in production code.
+
+verification:
+  status: passed
+  focused:
+    responses_input_compatibility: 19 passed
+    codex_backend_item_repair: 21 passed
+  observed_record_replay:
+    forwarded_id_present: false
+    encrypted_content_preserved: true
+    source_record_unchanged: true
+  documentation_metadata: 26 Markdown files passed
+  ruff: passed
+  diff_check: passed
+  full_suite:
+    unit: 1378 passed, 45 skipped
+    router: 1094 passed
+    channel: 406 passed, 80 skipped
+    runtime: 276 passed, 12 skipped
+    total: 3154 passed, 137 skipped
+  local_deployment:
+    install_path: C:\Users\djlov\.local\share\ciel-runtime
+    source_installed_sha256: B1AD0EB9798B2DFFF768C29AA8BD8AEB069A1958F10126E35BDC1FBF64A50103
+    installed_module_loaded_in_fresh_process: true
+    observed_record_forwarded_id_present: false
+    observed_record_encrypted_content_preserved: true
+    active_predeployment_processes_restarted: false
+
+follow_up_2026_09_04:
+  evidence_source: G:\OneCielDMSUI session 019fcac4-4538-7a53-91d1-565cd21a8c5e
+  finding: The same colon-ID reasoning records were rejected next for unverifiable encrypted content after their IDs were removed.
+  correlated_rejections:
+    logged_distinct_verdicts: 110
+    matched_session_records_before_repair: 118
+    matched_session_records_after_repair: 0
+  revised_policy:
+    forbidden_character_reasoning: Drop the whole hidden reasoning item.
+    router_synthesized_reasoning: Drop the whole hidden reasoning item.
+    legal_but_type_mismatched_reasoning: Preserve encrypted content and omit only its ID.
+    visible_messages_and_tool_history: Preserve and balance independently.

package/docs/journal/2026/09/04/diagnostics/workspaces/onecieldmsui/codex/replay-stall.okf

@@ -0,0 +1,100 @@
+task:
+  title: Diagnose Codex replay stalls in OneCielDMSUI
+  date: 2026-09-04
+  workspace: G:\OneCielDMSUI
+
+runtime:
+  ciel_cli_pid: 45892
+  ciel_server_pid: 67612
+  codex_pid: 71872
+  router_port: 9605
+  model: gpt-5.6-sol
+  session_id: 019fcac4-4538-7a53-91d1-565cd21a8c5e
+  codex_version: 0.153.0
+  process_started_local: 2026-09-04T01:43:12-05:00
+
+observed:
+  invalid_id_http_error_in_session: false
+  successful_latest_task_completions: 3
+  replayed_reasoning_items: 1700
+  forbidden_character_reasoning_ids: 257
+  router_synthesized_reasoning_ids: 14
+  router_log:
+    symptom: The upstream rejected encrypted reasoning one record at a time and the router retried after each verdict.
+    prefiltered_known_rejections_at_02_15: 105
+    concurrent_transport_failures:
+      - WinError 10054 connection reset
+      - WinError 10053 local connection abort
+      - SSLV3_ALERT_BAD_RECORD_MAC
+    confirmed_http_500_examples:
+      - 2026-09-04T02:05:59-05:00
+      - 2026-09-04T02:07:14-05:00
+
+correlation:
+  logged_verdicts_sampled_initially: 115
+  verdicts_matching_forbidden_character_ids: 101
+  verdicts_matching_router_synthesized_ids: 14
+  conclusion: The current stall is an encrypted-reasoning replay loop, not a recorded invalid-ID 400.
+
+repair:
+  boundary: ciel_runtime_support.responses_input_compatibility.repair_replayed_response_items
+  behavior: Drop hidden reasoning at replay when its ID is illegal or router-synthesized; retain visible messages and paired tool history.
+  hardcoded_session_or_model: false
+  source_session_mutated: false
+
+verification:
+  focused_tests:
+    responses_input_compatibility: 20 passed
+    codex_reasoning_rejects: 12 passed
+    codex_backend_item_repair: 21 passed
+  ruff: passed
+  actual_session_projection:
+    reasoning_before: 1700
+    forbidden_id_before: 257
+    router_synthesized_before: 14
+    forbidden_id_after: 0
+    router_synthesized_after: 0
+  actual_router_verdict_projection:
+    distinct_logged_verdicts_initial: 110
+    matched_records_before_initial: 118
+    matched_records_after: 0
+  full_suite:
+    unit: 1379 passed, 45 skipped
+    router: 1094 passed
+    channel: 406 passed, 80 skipped
+    runtime: 276 passed, 12 skipped
+    total: 3155 passed, 137 skipped
+  documentation_metadata: 26 Markdown files passed
+  local_deployment:
+    install_path: C:\Users\djlov\.local\share\ciel-runtime
+    source_installed_sha256: C2A3C308B8D518999A8C5BAA0FC1B8CC6F013DA58C5640D15D3A670A3B85DF03
+    installed_module_loaded_in_fresh_process: true
+    matched_rejected_before_at_final_check: 127
+    matched_rejected_after: 0
+    active_process_restarted: false
+  live_process_restarted_during_initial_verification: false
+
+recovery:
+  provider_transition_confirmed:
+    prior_turn_model: ciel-runtime-meta-muse-spark-1.3-contributor
+    resumed_model: gpt-5.6-sol
+    forbidden_id_records_attributed_to_prior_model: 257
+    evidence: Session turn_context records changed from Muse Spark Contributor to a Codex-routed GPT launch, and Codex displayed the same transition warning.
+  old_loop_exit_local: 2026-09-04T02:30:52-05:00
+  verified_resume:
+    selected_session_id: 019fcac4-4538-7a53-91d1-565cd21a8c5e
+    cwd: G:\OneCielDMSUI
+    model: gpt-5.6-sol high
+    prompt: State-free recovery marker request
+    router_http_status: 200
+    response: SESSION_RECOVERY_OK
+    unverifiable_reasoning_retries_after_restart: 0
+    invalid_id_errors_after_restart: 0
+  user_visible_runtime:
+    started_local: 2026-09-04T02:37:58-05:00
+    cmd_pid: 53748
+    ciel_cli_pid: 72436
+    ciel_server_pid: 80056
+    codex_pid: 41604
+    session_id: 019fcac4-4538-7a53-91d1-565cd21a8c5e
+    status: running

package/docs/journal/2026/09/04/implementation/observability/tool-calls/websocket-stream.okf

@@ -0,0 +1,49 @@
+schema: okf/1
+kind: task-journal
+title: CLI tool-call event WebSocket stream
+date: 2026-09-04
+
+request:
+  requirement: push each CLI runtime tool call to an externally consumable realtime stream
+
+change:
+  normalized_event_category: tool.call
+  sources:
+    - routed Anthropic-compatible stream tool calls
+    - routed Ollama stream tool calls
+    - routed OpenAI-compatible stream tool calls
+    - Claude CLI structured transcript tool_use records
+    - Codex CLI function_call, custom_tool_call, and local_shell_call records
+    - Muse Code runtime.session tool_calls records, including Windows WSL sessions
+  endpoint: GET /ca/events/ws?category=tool.call
+  protocol: RFC 6455 JSON text frames
+  access_control: existing router external-access authentication
+  cursor: after query parameter
+  duplicate_policy: suppress the transcript copy when the routed event has the same call_id
+
+verification:
+  focused_tests: 37 passed
+  live_websocket:
+    handshake: HTTP/1.1 101 Switching Protocols
+    frame_opcode: 1
+    published_count: 1
+    category: tool.call
+    source: cli-transcript
+    provider: codex
+    session_id: proof-session
+    call_id: call-live-proof
+    name: exec_command
+    arguments: '{"cmd":"rg TODO"}'
+  live_muse_transcript:
+    muse_version: 1.0.2
+    transcript_found: true
+    schema: runtime.session tool_calls
+    projected_call_type: muse_tool_call
+    projected_tool_name: bash
+  full_suite: 3167 passed; 137 skipped; 0 failed
+  lint: passed
+  local_install:
+    path: C:\Users\djlov\.local\share\ciel-runtime
+    verified_source_hashes: 18
+    hash_mismatches: 0
+  status: behavior_verified

