You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 567a6d4
Browse filesBrowse the repository at this point in the historyBrowse files
Copy file name to clipboardExpand all lines: structure/04_transports-and-sidecars.md
+17Lines changed: 17 additions & 0 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -437,6 +437,23 @@ pre-compaction checkpoint is not persisted for later carry-forward.
437
437
- 장점, 단점 및 영향: Active-context reporting stays monotonic within an uncompacted Cursor conversation; no-checkpoint turns remain estimated; a process restart loses the numeric cache, and when neither a checkpoint nor a carry-forward is available the turn reports a request-local estimate derived from the same pruned payload sent to Cursor (#373 — reporting output-only usage made Codex read the context as nearly empty). Estimates are never persisted or promoted into checkpoint carry-forward; only live checkpoint frames update the cache.
438
438
```
439
439
440
+
## Google thought-text visibility boundary
441
+
442
+
Google-family responses may represent model-internal reasoning as a text-bearing part with
443
+
`thought: true`. The Google adapter maps that text to the internal `reasoning_raw_delta` event;
444
+
only text without the marker becomes visible `text_delta`. Streaming SSE and buffered JSON share
445
+
one classifier so transport selection cannot change whether provider-declared reasoning is shown
446
+
as assistant output. Thought-signature observation still runs on the original parts before text
447
+
classification, preserving the opaque continuation state independently of display semantics.
448
+
449
+
[Decision Log]
450
+
- 목적과 의도: Prevent provider-marked internal reasoning from appearing as ordinary assistant text while preserving reasoning and tool-call continuation.
451
+
- 기존 구현 및 제약 조건: Both Google response paths emitted every non-empty `Part.text` as visible text; function calls, inline images, and Antigravity/Vertex thought-signature replay already depended on the original part ordering.
452
+
- 검토한 주요 대안: Drop thought text; classify it separately in each parser; remove the marker and keep visible text; use one shared classifier without mutating the provider parts.
453
+
- 선택한 방식: Map `thought: true` text to `reasoning_raw_delta` through one helper used by streaming and buffered parsing, leaving part order and signature observation unchanged.
454
+
- 다른 대안 대신 이 방식을 선택한 이유: Dropping the text loses reasoning replay/display policy input, while duplicated parser rules can drift and exposing marked thoughts violates the provider's visibility boundary.
455
+
- 장점, 단점 및 영향: Internal reasoning no longer leaks into normal answers and both transports stay consistent; downstream reasoning policy still decides whether raw reasoning is rendered or only preserved, and malformed non-boolean markers remain ordinary text rather than broadening hidden-content inference.
456
+
440
457
## Google tool-call thought-signature replay
441
458
442
459
Gemini may attach an opaque `thoughtSignature` to a `functionCall` and requires that exact value on
0 commit comments