Skip to content

fix(responses): consolidate tool history and continuation boundaries - #5560

Closed
luvs01 wants to merge 11 commits into
devfrom
stack/native-continuation-boundaries
Closed

luvs01 wants to merge 11 commits into
devfrom
stack/native-continuation-boundaries

Conversation

@luvs01

@luvs01 luvs01 commented Sep 22, 2026

Copy link
Copy Markdown
Collaborator

Summary

Preserve tool history and current user scope across external-provider continuations by combining the unlanded changes in #5350 and #5420 with the stateful output-only corrections carried in 82a5f6da81807eee582d744ce0086f22b19bd021 and aac783fe8d4fd90667ccb8ea815d31cc014e798a.

  • Keep Cursor continuation scope on the latest actual user input through compaction, preserve evidence roles and bounded checkpoint fallback, and make code-mode guidance distinguish emitted output from an empty completed cell.
  • Lower valid historical custom calls for destinations that explicitly deny native custom tools without expanding the live tool catalog. Reject malformed, duplicate, mismatched, or colliding identities before dispatch.
  • Preserve stateful output-only deltas when their calls remain upstream, while keeping replay-miss reasoning cleanup independent. Repair stable xAI custom item IDs only on the native-compatible path.
  • Preserve current dev's stronger empty-catalog tools: [] authorization boundary. Its already-landed selector repair and role-fixture fixes are not reapplied. Source follow-ups, author identities, and test placement remain accounted for.

This is an integration proposal. The other contributors' original PRs remain open and retain their own review and ownership.

Verification

  • Prepared head 2ec0cd12f5bb1567b3723dbf13555046604aa083 on dev 41ec40f7e31969d1ca7140e0d510077f74f4cbe3, with all three source heads rechecked before integration.
  • New combined adapter matrix: 10 passed / 56 assertions, covering stateful outputs, independent reasoning cleanup, capability-driven lowering, missing-result synthesis, pairing identity, and deny-all preservation.
  • Existing focused suites: 366 passed, with one file-size scan timing out at the unchanged 5-second limit. The unchanged standalone ratchet then passed 9/9 at its normal limit. The first failed run is retained in the validation record.
  • Typecheck, structure SSOT, privacy scan, and git diff --check passed. Frozen-lockfile documentation install and the actual documentation build passed: 497 pages.
  • Sixteen key implementation/regression files match their source or the protected dev implementation byte-for-byte. Seven carried commits have identical stable patch IDs; the latest Cursor follow-up differs only in patch context. JSON layout-marker cleanup is incorporated into the integration resolution.
  • The broader local changed-import-graph run did not complete: exit 124 at its unchanged 900-second limit, with Windows cleanup/ACL/timeouts and an unresolved assertion in existing server tests. Its 224 failure lines are partial output, not a completed total or proven baseline failures. No named failure matched the focused passing cases, but broader attribution remains unresolved.
  • Exact-head Cross-platform CI run 35740377784 passed, including four Linux test shards, both macOS shards, desktop Rust checks, documentation and the aggregate ci check. Optional Windows diagnostic shards and unsharded macOS control were skipped; this is not a full Windows-suite pass. Local broader validation, live-provider acceptance and independent security review remain pending. Keep this proposal draft.

Checklist

  • Scope stays focused and avoids unrelated cleanup.
  • Docs or release notes were updated when needed.
  • Security-sensitive changes were reviewed for secrets, auth, and unsafe defaults.

Co-authored-by: Yeonwoo Choi 32544727+twoimo@users.noreply.github.com
Co-authored-by: maosisheng maosisheng@gmail.com
Co-authored-by: Cursor cursoragent@cursor.com
Co-authored-by: Epinephrine luvs01@hanmail.net
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

twoimo and others added 11 commits September 22, 2026 22:30
…ontinuations

xAI rejects tool_choice auto/none when no tools are declared. Cursor Grok 4.6
tool hops also drifted off the current user request, re-echoed exec output, and
failed closed on missing custom_tool_call ids or an exhausted checkpoint.

Keep the existing request shape: drop only auto/none with an empty catalog,
scope external continuations to the current user request, repair ctc_ ids for
xAI, and rebuild a full replay when a checkpoint no longer fits the envelope.
Invalid or unreadable inputs now fail closed to the previous safe default with
debug diagnostics instead of throwing out of request assembly.

(cherry picked from commit 669992f)
…tination denies them

Routed lowering collected only current custom declarations, so a compacted or replayed custom_tool_call leaked to xAI-like gateways as the native item type and came back as a misleading 422 missing id. Convert protocol-history items from the top-level input without expanding the live catalog, request full replay for orphan results, and fail closed before serializing leftovers.

Co-authored-by: Cursor <cursoragent@cursor.com>
(cherry picked from commit 5da2883)
… passthrough ratchet cap

openai-responses-passthrough.test.ts is already at its 4809-line ceiling. Keep the new wire fixtures in a responses-prefixed file so the layout seed resolves it without raising a cap.

Co-authored-by: Cursor <cursoragent@cursor.com>
(cherry picked from commit 6f43794)
…pair

Co-Authored-By: Epinephrine <luvs01@hanmail.net>
(cherry picked from commit 67ccd8d)
Exercise stateful output-only deltas, independent replay-miss reasoning cleanup,
capability-driven historical lowering, placeholder ordering, native item-ID
repair and preservation of the existing empty-catalog denial. Record the
combined history contract and register the carried and new regression files.

The layout-marker cleanup from e8e179f was
completed while resolving its preceding source commit onto the current map.
The existing dev selector normalization and role-fixture corrections remain
authoritative and are not replaced by weaker or duplicate source changes.

Co-authored-by: Yeonwoo Choi <32544727+twoimo@users.noreply.github.com>
Co-authored-by: maosisheng <maosisheng@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Epinephrine <luvs01@hanmail.net>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
@coderabbitai

coderabbitai Bot commented Sep 22, 2026

Copy link
Copy Markdown
Contributor

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

  • Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

Comment @coderabbitai help to get the list of available commands.

@github-actions

Copy link
Copy Markdown
Contributor

Deterministic PR hygiene checks passed.

@lidge-jun

Copy link
Copy Markdown
Owner

리뷰 · 우선순위 58 / 80

이 PR은 Cursor·xAI로 이어지는 도구 대화가 끊기거나 잘못된 요청을 다시 잡으려다 깨지는 문제를 한곳에 모은 통합 수정입니다. 예전에 열려 있던 #5350(요청 범위·도구 결과 이어가기)과 #5420(과거 custom 도구를 목적지가 거부할 때 낮춰 보내기)의 미머지 내용에, stateful 출력만 오는 이어가기와 재현(replay) 놓침 때 추론 정리를 서로 섞지 않는 수정을 얹었습니다. 베이스는 dev입니다.

구체적으로는 (1) Cursor 외부 모델 이어가기에서 최신 실제 사용자 요청만 활성 범위로 두고, 요약·브라우저 주변 문맥은 역사로만 두며, Grok 4.6 코드 모드에서 빈 완료 셀을 “이미 출력했다”고 오해하지 않게 안내하고, 체크포인트가 봉투 한도를 넘으면 한 번 전체 재현으로 다시 짜고, (2) custom 도구를 거부하는 목적지에는 살아 있는 도구 목록을 키우지 않고도 과거 custom 호출·결과를 function 형태로 낮추며, 잘못된·중복·충돌 identity는 보내기 전에 막고 HTTP 400으로 돌리며, (3) stateful 목적지에서 호출이 서버 쪽에만 있을 수 있는 출력만 오는 델타를 지우지 않고, 빈 카탈로그의 tools: [] 거부 경계는 현재 dev 쪽을 지킨다고 합니다. 포커스 테스트와 합친 매트릭스는 통과했다고 적혀 있으나, 넓은 import-graph 검증은 900초에서 끊겼고, 보안 체크박스와 전체 CI·라이브 수락은 아직 비어 있습니다. Draft로 두는 판단은 본문과 맞습니다.

src/responses/custom-tool-compat.ts (rewriteHistoricalCustomItems / validateFinalCustomToolCompatibility) - 거부 목적지에서 과거 custom을 낮추고, 남는 native custom은 최종 가드에서 막습니다. 방향은 맞지만, 잘못 생긴 호출은 500이 아니라 400으로 떨어집니다. 클라이언트가 “재시도하면 고쳐진다”고 오해할 수 있어, 어떤 입력이 영구 거부인지를 응답 메시지·문서에 더 짧게 적어 두는 편이 좋습니다.

src/adapters/openai-responses/tool-output-recovery.ts (repairOrphanedInputItemsrepairOrphanOutputs) / passthrough.ts - stateful + 미전개 previous_response_id일 때 orphan 출력을 사용자 메시지로 바꾸지 않습니다. 의도(서버에 호출이 있을 수 있음)는 분명합니다. 반대로 전개 실패가 “진짜로 호출이 없는” 경우와 구분되는지는 넓은 스위트·실기기에서만 드러납니다. 지금 기록된 넓은 실행은 타임아웃이라 그 증거가 아직 없습니다.

src/adapters/cursor/protobuf-request.ts (체크포인트 envelope_exhausted 시 재호출) - 체크포인트 필드를 지우고 한 번 다시 짜는 형태라 무한 재귀는 막혀 보입니다. 다만 재현 경로에서도 같은 한도에 걸리면 그대로 에러입니다. 그 실패가 사용자에게 “이어가기 불가”로 읽히는지, 아니면 다른 복구로 넘어가는지는 이 diff만으로는 확인되지 않습니다.

본문 Verification - 포커스 366통과·합친 매트릭스 10통과는 적혀 있고, changed-import-graph는 exit 124·부분 실패 224줄로 미완료입니다. Cross-platform CI도 아직 큐/진행 중인 항목이 있습니다. Draft를 Ready로 올리기 전에 “넓은 실행이 무엇을 증명했는지”가 본문 주장과 같아야 합니다.

메인테이너의 판단이 필요한 지점

#5350과 #5420은 여전히 열려 있습니다. 이 통합 PR이 dev에 들어가면 원본 PR을 superseded로 닫을지, 원본만 남기고 이 PR을 닫을지, 아니면 일부를 되돌릴지 소유권 정리가 필요합니다. 같은 계약을 두 갈래로 머지하면 충돌·중복 수정이 납니다. 또한 빈 카탈로그 tools: [] 거부 경계를 “지킨다”는 주장과, historical lowering이 과거 이름을 현재 권한으로 키우지 않는다는 주장은 보안 리뷰 한 번의 대상입니다. 체크리스트의 보안 항목이 아직 비어 있습니다.

너의 추천

Draft는 유지한 채, Cross-platform CI와(가능하면) 끊긴 import-graph를 끝내거나 실패를 이번 변경과 무관하다고 이름 단위로 가른 뒤에만 Ready로 옮기세요. 머지 시 #5350·#5420은 이 PR에 흡수된 범위만큼 닫아 중복을 막는 편이 낫습니다. 보안은 custom-tool 낮추기·최종 가드·빈 카탈로그 거부 세 줄만 짧게 독립 확인하면 됩니다. 포커스 테스트 방향은 타당해 보이며, 넓은 검증이 비어 있는 상태에서는 머지하지 않는 것이 맞습니다.

이 댓글은 grok-bot이 작성했습니다

@lidge-jun

Copy link
Copy Markdown
Owner

Carried in #5608: nine commits are cherry-picked with their authors kept, and the first two land as their net change because the xAI empty-catalog selector part is already on dev in b20acc7 (#5376). A follow-up pins the exact host-wrapper classification in continuation scope. The source PRs #5350 and #5420 are credited in #5608 and remain with their authors. Closing as superseded. Thank you @luvs01.

@lidge-jun lidge-jun closed this Sep 22, 2026
lidge-jun added a commit that referenced this pull request Sep 23, 2026
…y work (#5608)

* docs(devlog): plan the L5 luvs01 Responses and usage bundle

Record the disposition, provenance and ordered carry recipe for #5474, #5305, #5434, #5560, #5542, #5553, #5562 and #5556.

* fix(usage): verify OAuth rotation before withheld attribution

(cherry picked from commit 1ac1ba0)

* test(oauth): cover the eligible-target positive case and gate assertion

(cherry picked from commit f6778bf)

* fix(oauth): require a live two-account roster before withheld attribution

The eligibility probe added for #5434 could report an alternate when the failed account had been removed and one stored account remained, because it skipped the roster-size guard that rotation applies. Read the roster once, apply the same guard, and describe the probe as applying no cooldown rather than as non-mutating.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* fix(adapters): close Grok continuation and normalized tool catalog gaps

Keep the existing request shape, scope external Cursor continuations to the
current user request, repair ctc_ ids for xAI, and rebuild a full replay when
a checkpoint no longer fits the envelope. Invalid or unreadable inputs fail
closed to the previous safe default with debug diagnostics.

This is the net change of two source commits. Their empty-catalog selector
edit to src/adapters/xai-web-search.ts is omitted because dev already carries
that rule (b20acc7, #5376), and the temporary test additions to capped files
are replaced by the final extracted test files.

(cherry picked from commit 71a9fe5)
(cherry picked from commit ffd50f4)
Refs #5350, #5560

* fix(cursor): preserve continuation scope and avoid false repetition recovery

(cherry picked from commit 5a99d4d)
(cherry picked from commit 31f21f0)

* fix(responses): lower undeclared historical custom tools when the destination denies them

Routed lowering collected only current custom declarations, so a compacted or replayed custom_tool_call leaked to xAI-like gateways as the native item type and came back as a misleading 422 missing id. Convert protocol-history items from the top-level input without expanding the live catalog, request full replay for orphan results, and fail closed before serializing leftovers.

Co-authored-by: Cursor <cursoragent@cursor.com>
(cherry picked from commit 5da2883)
(cherry picked from commit a326b67)

* test(responses): split historical custom-tool replay coverage off the passthrough ratchet cap

openai-responses-passthrough.test.ts is already at its 4809-line ceiling. Keep the new wire fixtures in a responses-prefixed file so the layout seed resolves it without raising a cap.

Co-authored-by: Cursor <cursoragent@cursor.com>
(cherry picked from commit 6f43794)
(cherry picked from commit 7e8fb09)

* fix(responses): reject malformed historical custom calls

(cherry picked from commit 52f7448)
(cherry picked from commit 57407be)

* fix(responses): bind historical outputs to custom calls

(cherry picked from commit 5654b41)
(cherry picked from commit c7781bf)

* fix(responses): reject duplicate historical call ids

(cherry picked from commit dc948dc)
(cherry picked from commit fbecefa)

* fix(xai): preserve stateful tool output continuations

(cherry picked from commit 4edc411)
(cherry picked from commit 82a5f6d)

* fix(xai): keep replay-miss reasoning cleanup independent of output repair

Co-Authored-By: Epinephrine <luvs01@hanmail.net>
(cherry picked from commit 67ccd8d)
(cherry picked from commit aac783f)

* test(responses): verify combined continuation boundaries

Exercise stateful output-only deltas, independent replay-miss reasoning cleanup,
capability-driven historical lowering, placeholder ordering, native item-ID
repair and preservation of the existing empty-catalog denial. Record the
combined history contract and register the carried and new regression files.

The layout-marker cleanup from e8e179f was
completed while resolving its preceding source commit onto the current map.
The existing dev selector normalization and role-fixture corrections remain
authoritative and are not replaced by weaker or duplicate source changes.

Co-authored-by: Yeonwoo Choi <32544727+twoimo@users.noreply.github.com>
Co-authored-by: maosisheng <maosisheng@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Epinephrine <luvs01@hanmail.net>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
(cherry picked from commit 2ec0cd1)

* fix(responses): normalize wrapped MCP tool names

(cherry picked from commit b4c839b)
(cherry picked from commit f68b32646069d3e385748fc6c36848765c68e02e)
(cherry picked from commit e555e73)

* fix(responses): unwrap default apply patch aliases

(cherry picked from commit 453df76)
(cherry picked from commit d63ff542415895fa25f5599aad9dc632c80de92d)
(cherry picked from commit 7cbbf44)

* docs(structure): record default.-prefixed alias recovery in the freeform contract

The responses owners still claimed only bare exec/apply_patch calls accept alternate-field or outer-fence recovery; compileCodeModeHelperInput now also unwraps provider-invented default. aliases. Update runtime.md, transports/responses.md, providers/{chat-compat,kiro,xai-grok}.md.

Co-Authored-By: Epinephrine <luvs01@hanmail.net>
(cherry picked from commit 165ddd0)
(cherry picked from commit d0776e703681729e2ea447775e77397faab16f29)
(cherry picked from commit 19a2005)

* fix(responses): compile structured shell payloads sent to code-mode exec

(cherry picked from commit 9430bbd)
(cherry picked from commit 59bc70bc30f954eea5a63f8a6fcf4de4dcbfd865)
(cherry picked from commit 9662528)

* test(responses): verify combined tool normalization boundaries

Verify that default.apply_patch names and patch/content wrappers emit the same executable input through JSON and fragmented SSE. Move the carried decision record to unused ADR-0099 because current dev already owns ADR-0097; preserve both the structured code-mode shell and existing schema-bound flat-shell contracts.

Co-authored-by: kosta <kosta963@gmail.com>
Co-authored-by: Flowershangfromthebranches <152056395+Flowershangfromthebranches@users.noreply.github.com>
(cherry picked from commit 3da6366a60ac5964e8a8de43486125d72729f8c8)
(cherry picked from commit b57d7c5)

* fix(responses): refuse ambiguous OpenCode Go resets

(cherry picked from commit 62ac159)
(cherry picked from commit b8f9a45)

* test(responses): prove OpenCode Go pre-answer reset refusal by execution

Per review on #5446: replace the replaySafe source-string count with an execution test that drops the connection before the answer on an opencode.ai/zen/go destination and asserts the 429 upstream_reset_replay_refused with exactly one send.

(cherry picked from commit 92b74ec)
(cherry picked from commit 808dd85)

* fix(routing): isolate policy retry body snapshot

(cherry picked from commit 8e2a0fe)
(cherry picked from commit db854bf)

* test(routing): pin the retry snapshot against nested input mutation

(cherry picked from commit a240fc9)
(cherry picked from commit b037810)

* fix(responses): isolate policy compaction state

(cherry picked from commit 714119e)
(cherry picked from commit e6f9339)

* fix(tests): bound the cold-spawn warm-up child on a live event loop

The warm-up child was waited on through Bun.spawnSync, which made the spawn's
own timeout the only bound it could honour — and no bound at all when the
child or the primitive wedged: while a synchronous spawn blocks, the event
loop is dead, so the hook budget and the per-test timeout freeze inside the
same wait and nothing reports anything. Run 35511743422's macos 2/2 leg held
that shape for eighteen silent minutes inside client-connect.test.ts before
the job ceiling cut it and reported cancelled, which the ci gate reads as
failure.

The bound now lives on the parent's live loop: an asynchronous spawn, SIGKILL
at the existing derived deadline, a short reap grace, and the call settles
with or without the child's exit or EOF — so a descendant holding the pipes
or a child that outlives its kill cannot turn a warm-up into an unbounded
wait. A timed-out child now fails the warm-up by name instead of hanging the
job.

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
(cherry picked from commit aa9b889)
(cherry picked from commit f86a534)

* fix(responses): fail closed on synthetic or stale compaction source selectors

compactionRoutingKeepsProviderIdentity evaluated override.sourceModel as the
raw client selector, so a synthetic --fast/--effort form of a policy or combo
selector (ocx/primary--fast) missed resolvePolicyProfileId/resolveComboId and
fell through routeConcreteModel to the default provider. The same fallthrough
swallowed policy aliases renamed or deleted mid-conversation, since
config.routingProfiles is mutated in place. Both cases could wrongly report
identity match and let provider-private compaction state or caller credentials
cross a backend boundary.

Strip synthetic-row suffixes via parseSyntheticRowId before the identity
checks, and treat a source that only routes through the default provider as
unproven: it can never match a concrete identity.

Co-Authored-By: Epinephrine <luvs01@hanmail.net>
(cherry picked from commit 7e59315)
(cherry picked from commit 76b40f9)

* fix(codex): bind scoped quota suppression to alternate

(cherry picked from commit 23a3694)
(cherry picked from commit 385f338)

* fix(ci): restore core.ts to file-size ratchet cap

The scoped-quota re-export grew src/server/responses/core.ts past its committed 210-line cap (213). Collapse the two-name re-export back to one line; the file's export list already carries longer single-line statements.

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
(cherry picked from commit 559db16)
(cherry picked from commit feb0c16)

* fix(codex): record wrapped quota on suppressed moves and bind caller main

- record the normalized 429/402 outcome before returning no-alternate on a
  suppressed same-workspace move, so a 5xx-wrapped quota refusal still cools
  the refused account instead of reading as transient
- bind a request-owned `main` alternate by the caller credential's own
  workspace id (chatgpt-account-id header, else the bearer token's account
  claim) via callerCodexWorkspaceAccountId
- apply the same scoped-quota workspace gate to the single bounded alternate
  send in the native /responses/compact path
- cover all three in tests and update the transport doc

Co-Authored-By: Epinephrine <luvs01@hanmail.net>
(cherry picked from commit 5267593)
(cherry picked from commit 466c75c)

* test(server): move scoped-quota auth cases into a sibling file under the size cap

server-auth.test.ts grew to 4684 against a 4589 baseline cap, so the
file-size ratchet failed shard 3/4. The three scoped-quota suppression
cases move byte-for-byte into server-auth-scoped-quota.test.ts, and the
pool-retry harness they share is extracted to
tests/helpers/pool-retry-harness.ts (per-run OPENCODEX_HOME dir, so each
importing file keeps its own module state under bun test --isolate).

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
(cherry picked from commit b391c99)
(cherry picked from commit 9050722)

* fix(codex): re-check abort after the scoped-quota body read

The workspace classification in shouldRetryCodexScopedQuotaOnAlternate
reads the first response body asynchronously, so a client disconnect can
land after the earlier abort check but before the branch records the
first account, cancels its body, and sends the alternate. Re-check the
abort signal immediately after the await in both paths: compact returns
the 499 client_cancelled response after releasing the alternate lease,
and the regular path releases its permit and lease and returns
no-alternate while still recording the first account's real outcome.

Co-Authored-By: Epinephrine <luvs01@hanmail.net>
(cherry picked from commit c52b64b)
(cherry picked from commit b2eda92)

* fix(codex): release the discarded compact rejection body on abort

The 499 exits around the scoped-quota classification return a fresh
response while the first rejection's body is still open; cancel it so
the abandoned upstream connection and tee resources are released.

Co-Authored-By: Epinephrine <luvs01@hanmail.net>
(cherry picked from commit 813efcc)
(cherry picked from commit 1069b54)

* fix(codex): keep compact abort cleanup off the return path

upstream.body.cancel() can wait on a custom or stalled source; awaiting
it at the abort checkpoints would park the 499 reply on cleanup. Fire
it with the request's abort reason and swallow rejection, the same
best-effort shape bufferCompactResponse already uses.

Co-Authored-By: Epinephrine <luvs01@hanmail.net>
(cherry picked from commit 2d1ee69)
(cherry picked from commit 37a006e)

* test(responses): preserve terminal refusal across recovery boundaries

Exercise both operator-granted and default-denied reset paths through policy fallback and the alternate-account eligibility gate. Clarify that retries serialize the original body snapshot while identity metadata is established per attempt, and document synthetic compaction identity and transient replacement refusal.

(cherry picked from commit be1fee9)

* test(cursor): pin exact host-wrapper classification in continuation scope

The wire carries no provenance for compaction summaries or ambient browser state, and the Codex client detects stored summaries by the same exact prefix. Pin that a user message that is itself an exact wrapper stays in history while the preceding real request remains the labeled one, and document the rule next to the continuation contract.

Co-authored-by: Yeonwoo Choi <32544727+twoimo@users.noreply.github.com>

* test(server): prove a suppressed same-workspace alternate is never sent

Both credentials in the 5xx-wrapped scoped-quota case carry the same workspace header, so a negative check on the other account id could not fail. Assert the exact one-send sequence, and narrow the transport contract: suppression covers the in-request move, while later requests still select by per-account health.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* docs(devlog): record the L5 responses sequence outcome

* docs(devlog): amend the L5 search and usage plan after review

* fix(responses): isolate Cursor combo shadow calls

(cherry picked from commit 3e5a4dd)
(cherry picked from commit 7f45883)

* fix(responses): repair terminal-less bridged search legs

(cherry picked from commit 1fb6005)
(cherry picked from commit c4fa8c8)

* fix(responses): repair continuation legs and prove behavior in tests

(cherry picked from commit 3948970)
(cherry picked from commit 8d46989)

* fix(responses): close shadow combo intersection and continuation coverage

Apply the existing source-target non-intersection rule before early combo interception, and add production-path behavioral coverage for terminal-less continuation repair.

(cherry picked from commit 4bfcc0a)
(cherry picked from commit 3f3fdf1)

* test(web-search): preserve complete but open bridge-leg coverage

Carry the end-to-end handleResponses regression and transport contracts for repairing both the first and continuation search legs. The corresponding production changes are already preserved by the earlier terminal-repair carries; keep this broader integration coverage without applying that implementation twice.

Source commit: b1044e7
Co-authored-by: Epinephrine <luvs01@hanmail.net>

(cherry picked from commit 973a4ac)

* test(web-search): keep repaired replay within caller and serving scope

(cherry picked from commit bb49c9f)

* test(server): reap fixture ACL workers before removing failover homes

(cherry picked from commit ae52669)

* fix(responses): keep combo shadow interception on the dispatch pick

The carried early intersection check sampled a combo target with routeModel before dispatch. The combo loop then picked again, so the interception decision could follow a different target than the one sent, and a random or round-robin combo advanced its selection state for a probe. Restore the existing combo rule (a shadow call rewritten to a combo enters the combo) and its regression test, and keep the combo-child isolation marker.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* test(web-search): bind repaired-leg replay to a keyed caller principal

The repaired-leg replay assertion expected a shared keyless loopback principal, which the replay cache refuses by design: keyless loopback callers get no retained replay. Configure an inbound API key, derive the expected principal with the production resolver, pass the full loopback admission, and assert that a keyless caller cannot form a scope.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* fix(usage): restrict legacy -main provider collapse to Codex labels

A -main suffix was stripped from ANY provider name, so a configured provider like openrouter-main silently merged into openrouter in usage summaries and pricing fallback. Restrict the compatibility mapping to the known Codex provider labels (openai/chatgpt/openai-multi-main), matching codexAccountProviders.

(cherry picked from commit 89a9272)
(cherry picked from commit 0f0ef96)

* fix(cli): mark access-key usage unavailable when attribution is absent

Without attributionSince the server is reporting an empty or unreadable usage log, but the table still rendered 0 totals and never-used markers that are indistinguishable from real data. Show an unavailable marker spanning the usage columns instead, matching the ambiguous-union precedent.

(cherry picked from commit 3c3f14b)
(cherry picked from commit 83514c3)

* fix(cli): treat a malformed attributionSince as unavailable

typeof === 'string' accepted any value, so a malformed attributionSince made
usageAvailable true and printed usage cells plus an 'attribution since'
footer. Parse once, derive availability from the validated result, and reuse
it for the footer. Covers the malformed-string regression.

(cherry picked from commit e7584c3)
(cherry picked from commit 1380693)

* fix(gui): keep the app-server status read alive across Models tab switches

(cherry picked from commit 9e6870d)
(cherry picked from commit 5563577)

* fix(usage): bound persisted requestedModel selectors

Selectors are not length-bound at admission, so persist over-long values as a bounded prefix + sha256 digest of the full selector and encode the request-history requested_model filter input identically. Two selectors sharing the bounded prefix stay distinguishable instead of colliding on a truncated prefix.

(cherry picked from commit e3d0f9b)
(cherry picked from commit c8a9d1a)

* fix(spend): enforce ceilings on native chat sends

(cherry picked from commit 347204f)
(cherry picked from commit 823a7d2)

* fix(usage): preserve observation and native reservation boundaries

(cherry picked from commit 22ee516)

* test(usage): isolate native spend regressions within file caps

(cherry picked from commit 1a8d5f7)

* test(usage): drain fixture owners before removing Windows homes

(cherry picked from commit 2241d03)

* test(usage): batch canonical pagination fixture writes

(cherry picked from commit ddfef13)

* fix(cli): accept only an ISO-8601 UTC attributionSince

Date.parse accepts strings such as "0" and rolls impossible dates over, so a malformed attributionSince could still switch the key table to numeric usage and never-used cells. Require the ISO-8601 UTC shape the server emits and an instant that round-trips to the same second.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* docs(usage): state the aliasing limit of the idempotent selector encoding

Encoded requestedModel values must map to themselves because rows are normalized again on read, so a literal selector equal to another selector's persisted form shares its identity. Say so where the encoding is defined, and pin the behavior in the request-history test so the limit is deliberate rather than silent.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* docs(devlog): record the L5 search and usage outcome

* docs(devlog): record the L5 pull request

* docs(structure): keep runtime.md within its 600-line budget

The default.-prefixed freeform wording wrapped one extra line and pushed structure/runtime.md over the structure SSOT budget. Reflow the paragraph; the wording is unchanged.

* refactor(cursor): move current-request selection out of protobuf-request

The carried continuation-scope helpers brought src/adapters/cursor/protobuf-request.ts to 2,000 lines, the file-size threshold for files without a recorded cap. Move the current-request selection and its host-wrapper classification to a sibling module unchanged; the caller passes its text extractor.

Co-authored-by: Yeonwoo Choi <32544727+twoimo@users.noreply.github.com>

* docs(devlog): record the #5553 re-pin in the L5 plan

* docs(devlog): record the L5 original closures

* fix(usage): keep the native Chat spend estimate out of recorded usage

The native Chat spend reservation reused usageLogInputTokens for its input estimate. That field also feeds attempt usage, so native Chat attempts started recording an estimate-inflated input instead of the provider-reported counts (hosted CI: server-xai-oauth-401-replay and server-key-failover-e2e). Reserve the estimate through a spend-only spendInputEstimateTokens field; the reservation still charges it before each physical send.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* test(server): identify the suppressed alternate by its credential

Hosted CI showed the send layer retrying the refused account's transient 502 on the same credential, so an exact one-dispatch count was the wrong oracle. Both accounts share a workspace header; record the credential each physical send presents and assert the suppressed alternate's credential never appears.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* fix: close L5 retry, continuation, and history review gaps

Classify spend refusals on transient retry legs, keep Cursor retry provenance in checkpoint actions, and rebuild legacy history selectors into the bounded projection. Scope policy snapshots and correct warm-up deadline handling, with focused regressions and documentation.

* fix: retain policy fallback for decorated selectors

Normalize supported Fast and effort policy selectors before snapshot gating, restore a real policy-hop fixture, and align Cursor wire assertions with retry provenance guidance.

---------

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>
Co-authored-by: Yeonwoo Choi <32544727+twoimo@users.noreply.github.com>
Co-authored-by: twoimo <twoimo@twoimoui-MacBookPro.local>
Co-authored-by: maosisheng <maosisheng@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Epinephrine <luvs01@hanmail.net>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: kosta <kosta963@gmail.com>
Co-authored-by: Flowershangfromthebranches <152056395+Flowershangfromthebranches@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

bug Something isn't working

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants