Skip to content

[Status 06] InferRSI evolution: 8150ee04 → f0e92e8d (2026-11-20) #538

Open

Description

@Dong1017

Weekly edition for Friday, November 20. The nominal obligation was 17:00 Asia/Shanghai; this check ran after a delayed trigger at 17:20:39. Status 05 was published at 17:25:31 on November 13, so its seven-day edition window would have ended at 17:25:31 today. This Friday report uses the earlier obligation and the actual window from that publication through this check. It does not issue a second edition.

1. Snapshot

  • Previous edition Status 05, #423.
  • Observation November 20, 2026, 17:20–17:29 Asia/Shanghai. This check consolidates changes since the prior edition's publication at November 13, 17:25:31. It carries forward the in-week records and adds the later main deliveries. Nov 13 evening · Nov 14 · Nov 15 · Nov 16 · Nov 17 · Nov 18 · Nov 19
  • Published body baseline 8150ee048af6503070fa71768fbf64dffcc1a77e.
  • Last published delta main c573f40a43ceef23ef08507fca662c95ffb65390.
  • Pinned current main f0e92e8d4992746ac62e89f16347ecab6141de67, 67 commits after the published body and eight after the last delta. Full comparison · Since last delta
  • Reporting window [2026-11-20 17:29:58, 2026-11-27 17:29:58) Asia/Shanghai, start inclusive and end exclusive: from this issue's GitHub publication time (2026-11-20T09:29:58Z) through seven days later. The #61 index receipt records the same bounds.

Apps workflows and the Cannbot adapter are now under apps.workflows and apps.integrations, and Benchmark exposes a frozen P1 preparation port. Inference knowledge is unified in one skill_library with seven task routes. Those deliveries are not S5 implementation acceptance, not #490 step-5 combination acceptance, and not a library-wide retrieval or knowledge-efficacy result. #371, #429, #430, #434 and #490 remain open.

2. What changed

Already recorded in the Status 05 in-week updates, and still on the ancestry of this main: the Apps prepare/bind and later layout slices; K02, B01, OBS, S01 and the formal v0.2.0 publication; shared contracts, root-shell and loader retirements; Find-unified search; service-reuse software #480; the knowledge-compare entry #501; and the bounded Issue closes recorded in those updates, including first-wave delivery closes, the #387 scope decision, and the November 19 process closes of #487, #467, #497, #447, #333 and #491. Those earlier records stay authoritative for their intervals; this edition does not copy them in full.

The eight commits after the last published delta, all now on the pinned main, are:

Change Delivery and acceptance boundary
Group Apps workflows under apps.workflows #517 merged November 19 23:02:11 Asia/Shanghai by zengchuang-hw, merge commit 493416520f537590eb8f2a4e74f276f82e488307, head 15458a4bf837880a56feb76d9d0cab4191d94507 (62 files, +1443/−971). Refs #490 and #371. Exact-head approvals: Dong1017 and cursor[bot]. This is #490 step 3 of 5 and succeeds closed unmerged #504. Old Python imports remain as explicit shims. Author-recorded test counts and wheel checks were not independently replayed here. #490 and #371 remain open. Unmerged MindIE step 4b #537 and the step-5 combination acceptance are not on this main.
Make profiling optional in service-reuse campaigns #521 merged November 20 10:03:07 Asia/Shanghai by jiangkuaixue123, merge commit b5756cd561e8b12b68899d9dfa6411feab4669f5, head 47239e4864b3a3694dd7320d50852922bdd75233 (17 files, +2931/−470). Refs #513. Exact-head approvals: zengchuang-hw and jiangkuaixue123. The PR makes formal performance and quality determine round acceptance, reports profile availability, and lets operators collect a profile after the round. Author-recorded checks were 109 tests and 43 subtests on WSL with a synthetic service. The PR body assigned the #513 repository acceptance decision to Dong1017 and did not claim closure. lirx-pd closed #513 as completed at 10:03:53, with no close comment at the close itself. That close is a process record, not independently replayed formal performance testing or real-model qualification on this pin.
Unify inference knowledge in one skill_library #516 merged November 20 10:31:59 Asia/Shanghai by amy-why-3459, merge commit 1b826b5f9666a8b734000e1eb2990e582cdef876, head c3f8d7fcab53a52acbecdc3f2f7d238257cb8204 (1096 files, +15260/−23846). Refs #515. Exact-head approvals: zengchuang-hw, Dong1017 and amy-why-3459. The slice keeps one physical library, seven task-entry skills, authored record provenance, and default Find over the catalog. Reviewer-recorded Windows offline Find recomputations, including the retained 28-pair median 2.028378 → 1.477994 s (ratio 0.72866) in 7028508772, are receipts at their recorded revisions (49341652 → 7c199747 in that comment, not a rerun of the merge head). 7033469690 treats those figures as a layout/migration regression bar, not routing or end-to-end retrieval benefit. #515 remains open. Those receipts were not independently replayed here and are not library-wide knowledge-efficacy PASS.
Add area:benchmark and area:apps labels #532 merged November 20 12:30:54 Asia/Shanghai by Dong1017, merge commit 4eaf7fb9f1033a355231002f5bfe353455d66b8d, head 829945bf0cf734d8258eb569f30dd56d17cddb8b (1 file, +10/−0). Exact-head approvals: cursor[bot] and zengchuang-hw. This is a checked-in label-manifest change. The two repository labels now exist with those names. The PR states that no Issue is closed by it, and that #430 and #530 are not its acceptance owners.
Mark source-built plugins as 0.0.0+source #525 merged November 20 13:49:02 Asia/Shanghai by Dong1017, merge commit 0697f08e0c834dba6170d496c188c94938bebcad, head 1600322a7bd14060139872b5382e1418a2f9f6c0 (9 files, +169/−4). Refs #511. Exact-head approvals: zengchuang-hw and cursor[bot]. Source builds of the Codex, WorkSwarm and DSH plugins report 0.0.0+source; formal Codex and WorkSwarm builds keep an explicit release version. The PR did not claim closure. Dong1017 closed #511 as completed in 7032171149, recording that native-host import or display was not measured and that resolving a clean tag checkout to its tag version stays a separate question. #152 remains open.
Gate RSI search on measured bottleneck diagnosis #519 merged November 20 15:44:10 Asia/Shanghai by Dong1017, merge commit b59300836df9b62b9035c7ada05e0e86c9f3cfaa, head 20530c5b3996964c10d5252f0c11d88ba93a28d2 (18 files, +399/−38). Refs #337 and #501. Exact-head approvals: cursor[bot], zengchuang-hw and Dong1017. New sessions must register a diagnosis before managed knowledge search or a candidate round; existing frozen sessions keep their plans. Author-recorded checks were 129 tests, 1 skipped and 43 subtests, and were not independently replayed here. #337, #502 and #509 remain open. This is a Harness sequencing gate, not acceptance of the knowledge-compare campaign and not a measured-effect PASS.
Move the Cannbot adapter under apps.integrations #528 merged November 20 16:13:55 Asia/Shanghai by zengchuang-hw, merge commit aa843235b8676df801f1460ae110a486bbd27325, head 2045c0fe315ebb1ab0c6e5203b502b2ddf4d499d (15 files, +1009/−779). Refs #490 and #371. Exact-head approvals: Dong1017, cursor[bot] and amy-why-3459. This is #490 step 4a only. MindIE remains under apps/mindie/ until unmerged 4b #537. The PR and 7034807242 keep #490 and #371 open and leave the step-5 out-of-tree combination acceptance, plus shim retirement, for later work. Author-recorded test counts were not independently replayed here. A real external Cannbot index is outside the recorded verification scope.
Add the frozen P1 Benchmark preparation port #530 merged November 20 16:14:13 Asia/Shanghai by zengchuang-hw, merge commit f0e92e8d4992746ac62e89f16347ecab6141de67, head 0a82e7b5ad7115f1b2e81818bcf227129208e4ec (5 files, +888/−1). Refs #430 and #371. Exact-head approvals: Dong1017, cursor[bot] and amy-why-3459. The port freezes and validates a configured P1 preparation; it does not install or register the worker on a Harness target and does not execute it. The PR records T36 coverage for this prepare slice; T37/T40 still depend on #429, and T35/T39/V23 still depend on #434. Shared CPU production wiring waits for all three ports. #430 and #371 remain open. This is not model- or service-gain evidence.

3. Architecture delta

flowchart LR
  root["src/inferrsi root<br/>__init__.py + __main__.py only"]
  apps["apps/<br/>cli, storage/, handoff/, workflows/,<br/>integrations/cannbot, host/,<br/>views/, delivery/, mindie/"]
  knowledge["knowledge/<br/>one skill_library + Find + seven task routes"]
  harness["harness/<br/>serving/ + diagnose-before-search"]
  benchmark["benchmark/<br/>metrics, validators, frozen P1 prepare"]
  infra["infra/<br/>artifacts.py only"]
  root -->|"__main__ delegates"| apps
  apps --> knowledge
  apps --> harness
  apps --> benchmark
  harness --> infra
Loading

The diagram is the implemented layout on the pinned main after #517, #516, #519, #528 and #530. MindIE is still under apps/mindie/. infra/ is still present. It is not a module import graph, and it is not a claim that #490 or #498 is complete.

4. Capability status

| Workstream | Code delivered | Goal acceptance / measured effect |
|---|---|
| Apps layout #490 / #371 | #517 (workflows) and #528 (Cannbot integrations) are on the pinned main, after earlier storage #493 and handoff #496 | #490 and #371 remain open. Step 4b MindIE #537 and step-5 combination acceptance are not on this main. Design acceptance is unchanged; this is not S5 implementation acceptance. |
| Production ports #429 / #430 / #434 | #530 lands the frozen P1 prepare/encode/validate slice | #429, #430 and #434 remain open. The port does not execute a worker. Shared CPU production wiring still waits for execute and offline assess. |
| Service-reuse profiling #513 / #467 | #521 is on the pinned main | lirx-pd closed #513 as completed, with no close comment. That close is not formal performance testing or real-model qualification on this pin. |
| Knowledge layout #515 / #516 | #516 is on the pinned main | #515 remains open. Reviewer-recorded Find timings are a layout/migration regression bar, not routing or end-to-end retrieval benefit. |
| Plugin source version #511 / #152 | #525 is on the pinned main | Dong1017 closed #511 in 7032171149, with native-host display still unverified. #152 remains open. |
| Diagnose-before-search #519 / #337 / #502 / #509 | #519 is on the pinned main | #337, #502 and #509 remain open. The gate is not knowledge-compare acceptance and not a measured-effect PASS. |
| #498 package retirement | No new slice in this eight-commit range; #503 and #510 remain on the ancestry | #498 remains open. Remaining work is still slices 3b, 4 and 5, plus the remaining model/ providers. |
| #476 / #455 / #379 / #350 / #348 / #38 / #520 | No new main delivery that closes them | All seven remain open. Unmerged #518, #482 and #492 are not main deliveries. #520 remains a historical INCONCLUSIVE record, not a main result. |

5. Evidence and limitations

This check read the latest Status search, the Status 05 body and its in-week updates, the #61 index, current main ancestry, selected changed paths, scoped acceptances, reviews, and hosted check metadata. The mermaid diagram above reflects implemented paths on the pinned main; it is not a module import graph.

On the pinned main, merge-lane and knowledge declaration policy succeeded on the push at November 20 16:14:16 Asia/Shanghai. That is not a full-suite or all-CI-green claim for this SHA. Full-suite software evidence cited above belongs to the reviewed pull-request heads. Green software checks do not establish measured effect, knowledge efficacy or campaign acceptance.

Author-recorded and reviewer-recorded test counts, Find timings and campaign figures at their recorded revisions were not independently replayed here. This report does not publish internal host or account paths.

No tests, installations, model or provider calls, workflows or experiments were started for this report. No code, labels, settings or other workstream state was changed by this reporting run, except the authorized Status edition, #61 index and Status 05 rollover once those writes are verified.

6. Next decisions

  1. Under task: Apps 应用层详细设计与迁移计划(CLI / Python / Plugin) #371 / task: Harness evaluation-mode CPU execute for Apps #429 / task: Benchmark production prepare and frozen resource closure #430 / task: Benchmark production offline assess for Apps #434 / task: 按职责整理 Apps 子模块目录(handoff / storage / workflows / integrations) #490, keep the merged workflows, Cannbot integrations and P1 prepare port separate from S5 implementation acceptance, from production execute/assess, and from task: 按职责整理 Apps 子模块目录(handoff / storage / workflows / integrations) #490 step-5 combination acceptance. Keep unmerged refactor(apps): move MindIE adapter into integrations #537 off the main-delivery board.
  2. Keep the task: retire the infra/ and model/ packages into their owning modules #498 slices that already landed in refactor(benchmark): move the metric contract and validators out of model #503 and refactor(harness,apps): move serving helpers and host bindings, remove the campaign facade #510 separate from retirement completion. Remaining work is still slices 3b, 4 and 5, plus the remaining model/ providers.
  3. Under task: Complete formal performance testing and decouple profiling in service-reuse campaigns #513 / task: Support service-reuse and multi-scenario optimization in managed campaigns #467, keep the merged Make profiling optional in service-reuse campaigns #521 software slice and the uncommented task: Complete formal performance testing and decouple profiling in service-reuse campaigns #513 close separate from formal performance testing and real-model qualification on the current pin.
  4. Under Organize inference knowledge in one skill_library and validate task-oriented retrieval #515 / Unify inference knowledge in skill_library with seven task routes and direct search #516 / feat: benchmark RSI knowledge with paired optimization runs #501 / task: connect engineering tasks to existing Agent execution and evaluation #337 / acceptance: eight-round RSI managed Knowledge comparison #502 / investigation: 8 轮 Knowledge on/off 对照暴露的 RSI 框架问题 #509, keep the merged skill_library layout and the diagnose-before-search gate separate from routing or end-to-end retrieval benefit, from acceptance Issue acceptance: eight-round RSI managed Knowledge comparison #502, and from any library-wide knowledge-efficacy claim.
  5. Keep rfc: 全量接入 cannbot-knowledge:MCP 常驻查询通道设计 #476, task: add GPU measurement-validity and profiler-attribution guidance to inference-benchmark-design #455, tracking: Skill Indexer 检索效率、准确性与逐轮采用 #379, task: 细化 vLLM Ascend 知识蒸馏与知识库整理 #350, task: anchor L1 weight demand against the checkpoint in performance-ceiling-analysis #348, task: add MLX backend discovery and seed source-grounded MLX knowledge for Apple Silicon #38 and [RFC] Agent plugins: a consistent InferRSI workflow in Codex, DSH and openJiuwen #152 open until their remaining recorded criteria are on main. Keep the bug: plugins built from a source checkout report version 0.1.0 instead of the release tag #511 close visible with its recorded native-host gap. Treat record: MiniMax-H3 Ref2VA step reduction, Cache-DiT, and sparse-attention experiments on Ascend 950PR #520 as a historical INCONCLUSIVE record, not a main result. Unmerged Add a MUSA backend, the vllm_musa repository pack and a vllm-musa skill #518, feat: permit unverified feature status and add boundary retrieval tests (Refs #350) #482 and fix(knowledge): guide Ascend tasks through external setup #492 are not main deliveries.

Activity

  1. added
    kind:statusPeriodic status snapshot
    area:processRepo-wide process, governance, privacy and security rules
    on Nov 20, 2026
  2. Dong1017 commented on Nov 20, 2026

    @Dong1017
    CollaboratorAuthor

    Posted by Dong-bot, an automated agent acting on behalf of @Dong1017 (always-on-bot, model claude-sonnet-5-5).

    Triage — Status 06 (#538)

    Assessment. This is the weekly evolution edition for the 8150ee04 → f0e92e8d span, and it is a record publication, not a decision request: no acceptance box is asked to be checked, and the only implicit ask is that the snapshot be accurate and that the acceptance boundaries hold. The boundaries do hold as written — every row pairs merge commit, head SHA, exact-head approvals and an explicit "not acceptance" clause, and §4/§6 keep #371 / #429 / #430 / #434 / #490 / #515 / #513 open rather than letting merged code imply closure. That is the same discipline Status 05 (#423) used and it makes the edition auditable. Findings below are about the parts a reader could mis-take or that will not age well, not about the merge list.

    Gaps worth fixing before the next edition

    1. Interval notation is ambiguous at the seam. §1 says the check consolidates "since the prior edition's publication at November 13, 17:25:31" and the observation runs 17:20–17:29 on November 20, while [Status 05] InferRSI evolution: 1246348f → 8150ee04 (2026-11-13) #423 declared its edition window as [2026-11-13 17:25:31, 2026-11-20 17:25:31). The 17:20:39–17:25:31 band on November 20 therefore sits inside [Status 05] InferRSI evolution: 1246348f → 8150ee04 (2026-11-13) #423's declared window and inside this edition's consolidation. In practice no in-week Status 05 delta was published on November 20 (last was the November 19 update c573f40a), so nothing is double-counted — but the edition should state the closed interval explicitly, in [a, b) form, and let the [Roadmap] InferRSI: reusable knowledge, verifiable inference optimization #61 index receipt carry the bounds, so the next reader does not have to reconcile two definitions.

    2. The ~20 minute trigger delay is now a pattern, not an incident. Status 05 fired at 17:20:48 against a 17:00 obligation; this one at 17:20:39. Narrating the delay each edition is honest but does not fix it. Either the scheduled trigger is offset by ~20 minutes (then the stated obligation should be corrected to match reality) or the trigger is unreliable (then the schedule needs repair). Worth one line in §6 naming which.

    3. src/inferrsi root claim vs the model/ remainder. The §3 diagram asserts the root holds "__init__.py + __main__.py only", but §4 simultaneously records that task: retire the infra/ and model/ packages into their owning modules #498 retirement still has slices 3b, 4, 5 and "the remaining model/ providers" outstanding. If those providers are still src/inferrsi/model/…, the diagram overstates the current tree. Please either scope the diagram label to the modules it actually covers or move the model/ remainder into it. The diagram's own caveat ("not a module import graph") does not resolve a path-existence claim.

    4. "Eight commits" vs the eight-row table. The rows are eight PRs; they equal eight commits only if every commit in c573f40a..f0e92e8d came from those eight merges. Say so explicitly, or say "eight commits, all from the merges below", so the count and the table cannot drift apart unnoticed.

    5. Shim retirement has no owner. refactor(apps): group application workflows in a workflows package #517 and refactor(apps): move Cannbot adapter into integrations #528 are recorded as landing with "old Python imports remain as explicit shims", and step-5 plus shim retirement are deferred. Deferred-with-no-owner is how permanent shims happen. Name the tracking issue or the acceptance owner for shim retirement alongside the task: 按职责整理 Apps 子模块目录(handoff / storage / workflows / integrations) #490 step-5 record.

    One record that will not age well

    1. task: Complete formal performance testing and decouple profiling in service-reuse campaigns #513 closed by another account with no close comment, 42 seconds after merge. The Make profiling optional in service-reuse campaigns #521 body assigned the task: Complete formal performance testing and decouple profiling in service-reuse campaigns #513 repository acceptance decision to the project owner; lirx-pd closed it as completed at 10:03:53 with no comment. That close is a process record and the edition says so correctly, but it leaves the assignment formally undischarged. Recommended fix, matching what was done for bug: plugins built from a source checkout report version 0.1.0 instead of the release tag #511 (7032171149), task: Support service-reuse and multi-scenario optimization in managed campaigns #467 and bug(contracts): scalar schema patterns accept a trailing newline #497: one close comment on task: Complete formal performance testing and decouple profiling in service-reuse campaigns #513 quoting the Make profiling optional in service-reuse campaigns #521 author-recorded checks (109 tests, 43 subtests, WSL, synthetic service) and stating explicitly that formal performance testing and real-model qualification on the current pin were not run. Otherwise every future edition has to re-explain this row.

    Clarifying questions

    Suggested next steps

    1. Use explicit [start, end) interval notation in §1; confirm the [Roadmap] InferRSI: reusable knowledge, verifiable inference optimization #61 index receipt and the Status 05 rollover landed before treating this edition as published, as §5 already conditions.
    2. Correct or scope the §3 root label relative to the model/ remainder.
    3. Add owner names for step-5 combination acceptance and shim retirement to §6 item 1, and for the blocked test groups (T37/T40 → task: Harness evaluation-mode CPU execute for Apps #429, T35/T39/V23 → task: Benchmark production offline assess for Apps #434) to the feat(benchmark): add frozen P1 preparation port #530 row.
    4. Land the task: Complete formal performance testing and decouple profiling in service-reuse campaigns #513 close comment above, or record in §6 item 3 that it remains owed.

    Not verified here

    I read this issue's body and #423's body plus its comment thread; I did not check the compare endpoints, commit counts, merge/head SHAs, review links, action-run IDs or the #61 index — commit ancestry, review state and CI scope in this edition remain as the author recorded them. No tests, installs or provider calls were run. This comment does not modify, close, label or accept any workstream state.

    Dong-bot (authorized delegate)

  3. amy-why-3459 commented on Nov 20, 2026

    @amy-why-3459
    Collaborator

    Thanks for Status 06. I checked the c573f40a...f0e92e8d compare: it is exactly eight commits, and they are the eight merge commits in the §2 table. Each row still pairs the merge commit, head SHA, exact-head approvals, and an explicit acceptance boundary, and §4/§6 keep #371 / #429 / #430 / #434 / #490 / #515 / #513 open rather than letting merged code imply closure. That is the same discipline as Status 05 and makes the edition auditable.

    A few notes after reading the body and checking the pinned tree at f0e92e8d:

    1. The §3 diagram omits packages that still exist on this pin. The root of src/inferrsi does contain only __init__.py and __main__.py as files, but the tree also still has model/, contracts/, and resources/. model/ is not empty: it still contains anthropic_agent.py, anthropic_edit_agent.py, and request_guard.py, which matches §4's remaining task: retire the infra/ and model/ packages into their owning modules #498 model/ providers. Please add those packages to the diagram, or state that it is a partial view, so the "implemented layout" cannot be read as the full tree.

    2. The reporting interval is ambiguous at the Status 05 seam. §1's observation is 17:20–17:29 on November 20, while [Status 05] InferRSI evolution: 1246348f → 8150ee04 (2026-11-13) #423 declared its window as [2026-11-13 17:25:31, 2026-11-20 17:25:31). The 17:20:39–17:25:31 band therefore sits in both. There was no Status 05 in-week delta today, so nothing appears double-counted, but an explicit [start, end) in §1 — and the same bounds on the [Roadmap] InferRSI: reusable knowledge, verifiable inference optimization #61 index receipt — would remove the ambiguity. I agree with 7035937736 on this point.

    3. The uncommented task: Complete formal performance testing and decouple profiling in service-reuse campaigns #513 close should be closed out as a process record. Make profiling optional in service-reuse campaigns #521 assigned the repository acceptance decision to the project owner; lirx-pd closed task: Complete formal performance testing and decouple profiling in service-reuse campaigns #513 as completed 46 seconds after the merge, with no close comment. The edition already treats that close as a process record, not formal performance testing or real-model qualification on this pin — that boundary is right. A close comment matching what was done for bug: plugins built from a source checkout report version 0.1.0 instead of the release tag #511, or an explicit §6 note that the comment is still owed, would keep later editions from having to re-explain the row. This is the same request as in 7035937736.

    4. Deferred shim and step-5 work has no named owner. The refactor(apps): group application workflows in a workflows package #517 / refactor(apps): move Cannbot adapter into integrations #528 rows and §6 item 1 correctly leave shim retirement and the task: 按职责整理 Apps 子模块目录(handoff / storage / workflows / integrations) #490 step-5 combination acceptance off this main, but they do not name who owns those records. Naming the acceptance owner next to the step-5 item would help the shims not become permanent by default.

    I listed src/inferrsi and src/inferrsi/model on f0e92e8d and counted the eight commits after c573f40a. I did not independently re-check every merge SHA, approval link, or the cited action-run IDs, and I did not replay author- or reviewer-recorded tests, Find timings, or campaign figures. This comment does not modify, close, label, or accept any workstream state.

  4. zengchuang-hw commented on Nov 20, 2026

    @zengchuang-hw
    Contributor

    NO PASS

    The main history checks out: f0e92e8d is 67 commits ahead of 8150ee04 and eight ahead of c573f40a. The brief also separates merged changes from outstanding acceptance. The remaining blocker is the reporting window: the #61 template requires explicit start and end timestamps and a timezone. §1 points to the index receipt, but the receipt lists only revisions and commit counts. Please state the actual edition window in §1 and the receipt: Nov 20, 2026 17:29:58 inclusive to Nov 27, 2026 17:29:58 exclusive, Asia/Shanghai. Keep the earlier evidence-check time separate so it is clear how late changes carry into the next update.

    Checked the GitHub publication timestamp, index receipt, template, and both pinned comparisons. This is a review of the status brief, not acceptance of the reported workstreams or a rerun of their experiments.

  5. Dong1017 commented on Nov 20, 2026

    @Dong1017
    CollaboratorAuthor

    In-week developer update — 2026-11-20 evening

    1. Snapshot

    • Change coverage interval: previous Status 06 publication at 2026-11-20 17:29:58 Asia/Shanghai through this check. Actual observation for this check: 19:18–19:23 Asia/Shanghai. The Friday 19:00 obligation fired at 19:18:07, after the 17:00 weekly edition. This check records later changes only and does not issue a second weekly edition. The next weekly obligation remains 2026-11-27 17:00 Asia/Shanghai.
    • Edition window: [2026-11-20 17:29:58, 2026-11-27 17:29:58) Asia/Shanghai. This is one in-week update on Status 06, not a new Status number.
    • Previous pin: f0e92e8d4992746ac62e89f16347ecab6141de67.
    • Pinned main: 6bdc86ee339c3007f15833434065adf893bff953, one commit after the Status 06 pin. Comparison

    One pull request landed on main after the weekly edition. #522 remains open for requester acceptance. Unmerged #539, opened at 18:02:27, is not a main delivery.

    2. What changed

    Change Record
    Trace CANNBot round knowledge in managed optimization #523 merged November 20 17:40:18 Asia/Shanghai by jiangkuaixue123, merge commit 6bdc86ee339c3007f15833434065adf893bff953, head cc552e47bc5a707f715fcb36da91b7d42525e42f (21 files, +1090/−76). Refs #522, which remains open. Exact-head approvals: Dong1017 and amy-why-3459. A same-head comment recorded DO NOT PASS only because required CI on cc552e47 had not yet finished; amy-why-3459 recorded that Ready CI 44709987715 later succeeded and cleared that item. The slice adds opt-in CANNBot search and pinned read receipts beside local captures, shows metadata and explicit citations in the Dashboard, keeps native local capture verification, and reserves final evaluation time when the external worker times out. Both exact-head approvals and the PR body keep requester acceptance of #522, and production CANNBot grant plus Ascend workload verification, outside the code verdict. The author-recorded 214 tests and Ruff result were not independently replayed here.

    3. Architecture delta

    flowchart LR
      start["start --cannbot-platform<br/>opt-in, default off"]
      knowledge["knowledge: local and CANNBot results stay separate"]
      read["knowledge-read: verify revision and digest"]
      cite["round --knowledge-ref<br/>delivered local item or pinned receipt only"]
      dash["Dashboard: metadata and citations<br/>no card content"]
      start --> knowledge
      knowledge --> read
      read --> cite
      cite --> dash
    
    Loading

    The diagram is the managed-round path landed by #523 on this pin. It is not a module import graph, and it is not #522 acceptance or a measured retrieval-benefit result.

    4. Capability status

    | Workstream | Code delivered | Goal acceptance / measured effect |
    |---|---|
    | CANNBot round knowledge #522 | #523 is on the pinned main | #522 remains open. The requester retains completion and closure authority. No post-merge requester acceptance is recorded on #522. This is not production-grant or Ascend-workload verification, and not a knowledge-efficacy result. |

    5. Evidence and limitations

    This check read the latest Status search, the Status 06 body and its comments, the #61 index, current main ancestry, #523 / #522 records and exact-head reviews, and hosted check metadata. The mermaid diagram above reflects the managed-round path on the pinned main; it is not a module import graph.

    On the pinned main, merge-lane and knowledge declaration policy succeeded on the push at November 20 17:40:21 Asia/Shanghai. That is not a full-suite or all-CI-green claim for this SHA. Full-suite software evidence cited above belongs to the reviewed pull-request head 44709987715. Green software checks do not establish measured effect, knowledge efficacy or campaign acceptance.

    Author-recorded test counts at their recorded revisions were not independently replayed here. This report does not publish internal host or account paths.

    No tests, installations, model or provider calls, workflows or experiments were started for this report. No code, labels, settings or other workstream state was changed by this reporting run, except the authorized in-week Status comment and #61 index once those writes are verified.

    6. Next decisions

    1. Under task: Trace CANNBot knowledge through optimization rounds and the Dashboard #522, keep the merged Trace CANNBot round knowledge in managed optimization #523 software slice separate from requester acceptance, from production CANNBot grant verification, and from any Ascend-workload or knowledge-efficacy claim. Keep task: Trace CANNBot knowledge through optimization rounds and the Dashboard #522 open until that recorded acceptance exists.
    2. Unmerged feat: lead benchmark reports with versioned metrics #539, opened after the weekly edition and still open under task: 定义 Benchmark metrics:单跑基础指标与 comparison 派生指标 #536, is not a main delivery. Keep task: 定义 Benchmark metrics:单跑基础指标与 comparison 派生指标 #536 open until its acceptance or close is recorded.
    3. The Status 06 decisions on task: Apps 应用层详细设计与迁移计划(CLI / Python / Plugin) #371 / task: Harness evaluation-mode CPU execute for Apps #429 / task: Benchmark production prepare and frozen resource closure #430 / task: Benchmark production offline assess for Apps #434 / task: 按职责整理 Apps 子模块目录(handoff / storage / workflows / integrations) #490, task: retire the infra/ and model/ packages into their owning modules #498, task: Complete formal performance testing and decouple profiling in service-reuse campaigns #513 / task: Support service-reuse and multi-scenario optimization in managed campaigns #467, and rfc: 全量接入 cannbot-knowledge:MCP 常驻查询通道设计 #476 remain in force; this increment does not change them.
  6. zengchuang-hw commented on Nov 20, 2026

    @zengchuang-hw
    Contributor

    NO PASS

    本轮周内更新已写出正确的 [2026-11-20 17:29:58, 2026-11-27 17:29:58)(Asia/Shanghai),并把观测时间和版本窗口分开。我核对了 GitHub 原始发布时间,起点一致。

    上轮剩下的是落点:原报告 §1 仍没列这两个时间,并指向一个未记录它们的 #61 初始回执。请把周内更新这行补到原报告 §1,并在 #61 的初始回执注明同样的边界或追加更正。无需改动主干提交数据。

  7. Dong1017 commented on Nov 20, 2026

    @Dong1017
    CollaboratorAuthor

    Posted by an agent on behalf of @Dong1017 (incremental review only; items already raised by others are not repeated).

    Agreeing with the NO PASS above, and adding the exact text so the two landing points are unambiguous.

    On the report side, #538 §1 currently ends its "Reporting window" bullet with "The #61 index receipt records the exact bounds after publication." That pointer is what is dangling: the receipt it points at — issuecomment-7038287968, posted 11:23 — records only the commit range (f0e92e8d → 6bdc86ee, one commit), with no bounds. §1 therefore needs to carry the bounds itself rather than delegate them: [2026-11-20 17:29:58, 2026-11-27 17:29:58) (Asia/Shanghai, start inclusive, end exclusive). This is the only open item on the Dong1017 side after the in-week update; it is a body edit, not something this comment can land, and it is not a trunk-data change.

    One thing worth keeping from the in-week update when §1 is revised: that update deliberately separates the observation time (17:20–17:29) from the edition window, and §1 should copy that split rather than merge them back into one interval — §1's single "Observation November 20, 2026, 17:20–17:29 Asia/Shanghai" line sits next to a seven-day window, which is where the earlier misreading came from. ^

    On the #61 side, a correction comment is posted on the initial receipt with the same bounds.

    ^ Applies only if the §1 revision is made by re-reading the in-week update; I have not checked whether the body-revision process already carries that phrasing.

    检查范围:read this thread's full comment history, the #538 body §1–§4, and the Status 06 in-week update and #61 receipt as they stand. Did not re-verify the commit list in §2 or run any repository checks.

  8. zengchuang-hw commented on Nov 20, 2026

    @zengchuang-hw
    Contributor

    NO PASS

    #61 更正回执已经补齐正确的起止时间。一个小但实质的来源更正:17:29:58 是原 Issue #538 的 GitHub 发布时间;周内更新的发布时间是 19:22:54,均为 Asia/Shanghai。区间不需要改,只需把回执的时间来源改成原 Issue,并在原报告 §1 写明精确区间。

  9. Dong1017 commented on Nov 21, 2026

    @Dong1017
    CollaboratorAuthor

    Thanks. You are right: 17:29:58 is the publication time of this issue, and the in-week update was published at 19:22:54 (both Asia/Shanghai). The bounds do not change. Two edits, both with an edit note where the text was a comment:

  10. zengchuang-hw commented on Nov 21, 2026

    @zengchuang-hw
    Contributor

    PASS(只针对报告窗口更正)

    复核了 上轮 NO PASS 的两点:#538 §1 与 #61 更正回执 现在一致,都以 #538 的 GitHub 发布时间 2026-11-20T09:29:58Z 为起点,而不是日内更新。换算并加七天后,区间为 [2026-11-20 17:29:58, 2026-11-27 17:29:58)(Asia/Shanghai,左闭右开)。上轮时间来源和报告正文缺精确区间的问题均已解决。这不是对整份状态报告或报告涉及的性能证据作验收。

  11. amy-why-3459 commented on Nov 21, 2026

    @amy-why-3459
    Collaborator

    Saw the §1 Reporting window update. It now states an explicit [2026-11-20 17:29:58, 2026-11-27 17:29:58) (Asia/Shanghai, start inclusive, end exclusive), taken from this issue's GitHub publication time (2026-11-20T09:29:58Z). I checked that conversion and the seven-day end. The linked #61 correction receipt records the same bounds and the same source (the original issue publication, not the later in-week update). Observation (17:20–17:29) remains a separate bullet, which keeps the window from being reread as the check interval.

    This resolves the interval-notation point from my earlier comment. I am not restating the other notes (diagram packages, the #513 close record, and naming the deferred-work owner).

  12. Dong1017 commented on Nov 21, 2026

    @Dong1017
    CollaboratorAuthor

    In-week developer update — 2026-11-21 evening

    1. Snapshot

    • Change coverage interval: previous in-week update at 2026-11-20 19:22:54 Asia/Shanghai through this check. Actual observation for this check: 19:10–19:18 Asia/Shanghai. The Saturday 19:00 obligation fired at 19:10:16. This check records later changes only and does not issue a weekly edition. The next weekly obligation remains 2026-11-27 17:00 Asia/Shanghai.
    • Edition window: [2026-11-20 17:29:58, 2026-11-27 17:29:58) Asia/Shanghai. This is one in-week update on Status 06, not a new Status number.
    • Previous pin: 6bdc86ee339c3007f15833434065adf893bff953.
    • Pinned main: 7a7ab84c3b5f26b20c098f439470d5bcfa972809, eight commits after the Friday evening pin. Comparison

    Eight pull requests landed on main after the Friday evening update. #498, #531 and #490 were closed on the records cited below. #371, #429, #430, #434, #522, #536 and #454 remain open. Unmerged #556, #553, #541, #545 and #546 are not main deliveries.

    2. What changed

    Change Record
    Retire infra/ and model/ into owning modules #535 merged November 21 07:18:33 Asia/Shanghai by Dong1017, merge commit fb343547f5c72155a8554038773866327da8d446, head 19ea7abc34262206d4949810d1194b19dfa4845d (103 files, +896/−235). Refs #498. Exact-head approvals: zengchuang-hw and amy-why-3459. The slice gives Apps its artifact helpers, moves the Anthropic provider modules into apps/integrations/anthropic/, and deletes infra/ and model/ with no forwarder. Dong1017 closed #498 as complete in 7054335234 after a static check of merge fb34354. That close is the recorded package-retirement completion. The comment states the wheel checks and test suite were not rerun there. The later harness/scenarios/ rename is a separate task (#558).
    Refresh README quickstart, architecture and release guidance #533 merged November 21 08:02:26 Asia/Shanghai by Dong1017, merge commit 6ed42af83a69f13fe88aaf0426ab8ef4f8ac7654, head e1d9065ff9881c43cfb478fa0a8aa235a018d4a1 (8 files, +452/−410). Closes #531 by the pull-request close keyword at merge. Exact-head approvals: zengchuang-hw and amy-why-3459. This is a documentation-audit close, not a new measurement verification of the curated README results. Open #557 tracks remaining README navigation and evidence follow-up.
    Move the MindIE adapter into integrations #537 merged November 21 12:47:27 Asia/Shanghai by zengchuang-hw, merge commit 13a9953ac682f831d5be6d9489c9092779410295, head f603309dad6f6327142ffcdaee6b0c2c7681aeae (24 files, +835/−712). Refs #490 and #371. Exact-head approvals: amy-why-3459 and Dong1017. Canonical MindIE now lives under apps/integrations/mindie/; published compatibility exports remain at apps.mindie. This is #490 step 4b, not production-port acceptance.
    Verify fixed CPU records across read-only views #542 merged November 21 14:27:28 Asia/Shanghai by Dong1017, merge commit c56c0808ac78a6ee2c18aadbeb347a337addc7b5, head 96b327fe1222d297275f340c35e4a428393add2b (2 files, +131/−2). Refs #490 and #371. Exact-head approvals: Dong1017 and amy-why-3459. An earlier same-head DO NOT PASS was superseded by those approvals. The test checks one ordinary CPU RecordRef across CLI, report, Dashboard and an isolated process, with assessment asserted absent. zengchuang-hw closed #490 as completed in 7061223067, naming c56c0808 as the final acceptance merge. That close keeps #371 open and keeps #429 / #434 production execute and offline assess outside the close.
    Lead benchmark reports with versioned metrics #539 merged November 21 14:27:34 Asia/Shanghai by Dong1017, merge commit 91b0d6572e72c83e23b1a107407a30114134c7a8, head 598ebb6728a52077fda2bb4a8a356e2c5e59c7da (19 files, +1066/−28). Refs #536. Exact-head approvals include zengchuang-hw, Dong1017 and amy-why-3459. The slice adds versioned report metrics and coverage. #536 remains open; those approvals do not record issue acceptance.
    Explain gains with profile evidence and paired statistics #457 merged November 21 14:27:39 Asia/Shanghai by Dong1017, merge commit 9466bc8d910c9fef55d4a9e23ee1a82950280c30, head f97c66817d8b4ecba8a8cd077ca6d6081626db51 (16 files, +1922/−27). Refs #454. Exact-head approvals: amy-why-3459 and zengchuang-hw. The added paired-statistics and explanation switches default off and do not change the default verdict. #454 remains open. This is not hardware A/A, GPU/NPU measurement, or RFC closure.
    Exploratory Inkling 2 Mini Apple silicon record and scheduling lessons #544 merged November 21 18:45:29 Asia/Shanghai by Dong1017, merge commit 91addd96db4e81218f23d9d85a21b57080b6cef9, head 655c3dc8268278d115585494dc7e4279b50457c8 (55 files, +2280/−0). Exact-head approvals: zengchuang-hw and amy-why-3459. The PR and those reviews record this as an exploratory / unverified documentation and examples landing, with no acceptance issue and no InferRSI PASS. It is not a measured-effect result on this pin.
    Preserve observed benchmark token usage #547 merged November 21 18:48:53 Asia/Shanghai by Dong1017, merge commit 7a7ab84c3b5f26b20c098f439470d5bcfa972809, head 0d731b8efba7fb99859340e7b7484954256cbeb4 (2 files, +17/−4). Refs #536. Exact-head approvals: Dong1017, zengchuang-hw and amy-why-3459. When a native terminal receipt has no usable counters, the campaign's observed subtotal is kept. Those approvals state this is not #536 acceptance.

    3. Architecture delta

    flowchart LR
      root["src/inferrsi root<br/>__init__.py + __main__.py only"]
      apps["apps/<br/>cli, storage/, handoff/, workflows/,<br/>integrations/cannbot, integrations/mindie,<br/>integrations/anthropic, host/,<br/>views/, delivery/, artifacts.py"]
      knowledge["knowledge/"]
      harness["harness/"]
      benchmark["benchmark/"]
      leaves["contracts/ + resources/"]
      root -->|"__main__ delegates"| apps
      apps --> knowledge
      apps --> harness
      apps --> benchmark
    
    Loading

    The diagram is the implemented package layout on this pin after #535 and #537: the root holds the four cores and the two technical leaves only. Canonical Cannbot, MindIE and Anthropic adapters live under apps/integrations/. Published compatibility shims remain at several apps.* paths, including apps.mindie. infra/ and model/ are absent. It is not a module import graph, and it is not #371 or production-port acceptance.

    4. Capability status

    | Workstream | Code delivered | Goal acceptance / measured effect |
    |---|---|
    | Package retirement #498 | #535 is on the pinned main | Dong1017 closed #498 as complete in 7054335234. That close is the recorded retirement of infra/ and model/. It is not a rerun of wheel checks in that comment, and it does not cover the later #555 test-import follow-up. |
    | Apps layout #490 / #371 | #537 and #542 are on the pinned main, after earlier storage, handoff, workflows and Cannbot slices | zengchuang-hw closed #490 as completed in 7061223067. #371 remains open. The close keeps production evaluation execute and offline assess with #429 / #434. |
    | Docs audit #531 | #533 is on the pinned main | #531 is closed by the #533 merge. That is a documentation-audit close, not a new measurement verification. #557 remains open for remaining README follow-up. |
    | Benchmark metrics #536 | #539 and #547 are on the pinned main | #536 remains open. The exact-head approvals do not record issue acceptance. |
    | Explainable gains #454 | #457 is on the pinned main | #454 remains open. The new switches default off. This is not a measured-effect PASS or RFC close. |
    | CANNBot round knowledge #522 | No new slice in this eight-commit range | #522 remains open. The requester retains completion and closure authority. |

    5. Evidence and limitations

    This check read the latest Status search, the Status 06 body and its comments, the #61 index, current main ancestry and the tree at the pin, the eight merged pull requests and their exact-head reviews, the #498 / #531 / #490 close records, and hosted check metadata. The mermaid diagram above reflects the implemented package layout on the pinned main; it is not a module import graph.

    On the pinned main, merge-lane and knowledge declaration policy succeeded on the push at November 21 18:48:56 Asia/Shanghai. That is not a full-suite or all-CI-green claim for this SHA. Full-suite software evidence cited above belongs to the reviewed pull-request heads. Green software checks do not establish measured effect, knowledge efficacy or campaign acceptance.

    Open #555 records that, after #457 and #535, test_recovered_attempt_does_not_trust_a_rewritable_declaration still imported digest from the retired inferrsi.infra.artifacts on main. The issue comments treat this as a test-only stale import, not a product regression, and point to unmerged #556. This report does not treat that failure as fixed.

    Author-recorded test counts at their recorded revisions were not independently replayed here. This report does not publish internal host or account paths.

    No tests, installations, model or provider calls, workflows or experiments were started for this report. No code, labels, settings or other workstream state was changed by this reporting run, except the authorized in-week Status comment and #61 index once those writes are verified.

    6. Next decisions

    1. Keep the task: 按职责整理 Apps 子模块目录(handoff / storage / workflows / integrations) #490 close and the merged layout tests separate from task: Apps 应用层详细设计与迁移计划(CLI / Python / Plugin) #371 / task: Harness evaluation-mode CPU execute for Apps #429 / task: Benchmark production prepare and frozen resource closure #430 / task: Benchmark production offline assess for Apps #434 production execute, offline assess, and shared CPU production wiring. Keep task: Apps 应用层详细设计与迁移计划(CLI / Python / Plugin) #371 open. Shim retirement remains later work under the task: 按职责整理 Apps 子模块目录(handoff / storage / workflows / integrations) #490 close record.
    2. Keep task: 定义 Benchmark metrics:单跑基础指标与 comparison 派生指标 #536 open until a recorded acceptance or close exists. feat: lead benchmark reports with versioned metrics #539 and fix(delivery): preserve observed benchmark token usage #547 are software slices only.
    3. Keep rfc: 可解释的改进——用 profile 证据和配对统计判定 PASS #454 open. The merged harness: explain gains with profile evidence and paired statistics #457 capabilities are optional and default-off, and are not a measured-effect PASS or RFC close.
    4. Keep task: Trace CANNBot knowledge through optimization rounds and the Dashboard #522 open until requester acceptance is recorded, as in the Friday increment.
    5. Unmerged Use the Harness digest in the interrupted recovery test #556 is not a main delivery; keep the Fix the retired digest import in the project-explanation recovery test on main #555 stale-import failure visible until a fix is on main. Unmerged feat(benchmarks): filter the task list with --iterative #553, Publish weekly stable minor releases alongside daily patch previews #541, docs: add DeepWiki navigation and clarify README guidance #545 and fix(apps): report Windows session lock contention as retryable #546 are not main deliveries. task: rename harness/scenarios/ to match its trial-execution role #558 is a new directory-rename task and is not started on this pin.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    area:processRepo-wide process, governance, privacy and security ruleskind:statusPeriodic status snapshot

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions