Repository navigation
[Status 04] InferRSI evolution: 68108700 → 1246348f (2026-11-06) #366
Description
Weekly edition for Friday, November 6, consolidated after the 17:00 Asia/Shanghai reporting deadline. This snapshot records the evidence available at the evening check.
1. Snapshot
- Previous edition Status 03, #286.
- Observation November 6, 2026, 19:00–19:09 Asia/Shanghai. This check consolidates changes since the prior edition's publication at October 31, 14:40:09. It carries forward the November 1–5 in-week records and resolves their later status changes. In-week record · Last delta
- Published body baseline
6810870046b4f493d39b6eaede0f1a4004369235. - Last published delta main
e843b9d31cb86e1fd12d00890c89798b0bedd1fe. - Pinned current main
1246348f252dbc7e473efb1752c2d81da15e23fd, 31 commits after the published body and seven after the last delta. Full comparison - Reporting window The actual GitHub publication timestamp on November 6 through that timestamp plus seven days, start inclusive and end exclusive, in Asia/Shanghai. The [Status 01] InferRSI evolution: initialization → ac702729 (2026-10-20) #41 index receipt records the exact bounds after publication.
Developers can now inspect recorded optimization work in the Dashboard, including explicitly adapted Ascend records, and install from two host-specific preview archives. Review and trial evidence have clearer contracts. Several trackers closed with scoped acceptance or a recorded stop decision.
2. What changed
| Developer takeaway | Integrated behavior and where to look |
|---|---|
| Recorded work is easier to inspect and qualify. | inferrsi dashboard adds offline HTML and a read-only loopback live view. The caller records context and campaign evidence. harness/dashboard_external.py verifies an explicit candidate-bound external record and its source hashes, then displays per-tier measurements and separate performance and quality verdicts. Disconnected input, absent history and declared nonapplicability have distinct states. External verdicts stay out of the verified gain ranking. In the execution path, --require-trial-check freezes opt-in keys in runspec.py. Missing, malformed or false checks on either measured arm yield INCONCLUSIVE and withhold qualified gain and its chart. #319 · #358 · #341 |
| Daily preview publication now serves Codex and WorkSwarm. | .github/workflows/release.yml builds and validates packages, optionally generates notes through tools/generate_release_notes.py, and publishes tags and assets in the separate publisher. The native WorkSwarm v0.2.7 adapter packs the canonical runtime into runtime/payload.zip and verifies a content-addressed per-user cache to avoid deep import paths on Windows. The first dual-host preview, v0.1.4, published at November 6, 07:08:32 Shanghai, from 5c86eb57. It includes the wheel, sdist, both host ZIPs and SHA256SUMS. The scheduled run succeeded and the published body includes AI notes and the GitHub changelog. Its start at 07:01:30 was later than the nominal 02:00 trigger. #239 · #313 · #327 · Release · Run |
| Knowledge exposes applicability, exclusions and source boundaries. | Canonical metadata.json sidecars describe conditions and advice. tools/build_skill_coverage.py separates reasoned N/A, unknown, supplementary guidance and positive mappings. The methods/hardware/framework batches preserve source and efficacy boundaries. The reverted bulk PyTorch/TorchNPU delivery was replaced by clean #312. The current repository/framework view has advice on 29 of 31 roots, with the two newer PyTorch/TorchNPU roots outside the frozen advice batch. Husky's research row remains unknown for resource coverage. The repositories/vllm-ascend/ skill adds v0.27.1rc1 model/fusion/communication history with version and dispatch checks. #282 · #312 acceptance · #264 disposition · #354 |
| Evaluation and review records preserve their separate decisions. | tools/skill_eval_pr.py makes nonpaid plans actionable, and author imports preserve unknown command counts. .agents/skills/review-pr/SKILL.md binds PASS/DO NOT PASS to the reviewed revision and documents the exact LGTM approval token. tools/check_labels.py reads explicit external label ownership alongside the managed manifest. #357 removes the retired synthetic KV-specific tooling and ready-lane job while retaining the four evidence bundles and the general editing adapter. #292 · #300 · #346 acceptance · #284 closure · #357 |
| Campaign archives preserve scoped improvements and missed goals. | The five-second eight-910B1 MiniMax-H3 archive records a historical lossy complete-HTTP comparison of 407.269 → 108.120 s, with 14 frozen media comparisons passing and 3.10% higher sampled peak HBM. The separate lossless work retains its own quality gates. Drama-3 FULL decode graphs improved one-stream throughput on two 910B1 cards, with the requested RTF targets still unmet and the archived verdict INCONCLUSIVE. The #353 DeepSeek campaign has a current three-tier target and retained intermediate results. The latest draft-attention attempts did not activate the new service path and remain INCONCLUSIVE. H3 #294 · Lossless #293 · Drama-3 #132 · DeepSeek record |
3. Architecture delta
The maintained structure remains Entry/Knowledge → Harness → Model/Infra. The Dashboard consumes stored artifacts and explicit external records. Trial qualification is part of the existing Harness, while host packaging and optional release-note generation run outside experiment execution. Reference and SearchSpace remain the authorities for source and coverage. Maintained architecture · External record contract
Proposed. #362 discusses Knowledge, Workbench and Benchmark boundaries and four public handoffs. #363 proposes one speculative successor depending on an unverified parent, with failure propagation and separate frozen validation. #365 proposes one YAML input for CLI and agents. The current source tree still uses its established modules and contracts. The three proposals await scope and interface decisions. #362 · #363 · #365
Documentation drift. The architecture directory map predates the Dashboard entry points. #238's lower deliverables line still calls #313 draft, while the current release record shows its merged path running. The legacy changelog label categories still differ from the managed manifest. The Drama-3 plan keeps placeholder baseline and candidate SHAs, limiting revision-bound replay. Release tracker · Drama-3 archive guide
4. Capability status
| Workstream | Delivery and existing owner | Goal acceptance and validation boundary |
|---|---|---|
| Dashboard and trial qualification | #319/#358/#341 merged. david6666666 retains #314. jiangkuaixue123 authored #358 and amy-why-3459 authored #341. | #356 is closed after the scoped repair, with two approvals at aeb2c7f7. The contributor reports local browser inspection of real recorded data. #314 remains open. #340 accepts the bounded software delivery, including the opt-in evidence gate. The booleans are adapter attestations. Dashboard review · #340 acceptance |
| Release and host lifecycle | #239/#313/#327 merged. zengchuang-hw retains #238 and the pending #330 integration. | Actual dual-host publication is verified. The author's native WorkSwarm installation receipt belongs to prior head b8a2cb54. The merged integration artifact has software checks. Client lifecycle, Codex import, schedule timeliness and the full three-host #152 goal remain open. WorkSwarm evidence · #152 scope |
| SkillEval, Indexer and KV pilot | Dong1017 recorded the #253/#255 closures and authored #357. #248 remains open and unassigned. | #253 accepts the bounded artifact path, with a 19/19 vs 19/19 tie on one reconstructed CPU fixture. No judge was adopted and corpus scoring was retired unachieved. #255 accepts 9.434% lower main-path latency on its frozen 120-pair CPU study, with wide tail uncertainty and the accuracy objective retired unachieved. Both KV attempts remain INCONCLUSIVE, with the second stopped at HTTP 403 and actual cost unknown. The retirement preserves those receipts. #253 decision · #255 decision · #248 criteria |
| Governance and evidence | #346/#284 closed with integrated review and label changes. #273 closed the R1–R6 decision record. Rohank143 retains open #328. | The first natural post-fix label run succeeded at 01c73521. R6's 512 KiB per file / 2.0 MiB net per PR thresholds are advisory. The proposed checker and external asset metadata validation await integration. Nightly · R6 decision · #328 |
| Knowledge and device work | Dong1017 retains #91. Rohank143 accepted the separate #350 Batch 1 scope in discussion. jiangkuaixue123 reports #353. | Metadata and source delivery have scoped dispositions. The #350 readset and acceptance reviewer still need confirmation. The latest #353 retained C1/C4/C16 values are 33.2732 / 35.6156 / 45.5526 ms/token, with 295/300 GSM8K in the contributor's record. All three tiers remain above the current ≤25 target. Earlier C32 results retain their original scope. #350 handoff · #353 retained result and failed dispatch record |
5. Evidence and limitations
This check read the full edition and relevant comments, #41 template and index, publication search, main ancestry, selected changed paths, scoped acceptances, reviews, and release/tag/run metadata. The embedded SVG is byte-identical to the previously rendered overview. The report carries hosted checks and contributor measurements at their original revisions and protocols. It does not independently reproduce device results, verify package installation or audit AI-provider identity, usage or billing. This reporting run performed no tests, device experiments, provider calls or workflow dispatches.
6. Next decisions
- Continue [Task] RSI Dashboard: Optimization Workflow Visualization and Trace Capture #314, task: configure a GitHub Actions release bot #238 and [RFC] Agent plugins: a consistent InferRSI workflow in Codex, DSH and openJiuwen #152 through their existing owners. Use the external record and host archive contracts when checking deployment and lifecycle acceptance.
- Keep task: measure whether InferRSI helps Claude Code and Codex optimize inference tasks #248's unfinished criteria, the [Tracking] SkillEval evolution: local evidence, PR review and optional LLM judge #253 tie, [Tracking] Skill Indexer: retrieval efficiency and accuracy #255's unachieved accuracy, and acceptance: DeepSeek V4.1 Flash 910C 8K/1K TPOT ≤25 ms/token at C1/C4/C16 #353's unmet target visible. Confirm task: 细化 vLLM Ascend 知识蒸馏与知识库整理 #350's source and review boundaries before expanding its slice.
- Review ci: enforce evidence byte budgets and external asset metadata (Refs #273, R6) #328 against the advisory R6 decision. Keep discussion: Knowledge / Workbench / Benchmark 三模块目标架构——现状、原语、交接时序与遗留问题 #362/[RFC] RSI 投机验证：覆盖推理引擎优化与 Agent/Skill 自我改进 #363/task: support shared YAML task input for the InferRSI CLI and agents #365 as design discussions until maintainers record ownership and interface decisions.
Activity
amy-why-3459 commentedon Nov 6, 2026
Thanks for publishing Status 04 — the main advance
68108700→1246348f(31 commits) is easy to follow, and the Dashboard / dual-host preview / knowledge-boundary sections make the week’s integrated behavior concrete.A few notes on the snapshot and the §6 next decisions:
Evidence contracts look aligned. Offline Dashboard + explicit external Ascend records (with disconnected / absent / N/A states kept distinct) and
--require-trial-checkfreezing opt-in keys are the right shape for “inspectable, non-overclaimed” gains. Worth keeping external verdicts out of the verified gain ranking as stated.Release path. Actual dual-host
v0.1.4publication is useful progress on ci: automate daily releases with Python and Codex plugin packages #239/ci: summarize release notes with a configurable AI API #313/feat: add native WorkSwarm plugin and release artifact #327. The remaining gap called out here — client lifecycle, Codex import, schedule timeliness, and the full three-host [RFC] Agent plugins: a consistent InferRSI workflow in Codex, DSH and openJiuwen #152 goal — still looks like the right acceptance boundary for task: configure a GitHub Actions release bot #238/docs: add an English Codex plugin installation guide #330 owners.Unmet / inconclusive work should stay visible. Agree with keeping task: measure whether InferRSI helps Claude Code and Codex optimize inference tasks #248’s unfinished criteria, the [Tracking] SkillEval evolution: local evidence, PR review and optional LLM judge #253 tie, [Tracking] Skill Indexer: retrieval efficiency and accuracy #255’s unachieved accuracy, and acceptance: DeepSeek V4.1 Flash 910C 8K/1K TPOT ≤25 ms/token at C1/C4/C16 #353’s ≤25 ms/token miss on the board rather than folding them into narrative success. Same for confirming task: 细化 vLLM Ascend 知识蒸馏与知识库整理 #350’s source/review boundaries before expanding that slice.
Design discussions stay discussions. discussion: Knowledge / Workbench / Benchmark 三模块目标架构——现状、原语、交接时序与遗留问题 #362 / [RFC] RSI 投机验证：覆盖推理引擎优化与 Agent/Skill 自我改进 #363 / task: support shared YAML task input for the InferRSI CLI and agents #365 are correctly parked until maintainers record ownership and interface decisions; no migration or merge-precondition should be inferred from this edition alone.
Doc drift called out in §3. The architecture directory map vs Dashboard entry points, task: configure a GitHub Actions release bot #238 still labeling ci: summarize release notes with a configurable AI API #313 as draft after the merged release path, and Drama-3 placeholder SHAs are good follow-ups — small doc PRs would reduce confusion for readers of this status series.
Happy to help triage any of the open owner tracks (#314, #238/#152, #328 vs advisory R6) when the next scoped PR lands.