[Kendrick-Song]

Summary

EverOS 1.6.0. Two things land together because the second was designed for the first:

1. The everalgo 2026-10 contract line. Every algorithm package moves to the LLMClient.achat contract (everalgo-core 0.6.0, boundary 0.4.1, user-memory 0.9.0rc4, agent-memory 0.6.0rc3, rank 0.5.0rc3, knowledge / clustering / parser 0.3.0rc2), pinned exactly. The six release candidates stay as pinned; pip install -U everos resolves them without a pre-release flag.
2. The target architecture. Storage ports owned by the domain, one composition root (everos.service.build_application), three facades (MemoryService, KnowledgeService, OpsService), strategies as (event, deps) functions, cascade and OME as domain-agnostic engines, one layout convention across infra (facade + private _x/ directories), and 14 import-linter contracts guarding all of it. Scored 94 / 100 against the terminal-architecture rubric (from 48 on main).

What changes for a running deployment

Upgrade is one command and a restart; the first start migrates system.db (three nullable memcell columns). Markdown, the index and the OME jobstore are untouched; a root written by a 2026-08 release reopened on this branch with no manual step (see Verification).

Breaking, hence feat!: the EVEROS_DISABLE_CASCADE* / EVEROS_OME_* environment variables are replaced by [cascade] / [ome] settings and no longer read; [llm] is required to build the application, everos cascade … included; GET /health drops capabilities.llm; prompt slots are replaced by config/custom_instructions/<operator>.yaml; library imports of everos.service.memorize / search / get move to build_application and the facades. Additive: entry_id on the agent-case items of /search and /get. Extraction output changes with the everalgo prompts, and a profile's language is frozen from its first extraction after the upgrade. The CHANGELOG Upgrade section is the operator-facing version of this paragraph and is the release page text.

Area

• Architecture method
• Benchmark
• Use case
• Documentation
• Developer experience
• CI, build, or release

Verification

CI at e315967: lint, unit (3.12 / 3.13 / Windows), integration (3.12 / 3.13 / Milvus 2.6), package build.
Hermetic on the final tree (EVEROS_ROOT=/tmp/..., every provider key empty):
  make lint        -> ruff clean, import-linter 14 kept / 0 broken, datetime / CJK / openapi gates OK
  pytest tests/unit         -> 3091 passed, 4 skipped
  pytest tests/integration  -> 200 passed, 9 skipped
  pytest tests/e2e          -> 64 passed (non-live subset)
Live, local only (real LLM + embedding + rerank + multimodal credentials):
  pytest tests/e2e -m 'live_llm or slow' -> 22 / 22 passed, incl. 419 LoCoMo messages end to end
Upgrade smoke: a copy of a real root written 2026-08-04 (4 md files, 4 memcells, FTS v2, no case_* columns)
  reopened on this branch: migration on first command, old episodes searchable, new session
  add / flush / search / get OK, cascade drained, zero errors.

Checklist

• I kept the change scoped to the relevant area.
• I am opening this from a separate branch, not pushing directly to main.
• I updated docs, examples, or setup notes when behavior changed.
• I added or updated tests when the change affects behavior.
• I did not commit secrets, .env files, dependency folders, or generated output.
• Active relative links in Markdown files resolve.

Notes for Reviewers

• 33 commits, in three groups. 16 stages (squashed from 88; contract layer → ports → strategies → bootstrap → singletons → user track → agent track → search / knowledge → entrypoints → fix-ups), each self-contained. Nine fixes for open issues, listed under Closes. Eight pure-move commits for the layout convention plus their CHANGELOG entry; git log --stat shows them as renames and no public facade changes.
• Behaviour differences from main are enumerated and classified in the CHANGELOG Changed section.
• Deferred on purpose, with designs on file: write-path reliability (idempotent memorize, resumable extraction, honest failure status; issues Client timeout + retry can duplicate memcells for identical session/messages payload #274 / memorize(): make delivery resumable per memcell so retries and replays are idempotent #516 / [Bug]: memorize(): a failure between memcell insert and buffer rewrite leaves an orphan memcell and duplicates it on the next add #539 / 写入失败被上报为成功:cascade worker 不可恢复时 flush 仍返回 extracted;且失败无重试上限 #364 / flush returns 500 when SQLite write hits the busy timeout #498) and upstreaming the hand-written search algorithms to everalgo-rank.
• Squash-merge with this title; the CHANGELOG ## [1.6.0] section is the release page.

Closes

• Closes Decouple index record models from LanceModel / Vector #476 — index records are plain pydantic (IndexRecord + DenseVector); the LanceDB adapter owns the Arrow mapping.
• Closes Agent-pipeline e2e is the only coverage of the skill chain and CI never runs it #397 — the skill chain runs in CI without credentials: tests/integration/test_ome_strategies_integration.py::test_skill_chain_e2e plus the canned agent-track unit tests.
• Closes [Bug]: English timestamped messages produce Chinese/German profile output #402 — the cell's language is detected once and passed down; the profile stores primary_language and never re-judges. Verified locally on qwen/qwen3-235b-a22b-2507 (the only model that reproduced the drift): 3/3 English-burst runs and 2/2 dialogue profiles came back English.
• Closes case→skill bridge matches index row ids against md-side source_case_ids #478 — the case→skill bridge is keyed by case entry id (the shape source_case_ids holds), not the index row id; SearchAgentCaseItem / GetAgentCaseItem gain entry_id (additive).
• Closes First /add can fail when tiktoken downloads o200k_base at runtime #277 — Application.start warms tiktoken's o200k_base encoding before opening storage; a host that cannot load it fails to start with TokenizerAssetError instead of a generic 500 on the first agent /add.
• Closes [Bug]: OME crash recovery skips RUNNING rows younger than 30 min and never revisits them; the run is lost #540 — OME crash recovery resumes every RUNNING row of the previous process; the 30-minute age filter lost runs cut short by a restart.
• Closes [Bug]: cascade never re-indexes a file restored with its old mtime after the worker folded it into a deletion #541 — the cascade worker records a folded delete (deleted) and the reconciler re-emits a present file whose row says so, whatever its mtime.
• Closes radius is documented and validated but has no effect on any search lane #394 — radius is a pre-fusion cosine floor on every vector route (vector, hybrid, agentic); the dead RankInput.radius plumbing is gone.
• Closes IndexRepository: retire scope-blind read methods (find_by_owner, …) #480 — the scope-blind IndexRepository.find_by_owner is removed (no callers).
• Closes [Bug]: add/flush returns 500 every time when an extraction reply quotes code with an unmatched } (everalgo brace-counting JSON extractor) #515 — the everalgo 2026-10 line parses replies with model_validate_json (no brace counting); timeout_seconds is now listed in the [llm] configuration reference.

🤖 Generated with Claude Code

[gloryfromca]

No blockers; this can merge as far as I am concerned.

Reviewed the full 58-commit diff and history against main, including CLAUDE/architecture rules, EverAlgo LLM/boundary/user/agent/search contract adaptation, composition-root and facade callers, storage/schema migration and lifecycle paths, API/CLI backward compatibility, and test changes. The OpenAPI contract remains unchanged except for the release version, all import-linter architecture contracts hold, and I found no test weakening.

Local verification passed: ruff check/format; all 13 import-linter contracts; repository, datetime, singleton, OpenAPI, docs, and CJK gates; unit (3050 passed, 4 skipped); integration (199 passed, 9 skipped, 7 deselected); and e2e (64 passed, 22 deselected). GitHub lint, package, links, commit-message, title, and Linux integration checks were green; remaining platform/Milvus/unit jobs were still running when posted.

[gloryfromca]

No blockers; this can merge as far as I am concerned.

Reviewed the change since 18b45d4 and rechecked it in the context of the full diff against main. I covered CLAUDE/architecture rules, the dependency-injection diff, affected route callers and history, HTTP/OpenAPI compatibility, test weakening, and the entrypoint/service architecture boundary. The refactor preserves route behavior and the published API while making the application and request ID overrideable through FastAPI dependencies; no test configuration or coverage was weakened.

Verification: all 128 API unit tests passed; ruff check and format-check passed; all 13 import-linter contracts held; OpenAPI matched the committed schema; datetime and singleton gates passed. The repository's make wrapper was unavailable in this environment, so I ran these constituent gates directly.