package/docs/journal/2026/09/04/implementation/web-chat/input/raw-injection.okf

@@ -0,0 +1,42 @@
+schema: okf/1
+kind: task-journal
+title: Exact-message Web Chat injection
+date: 2026-09-04
+
+request:
+  endpoint:
+    - POST /ca/chat/messages
+    - POST /ca/channel/messages
+  requirement: inject only the supplied message without descriptive or response instructions
+
+change:
+  option: raw_injection=true
+  guarantees:
+    - model input equals the admitted message string
+    - no web envelope or correlation marker
+    - no attachment projection
+    - no web ACK/reply contract
+    - no MCP response hint
+    - raw messages are isolated from batching
+  independent_axes:
+    - input_transport
+    - response_mode
+
+verification:
+  focused_tests: 31 passed
+  live_http:
+    endpoint: POST /ca/channel/messages
+    status: 200
+    runtime_queue_recorded: true
+    expected_characters: 38
+    web_chat_prompt_exact: true
+    router_body_exact: true
+    expected_sha256: 2630b5fd9533735276f479f424f07db99aabb7d755b856d9c5e16f1de8b9f7da
+    actual_sha256: 2630b5fd9533735276f479f424f07db99aabb7d755b856d9c5e16f1de8b9f7da
+  full_suite: 3167 passed; 137 skipped; 0 failed
+  lint: passed
+  local_install:
+    path: C:\Users\djlov\.local\share\ciel-runtime
+    verified_source_hashes: 18
+    hash_mismatches: 0
+  status: behavior_verified

package/docs/journal/2026/09/05/diagnostics/tty/claude-raw-idle-render.okf

@@ -0,0 +1,25 @@
+title: Claude idle raw TTY prompt render timeout
+date: 2026-09-05
+evidence:
+  instance:
+    router: 9471-81e6fa259d15
+    request: 278
+    accepted: 2026-09-05T13:18:48
+    timeout: 2026-09-05T13:19:18
+    error: prompt_render_timeout -> prompt_not_submitted
+    later: 2026-09-05T13:19:47 submit_confirmed attempt=1
+  installed:
+    path: C:/Users/djlov/.local/share/ciel-runtime/ciel_runtime_support/windows_conpty.py
+    algorithm: prompt[:48] UTF-8 bytes searched directly in VT output
+  reproduction:
+    command: python scripts/probe_claude_raw_prompt_render.py
+    isolated_cli: Installed Claude; draft-only input; no Enter or API request
+    bytes: 187
+    observed: Claude draws spaces using ESC[1C cursor movement
+    installed_detected: false
+    working_tree_detected: true
+conclusion:
+  confirmed_defect: VT output cannot be compared directly with plain prompt bytes. The readiness gate returns before sending Enter.
+  existing_fix: Working tree normalizes VT text; it is absent from the inspected installation.
+  scope: Demonstrates this failure path, not every reported wakeup failure.
+  deployment: Not performed; active user sessions unchanged.

package/docs/journal/2026/09/05/implementation/events/search-response-urls.okf

@@ -0,0 +1,23 @@
+title: Web search response URLs over the existing tool.call event stream
+evidence:
+  claude:
+    source: Local Claude transcript 81b0504c-4826-42ef-91f5-397369ca6aef
+    call: toolu_018Firb4ystzFgqgJ73msgF8
+    format: tool_use WebSearch -> tool_result content string with JSON Links list
+  codex:
+    source: https://raw.githubusercontent.com/openai/codex/main/codex-rs/protocol/src/models.rs
+    finding: WebSearchCall can contain only id/status/action; Search action has query/queries, not result URLs.
+    api_source: https://developers.openai.com/api/docs/guides/tools-web-search
+    supported: Structured sources/results/citations when retained; correlated search function outputs.
+    limitation: No fabricated result URLs for metadata-only native rollouts.
+implementation:
+  transport: Existing event bus and /ca/events/ws?category=tool.call, SSE and recent
+  data: phase=result, call_id, name, runtime, urls; no query/title/body
+  persistence: Per-transcript cursor with bounded 512-call and result-hash state
+verification:
+  live_loopback:
+    command: python scripts/probe_web_search_result_stream.py <transcript> <call_id>
+    observed: HTTP/1.1 101 Switching Protocols; nine real Claude result URLs received in a WS text frame.
+    safety: Isolated EventBus and loopback port; no writes to live user transcripts or external destinations.
+  tests: Parser, restart/partial-line integration, transcript delivery, existing WS adapter and tool-call projection.
+  deployment: Not requested in this turn; no deployment or push performed.

package/docs/journal/2026/09/05/implementation/tools/native-injection.okf

@@ -0,0 +1,19 @@
+title: Native versus non-native Ciel tool injection
+date: 2026-09-05
+scope:
+  ownership: Ciel-generated definitions only; user MCP configuration unchanged.
+  policy:
+    injection_mode: always | native | non_native
+    default: always (channel/runtime integration remains available)
+  launches:
+    claude: Anthropic direct and routed launches exclude generated DDG/fetch even with a search override.
+    codex: Native provider classification passed to workspace MCP projection, including app-server.
+    existing_codex_search: No automatic DDG generator call exists in the inspected Codex launch path.
+  verification:
+    method: Execute tests and read actual temporary MCP JSON/argument projections across non-native/native/non-native launches.
+    boundary: No live authenticated model query or deployment requested/performed.
+    results:
+      workspace_mcp: 18 tests passed; actual generated JSON/argument projections verified.
+      policy: 2 tests passed.
+      claude_launch: 1 test passed, including native/non-native explicit-search command assembly.
+      codex_full_suite: Interrupted; opened an interactive resume picker. Only the identified test process tree was terminated. Not claimed as passing.

package/docs/journal/2026/09/05/implementation/windows/conpty/prompt-delivery-lifecycle.okf

@@ -0,0 +1,43 @@
+knowledge:
+  task:
+    title: Windows ConPTY prompt delivery lifecycle hardening
+    date: 2026-09-05
+    scope:
+      - ANSI and Korean prompt-render observation
+      - no-repeat submission failure behavior
+      - runtime input lifecycle API and events
+      - strict delivery ordering
+      - private queue cursor accounting
+      - resize failure and redraw observation
+  evidence:
+    source:
+      prompt_detection: ciel_runtime_support/windows_conpty.py
+      injection_policy: ciel_runtime_support/channel_injection.py
+      ordering: ciel_runtime_support/channel_pending_injection.py
+      lifecycle: ciel_runtime_support/runtime_input_status.py
+      api: ciel_runtime_support/chat_http_controller.py
+    confirmed_behavior:
+      - ANSI CSI, OSC, SGR, and cursor-right output is projected only for detection; input text is not decorated or rewritten.
+      - prompt-render timeout requests one draft clear, records whether redraw confirmed it, then disables automatic retyping.
+      - an unconfirmed submission is sticky failed state and blocks later private inputs at the same cursor.
+      - TTY and session-socket drafts are limited to one request; router-body batching remains independent of TUI drafts.
+      - backlog scan_tail is composed with _runtime_input_scan_max_id, not the public chat repository.
+      - ConPTY resize logs HRESULT or exception on failure and redraw observed or timeout after success.
+      - no ASCII sentinel or automatic router-body fallback remains in the Windows terminal proxy.
+  verification:
+    focused:
+      windows_conpty: 33 passed
+      channel_injection: 13 passed
+      runtime_input_status: 2 passed
+      channel_bridge: 216 passed, 80 skipped
+    native_windows:
+      transport: real CreatePseudoConsole child process
+      utf8: ANSI-fragmented Korean prompt observed and submitted; child returned TURN_STARTED
+      resize_redraw: Claude 2783 bytes and Codex 2233 bytes after 120x30 to 123x32 resize; both logged redraw=observed
+    http_probe: POST returned queued for request 73; GET returned replied; events were queued, submitted, replied
+    full_suite: 3174 passed, 137 skipped, 0 failed
+    lint: passed
+    docs: passed
+    diff_check: passed
+    local_deployment: not requested in this task
+    nightly_push: not requested in this task

package/docs/journal/2026/09/05/release/nightly/conpty-native-tools.okf

@@ -0,0 +1,13 @@
+title: Local and nightly release of ConPTY delivery and managed tool policies
+date: 2026-09-05
+authorization: User requested local and nightly deployment.
+scope:
+  changes: Previously implemented ConPTY input lifecycle and native tool injection changes, with regression tests and diagnostic probes.
+verification:
+  static: Ruff, documentation metadata and git diff whitespace checks passed.
+  cli_probe: Installed Claude draft output reproduces old detector failure and working-tree detector success without submitting a model request.
+deployment:
+  local_target: C:/Users/djlov/.local/share/ciel-runtime
+  remote_branch: nightly
+  mechanism: Repository Windows installer; GitHub npm publish workflow.
+  active_sessions: Do not terminate or restart user sessions; loaded Python code requires runtime restart.

package/docs/journal/2026/09/05/release/nightly/search-result-urls.okf

@@ -0,0 +1,14 @@
+title: Local and nightly deployment of URL-only search result events
+date: 2026-09-05
+authorization: User requested local and nightly deployment.
+scope:
+  endpoint: Existing /ca/events/ws?category=tool.call plus SSE/recent
+  payload: phase=result and urls with call correlation; no search response prose
+  limitation: Codex metadata-only rollouts cannot supply absent result URLs.
+verification:
+  preflight: Ruff, documentation metadata and whitespace checks passed.
+  functionality: Real Claude search replay produced nine URLs over a loopback WebSocket; 23 targeted tests passed.
+deployment:
+  local_target: C:/Users/djlov/.local/share/ciel-runtime
+  remote: nightly branch through the repository npm publish workflow
+  sessions: User processes remain running; restart required for already-loaded modules.

package/docs/journal/2026/09/06/implementation/codex/inherited-web-mcp.okf

@@ -0,0 +1,22 @@
+title: Disable inherited replacement web MCPs for native Codex launches
+date: 2026-09-06
+scope:
+  requested: Implement native-only overrides and deploy locally.
+  project: G:/ciel-bridge-service
+  cause: Global Codex config declares duckduckgo and web_fetch; prior Ciel injection filtering did not suppress inherited entries.
+changes:
+  cli: Append enabled=false overrides after workspace MCP projection.
+  app_server: Apply identical overrides for native providers, direct or routed.
+  non_native: No override emitted.
+verification:
+  actual_cli:
+    command: codex [-c mcp_servers.duckduckgo.enabled=false -c mcp_servers.web_fetch.enabled=false] mcp get <server> --json
+    cwd: G:/ciel-bridge-service
+    baseline: Both enabled=true.
+    native_override: Both enabled=false.
+    global_config_hash: Unchanged before/after inspection.
+  tests: Runtime group 276 tests completed successfully with 12 skipped; policy tests passed.
+deployment:
+  target: C:/Users/djlov/.local/share/ciel-runtime
+  sessions: No existing user process terminated; restart required.
+  nightly: Not requested for this turn.

package/docs/journal/2026/09/07/integration/main/nightly-merge.okf

@@ -0,0 +1,13 @@
+title: Merge current nightly development into local main
+date: 2026-09-07
+authorization: User requested merging the current version into main.
+scope:
+  source: nightly, including the previously uncommitted native Codex MCP suppression fix
+  target: local main
+  excluded: Unrelated Ciel Bridge discussion; no bridge code changes
+verification:
+  remote_preflight: Local main matches origin/main after fetch.
+  checks: Ruff and documentation metadata passed; runtime group completed 276 tests with 12 skipped.
+  git: Verify source is an ancestor of main and merged tree matches source.
+publication:
+  policy: No remote push or npm publication; main push triggers stable publish.

package/docs/journal/2026/09/07/release/stable/0.2.38.okf

@@ -0,0 +1,10 @@
+title: Publish stable 0.2.38 so existing clients can detect the update
+date: 2026-09-07
+authorization: User requested deployment after identifying unchanged published version.
+cause:
+  previous: main advanced but package version stayed at already-published 0.2.37.
+  correction: Synchronize npm package and Python runtime versions at 0.2.38.
+deployment:
+  channel: npm latest through main branch publish workflow
+  verification: Check CI, publish result, npm latest version/gitHead, and downloadable package.
+  exclusions: No user session restart or configuration changes.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@oneciel-ai/ciel-runtime",
-  "version": "0.2.37",
+  "version": "0.2.38",
   "description": "Universal AI coding-agent runtime and model-routing layer for Claude, Codex, AGY, ZCode, and compatible runtimes.",
   "license": "MIT",
   "author": "One Ciel LLC",