Skip to content

feat: bring memory and middleware credential rotation into 3.1.0-beta2 - #2716

Merged
likaiaki merged 166 commits into
feat/3.1.0-beta2from
feat/3.1.0-beta2-lkk
Nov 17, 2026
Merged

likaiaki merged 166 commits into
feat/3.1.0-beta2from
feat/3.1.0-beta2-lkk

Conversation

@likaiaki

Copy link
Copy Markdown
Collaborator

Summary

  • Brings this branch's long-term memory work and the original-installer middleware password rotation onto feat/3.1.0-beta2.
  • Merged the current target first. The 20 content conflicts are resolved in f57572670b, and a fresh merge against feat/3.1.0-beta2 is clean.
  • Request scopes from the target are kept. Memory routes use the tenant scope, and the anonymous browser-bucket check uses the public scope. Workflow chat keeps both the target's browser execution scope and this branch's memory registration.
  • Image-generation error 15033 stays with the target. Tool egress errors move to 15037–15040 so the two features do not share a code.

Test plan

  • Confirm the PR is mergeable and GitHub reports no conflicts with feat/3.1.0-beta2
  • Run the tool egress and image-generation tests that assert 15033 and 15037–15040
  • Open a workflow chat and a memory page on a build from this branch
  • Run bash docker/install.sh against an existing deployment and confirm password rotation still reaches the ready check

Made with Cursor

likaiaki and others added 30 commits November 10, 2026 22:02
Disable proxy buffering on the assistant SSE response and emit an empty delta while a tool call produces no tokens, so the API no longer arrives in bursts or looks idle.

Co-authored-by: Cursor <cursoragent@cursor.com>
- Restrict /api/v1/tool/test and URL-based /api/v1/tool/schema to global
  superusers; content-only schema parsing is unchanged.
- Route both entries through a pinned public-only transport: all DNS/IP
  answers validated, connection fixed to validated IPs, original Host/SNI
  and TLS verification kept, redirects and proxies refused, identity
  encoding, 1MiB/2MiB body limits and 30s deadline.
- Normal OpenApiTools/ToolExecutor/BoundedOpenApiTools execution keeps
  HEAD behavior; no private-origin allowlist is provided.
Adds the two research write-ups and the PRD covering daily chat,
deep research, assistants and workflows.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Registers the memory domain objects and candidate invariants
INV-44 to INV-47 in the v3.0.0 release contract.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Splits acceptance criteria into test-verified and eval-verified,
keeps memory on user disable and tenant change, and rewords the
candidate invariants accordingly.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…ment entrypoints

Externalize encryption keys, encrypt middleware configuration, and guide
fresh installation and legacy adoption through the existing install command.
Add native credential adapters, resumable database rewrap, deployment tests,
and HTML governance, operations, and local acceptance documentation.

Validated 231 focused tests, 9 live component tests, original-command retry,
key recovery, legacy migration, five-component rotation, and paired rollback.
DM and remaining full-release acceptance gates remain explicitly open.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Adds entry-point eligibility registration, explicit tenant+user
filtering, tool-event redaction, call-time injection and soft
delete; syncs spec AC-71 and edge cases.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Registers eligibility only at logged-in endpoints, completes the
tool-event redaction points, and adds a verify-first list.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Consolidates deep-research turns by turn record, injects the
registration callback from logged-in endpoints, and reads the
workflow switch from the online version.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Keeps the per-user, assistant and workflow switches; the
workstation config only carries the consolidation model.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Browser-facing MinIO paths become /object/{bucket}/{key}. Nginx proxies
the fixed /object/ prefix and asks the backend (auth_request) whether the
bucket is the current public or temp bucket, so renaming buckets no
longer requires nginx changes. Stored legacy /{bucket}/... paths are
rewritten on read; no data migration.

Co-authored-by: Cursor <cursoragent@cursor.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e 121

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
jieyuhuayang and others added 28 commits November 14, 2026 03:10
…rding

- Redacted memory tool outputs keep the content-free ok / reason_code, so a
  step can say "已记下 1 条记忆" or "记忆写入被拒绝:内容里有相对时间"
  instead of a bare "记忆操作" (daily chat timeline, deep research steps,
  assistant run cards share describeMemoryStep). Legacy events: a write with
  no item id reads as rejected.
- The receipt under an answer lists one memory per row (operation only when
  mixed, content, undo); a correction shows the old text struck through.
- "常驻" no longer appears anywhere in the UI (user ruling 2026-11-14):
  attribution is one list, the settings list drops the tag; spec AC-87/88.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ory list

MessageFeedbackButtons renders display: contents, so placing it directly in
the flex-col wrapper added for the memory notice stacked the two buttons.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Resident memory is carried into every turn, so listing it made unrelated
items appear under every answer (a sorting question showed a meeting date).
has_attribution now means the turn retrieved items, in the daily, assistant
and deep research signals and in history flags; the list shows retrieved
items only. Revises AC-87 and design decision 16.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…emory switch

Templates saved before the switch existed carry no switch, so the editor's
legacy upgrade added it switched off and template-created workflows ignored
the new-workflow default. Creating from a template now sets the switch to the
default (or keeps it on when the template already turned it on).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…tion

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The item text already states the date, so the meta line repeated it. The
event date still drives resident expiry and stays editable in the item
editor. Revises AC-88.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… the action row

- The items a search returned now open from its step in the thinking process
  (daily, assistant, deep research), each deletable; the separate
  'used memory' entry under the answer is gone. Back-to-back searches (often
  fired in parallel) read as one step with distinct items counted.
- The write receipt sits in the answer's action row with like/dislike.
- Memory tool descriptions, search results and the resident block say memory
  is not a numbered source: a model had cited a memory write as [S1], which
  showed as literal text.
- @bisheng/chat: optional runDetail slot for a run's expanded body.
Revises AC-87 and design decision 16.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ies have details

- The receipt entry in the action row shrank to one character per line (CJK
  text has no min width in a flex row); it no longer shrinks or wraps.
- Search results mark items that carry a detail and say how to fetch it with
  detail_ids; fetched details are labelled as the full text. Before this the
  model could not know detail_ids was worth using.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
memory_search with detail_ids now returns those items' full detail text
directly (SearchService.get_details: the caller's own active items only).
Before, details were read only for items the same query happened to hit
again, so the model had to repeat its query to get one. query is optional
when detail_ids is given; the 'has detail' marker says how to fetch it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…p it

- The memory page folds a long memory's full text under its line
  ('Show full text'); it never says 'detail'. New owner-only endpoint
  GET /api/v1/memory/items/{id}/detail (12113 when the text cannot be read).
- A hand edit used to replace the item without its full text, silently
  dropping it. The edit now copies the full text to the new item, and fails
  instead of dropping it when the text cannot be read.
Revises AC-88.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…to change it

- Opening a memory shows its own view: last updated, summary, details
  (markdown), type and source, with edit and delete, and a box at the bottom
  to tell the AI how to change it. The AI change applies at once and can be
  undone (edits it back to the previous version). Replaces the in-row
  'show full text' toggle.
- Add and edit take the details too: left out keeps them, empty removes them.
  The edit dialog loads them, and leaves them alone if they could not be read.
- POST /api/v1/memory/items/{id}/revise: the memory model rewrites summary,
  details, type and event date through the edit path (audit
  memory.item.revise); unchanged answers write nothing; failures write
  nothing (new 12114).
Revises AC-88, AC-89; adds AC-107.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…l; details at body size

- After an AI change or an edit, the list reloads before it holds the new
  version; the view briefly fell back to the list and lost its 'changed ·
  undo' note. It now stays on the old version until the new one arrives.
- Restore memory.event_time_value (still used on the memory view).
- Details render at the summary's size instead of prose size.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…asking the AI

- The opened memory shows its line as the heading and its details as the
  body, with type, source, event date and update time in one quiet row: no
  'summary'/'details' labels to learn.
- The ask-the-AI box is pinned to the bottom of the pane like a chat
  composer (it was easy to miss under the content).
- No hand editing: the edit button, the list's pencil and the edit dialog are
  gone; adding a memory is a single box. The edit endpoint stays for the AI
  change and its undo.
Revises AC-88, AC-89.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A memory without details showed a single line and nothing else. The view now
follows Claude's memory page: last updated, summary, details, and says when
there are no details yet (the AI can add some). Revises AC-88.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…es resident room

Checked the consolidation prompt against the memory research (§7):
- #7 / AC-20: candidates carry a 'detail' (the specifics behind the one-line
  conclusion, as a markdown list). Before, consolidated memories never had
  one. A detail that fails validation is dropped on its own; a tense rewrite
  carries the old detail over.
- #18: duplicate_of / contradicts only when sure; when unsure, a new item.
- #12: the prompt states the resident limit, how many are used and how many
  are left.
Revises AC-48.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
'本周结论', '下周计划', '本月总结' name parts of a report or plan, but the
relative-time check rejected them, so a consolidated weekly-report preference
was dropped on 104. Period words right before such section nouns are no
longer rejected; '本周在北京出差', '这个月要完成', '下周三之前' still are.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… in beta1)

beta2 branched before the knowledge unification / F118 merge; this brings
all of beta1 in while keeping beta2's own commits (middleware credentials,
F009 browser object paths, upload/endpoint/SSRF hardening, open API streams).

Resolution notes:
- Helm: beta1's restructured chart is the base; beta2's credential keyring is
  wired in (BS_SECRET_KEYS_FILE in every backend container, generated keyring
  Secret, existingKeyringSecret for upgrades, extraEnv guard). The settings
  renderer now writes MinIO/Milvus/ES secrets as encrypt(<token>), which
  beta2's strict loader requires. F009 nginx object path ported.
- Config: unknown top-level keys warn and are ignored (beta1 rule) in both
  config_service and the credential preflight; plaintext credentials are still
  rejected (beta2 rule).
- Knowledge: beta2's trusted-upload checks ported onto beta1's staged/owned
  upload flow; the legacy /process path into SPACE knowledge, which beta1
  routed around them, now resolves references too.
- Route login guard: five routes with their own auth (DSH mobile device /
  refresh token, storage browser-bucket auth_request) added to the frozen list.
- Removed beta2's filesUpload.process.test.tsx (tests the page beta1 replaced).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Brings the workflow/assistant/daily/task-mode long-term memory feature
(feat/v3.0.0/132-long-term-memory) onto the 3.1 beta2 line.

Conflict resolution keeps 3.1's structure and adds the memory hooks:
- workflow publish keeps validate_version and runs the memory node check
  when going online; chat WS passes memory_register next to the lease args
- assistant agent keeps 3.1's admission/login_user and adds memory turns
- llm.py keeps 3.1's reference handling; memory_llm joins
  WORKBENCH_MODEL_OBJECTS and the registry-only system-model autofill
- daily/task-mode chat keeps 3.1's leases, generated files and failure
  info, plus memory turn binding, tools and redaction
- alembic merge revision joins f132_long_term_memory onto the 3.1 head
- F132 tests adapted to 3.1 fixtures (execution admission, publish
  transition, content-violation marking, model reference repository)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Editing or AI-revising a memory keeps its source scene and only switches
origin to manual, but the label treated origin=manual as "added by hand",
so a chat-born item read 手动添加 after any change. The label now follows
the scene alone; items added by hand still carry scene=manual (AC-88).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Keep application credentials encrypted with an external key, mount native middleware files as restricted plaintext, and handle legacy Elasticsearch authentication in the same install command. Helm stays outside this change.

Co-authored-by: Cursor <cursoragent@cursor.com>
Keep memory and credential work while taking the target line's request scopes, conversation ownership, and image-generation error code.

Co-authored-by: Cursor <cursoragent@cursor.com>
The backend health interval is 90 seconds, so a two-minute readiness check expires while startup is still in progress.

Co-authored-by: Cursor <cursoragent@cursor.com>
Compose config now requires openfga.env, and a non-interactive install refuses an image that does not report plaintext-native-v1. The structural checks and command-boundary fixture follow those rules, and Gateway remains refused until its credential acceptance is done.

Co-authored-by: Cursor <cursoragent@cursor.com>
The dev servers proxy stored files through /object. The config test still looked for the old /bisheng prefix, so the route check failed.

Co-authored-by: Cursor <cursoragent@cursor.com>
ProgressItem only has three console.log lines left, so the suppression ledger still counting four failed platform lint. The memory notice test imported within without using it.

Co-authored-by: Cursor <cursoragent@cursor.com>
Keep the /object storage proxy and the target's legacy admin-scope proxy flag. Keep both F009 object-path and F016 credit-pricing rows in the beta1 release contract.

Co-authored-by: Cursor <cursoragent@cursor.com>
@likaiaki
likaiaki merged commit 461433b into feat/3.1.0-beta2 Nov 17, 2026
6 of 7 checks passed
jieyuhuayang added a commit that referenced this pull request Nov 17, 2026
Bringing beta1 into beta2 (#2716) left two heads, so alembic upgrade head
refused to run. The merge revision has no schema change.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants