@oneciel-ai/ciel-runtime 0.2.22 → 0.2.23

package/docs/journal/2026/08/21/wing-repeated-tool-guard-finalization.okf

@@ -0,0 +1,81 @@
+okf_version: "1.0"
+knowledge:
+  task:
+    id: wing-repeated-tool-guard-finalization-20260821
+    date: "2026-08-21"
+    objective: >-
+      Repair the WING routed Codex turn that exposes the repeated-tool guard as
+      the assistant's final answer instead of completing the user's request.
+
+  evidence:
+    host: WING
+    workspace: "C:\\Users\\daniel.yun.WING"
+    runtime_version: "0.2.22-nightly.20260821-213321.46d0e4e"
+    session: "01a02639-b0aa-7530-9115-8084007b2754"
+    transcript: >-
+      C:\Users\daniel.yun.WING\.codex\sessions\2026\08\21\
+      rollout-2026-08-21T16-28-30-01a02639-b0aa-7530-9115-8084007b2754.jsonl
+    router_log: >-
+      C:\Users\daniel.yun.WING\AppData\Roaming\ciel-runtime\
+      router-instances\9140-f895679997f4\router.log
+    observed_sequence:
+      - "20:18:18 the new Korean user question was present in the transcript"
+      - >-
+        20:18:21 the upstream reasoning correctly identified that the user asked
+        whether a message had arrived from Samuel
+      - >-
+        20:18:21 and 20:18:25 the same exec_command was emitted and returned the
+        same HTTP 400 JSON-RPC result
+      - >-
+        20:18:29 the third identical call was dropped with completed_repeats=2
+      - >-
+        transcript ordinals 126-129 projected the runtime guard notice as an
+        assistant message and immediately recorded task_complete
+    excluded_causes:
+      - "The user's new question was not missing from the Codex transcript."
+      - "The upstream model did not misidentify the user's requested subject."
+      - "The duplicate-call guard did prevent a third identical execution."
+    security_note: >-
+      The captured command arguments contained an authentication value in the
+      transcript. It is redacted from this journal and all reports.
+
+  confirmed_cause: >-
+    project_ollama_response converts the internal repeated-tool guard into a
+    normal assistant text block. recover_preamble_only_turn did not classify
+    that runtime-owned control notice as recoverable, so the Responses adapter
+    emitted it with end_turn and Codex committed task_complete.
+
+  correction:
+    status: implemented_and_locally_verified
+    behavior:
+      - >-
+        Recognize only the exact runtime-owned repeated-tool notice and perform
+        one bounded upstream continuation.
+      - >-
+        Do not replay the control notice to the model. Supply a marked internal
+        continuation directing it to use the existing result, choose a distinct
+        action, or give the concrete answer.
+      - >-
+        The internal continuation does not reset repeated-call history, so the
+        blocked call cannot execute again during recovery.
+      - >-
+        Accept either a distinct tool call or a visible final answer from the
+        recovery request.
+      - "No host, user, model, provider, agent, or command-specific exception."
+
+  verification:
+    focused:
+      ollama_response_projection: "49 passed"
+      codex_turn_recovery: "32 passed"
+      tool_repeat_dedupe: "10 passed"
+      ruff: passed
+      documentation_metadata: passed
+      git_diff_check: passed
+    full_test:
+      unit: "1121 passed, 44 skipped"
+      router: "906 passed"
+      channel: "381 passed, 80 skipped"
+      runtime: "252 passed, 12 skipped"
+      total: "2660 passed, 136 skipped"
+    wing_live: pending
+    deployment: pending

package/docs/journal/2026/08/21/wing-working-directory-memory-deployment.md

@@ -0,0 +1,110 @@
+# Wing working-directory Remote Memory deployment
+
+## Task
+
+Store each remotely synchronized memory tree below the runtime's actual launch
+workspace, inject its verified index address into routed model instructions, and
+publish/install the result as a nightly build for an end-to-end Wing check.
+
+## Evidence before the correction
+
+- Wing runs Ciel from `C:\Users\daniel.yun.WING`.
+- Wing configuration explicitly sets `remote_memory.directory=.ciel/memory`.
+- The installed `0.2.22-nightly.20260821-072901.5e5bdec` build instead mapped
+  that value to its AppData workspace-state directory.
+- A production-pipeline capture on Wing showed the AppData index address in
+  both the translated Anthropic `system` block and the first Ollama `system`
+  message. This proves the address was not dropped by protocol conversion; it
+  also proves the destination did not match the requested working-directory
+  placement.
+- A real routed model turn nevertheless claimed that no memory index was
+  present. That answer contradicts the captured wire payload and is not used as
+  evidence that the router dropped the prompt.
+- After the first working-directory nightly was installed, the real model turn
+  again denied seeing the pointer, searched the home tree, then found and read
+  `.ciel/memory/index.md`. Source inspection showed that the pointer was placed
+  before Codex's later developer context and Ciel's execution reminder, not at
+  the requested bottom of the final system text.
+
+## Correction
+
+- Interpret `remote_memory.directory` relative to the resolved launch
+  workspace and restore `.ciel/memory` as the default.
+- Permit a user-home launch workspace while retaining path traversal and
+  workspace-boundary checks.
+- Keep synchronization metadata in the per-workspace Ciel state directory.
+- Record the resolved workspace in that metadata and refuse to inject an index
+  when the record belongs to a different launch workspace.
+- Continue removing obsolete Ciel-managed pointer blocks from native runtime
+  instruction files; no new `AGENTS.md`, `CLAUDE.md`, or `GEMINI.md` pointer is
+  created.
+- During Responses-to-Anthropic conversion, remove any earlier managed pointer
+  and append it after all converted developer context. During OpenAI/Ollama wire
+  projection, move it behind execution reminders and state messages so it is
+  the final text in the system message.
+
+## Local verification
+
+- Remote Memory suites: 17 tests passed, including loopback HTTP downloads,
+  overwrite behavior, home-workspace placement, cross-workspace rejection,
+  protocol-family injection, and Responses-to-Ollama wire projection.
+- Event settings suite: 25 tests passed.
+- Configuration repository suite: 7 tests passed.
+- Ruff and `git diff --check` passed for the changed implementation and tests.
+- The full suite passed: `Ran 2633 tests in 290.487s`,
+  `OK (skipped=136)`.
+- `npm pack --dry-run --json` succeeded with 385 package entries and included
+  the changed runtime, Remote Memory implementation, documentation, tests' task
+  journal, and package metadata.
+
+## Deployment verification
+
+- Commit `66483657e033943a8388b805c90a468d1104c83a` published as
+  `0.2.22-nightly.20260821-075242.6648365`; npm `gitHead` matched.
+- GitHub npm publish run `32460413577` completed successfully, including its
+  published-tarball verification.
+- Wing installed that exact version and synchronized 28 files (50,689 bytes)
+  to `C:\Users\daniel.yun.WING\.ciel\memory`; the index existed and its first
+  line was `---`.
+- The first real model test only reached the correct two-line result after a
+  broad filesystem search. That is recorded as a failed prompt-visibility
+  check, not as successful instruction use.
+- After correcting final-system placement, the focused Remote Memory and prompt
+  injection suites passed 25 tests, Ruff passed, and the second full suite
+  passed: `Ran 2633 tests in 292.956s`, `OK (skipped=136)`.
+- Commit `21872ed694428e46d4246e4bc08389f84ae462de` published as
+  `0.2.22-nightly.20260821-080802.21872ed`; npm `gitHead` matched and GitHub CI
+  run `32461589362` completed successfully.
+- Wing installed that exact version and again synchronized 28 files (50,689
+  bytes) below the launch workspace. The index existed and its first line was
+  `---`.
+- A real no-search model turn still denied receiving the index address. A
+  loopback Ollama wire capture then measured zero Remote Memory markers and no
+  index path in the complete upstream JSON.
+- A temporary, non-published probe of the same installed source recorded
+  `marker_after=True` for both the `openai_responses` injection and the later
+  `anthropic_messages` injection. The package and provider configuration were
+  restored immediately afterward.
+
+## Final wire truncation correction
+
+- The measured transition is therefore: marker present after Anthropic
+  injection, marker absent in Ollama wire projection.
+- Source inspection identifies the destructive transition in
+  `coalesce_ollama_system_messages`: it applies
+  `compact_message_text_for_prompt` to the combined Anthropic system text.
+  That helper has a fixed `PROMPT_MESSAGE_TEXT_LIMIT=20000`, so a pointer at the
+  tail of Wing's longer Codex system prompt is removed before the existing
+  post-projection placement helper runs.
+- `move_memory_pointer_to_system_end` now accepts a verified fallback pointer.
+  The final Ollama and OpenAI chat projection wrappers obtain that pointer from
+  the current workspace state and restore it only when the managed block was
+  removed by an intermediate projection.
+- The regression test uses a 25,000-character system instruction and verifies
+  that the ordinary system text is truncated while the workspace memory index
+  is present exactly once at the final system-message tail.
+- Focused Remote Memory and prompt-injection verification passed 25 tests;
+  Ruff and `git diff --check` also passed. The full suite passed 2,633 tests
+  with 136 skips, and `npm pack --dry-run --json` reported 385 entries.
+  Nightly publication and a new Wing wire/model check remain required before
+  closure.

package/docs/journal/2026/08/21/wing-workspace-state-memory-prompt-injection.md

@@ -0,0 +1,88 @@
+# Wing workspace-state Remote Memory correction
+
+## Task
+
+Correct the 0.2.22 Remote Memory behavior for Wing, where the runtime launch
+directory is the user's home directory and no separate project directory
+exists.
+
+## Contradiction found
+
+- The deployed 0.2.22 implementation rejected a launch workspace equal to the
+  user home.
+- Wing's recorded router workspace is
+  `C:\Users\daniel.yun.WING`; there is no separate project checkout.
+- Therefore the rejection prevented every post-deployment synchronization and
+  could not satisfy Wing's actual deployment model.
+
+## Corrected boundary
+
+- Downloaded memory is owned by the Ciel workspace state directory, not by the
+  process current directory.
+- The default target is `<workspace-state>/memory`.
+- The saved index address is verified to resolve below that state directory
+  before it is exposed to a prompt.
+- Routed OpenAI Responses, OpenAI Chat, and Anthropic Messages requests receive
+  the same managed memory-index block through the protocol-aware prompt
+  injector.
+- No Remote Memory pointer is created in home `AGENTS.md`, `CLAUDE.md`, or
+  `GEMINI.md`. A block written by 0.2.22 is removed without altering other user
+  text.
+- The legacy `.ciel/memory` directory setting maps to the new state-local
+  `memory` directory for upgrade compatibility.
+
+## Verification evidence
+
+- `python -m unittest discover -s tests -p "test_remote_memory*.py"`
+  passed 15 tests, including a real loopback HTTP manifest/file download.
+- `python -m unittest discover -s tests -p "test_prompt_injection.py"`
+  passed 8 protocol transformation tests.
+- `python -m unittest discover -s tests -p "test_event_settings_cli.py"`
+  passed 25 configuration CLI tests.
+- Remote Memory integration tests assert injection into OpenAI Responses,
+  OpenAI Chat, and Anthropic Messages, plus idempotence.
+- The first full-suite run found seven failures: one architecture line-budget
+  failure and six object-identity regressions when memory was disabled. The
+  injector was moved to its bounded module and changed to return the original
+  request object when no prompt exists. Focused architecture, channel bridge,
+  and Codex compatibility suites then passed.
+- The verbose Remote Memory integration run passed all 15 tests. Its console
+  output explicitly records the loopback HTTP download, home-launch/state-local
+  storage, three protocol-family injections, and idempotence cases as `ok`.
+- The final full suite passed: `Ran 2631 tests in 294.411s`, `OK (skipped=136)`.
+- Targeted Ruff validation passed after removing one unused test variable.
+- `npm pack --dry-run --json` succeeded for
+  `@oneciel-ai/ciel-runtime@0.2.22`, reported 384 package entries, and included
+  the changed `ciel_runtime.py`, `remote_memory.py`, documentation, and journal.
+- `git status --short` confirms the pre-existing August 20 journal changes are
+  still present as unstaged user changes; this task did not overwrite them.
+
+## Deployment state
+
+The correction is implemented and verified in the local worktree only. No git
+commit, remote push, package publish, installation, or Wing process restart was
+performed by this task.
+
+## Wing load check after install command
+
+- The agent screenshot reports loading `memoir` from
+  `C:\Users\daniel.yun.WING\AGENTS.md`; it does not report the Ciel workspace
+  state memory index.
+- SSH inspection shows Wing still has
+  `0.2.22-nightly.20260821-051020.13f5d91` installed.
+- The installed `remote_memory.py` has the home-rejection message and has
+  neither `DEFAULT_DIRECTORY = "memory"` nor
+  `inject_current_memory_prompt`.
+- The new workspace-state `memory` directory does not exist and contains zero
+  files. The only state record is the stale 28-file record last written at
+  `2026-08-21T04:45:14.4014143Z`, with root `.ciel/memory`.
+- The latest launch logs at `02:21:12` and `02:21:31` record the same user-home
+  rejection for Claude and Codex. No later successful memory update exists.
+- Home instruction files contain zero Ciel Remote Memory pointer blocks, while
+  both `AGENTS.md` and `CLAUDE.md` contain the separate `memoir` content.
+- The npm registry confirms the `nightly` dist-tag still points to
+  `0.2.22-nightly.20260821-051020.13f5d91`.
+
+Conclusion: the newly corrected memory implementation is not deployed, and
+Wing has not loaded it. Installing `@nightly` could only reinstall the old
+nightly because the dist-tag has not moved.

package/docs/journal/2026/08/22/alibaba-singapore-dual-endpoint-wire-alignment.okf

@@ -0,0 +1,142 @@
+okf_version: "1.0"
+knowledge:
+  task:
+    id: alibaba-singapore-dual-endpoint-wire-alignment-20260822
+    date: "2026-08-22"
+    objective: >-
+      Verify the two user-supplied Alibaba Model Studio Singapore API
+      references and apply both OpenAI Chat and Anthropic Messages wire
+      contracts without inferring the protocol from shared JSON fields.
+
+  official_sources:
+    anthropic_messages:
+      console_document_id: "2980295"
+      public_url: >-
+        https://www.alibabacloud.com/help/en/doc-detail/2980295.html
+      verified_base_url: >-
+        https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/apps/anthropic
+      verified_request_url: >-
+        POST https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/apps/anthropic/v1/messages
+      verified_auth: >-
+        Model Studio API key in either x-api-key or Authorization: Bearer.
+      verified_request_fields:
+        - max_tokens is required
+        - thinking is an optional Anthropic Extended Thinking object
+        - streaming emits thinking and text content blocks
+    openai_chat:
+      console_document_id: "3016807"
+      public_url: >-
+        https://www.alibabacloud.com/help/en/doc-detail/3016807.html
+      verified_base_url: >-
+        https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1
+      verified_request_url: >-
+        POST https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1/chat/completions
+      verified_auth: "Authorization: Bearer <Model Studio API key>"
+      qwen38_reasoning:
+        supported_efforts: [low, medium, xhigh]
+        documented_aliases:
+          max: xhigh
+          high: xhigh
+          minimal: low
+          none: enable_thinking_false
+        mutual_exclusion: >-
+          qwen3.8-max rejects reasoning_effort and thinking_budget when both
+          are present.
+        documented_budget_mapping:
+          low: 4096
+          medium: 16384
+          xhigh: 262144
+
+  pre_change_reproduction:
+    provider: alitoken
+    model: qwen3.8-max
+    selected_protocol: anthropic_messages
+    selected_base_url: >-
+      https://token-plan.ap-southeast-1.maas.aliyuncs.com/apps/anthropic
+    observed_wire_profile_thinking_policy: strip
+    observed_input:
+      max_tokens: 32768
+      thinking:
+        type: enabled
+        budget_tokens: 16384
+    observed_normalized_output:
+      thinking_present: false
+      max_tokens_present: false
+      max_completion_tokens: 32768
+    evidence_conclusion: >-
+      The endpoint selection was correct, but Alibaba request normalization
+      inferred OpenAI Chat from the shared messages field. It consequently
+      removed the native Anthropic thinking object and rewrote the required
+      Anthropic max_tokens field to the OpenAI max_completion_tokens field.
+
+  implementation:
+    architecture:
+      file: ciel_runtime_support/architecture.py
+      change: >-
+        Added a protocol-aware request-normalization hook. Existing adapters
+        retain their protocol-neutral behavior by default.
+    pipeline:
+      file: ciel_runtime_support/provider_policy.py
+      change: >-
+        Pass the already-resolved upstream wire protocol to provider request
+        normalization instead of requiring adapters to infer it from body
+        structure.
+    alibaba_provider:
+      file: ciel_runtime_support/providers/alibaba.py
+      changes:
+        - Mark the native Alibaba Anthropic contract as preserving Anthropic thinking.
+        - Preserve native Anthropic request fields and history blocks unchanged.
+        - Retain existing Qwen3.8 OpenAI Chat reasoning and token normalization.
+    compatibility_facade:
+      file: ciel_runtime.py
+      change: >-
+        Accept an optional MessageProtocol and project it through the provider
+        adapter facade without exceeding the monolith line-count ratchet.
+
+  post_change_verification:
+    native_anthropic_projection:
+      upstream_format: anthropic-messages
+      thinking_policy: preserve
+      thinking_preserved: true
+      historical_thinking_blocks_preserved: true
+      max_tokens_preserved: true
+      max_completion_tokens_added: false
+    openai_chat_projection:
+      reasoning_effort_medium_preserved: true
+      conflicting_thinking_budget_removed: true
+      max_tokens_projected_to_max_completion_tokens: true
+    automated_tests:
+      alibaba_provider: "34 passed"
+      architecture_budget: "2 passed"
+      unit: "1129 passed, 44 skipped"
+      router: "910 passed"
+      channel: "381 passed, 80 skipped"
+      runtime: "252 passed, 12 skipped"
+      aggregate: "2672 passed, 136 skipped"
+      ruff: passed
+      compileall: passed
+      diff_check: passed
+    live_singapore_requests:
+      configuration_scope: >-
+        Ciel workspace configuration for C:/Users/djlov/ciel-runtime; API key
+        presence was confirmed without printing the credential.
+      openai_chat:
+        url: >-
+          https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1/chat/completions
+        status: 200
+        elapsed_seconds: 2.469
+        finish_reason: stop
+        response_text: "OK."
+      anthropic_messages:
+        url: >-
+          https://token-plan.ap-southeast-1.maas.aliyuncs.com/apps/anthropic/v1/messages
+        status: 200
+        elapsed_seconds: 1.562
+        stop_reason: end_turn
+        response_text: "OK."
+
+  evidence_boundary: >-
+    The live probes establish that both configured Singapore endpoints,
+    bearer authentication, qwen3.8-max, and minimal non-streaming request
+    shapes were accepted. They do not benchmark comparative reasoning latency
+    or prove behavior for unrelated Alibaba models.

package/docs/journal/2026/08/22/delluhiold-claude-compaction-empty-tool-name.okf

@@ -0,0 +1,168 @@
+okf_version: "1.0"
+knowledge:
+  task:
+    id: delluhiold-claude-compaction-empty-tool-name-20260822
+    date: "2026-08-22"
+    objective: >-
+      Diagnose and repair the Claude /compact 400 error in the ai-net
+      workspace on delluhiold without attributing an unverified cause.
+
+  remote_evidence:
+    host: delluhiold
+    user: oneciel
+    workspace: /home/oneciel/ai-net
+    transcript: >-
+      /home/oneciel/.claude/projects/-home-oneciel-ai-net/
+      bd9b33a6-2be4-46ac-81bc-2bd5c5eed995.jsonl
+    transcript_scan:
+      lines: 42279
+      malformed_json_lines: 0
+      tool_use_blocks: 9005
+      invalid_empty_name_blocks: 3
+      invalid_lines:
+        - line: 42224
+          timestamp: "2026-08-22T05:21:37.787Z"
+          model: qwen3.8-max
+          tool_use_id: toolu_53d6b073b4cd405183a039cf
+          raw_arguments: "}"
+        - line: 42246
+          timestamp: "2026-08-22T05:36:06.176Z"
+          model: qwen3.8-max
+          tool_use_id: toolu_114f0f73eed54a8da8bda2f3
+          raw_arguments: ""
+        - line: 42248
+          timestamp: "2026-08-22T05:46:14.643Z"
+          model: qwen3.8-max
+          tool_use_id: toolu_0275f15ac8214d95be41982e
+          raw_arguments: "}"
+      paired_results: >-
+        Each invalid block is followed by a tool_result containing
+        "No such tool available" for the same tool_use_id.
+
+  endpoint_evidence:
+    configured_provider: alitoken
+    configured_base_url: >-
+      https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1
+    native_compat: true
+    resolved_claude_base_url: >-
+      https://token-plan.ap-southeast-1.maas.aliyuncs.com/apps/anthropic
+    resolved_messages_url: >-
+      https://token-plan.ap-southeast-1.maas.aliyuncs.com/apps/anthropic/v1/messages
+    conclusion: >-
+      Claude used the native Anthropic endpoint. The earlier OpenAI endpoint
+      attribution contradicted the provider configuration and source and was
+      withdrawn.
+
+  chronology:
+    corrupt_records_created_utc: "2026-08-22T05:21:37Z..2026-08-22T05:46:14Z"
+    current_nightly_installed_utc: "2026-08-22T06:06:04Z"
+    current_version: 0.2.22-nightly.20260822-030054.b70a228
+    current_output_reproduction: >-
+      Feeding the same empty-name Anthropic SSE block to the current runtime
+      emits no tool_use and downgrades the malformed tool_use stop to end_turn.
+    remaining_failure_before_patch: >-
+      The three historical invalid blocks remain in the resumed transcript and
+      are revalidated by compaction.
+
+  implementation:
+    file: ciel_runtime_support/anthropic_tool_turns.py
+    behavior: >-
+      Before provider-wire normalization, remove only assistant tool_use blocks
+      whose names are empty or whitespace and user tool_result blocks that
+      explicitly reference those invalid IDs. Preserve valid, unmatched, and
+      unrelated history.
+    pipeline: ciel_runtime_support/provider_policy.py
+    scope: all Anthropic Messages requests, including native Anthropic providers
+
+  verification:
+    focused_provider_wire_tests: "7 passed"
+    alibaba_provider_tests: "31 passed"
+    architecture_contract_tests: "261 passed, 42 skipped"
+    full_test_groups:
+      unit: "1129 passed, 44 skipped"
+      router: "907 passed"
+      channel: "381 passed, 80 skipped"
+      runtime: "252 passed, 12 skipped"
+      aggregate: "2669 passed, 136 skipped"
+    ruff: passed
+    diff_check: passed
+    deployment:
+      commit: d7d93697401503be7ac5d46c28b06339f31ec592
+      nightly: 0.2.22-nightly.20260822-062859.d7d9369
+      local_windows: verified
+      remote_prefix: /home/oneciel/.npm-global
+      remote_router_pid: 1871851
+      remote_router_source: >-
+        /home/oneciel/.npm-global/lib/node_modules/@oneciel-ai/ciel-runtime/
+        ciel_runtime.py
+      remote_provider_restored: anthropic
+    live_remote_request:
+      workspace: /home/oneciel/ai-net
+      provider_during_probe: alitoken
+      model: qwen3.8-max
+      input: >-
+        Included an assistant tool_use with name="" and its paired error
+        tool_result, matching the observed corrupt history shape.
+      status: 200
+      response_text: OK
+      router_activity_time: "2026-08-22T06:33:21Z"
+      router_activity_bytes: 303
+    evidence_boundary: >-
+      The original interactive Claude process had already exited before the
+      patched package was installed, so its /compact UI command was not rerun.
+      The live ai-net router and real Alibaba Anthropic endpoint accepted the
+      same invalid-history shape after request sanitization. The original
+      transcript file was not modified.
+
+  interactive_revalidation:
+    time_utc: "2026-08-22T06:44:07Z..2026-08-22T06:55:10Z"
+    prior_completion_claim_withdrawn: true
+    failed_launch_evidence:
+      executable: /usr/local/bin/ciel-runtime
+      installed_version: 0.2.22-nightly.20260822-063634.5223153
+      installed_sanitizer_sha256: >-
+        792c65dc1b7dc7edfbac5eb7f032edfa5455730aac0d8c643b6ee10dd3a296c5
+      active_provider: anthropic
+      active_mode: anthropic-native
+      router_requests_after_0641_utc: 0
+      result: >-
+        The failing resumed Claude process bypassed the router, so the request
+        boundary sanitizer was not invoked. The installed file itself was the
+        final patched version; deployment version mismatch was not the cause.
+    corrected_workspace_state:
+      workspace: /home/oneciel/ai-net
+      provider: alitoken
+      model: qwen3.8-max
+      mode: ciel-runtime-router
+      base_url: >-
+        https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1
+    actual_tty_compaction:
+      command: /usr/local/bin/ciel-runtime --continue
+      session_id: bd9b33a6-2be4-46ac-81bc-2bd5c5eed995
+      transcript_result: compact_boundary
+      trigger: manual
+      duration_ms: 303647
+      pre_tokens: 319439
+      post_tokens: 14108
+      compacted_at_utc: "2026-08-22T06:54:11.187Z"
+      precompact_hook: passed
+      postcompact_hook: passed
+      empty_tool_name_400_during_compaction: false
+    post_compaction_request:
+      prompt: "검증: OK라고만 답해"
+      upstream_model: qwen3.8-max
+      router_activity_time: "2026-08-22T06:55:10Z"
+      result: >-
+        The request reached the model and emitted valid named Bash tool calls;
+        it did not reproduce the empty-tool-name 400. The model ignored the
+        requested terse answer and resumed work from the compacted summary, so
+        the response was interrupted before further side effects and the TTY
+        session was exited normally.
+    final_state:
+      provider_preserved: alitoken
+      model_preserved: qwen3.8-max
+      active_test_processes: 0
+      router_after_exit: down
+      expected_next_launch: >-
+        The Ciel launcher starts and owns the workspace router when the user
+        next launches or continues Claude.

package/docs/journal/2026/08/22/remote-memory-relative-prompt-paths.okf

@@ -0,0 +1,79 @@
+okf_version: "1.0"
+knowledge:
+  task:
+    id: remote-memory-relative-prompt-paths-20260822
+    date: "2026-08-22"
+    objective: >-
+      Project remote-memory root and index locations into system context as
+      portable paths relative to the active workspace instead of machine-bound
+      absolute paths.
+
+  evidence:
+    before_change:
+      source: ciel_runtime_support/remote_memory.py
+      observed: >-
+        _current_memory_addresses validated the memory tree and returned
+        str(root) and str(path), causing Windows drive and user-profile paths to
+        be embedded in native instruction files and routed system context.
+    requested_example:
+      absolute_root: C:/Users/daniel.yun.WING/.ciel/memory
+      absolute_index: C:/Users/daniel.yun.WING/.ciel/memory/index.md
+
+  implementation:
+    validation: >-
+      The stored workspace, root, and index are still resolved to absolute
+      filesystem paths and rejected unless the directory and index exist below
+      the recorded active workspace.
+    projection:
+      root: .ciel/memory
+      index: .ciel/memory/index.md
+      format: portable POSIX-style workspace-relative path
+      guidance: >-
+        The model is explicitly told to resolve the paths from the current
+        workspace root before reading the index and related memory files.
+    configured_directory: >-
+      A CTR-configured directory such as agent-data/memory is preserved as
+      agent-data/memory and its nested index remains workspace-relative.
+    native_files: >-
+      Download-time projection and later pointer restoration both use the same
+      verified relative addresses, so copied CLAUDE.md, AGENTS.md, and GEMINI.md
+      files do not retain a source-machine user profile or drive letter.
+
+  verification:
+    status: passed
+    focused:
+      result: "37 tests run; all passed"
+      coverage:
+        - manifest download and native instruction projection
+        - protocol finalization and hard compaction
+        - custom memory directory projection
+    portability_assertions:
+      - "The prompt contains .ciel/memory and .ciel/memory/index.md."
+      - "The prompt does not contain the resolved active workspace path."
+      - "A custom agent-data/memory directory remains portable and relative."
+    full_suite:
+      total_run: 2691
+      passed: 2555
+      skipped_environment_dependent: 136
+      groups:
+        unit: "1148 run, 44 skipped"
+        router: "910 run"
+        channel: "381 run, 80 skipped"
+        runtime: "252 run, 12 skipped"
+    quality:
+      compileall: passed
+      ruff: passed
+      documentation_metadata: "passed; 22 Markdown files"
+      diff_check: passed
+      package_dry_run: "passed; 401 local entries"
+
+  deployment:
+    local:
+      status: passed
+      path: C:/Users/djlov/.local/share/ciel-runtime
+      evidence: >-
+        The installed remote_memory.py SHA-256 matched the tested source. An
+        installed-module execution projected agent-data/memory and its index,
+        included workspace-root guidance, and exposed no absolute workspace.
+    nightly:
+      mechanism: "git push origin HEAD:nightly triggers npm-publish.yml"