Repository navigation
knowledge: distill Splash with portable AREX workflow - #186
Conversation
Distill format/memory, serving/cache and DFlash2 qualification contracts with explicit native backend blocks and source-Python verification. Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (approved_arex_author; author and integrator) Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (runtime_memory_author; author) Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (serving_state_author; author) Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (qualification_author; author) Signed-off-by: WeiQing Chen <530634352@qq.com>
Generate source-bound abstracts and account for the one additional packaged knowledge record in the exact installation total. Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (approved_arex_author; generated integration and validation) Signed-off-by: WeiQing Chen <530634352@qq.com>
Skill evaluationPLANNED: 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. Evidence and logs · Retention: 30 days. Pipeline evidence only; offline fixed outputs do not establish skill efficacy. Plan bindings, budget and frozen examPlan SHA-256: Model/settings:
Uncovered / unsupported paths (130)
|
Package extensionless third-party notices and assert their installed attribution. Distinguish model-stop/output-limit exclusions from later stop-string cuts in accepted-proposal counters; reconcile the assertion inventory. Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (approved_arex_author; review fixes) Signed-off-by: WeiQing Chen <530634352@qq.com>
Bind generated abstracts to the corrected source snapshot; preserve notices and precise accepted-proposal counter semantics. Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (approved_arex_author; generated integration) Signed-off-by: WeiQing Chen <530634352@qq.com>
Dong1017
left a comment
There was a problem hiding this comment.
Scoped early draft review at ae11c88, base 4817f92, within #185's partial-delivery scope: no verified actionable blocker in the inspected helpers, graph/qualification/cache guidance, tests and packaging changes. Targeted comparisons of the budget parser and ABBA reduction with pinned upstream code agree; the extensionless notices packaging correction is present.
Ready CI 40491154617 passes all four full-check/installed-package jobs on Python 3.10/3.11. This is software/package evidence. The Skill Eval plan has zero selected comparisons; Metal/ANE/model/native-persistence gaps remain explicitly unverified and are valid limitations of the accepted partial draft, not new feature requirements.
Not a merge-ready LGTM: retain draft status, the prescribed ordered main/catalog reconciliation and separate informed candidate-admission prerequisite. Targeted static review only, not exhaustive upstream/hash verification; no PR code, model or hardware execution.
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
Preserve existing main and both candidate graphs; resolve only discovery, counts, navigation and isolated test paths. Native limits unchanged. Signed-off-by: WeiQing Chen <530634352@qq.com> Agent-assisted-by: Codex / approved_arex_author (gpt-6.1-sol, xhigh)
Signed-off-by: WeiQing Chen <530634352@qq.com> Agent-assisted-by: Codex / approved_arex_author (gpt-6.1-sol, xhigh)
Governing issue
Refs #185
User-approved native/source-module adaptation and partial candidate drafting. Final informed limited-candidate admission and issue closure were explicitly authorized by the user on2026-10-27T12:08:19Z; all candidate/backend-unverified markers and known failures remain. Strata is explicitly skipped. Merge order remains DwarfStar → Splash → Lily → Husky; before Splash integration, refresh against the resulting main/catalog after DwarfStar.
Delivered
Pinned Splash operating graph with runtime-planning, serving-cache and qualification routes; source hash/readset/license/notices, three offline helpers (budget syntax, request JSON, descriptive ABBA reduction),10 assertion-backed cases and AREX integration/verification records. Registry/catalog, two matrices and exact packaging expectations update together. Method is actual portable AREX meta-skill execution by a compatible agent, not a DisCo CLI/model API run.
Actual checks
-34 source Python LRU/protocol tests PASS (fake-peer/golden codec logic, not native C++/Metal cache/scheduler).
-50 source-parser/bundled-helper equivalence fixtures PASS (syntax only, not memory fit).
-Ruff and3 graph/helper tests PASS; helper help/positive/negative fixtures PASS.
-AREX recursive license gate valid across4 SKILL.md files. Actual gh query unavailable, so NO_LICENSE fallback remains warning only; source Apache-2.0 and THIRD_PARTY_NOTICES independently preserved.
-Full InferRSI suite before review fixes passed1332 tests,8 skipped,135 subtests. After the fixes, the stable-tree full suite passed1332 tests,8 skipped,135 subtests; the clean noneditable wheel check passed109 records and retained the unchanged third-party notices. Current exact-head CI remains authoritative. No independent GitHub approval is claimed.
Independent review fixes
The read-only review found extensionless THIRD_PARTY_NOTICES missing from the wheel and an ambiguous stop-counter sentence. Package-data now includes that exact basename with an installed attribution assertion; source notice bytes are unchanged. Accepted proposals cut by a later stop string still count, unlike model-stop/output-limit exclusions. The reference and assertion inventory are corrected, and the source-bound catalog regenerated. Independent report remains outside the repository; no public review COMMENT/APPROVE is claimed.
Required gaps
BLOCKED_REQUIRED_BACKEND: Apple Metal/ANE/native model/device and actual disk-state execution. No compatible Apple host/weights were available. Numerical/model quality, native recurrent/cache recovery, service performance and skill efficacy are unverified. No weight download, GPU/paid provider invocation or upstream change. CPU/synthetic checks do not clear native gates. Adaptive verify-row, image-cache salt and process-local Responses-history boundaries are explicit.
Artifacts:
tests/fixtures/repo_to_skill/splash/. Pinned source:39a97385f8560cf8a93972cec9b0a963c9195250; method:ac3fe1afa80fb9a09775ecfb2b6cc3ba850a2db6.Knowledge authorship
Merged candidate knowledge is not engine qualification, campaign PASS or efficacy acceptance. Issue closure remains separately assessable; Refs is intentionally non-closing.
Sequential integration
DwarfStar merged first at e013ba8. This branch integrates that main without changing either runtime graph; preserves current main tooling, combines navigation and registry, isolates the original Splash tests under test_splash_repo_to_skill.py, and regenerates from published source99fca181e71fab22c09bb1967335d0653852eff3. Counts62 roots/89 skill documents/110 records.23 focused graph/index/library tests and full ruff pass. Independent delta review and new exact-head Ready CI completing; no native qualification promotion.