[gloryfromca]

No blockers; this can merge as far as I am concerned.

Reviewed the two-commit delta since 6bdafe9 in the context of the full diff against main. I covered the project rules and architecture constraints, schema relocation and mapping callers, the table-driven exception-handler behavior and history, HTTP/Python compatibility, and test weakening. The published OpenAPI contract is unchanged, every prior application-error mapping retains its status/code/message policy, and the only test edits are import-path updates. No fourth-round blocker criteria are implicated.

Verification passed: 134 affected API and scope/path property tests; ruff check and format-check; all 13 import-linter contracts; OpenAPI drift; datetime discipline; and singleton checks.

[gloryfromca]

No blockers; this can merge as far as I am concerned.

Reviewed the delta since e1bae0b in the context of the full diff against main. I covered project rules and architecture, the exception hierarchy and all callers, response compatibility, history, and test weakening. The new exact-class mapping makes ordinary API requests that lack LLM configuration return the established 422 provider error without changing unrelated RuntimeError handling; the test change is additive. No late-round blocker criteria are implicated.

Verification passed: all 129 API unit tests; ruff check and format-check; all 13 import-linter contracts; OpenAPI drift; datetime discipline; and singleton checks.

[gloryfromca]

No blockers; this can merge as far as I am concerned.

Reviewed the two-commit delta since d990ff1 in the context of the full diff against main. I covered project rules and architecture, the EverAlgo boundary contract, pipeline/event/persistence/replay callers, backward compatibility with pre-language markdown and events, and test weakening. Explicitly complete tails are closed exactly as the upstream contract delegates; language pinning preserves the documented inference fallback when metadata is absent. The test changes add coverage while retaining the existing buffered-tail e2e path. No late-round blocker criteria are implicated.

Verification passed: targeted unit, integration, and e2e tests (117 passed, 1 deselected by existing suite selection); ruff check and format-check; all 13 import-linter contracts; OpenAPI drift; datetime discipline; and singleton checks.

[gloryfromca]

No blockers; this can merge as far as I am concerned.

Reviewed the two-commit delta since 1f55f74 in the context of the full diff against main. I covered project rules and architecture, public and queued service callers, HTTP/CLI compatibility, renamed-module history, and test weakening. The facade now exposes one public method per add/flush use case, queue delivery still targets the inline cycle, the HTTP/OpenAPI contract is unchanged, and the deliberate Python-facade rename is documented with updated in-repo callers. White-box tests moved to the private cycle while public orchestration remains covered. No late-round blocker criteria are implicated.

Verification passed: selected service, API, CLI, property, and integration tests (530 passed, 1 deselected by existing suite selection); CLI help; ruff check and format-check; all 13 import-linter contracts; OpenAPI drift; datetime discipline; and singleton checks.

[gloryfromca]

No blockers; this can merge as far as I am concerned.

Reviewed the five-commit delta since d44ba9e in the context of the full diff against main. I covered the project rules and architecture constraints, moved middleware/upload/backfill/cascade/Milvus/OME code and their callers, commit history, Python and persisted-job compatibility, and test weakening. The new module boundaries remove a service-to-infra exception while preserving the prior public cascade and Milvus imports and OME's persisted APScheduler callable paths; the test changes move or strengthen coverage rather than relaxing it. No late-round blocker criteria are implicated.

Verification passed: the complete unit suite (3091 passed, 4 skipped); focused affected unit and vector-backfill integration tests (230 passed); ruff check and format-check; all 13 import-linter contracts; OpenAPI drift; repository asset/file-size/deprecated-name/contributor-doc gates; datetime discipline; singleton checks; documentation links; CJK policy; and git diff checks.

[gloryfromca]

Blocking: The committed sphere golden fixture must be corrected so the new regression test passes.

Reviewed the four-commit delta since ee07484 in the context of the full diff against main. I covered project rules and architecture, the TUI sphere/app module extraction, CLI and README-media callers, commit history, Python import compatibility, removal reachability, and test weakening. The renderer extraction itself preserves output and the removed solving path was unwired, but the new golden fixture fails its own regression test; details are inline.

Verification: the affected TUI/CLI suite reported 104 passed and 1 failed; the isolated golden suite reported 1 passed and 1 failed. Ruff check and format-check, all 13 import-linter contracts, OpenAPI drift, datetime discipline, singleton checks, docs, CJK policy, and diff checks passed. This blocker is introduced by this revision, reached in the ordinary Python 3.13 unit CI job, and leaves no working merge path while that required test fails.

[gloryfromca]

No blockers; this can merge as far as I am concerned.

Reviewed the fix delta since d5bb7dd in the context of the full diff against main. I covered the project rules and architecture constraints, the golden serialization and regenerated fixture, affected TUI/CLI callers, commit history, cross-platform compatibility rationale, and whether the test was weakened. Rounding only the non-visual depth field to 1e-4 removes platform libm ULP noise while the regression still pins all rendered cells and markup; my prior blocker is fixed, replied to, and resolved. No late-round blocker criteria remain.

Verification passed: all 105 affected TUI/CLI tests; ruff check and format-check; all 13 import-linter contracts; OpenAPI drift; datetime discipline; singleton checks; and diff checks.

[gloryfromca]

No blockers; this can merge as far as I am concerned.

Reviewed the three-commit delta since f08d4f0 in the context of the full diff against main. I covered the project rules and architecture constraints, package init/facade requirements, the YAML-loader rename and all in-repo callers, the removed evolution package and history, Python import compatibility, documentation, and test weakening. The deleted evolution package was empty and unreferenced; the loader implementation is a pure move; and the deliberate everos.component.config rename is called out in the changelog with a direct migration path. No late-round blocker criteria are implicated.

Verification passed: all 46 affected loader, custom-instruction, and bootstrap tests; ruff check and format-check; all 13 import-linter contracts; OpenAPI drift; datetime discipline; singleton checks; documentation checks; CJK policy; and full-diff checks. The earlier golden-fixture thread remains resolved.

[gloryfromca]

No blockers; this can merge as far as I am concerned.

Reviewed the three-commit delta since 65c727b in the context of the full diff against main. I covered the project rules and architecture constraints, the provider/build lifecycle, memory extraction and search callers, cascade CLI behavior, health/OpenAPI and Python compatibility, commit history, and test weakening. Requiring the extraction LLM at assembly preserves the server's existing startup invariant while making downstream types honest; optional embedding/rerank/multimodal behavior is unchanged, the removed health flag was constant, and the API/Python migrations are documented. No late-round blocker criteria are implicated.

Verification passed: the complete unit suite (3090 passed, 4 skipped); targeted changed-path integration/e2e tests (115 passed, 3 deselected by existing suite selection); ruff check and format-check; all 13 import-linter contracts; OpenAPI drift; datetime discipline; singleton checks; documentation checks; CJK policy; and full-diff checks. The earlier golden-fixture thread remains resolved.

[gloryfromca]

No blockers; suggestions only, and they are marked inline.

Reviewed the two-commit delta through 2f1c5ef. The primary vector, hybrid, and agentic paths now apply radius to raw dense scores before fusion/rerank, and the dead RankInput.radius plumbing is removed. The scope-blind find_by_owner removal has no repository callers; scoped alternatives remain.

Named nonblocking follow-up: the public advanced llm_multiround method still calls dense_recall directly and its retrieval function accepts no radius, so the new “every method's vector route” wording has one exception. Either document that exception or plumb the floor into each per-sub-query dense result later. I am not holding this revision on it: the ignored behavior predates this delta, that method is absent from the documented SearchMethod table, and callers have working vector/hybrid/agentic routes. It therefore fails the late-round introduced/ordinary-use/no-way-out blocker bar.

Covered the repository rules and architecture constraints, the diff and affected callers, issue and implementation history, API/backward compatibility, score-stage semantics, and whether tests were weakened. Validation: focused search/persistence/E2E selection 371 passed (15 deselected); Ruff check/format, all 14 import contracts, docs/CJK, OpenAPI drift, and repository gates passed. GitHub lint, integration, links, title, and package-build checks are green; the commit-message job currently reports the mechanical 73-vs-72-character subject failure, and remaining unit/Milvus matrix jobs are still running. The previously opened golden-fixture thread remains resolved.

[gloryfromca]

No blockers; suggestions only, and they are marked inline.

No new findings in this revision. I reviewed the internal-package privatization, backfill relocation, and index-record facade cleanup, including callers, history, facade/backward compatibility, tests, and the repository architecture rules. The force-pushed 14006c4 tree is byte-for-byte identical to the reviewed aac8e10 tree, so the completed validation applies to the current head.

Validation: 3,091 unit tests passed with 4 skipped; 200 integration tests passed with 9 skipped and 7 deselected. Ruff check/format, all 14 import-linter contracts, repository assets/file sizes/naming/contributor-doc/datetime/singleton/OpenAPI gates, documentation checks, CJK policy, deterministic fixture regeneration, and diff checks passed. Public facade symbol sets remain unchanged except for the intentional added cascade constant, and no affected test was weakened. Current GitHub lint, integration, package, commit-message, link, and title checks are green; the unit and Milvus matrix jobs are still running.

The previously named llm_multiround radius follow-up remains the sole nonblocking item in my overall stance. The golden-fixture thread I opened remains resolved.

[gloryfromca]

No blockers; suggestions only, and they are marked inline.

No new findings in the 0675372 OME lock refactor. I reviewed the full PR diff plus this revision delta, the OmeLock and engine callers/history, lifecycle and failure cleanup, facade/backward compatibility, test changes, and the repository architecture rules. The engine now delegates its existing process lock lifecycle to OmeLock without changing the production ownership path.

Validation: 255 affected OME/service/integration tests passed; Ruff check and format passed; all 14 import-linter architecture contracts passed; a direct two-instance lock lifecycle smoke test confirmed acquire/contention/release/context behavior. GitHub lint, package, commit-message, link, and title checks are green; the wider unit/integration/Milvus matrix is still running. The previously named llm_multiround radius follow-up remains the sole nonblocking item in my overall stance, and the golden-fixture thread remains resolved.

[gloryfromca]

No blockers; suggestions only, and they are marked inline.

No new findings in the 76ca9fa E2E adjustment. I reviewed the full PR context plus this delta, the parser-to-canonical-message-to-memcell persistence path, affected callers and history, compatibility, whether the assertion was weakened, and the repository rules/architecture constraints. Reading either the session buffer or serialized memcell is valid because both carry the parser-derived CanonicalMessage.text; the assertions still require parsed HTML markers/content rather than merely the original request URI or encoded input.

Validation: supporting parser, boundary, SQLite memcell, and storage coverage passed 60/60; the changed live-LLM module collected 3 tests but all 3 were deselected by the default marker policy (no credentialed/token-burning run was attempted). Ruff check/format, diff checks, and all 14 import-linter contracts passed. GitHub lint, integration, package, commit-message, link, and title checks are green; remaining unit and Milvus jobs are still running. The previously named llm_multiround radius follow-up remains the sole nonblocking item in my overall stance, and the golden-fixture thread remains resolved.

[gloryfromca]

No blockers; suggestions only, and they are marked inline.

Reviewed the changelog-only delta against the repository rules, implementation and callers, prior environment-variable behavior/history, backward compatibility, tests, and the applicable architecture constraints. No tests were weakened in this revision. Documentation/CJK/diff checks passed, and the targeted config, cascade CLI, health API, SQLite schema, and custom-instruction tests pass (34/34 on Python 3.13.2). CI lint, integration, links, package build, commit-message, and title checks are green; the remaining matrix jobs are still running.

[gloryfromca]

No blockers; suggestions only, and they are marked inline.

The new revision only removes an accidentally committed synthetic fixture draft and its builder. I verified that neither file was referenced at the prior revision, and their removal does not change runtime behavior, architecture, backward compatibility, or test coverage. The existing changelog migration-table clarification remains a named nonblocking documentation follow-up.

Covered: repository rules, the full change and revision delta, fixture consumers and history, backward compatibility, test-strength preservation, and architecture constraints. Local unit result: 3,091 passed, 4 skipped on Python 3.13.2. Repository asset/file-size, deprecated-name, contributor-doc, datetime, singleton, OpenAPI, CJK, documentation, and diff checks passed. Current CI is green for lint, package build, Linux unit/integration jobs, links, commit messages, and PR title; Windows unit and Milvus 2.6 integration jobs are still running.