Repository navigation
fix: windows, assignment notes and lost information (2026-10-23 review, 4.101.0) - #173
Merged
Merged
Conversation
…-10-23 review (4.101.0)
Window (max length)
- A window learned from an overflow is used on the next turn: Stage 6 learns it under the names
the host resolved the window with (state.metadata), and provider aliases are folded (vllm runs
on the "custom" client, Azure on "azure_foundry" with an /openai/v1 URL).
- Azure reads its catalog rows (azure_foundry) and the row's underlying model name for the table.
- Catalog window values given as text count; true no longer means a window of 1.
- Overflow messages: vLLM 0.20 serving input count, vLLM V1 engine decoder prompt (the encoder
variant is not a window), llama.cpp exceed_context_size (n_ctx, n_prompt_tokens).
- The budget keeps room for the output cap the client really sends (thinking raises it), and a
max_tokens above half the window is lowered to fit (8k vLLM + 8192 was refused every time).
Assignment memory
- A rewrite writes the stored note (a loose name made a second copy), keeps the category, tags and
importance it does not give, and an assignment note stays one ("done" is added).
- Tags are cleaned (#, quotes, JSON array text). memory_read shows only the reader's own
conversations when the host answers visible_conversations(ids). The file backend keeps
written_in. Pinned-facts marker counts what is shown. Retired events name their release.
Lost information
- A tool result clipped to the window is recorded as the tool returned it; only the request is cut.
- A clipped Read (also in ToolBatch, and files attached to a cut reference block) is not a read,
and an Edit alone is not either: Write could replace a file whose middle was never seen.
- Checks that fail after a tool ran withhold the result instead of saying access was denied.
- The clamp notice goes after an [ERROR] answer, not before it.
- SSH timeouts keep the output; Grep stops inside a huge file and says where.
- Markers name only tools the model can call; omitted tool blocks of older turns, the earlier
conversation string, reference material and the question say how to read or what to do; the
memory node's earlier conversation is cut only under window pressure; ToolResultRead pages fit
the window; the "interrupted" tail no longer guesses the cause; the turn's text list resets.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
무엇을 고쳤나 (2026-10-23 전체 흐름 검토)
검토에서 확인한 문제를 고친다. 자세한 항목은 CHANGELOG 4.101.0.
창(최대 길이)
azure_foundry)과 그 행의 실제 모델 이름을 읽는다. 카탈로그 값"131072"를 받고true를 받지 않는다.assignment 메모리
memory_read는 호스트가 주면 읽는 사람의 대화만written_in에 보인다. 파일 저장소도written_in을 지킨다.정보 유실 방지
[ERROR]뒤에.시험
tests/unit/test_review_2026_10_23.py(핵심 동작 15개) + 바뀐 동작에 맞춘 기존 시험 3곳.--check.🤖 Generated with Claude Code