[Status 03] InferRSI evolution: 48e0d589 → 68108700 (2026-10-31 catch-up) #286

Activity

1. Dong1017

   added

   area:processRepo-wide process, governance, privacy and security rulesRepo-wide process, governance, privacy and security rules

   kind:statusPeriodic status snapshotPeriodic status snapshot

   on Oct 31, 2026
2. Dong1017 commented on Nov 1, 2026

   

   Dong1017

   on Nov 1, 2026 – with ChatGPT Codex Connector

   CollaboratorAuthor

   More actions

   In-week developer update — 2026-11-01

   1. Snapshot

   Observed catch-up: 2026-10-31 14:40:09 → 2026-11-01 19:12:14 Asia/Shanghai, from the edition's actual publication to this consolidated check. This does not reconstruct intermediate daily observations; the weekly body remains unchanged.

   • Edition window: [2026-10-31 14:40:09, 2026-11-07 14:40:09) Asia/Shanghai.
   • Published baseline main: 6810870046b4f493d39b6eaede0f1a4004369235.
   • Previous delta: none under this edition.
   • Pinned current main: 67ee3cceba7824eac758d78652a8fe80f71b0c12; seven-commit comparison.

   Developers get clearer nonpaid author instructions, descriptive method contracts and narrower contribution checks. The separate KV freeze and new hardware campaign retain their own execution/acceptance boundaries.

   2. Five material changes

   Change	Developer consequence and implementation	Delivery, acceptance and effect
   Actionable nonpaid SkillEval handoff	#292 updates tools/skill_eval_pr.py to lead PLANNED feedback with next action, blocker and required evidence. Canonical/index changes can get an actual ancestor-bound source pin for the existing index check; plan and publisher use matching full history. #291 adds CLI stage/capture/report controls.	Both merged. Commands/previews are unexecuted instructions, not zero-valued results or refreshed historical bot evidence. A mapped path still needs an explicit relevance decision. #253's real author execution → evidence → independent disposition remains unfinished.
   One additional KV-pair binding, with old evidence preserved	#287 adds the exact PR275-only USD 350 lifetime exception; generic PR caps remain USD 200. #290 binds the reviewed v2 freeze through tools/kv_live_authorization.json and existing admission/policy tools.	Merged software/data delivery, not a new result. The first unknown-cost USD 162.72 claim and INCONCLUSIVE run remain. The latest published window ended October 31 at 20:00 Shanghai, with admission cutoff 19:38. The #248 comments read here contain no completed second-pair efficacy receipt. Enabled locator data does not override expiry or authorize a third attempt.
   Method conditions become visible without promoting maturity	#263 fills applicability/recommendations for 17 methods under the accepted field semantics. All 20 registered methods now have applicability metadata; canonical sidecars feed generated catalog/coverage.	Merged. Input contract, produced artifact, failure modes and refusal point describe the procedure; source bodies, candidate status and empirical-validation boundaries are preserved. This is neither new source extraction nor measured retrieval/agent benefit; other batches remain separate.
   Contribution checks get explicit, bounded rules	#296 adds tools/check_pr_references.py to detect closing syntax embedded in PR-body sentences, including negation. #298 adds exact gpt-6.1-terra at xhigh/max to .github/knowledge-model-policy.json.	Both merged. #273 R1 accepts a non-required ready-lane check on existing events; later body edits wait for another push. It does not inspect commit messages/Development links or complete #273. The model allowlist checks declarations, not runtime identity, source truth or efficacy.
   A distinct 910C H3 campaign now has an existing owner	#297, owned by Dong1017, separates four logical 910C devices and 5s/15s × lossless/lossy tracks from the eight-device 910B1 evidence in open #294. The owner record records the selected quality/target direction.	Open campaign, not merged capability or optimization acceptance. Engine environments, aligned no-approximation baselines and exact automated quality contracts still need freezing. Reported preflight timings are not matched optimization A/B evidence. Reuse the existing H3 session; no parallel campaign is implied.

   3. Architecture delta

   

   The layer split is unchanged. Changes sit in Entry/Knowledge sidecars, the existing PR planning/feedback surface and external contribution checks; KV execution still belongs to the existing bounded controller. Reference/SearchSpace remain authoritative. A PLANNED summary is work instruction; it neither runs its suggested command nor changes the grader.

   Proposed: #295 Async Workbench discusses manager/worker contexts and recovery. It is a draft with no assignee or delivered scheduler; #249/#268's existing ownership/review remains separate.

   Documentation drift retained: the daily-admission guide still describes an unchanged checked-in schema-1 rollout. The generic architecture future-planning line concerns the engine/campaign lane; PR planning already exists. These historical descriptions must not override current configuration/focused contracts.

   4. In-flight work and existing ownership

   • Dong1017, author of the merged PRs above and coordinator of task: measure whether InferRSI helps Claude Code and Codex optimize inference tasks #248/[Tracking] SkillEval evolution: local evidence, PR review and optional LLM judge #253: preserve old attempts, reconcile actual author evidence and separately disposition remaining acceptance. This report grants no execution, budget extension or retry.
   • Dong1017, assigned [Campaign] MiniMax-H3 Ref2VA on 910C: 4 logical devices, 5s/15s × lossless/lossy optimization #297: continue its existing environment/baseline/quality-contract preparation. Qualify MiniMax-H3 Ref2VA five-second Ascend optimization #294 is separate and still open.
   • NumberWan, assigned task: parallel optimization rounds with orthogonal lanes for multi-agent harnesses #249: continue the existing open feat(harness): offline parallel-round coordinator #268 review. [Draft RFC / Experimental] Async Workbench：独立 manager 与 loop worker 的上下文和生命周期 #295 supplies discussion, not an accepted replacement.
   • [Investigation] Agent-led workflow gaps: closure, decision visibility, review provenance and focus #273 R6 / task: agent-maintained issue system: one label scheme, auto-labeling and a pending-decision view #284: raw-evidence publication limits and issue/decision visibility remain open; no overall acceptance or new owner is inferred.

   5. Evidence and limitations

   This session inspected pinned source, actual merge flags, complete Status comments, relevant issue/PR discussions and owner decisions, and locally rendered the pinned SVG. Software/CI claims are attributed to their linked authors/reviewers; they were not rerun here. #291's controls are synthetic in-process CLI checks, not actual local-agent or sandbox qualification.

   The workflow-specific manual-run query was unavailable; a sample of 30 recent repository runs is incomplete. No absence-of-execution, current ledger balance, billing settlement or second-pair effect is inferred from that sample or the missing outcome comment. The published v2 window remains historical.

   No tests, model/provider/evaluation clients, GPU runs, workflow dispatch or repository-setting changes were performed for this update.

   6. Next decisions

   1. Reuse [Tracking] SkillEval evolution: local evidence, PR review and optional LLM judge #253's existing author/reviewer path and require actual command/input/output/failure receipts; do not count displayed instructions as execution.
   2. Under task: measure whether InferRSI helps Claude Code and Codex optimize inference tasks #248, reconcile the additional-pair outcome/obligations and expired freeze through its existing owner. Keep missing evidence unknown and preserve the first interruption.
   3. Continue [Campaign] MiniMax-H3 Ref2VA on 910C: 4 logical devices, 5s/15s × lossless/lossy optimization #297's four-track preparation and [Investigation] Agent-led workflow gaps: closure, decision visibility, review provenance and focus #273/task: agent-maintained issue system: one label scheme, auto-labeling and a pending-decision view #284/task: parallel optimization rounds with orthogonal lanes for multi-agent harnesses #249's own decisions without converting proposals, declarations or CI into efficacy acceptance.
3. Dong1017 commented on Nov 2, 2026

   

   Dong1017

   on Nov 2, 2026 – with ChatGPT Codex Connector

   CollaboratorAuthor

   More actions

   In-week developer update — 2026-11-02

   1. Snapshot

   Observed catch-up: the previous publication, 2026-11-01 19:13:56 → 2026-11-02 19:28:08 Asia/Shanghai. This consolidates the actual check rather than reconstructing daily observations; the weekly body remains unchanged.

   • Edition window: [2026-10-31 14:40:09, 2026-11-07 14:40:09) Asia/Shanghai.
   • Published edition main: 6810870046b4f493d39b6eaede0f1a4004369235.
   • Previous delta main: 67ee3cceba7824eac758d78652a8fe80f71b0c12.
   • Pinned current main: 2baebf6ebb815ebf07d21c41e3b3754858c0798e; 11-commit comparison.

   Developer consequence: an actual downloadable preview now exists, operating knowledge has a clean accepted delivery, and H3 evidence distinguishes a bounded historical lossy result from still-unaccepted lossless/four-track goals.

   2. Five material changes

   Change	Developer consequence and code/contract	Delivery, acceptance and effect
   First natural release completed	#239 merged the existing-source → package validation → tag/Release pipeline in .github/workflows/release.yml and tools/check_release.py. The scheduled run succeeded; v0.1.0 Developer Preview was published November 2, 05:35:13 Shanghai, with wheel, sdist, standalone Codex ZIP and SHA256SUMS.	API readback confirms non-draft/prerelease, bot author, all four assets and tag source 6c9f90daae2d404915378b910f581b4dcff7e7da; that release is not the current main snapshot. The run started at 05:29 after the nominal 02:00 slot. #238 still awaits requester acceptance; actual account import and future timeliness are unverified. GPT-generated notes remain the open #313 extension.
   H3 evidence landed, with separate unresolved targets	#294 merged the eight-910B1, C1, 124-frame/50-step evidence under evidence/h3-ref2va-ascend-293/. Its frozen report records 407.269→108.120 s warm complete-HTTP latency, 73.45% lower, with 3.10% higher sampled peak HBM.	Historical lossy Cache-DiT evidence: A1/B1/B2/A2, two server means per arm, three measured requests/server after two warmups. Fourteen frozen media comparisons pass; this is not statistical-significance or semantic-quality certification. Fifteen seconds was deferred. The new #293 direction instead requires lossless 5s and 15s with no audio degradation; #318 is open, with no new accepted lossless speedup. Separately, draft #317, read at a0677a03, reports only 1.57% from one four-logical-910C screening pair, below its 50% target; overall INCONCLUSIVE, failed lossy profiles and incomplete final confirmation remain.
   PyTorch/TorchNPU operating graphs accepted after a clean resubmission	#312 supplies two root skills, 16 child procedures and 12 helpers under frameworks/pytorch/pytorch/ and frameworks/torch-npu/torch-npu/, plus registry/catalog/package integration. It follows the explicit oversized-delivery revert #311.	#309 acceptance accepts the requested clean replacement: 111 runtime files retained, 160 construction/process fixtures omitted, historical evidence linked rather than recopied. Source/software/package checks do not qualify CUDA, NPU, HCCL, source builds or skill efficacy. Count the final replacement, not both the reverted and replacement deliveries.
   Unknown author command counts stay unknown	#300 updates tools/skill_eval_local.py and skill_eval_compare.py: new unparsed author logs retain commands: null; legacy author-reported entries with unknown completed turns render an unknown observed command count.	Merged reporting correction. Explicitly parsed zero remains zero, mixed unknown aggregation stays unknown, and historical artifacts are not rewritten. The compatibility tool_calls field is not an authenticated count of all model tools. #253's broader effectiveness/author-evidence acceptance remains open.
   Issue navigation now has a shared label contract and drift report	#299 adds the 21-label .github/labels.json, status/priority conventions, template labels and read-only tools/check_labels.py in the existing nightly lane. The existing contributor's receipt records live label reconciliation.	Delivered rules and reported application; #284 remains open awaiting its existing completion/first-nightly verification. status:proposed means a named pending decision. Drift reporting does not auto-fix labels; label maintenance must not manipulate the protected paid-execution switch. No labels/settings were changed by this reporting session.

   3. Architecture delta

   

   The maintained layer split remains. New operating graphs enter through canonical knowledge/Reader; release tooling consumes packaged runtime/resources outside the experiment loop. H3 engine changes remain downstream contributions; retained campaign evidence does not add an InferRSI execution backend. Reference/SearchSpace remain authoritative. Proposed GPT notes and broader coordination work are separate from implemented publication and read-only checks.

   Configuration/documentation drift: release-note categories and the release guide still name legacy enhancement/bug/documentation, while the current label manifest uses prefixed names. The catch-all category remains present. Categorization needs reconciliation through existing owners; missing notes or publication failure are not inferred.

   4. In-flight work and existing ownership

   • zengchuang-hw, existing task: configure a GitHub Actions release bot #238 implementer: requester acceptance, actual Codex import evidence and the separate ci: summarize release notes with a configurable AI API #313 notes proposal.
   • amy-why-3459, Qualify MiniMax-H3 Ref2VA five-second Ascend optimization #294 author / existing eight-910B1 workstream: [Campaign] MiniMax-H3 Ref2VA Ascend NPU latency optimization with video/audio quality gates #293's lossless duration/audio contracts and open H3 Ref2VA lossless follow-up: validate audio coverage and scorer identity #318. Dong1017, assigned [Campaign] MiniMax-H3 Ref2VA on 910C: 4 logical devices, 5s/15s × lossless/lossy optimization #297: independent review of draft evidence: track H3 Ref2VA on four logical Ascend 910C devices #317 and remaining four-track disposition; its direction ceiling is reported exhausted, with further work requiring requester authority.
   • david6666666, accepted knowledge: complete AREX distillation of PyTorch and TorchNPU #309/knowledge: add PyTorch and TorchNPU skills without process artifacts #312 delivery: retained source/native limits remain explicit, without reopening completed packaging work.
   • Existing task: agent-maintained issue system: one label scheme, auto-labeling and a pending-decision view #284 acting contributor (Dong1017 account): first-nightly/closure evidence; Dong1017, fix: preserve unknown command counts in author imports #300 author: remaining [Tracking] SkillEval evolution: local evidence, PR review and optional LLM judge #253 evidence scope. This report takes over none of these tasks.

   5. Evidence and limitations

   This session inspected pinned source, full edition/comments, merge flags, specific acceptance/decision receipts, release/tag/asset metadata and the scheduled run record. The pinned SVG was rendered locally. Package binaries were not downloaded, installed or imported; API digests are metadata, not a new independent archive validation.

   H3 measurements/quality checks and software test results are carried from their exact author/reviewer records. Original media, weights and host logs remain external; no device replay or semantic-quality review was performed here. #317 is a moving draft, and its historical execution revisions are distinct from its publication head. A green CI job, accepted knowledge packaging or a published Release cannot establish campaign acceptance.

   No tests, provider/model/evaluation clients, GPU/NPU experiments, workflow dispatch, credentials or repository-setting changes were performed for this report.

   6. Next decisions

   1. Reconcile task: configure a GitHub Actions release bot #238's publication/import acceptance and timing expectation, and its label-category drift, without rewriting the existing public Release.
   2. Keep [Campaign] MiniMax-H3 Ref2VA Ascend NPU latency optimization with video/audio quality gates #293's new lossless requirements and [Campaign] MiniMax-H3 Ref2VA on 910C: 4 logical devices, 5s/15s × lossless/lossy optimization #297's four-track target separate from historical lossy results and primitive-level tests. Preserve failures and INCONCLUSIVE; use existing owners for further budget/acceptance decisions.
   3. Finish task: agent-maintained issue system: one label scheme, auto-labeling and a pending-decision view #284's existing verification/closure and [Tracking] SkillEval evolution: local evidence, PR review and optional LLM judge #253's actual author-evidence requirements; do not substitute labels, command-count fixes or source/CI coverage for efficacy.