Skip to content

refactor(apps): canonicalize CLI and retire root entries (A1) - #461

Merged
Dong1017 merged 6 commits into
mainfrom
impl/a1-cli-t381
Nov 16, 2026
Merged

Dong1017 merged 6 commits into
mainfrom
impl/a1-cli-t381

Conversation

@Dong1017

@Dong1017 Dong1017 commented Nov 16, 2026 •

Copy link
Copy Markdown
Collaborator

Scope

Refs #408

A1 only, following merged C1/C2/R1. This is one slice of the open Apps/architecture issues; A2/M1/R2/R3 are not implemented.

  • Make inferrsi.apps.cli:main the canonical CLI, with arguments in inferrsi.apps.cli.arguments; route package __main__ directly to that canonical entry.
  • Delete root cli.py and arguments.py, without permanent forwarding shells or sys.modules aliases. Retire the R1 temporary arguments-import exception and update active consumers, tools, plugins, tests, migration documentation and the exact manifest/root-inventory guard (10 → 8).
  • Keep the architecture boundary precise: only the necessary top-level __main__ → apps.cli main seam is allowed. Preserve the assertion that a clean scan does not imply core isolation, and preserve negative fixtures. This PR does not claim full isolation or relax broad banned prefixes.
  • Preserve Read fixed attempt status and archived claims in Apps #449/Connect report Dashboard and trace to fixed attempt records #450 CLI features and host/injection/error behavior; no unrelated refactor.

Main integration and provenance

Initial implementation baseline: aaa17a5dd333449d15fb052bb09b2012d55f2327.

Normal signed-off checkpoint 41491c966b2606a0f4cefc48eb703b9649fa0d2e, followed by ordinary merge of main f536288710c726a827bfa3e4357c2c911984cbc3 (#351), producing 18f189a8873da75b4fbcae2910711b2a37a8bbf3. No rebase, amend, force push or skipped hooks. All nine incoming main paths were preserved; the installed checker retains A1 import changes plus main's precise resource count 925.

All substantive A1 source, tests and documentation were authored in the original Claude Code session using the requested/reported claude-fable-5-2, effort max. This is client/session attribution, not a backend attestation. The coordinator performed independent review/acceptance, mechanical Git/snapshot/build operations and this PR metadata, not substantive product drafting. Manifest relocking and precision-guard changes still require formal architecture review; passing tests is not that approval.

Independent acceptance — initial implementation head 18f189a88

Evidence is retained in the T381 execution records and the owned CPU-only H200 isolation directory. No GPU, training or other users' services were used. Heavy tests/builds now run on H200, serially.

  • Scope/provenance and full diff reviewed. Changed 44 Python files passed Ruff; the final two guard files also passed. No claim of whole-repository Ruff cleanliness (the previously recorded unchanged harness/run_store.py F811 baseline remains outside this PR).
  • 4,681 tests collected at the recorded candidate point; four seam/document freshness checks passed. Subsequent exact guard additions were collected and validated separately. This was not a full test-suite run.
  • Nine real CLI command exit/stdout/stderr byte comparisons passed against both pre-A1 main and Connect report Dashboard and trace to fixed attempt records #450. Relevant CLI/parser/package-entry ASTs stayed unchanged afterward; incoming main changes were separately checked rather than repeating unchanged baselines.
  • Seven focused CLI runtime tests passed, including Read fixed attempt status and archived claims in Apps #449/Connect report Dashboard and trace to fixed attempt records #450 behavior, bad flags, SQLite revision handling, propagation of programming exceptions, and dashboard/live flags.
  • Exact root/leaf/retired-import/architecture negative checks passed in recorded focused groups (35 tests + 4 subtests, and 64 tests + 9 subtests; overlapping evidence is not added into a fabricated total). The final legal main seam was independently checked with 8 selected tests and 39 subtests. No scan, assertion or timeout was weakened.
  • Windows host run: 63 passed, 8 failed. Seven were actual 45-second isolated subprocess timeouts; the eighth was DSH parsing empty output. These failures remain recorded; Windows is not represented as green. A direct DSH helper comparison of candidate vs immutable main subsequently returned exit 0 for both (15.738s / 11.709s); the original transient failure's cause is not established.
  • H200/Linux: all seven Windows-timeout cases passed, 0 failed, 253.64s, under their existing tests and deadlines. The other unchanged 63 Windows passes and previous baselines were not rerun. After main integration, the affected tests/test_skill_coverage_view.py module passed (exit 0, 28.511s).

Real distribution / installed-entry acceptance on the merged head

  • Built a real sdist, then a real wheel from that sdist; created one fresh Linux venv, installed the wheel normally outside the checkout, and installed the declared skills PyYAML extra needed by the checker. No Windows venv was copied. Core metadata declares no runtime dependencies.
  • Version 0.1.11.dev22+g18f189a88, obtained first from the actual clean native Git checkout. The verified source archive uses the documented setuptools-scm pretend-version input for the sdist only; wheel metadata comes from the sdist PKG-INFO.
  • The actual installed project checker passed, including 925 installed resources, installed rather than source origins, and the retained knowledge/resource checks. Resource inventory SHA256: 8a59aef44a44385ae050d555b64ee33456622cdd2970ac42d618719b5637fd99.
  • Wheel root8, canonical entry-point metadata, exact root __main__, 13 retired import failures/no old CLI aliases, REPO_TYPES identity, all seven pack bytes and package version checks passed.
  • The real console and python -m inferrsi --help both exited 0 outside the checkout. An extra coordinator harness assertion initially and incorrectly required raw console/module help to be identical. That failed receipt is preserved. Argparse normally renders inferrsi vs __main__.py usage labels/wrapping; each complete captured output was then independently matched byte-for-byte to the unchanged installed parser with its own actual program name. No product/test/guard/timeout changes or rebuilds were used to hide the harness error.
  • All 3,066 candidate source-file SHA256s still match the transferred merged candidate after build/installation. Source content SHA256: cafce2876be0877bf4b0e8b4dee441dfce2bda9545bf244b9e5abb69e1ac986b.
  • sdist SHA256: b0c77fc9bd34d5354d17bce306acad49ffc357a67d3fe0d5cceda5d7fd450a2d.
  • wheel SHA256: 6872b5931ec93f802d3b55e4989fdd9b4d878c00eb4d73c7bfd439c1ba8602e5.

Full Linux/Python CI and independent architecture reviews remain required; no automatic merge or issue closure is requested. Existing old408 dirty worktree/evidence branches were not modified. Per the owner's handoff instruction, this workflow stops after this A1 PR closes; later queue work is not started here.

Review follow-up — head 7fd5bb1fe388c9046b9053acf61c5d393c0c2cd9

Follow-up to review5915579655, which assessed 18f189a8873da75b4fbcae2910711b2a37a8bbf3.

  • Both completed full-check jobs at that initial head had one identical failure, tests/test_campaign_record_injection.py::test_harness_records_load_no_knowledge_or_apps_module: a stale expectation of eleven retired roots despite the legitimate A1 total of thirteen. Each run also reported 4,657 passed, 27 skipped, 616 subtests passed; those initial failures are retained, not relabeled as success.
  • Original Fable session authored the precise repair: pin R1's historical eleven, the exact two A1 CLI/parser paths, disjointness and union, and total thirteen. Independent exact Edit replay confirmed authorship. The original subprocess isolation script, negative assertions and deadlines are AST-identical; no runtime/package/CI-policy change, dropped assertion or timeout increase.
  • H200 single-worker CPU acceptance on this fix: the entire affected campaign module: 20 passed in 6.26s; changed-file Ruff passed. The real installed-package checker also passed again outside checkout with 925 resources, using the previously built/fresh-installed 18f189a88 wheel. All package, resource, pyproject and checker bytes are unchanged; no new-head wheel build is falsely claimed. New GitHub installed-package jobs will build their own new-head distributions.
  • Full transferred fix candidate: 3,066 source files, content SHA256 294b9a744e1bc3efeee2cb3db5a089012165c395ed42ef5e9f75e21d29e9af36, all unchanged after acceptance. The earlier sdist/wheel hashes and version below describe the initial implementation distribution, not a rebuilt fix-head artifact.
  • Normal signed-off follow-up commit, direct child of 18f189a88; latest main remains f536288710c726a827bfa3e4357c2c911984cbc3. A real fast-forward push triggers fresh CI using the corrected standalone Refs body, rather than manufacturing a green check from the old failing PR-reference event.
  • Still open: successful current-head full CI and the Windows host-evidence gap identified in the review. The original Windows 63 passed / 8 failed record remains unchanged. No new Windows/macOS run or hosted job is claimed, and H200/Linux success is not proof of Windows behavior. Architecture approval remains pending.

Follow-up — current head cb59d3aa135d803d8dfba84d7bbd94a918551bc8

  • 960082badc9e0554d12b1563637fb0755dad0f83 adds one host-compatibility job to .github/workflows/ci-ready.yml. On windows-latest and macos-latest, one host at a time, it runs only the eight nodes from the recorded Windows run, then builds a wheel from the sdist, installs it with [skills] into a fresh venv outside the checkout, runs the unchanged tests/check_installed.py with -I, and compares the console script with python -I -m inferrsi. The original Claude Code session (claude-fable-5-2, effort max) authored the job; replaying its single recorded Edit on the parent reproduces the committed file byte for byte. No existing job, trigger, assertion or test deadline changed.
  • cb59d3aa135d803d8dfba84d7bbd94a918551bc8 is an ordinary signed-off merge of main 016c7c7df023944f0f46d8e777d6d31f1df1764c (Extend VLA observation-to-action benchmark guidance #342, Wire the Harness loop evaluation port through Apps #462) with no conflicts. The installed checker keeps the A1 import changes and takes main's resource count 927. Wire the Harness loop evaluation port through Apps #462 adds no reference to the retired root CLI modules.
  • Same-host Windows comparison of the eight nodes: comment. Unchanged main and the A1 candidate both gave 5 passed / 3 failed, with the same three isolated-subprocess commands exceeding the existing 45 s deadline; these are attributed to local resource limits on that machine.
  • Hosted Windows/macOS results come from this head's host-compatibility job and are not claimed here before the run finishes. Architecture approval remains pending.
{"schema_version":1,"head_sha":"cb59d3aa135d803d8dfba84d7bbd94a918551bc8","human_only":false,"material_ai_authors":[{"tool":"Claude Code","model_id":"claude-fable-5-2","effort":"max","role":"author"}]}

Refs #408

Agent-assisted-by: Claude Code / claude-fable-5-2 [effort=max] (author)
Signed-off-by: GUOGUO <xwdong1998@163.com>

Copy link
Copy Markdown
Collaborator Author

@amy-why-3459 @zengchuang-hw Please review A1, tracked by Issue408, at head 18f189a8873da75b4fbcae2910711b2a37a8bbf3 (includes main f536288710c726a827bfa3e4357c2c911984cbc3). This PR canonicalizes apps.cli:main, retires the root CLI/arguments modules without forwarders, removes the temporary arguments exception, and updates consumers plus the exact root8/leaf guards. Manifest relocking requires formal architecture review. The seven Windows-timeout host cases passed on H200; real sdist-to-wheel, fresh installed-resource925 and console/module entry checks passed. The original Windows 63-pass/8-fail result and coordinator harness correction are preserved in the PR body, along with the remaining CI/baseline limitations.

@Dong1017 Dong1017 added the kind:task Bounded delivery task with a delegated slice label Nov 16, 2026

@zengchuang-hw zengchuang-hw left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

PR review verdict — DO NOT PASS (verification remains incomplete; no code defect found in this review)

First independent review of A1 for #408. Reviewed base f536288 -> head 18f189a. Previous review: none. Request: #461 (comment). I reviewed the actual code and governing Issue, not other reviewers' opinions.

Architecture assessment:

  • The canonical apps.cli package and parser replace both root modules, without forwarders or registry aliases. Runtime CLI normalized AST is equivalent after resolving moved imports and accounting for import-order changes.
  • REPO_TYPES is read through the existing Knowledge catalog binding, not copied; the neutral leaf's bytes are unchanged. Removing the temporary importer, preserving its exact historical record and keeping the permanent Infra rule is consistent with A1.
  • The root main exception is restricted to the exact plain module-level import of main from apps.cli; negative fixtures still reject wider shapes/targets. Root inventory is exactly eight. Independently evaluated first_wave_report has no violations but still lists 12 Harness-to-Infra hits and five resource loaders. This is not complete isolation.

Executed locally at this head (Linux/Python 3.12, isolated checkout):

  • Six gate/manifest/root/facade/frozen-legacy modules: 151 passed, 274 subtests passed.
  • Apps assessment/CLI plus Codex, DSH and WorkSwarm plugin modules: 118 passed, 2 subtests passed, including external-workspace DSH bundle execution.
  • ruff check src tests: passed.
  • Nine real python-module command comparisons against base: identical exit code/stdout/stderr (help/version/skills help/benchmark list/fixed-record error paths/invalid option).
  • Active-consumer search found no old root CLI/parser imports outside intentional tests/history.
    The initial mixed test run had one subprocess import failure with a relative PYTHONPATH; rerunning with the absolute candidate source path passed. No product/test edits were used.

Before approval:

  1. Complete the current-head full-check jobs for Python 3.10 and 3.11. Both are still in progress; focused checks cannot replace the full suite.
  2. Obtain a successful pr-references rerun. The failed job rejected the negated closing phrase in the original body. The current body no longer matches that syntax, but the recorded check remains failed: https://github.com/InferMatrixLab/InferRSI/actions/runs/43812153335/job/133115992805.
  3. Close the host-matrix evidence gap required by #408: the reported Windows run remains 63 passed / 8 failed. