Repository navigation
knowledge: add source-grounded Lithos Metal candidate skills - #117
Conversation
Distill compiler fusion controls, GPU generation state and exact serving-cache ownership from the pinned source. Register the candidate and preserve source attribution and explicit model-validation limits. Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (author, integrator) Signed-off-by: jiangkuaixue123 <jiangxiaozhou111@163.com>
Keep AREX role, visibility and license declarations in local-skill-v1 sidecars without changing the repository reader contract; document the destination-specific normalization. Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (author, integrator) Signed-off-by: jiangkuaixue123 <jiangxiaozhou111@163.com>
Preserve the pinned unload sequence: release GPU engines, attempt bounded prefix persistence, then clear host snapshots. Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (author, integrator) Signed-off-by: jiangkuaixue123 <jiangxiaozhou111@163.com>
Regenerate registered skill abstracts against the source snapshot integrated with current main, retaining upstream skills and metadata. Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (author, integrator) Signed-off-by: jiangkuaixue123 <jiangxiaozhou111@163.com>
Update the existing ownership smoke expectation for the newly registered candidate. Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (author, integrator) Signed-off-by: jiangkuaixue123 <jiangxiaozhou111@163.com>
Dong1017
left a comment
Dong1017
left a comment
There was a problem hiding this comment.
Knowledge/content and packaging: LGTM at c1dc3b1, within #115's source-grounded candidate scope. One PR-description correction is needed before merging; see below.
| Reviewed area | Result |
|---|---|
| Compiler/fusion, generation/DSpark state, serving/cache guidance | Key contracts agree with pinned upstream 6653b7df; no actionable fidelity defect found |
| Generated projection | All 24 recorded source hashes match reviewed bytes and generation snapshot 9aec631; 58 existing-card changes only refresh source/registry provenance |
| Four-node graph | Effective metadata, navigation and resource references remain consistent; candidate / empirical_validation:not_run preserved |
| License/package resources | LICENSE and retained notices match pinned originals; package-data includes license and helper, and clean-wheel checks cover them |
| Remote CI | Full-check and installed-package lanes pass on Python3.10/3.11; current admission status is success (run) |
Before merge: remove the closing keyword in the PR description. The sentence 'this PR does not close #115' still contains GitHub's close #115 syntax despite the negation. This exact pattern already caused an unintended tracker linkage in #114/#91. Replace it with wording such as 'Issue #115 remains open for separate owner acceptance after this delivery.' Keep the ordinary Refs #115 reference. This is a metadata fix, not a request to change source or rerun hardware.
Review limits: independent source/metadata/hash inspection and CI verification, not a fresh full-suite run or reproduction of the author-reported CPU/Metal tests. No native workload, checkpoints, full-model continuation, DSpark numerical/quality or service-performance experiment was executed here. Those explicit unverified scopes do not block this knowledge-only candidate delivery. This review does not grant merge or issue-closure authority.
Dong1017
added
the
area:knowledge
ChatGPT Codex Connector
Dong1017
left a comment
Dong1017
left a comment
There was a problem hiding this comment.
Integration-only follow-up at unchanged head c1dc3b1: GitHub now reports mergeable=false / dirty against current main 4817f92. The earlier source-scope LGTM does not establish current integration readiness. Reconcile the branch with main, preserve other contributions, regenerate the packaged projection only as required by the existing source-binding contract, and validate the resulting head before merge.
The previously identified PR-body closing-keyword concern is also still present; it remains the same finding, not a new source defect. No unchanged full diff or hardware/model experiment was rerun in this integration check.
|
Integration follow-up for head This head also exhausted GitHub's |
Preserve current source assets, registry entries and matrix coverage while adding the Lithos graph. Normalize redundant sidecar defaults, retain diagnostic help without a module header, and align existing inventory assertions with the combined sources. The generated projection is refreshed separately from this committed source union. Agent-assisted-by: Codex / gpt-6.1-sol [effort=max] (material author, integrator) Signed-off-by: Dong1017 <55723162+Dong1017@users.noreply.github.com>
Refresh all skill abstracts and the reverse coverage view from the committed canonical source union. Preserve source grounding and candidate verification limits. Agent-assisted-by: Codex / gpt-6.1-sol [effort=max] (material author, integrator) Signed-off-by: Dong1017 <55723162+Dong1017@users.noreply.github.com>
Skill evaluationAutomatic execution / cumulative PR budget: Advisory: obtain missing evidenceRetained evidence availability: planned / unexecuted. No approval or merge authorization is issued.
PLANNED: execution is not authorized by knowing this hash. Selected comparisons: 0 · Estimated task calls: 0. Applicable only while these revisions and inputs match the current PR. Plan bindings, budget, frozen exam and advisory detailsUnchanged-correct is valid neutral evidence, not incremental benefit. Per-case outcomes (discovery excluded; regressions and gaps first): []Showing at most 20 of 0 pairs; the evidence link contains all outcomes, including omitted regressions/gaps. Plan SHA-256: Model/settings:
Provider/settings: Uncovered / unsupported paths (111)
|
Apply the current Python authoring rule to the three existing test files touched by this integration. Their executable ASTs are unchanged. Packaged source, source inputs and generated projections remain byte-identical. Agent-assisted-by: Codex / gpt-6.1-sol [effort=max] (material author, integrator) Signed-off-by: Dong1017 <55723162+Dong1017@users.noreply.github.com>
|
Integration-owner response for final head
The independent approval agent is assessing this complete head separately. Whole-Issue acceptance remains with the delegated Issue owner after review and integration. |
Dong1017
left a comment
Dong1017
left a comment
There was a problem hiding this comment.
LGTM 2709a51
Independent reviewer: separate review-only Codex agent, gpt-6.1-sol / ultra, acting through the shared Dong1017 account. I did not author or edit the contribution, generate its catalog, commit, or push.
Reviewed head: 2709a51. Assessed base: 7d1d82f. This recommendation covers Issue #115's accepted English, source-grounded candidate delivery.
- Independently checked compiler/fusion controls, StepState/DSpark commit and native-ring completion ownership, and serving/pack/prefix/session/idle contracts against lithos-ai/lithos-metal at 6653b7df5f2c4392ec7b038960f2796b1499923f. The four-node guidance graph and stated mechanism limits are consistent. LICENSE and both retained notices exactly match the pinned upstream blobs.
- Reviewed the genuine current-main integration and preserved existing consumers. Eighteen original owner files are byte-identical; the four sidecar default deletions preserve effective metadata, and the helper inspection function is unchanged. All 28 owner-local links resolve. I independently checked all 540 generated source inputs against actual committed source 26743ab and verified that all 76 preexisting discovery records retain their semantics. Candidate/not_run and locally_verified:false remain explicit.
- Personally ran the existing complete CPU knowledge-contract module: 13 cases passed, with relevant input bytes verified against the source commit before and after execution. I inspected the author's 27 focused-test and real noneditable-wheel/isolated installed-check receipts on eebd094; these remain author-executed results. I independently verified that the final three-file test-header delta preserves executable ASTs and the identical package/generated/source bytes.
- Fresh final-head Ready CI passes full-check and installed-package on Python 3.10/3.11. The latest declaration evaluation and required knowledge-model-policy status pass. This supersedes the earlier evaluation of stale PR-body metadata on this same head. Skill evaluation plan/publish completed; its conditional execute job is not applicable to this plan and was skipped, not passed.
The old conflict/body findings are addressed by the real integration and non-closing Issue wording. I read the current discussion, both historical reviews and the current advisory; there are no inline threads or remaining actionable findings.
Explicit disposition of the Skill evaluation advisory: zero comparisons are planned/unexecuted and its 111 uncovered paths provide no efficacy evidence. The reviewed source/content, provenance, discovery and package checks satisfy this accepted source-only candidate scope; provider trials are not an acceptance requirement here. No paid execution, efficacy improvement or promotion is implied.
The historical 153 upstream CPU / nine M4 synthetic Metal / 697-suite results remain earlier author reports, not my reruns. Full-model continuation, DSpark numerics/quality, native-backend qualification and service performance remain unverified. This LGTM is a content/integration recommendation, not campaign PASS or independent merge/Issue-closure authority.
Governing Issue and delivery slice
Refs #115
The accepted scope is English, source-grounded candidate guidance for Lithos Metal: compiler/fusion, GPU-resident generation and DSpark state, and serving/cache validation. This PR delivers one registered entry with three owner-local subskills, pinned source/readsets and notices, a bounded software/CPU diagnostic, both knowledge matrices, and packaged discovery. Native/full-model qualification remains outside this accepted knowledge-only delivery.
Sources and integration
create-repo-skillandverify-repo-skill, adapted to InferRSI ownership and candidate-state contracts. Effective license/role/visibility metadata stays in sidecars; raw DisCo-frontmatter compatibility and deployment are not claimed.7d1d82f55cda89842bd23ee2da91b521ee60eb08, preserving all 76 existing registry entries, matrix rows, and other source assets, including the HF and SGLang contracts. The combined inventory is 77 registered roots, 124 skill documents, and 61 registeredlocal-skill-v1profiles; the six existing typed cases remain intact.references: []fields and removes the helper's module header while retaining explicit CLI help; its inspection function is unchanged. Historical source provenance, license/notices, DCO and actual Sol/xhigh authorship remain intact.26743abcfae48cd6223bd757e74f193d97e3e5a7, then natively regenerated all 78 catalog outputs and the coverage view intoeebd09469ca9a28fd50dc6da7c4dd90485ce17b1(G). All 540 source inputs are bound to that actual committed source. On Windows, the stock per-file Git argument list is 53,515 characters; the private generation driver uses unchanged native hashing/rendering and coverage functions with exact source-set and full-library Git guards. The overlong stock CLI was not rerun. No product generator, mapper, admission policy or evaluator was changed.2709a51bee9f5e27328d24b0a07dc30c6e13fe27(T) only removes the inherited module descriptions from the three existing test files already touched by this integration. Their executable ASTs, all 540 skill inputs, package source tree and generated outputs are unchanged from G. No repeated generation or wheel build was needed for this test-only delta.Checks and evidence
tests/test_knowledge_contract.py26743abc; all new contract assertions retainedruff check src tests tools--help26743abc; all output bytes read back and source inputs preservedeebd0946(594.46 seconds); T only removes module descriptions with executable-AST equality verifiedpython -Ioutside the checkout, using only the existing actual PyYAML 6.0.3 distribution copied into the private environment with per-file hash equality verified. No system site packages or source checkout were available to imports. T retains the identical package source tree and executable checker AST; final-head Ready CI supplies its fresh installed-package runsLocal checks use Python 3.13.5 on Windows; repository Ready CI supplies Python 3.10/3.11. No dependencies were upgraded, no upstream runtime was installed, and no model, Metal/GPU or provider job was run in this integration.
The original contributor reported 153 selected upstream CPU contract tests, nine synthetic Metal ICB tests on Apple M4 Pro, and 697 InferRSI tests with 14 skips and 70 subtests across the earlier delivery revisions ending at
c1dc3b1c235005bbbc7efc6e7ba3d3f3d2c93669. Those are retained historical author reports. This integration author did not reproduce them; the earlier source-review recommendation also remains scoped to its original revision.All four nodes remain candidate, with empirical_validation: not_run. Full-model continuation, DSpark numerical/quality validation, native-backend qualification and end-to-end service performance remain unverified. Software/static/CPU/synthetic checks establish only their stated scope and no campaign PASS.
Remaining work and Issue acceptance
The authorized Issue owner assesses the accepted criteria after current-head CI, independent revision-scoped LGTM and installed-package evidence are complete. Issue #115 remains open for separate owner acceptance. No automatic closing relationship is requested.
Knowledge authorship declaration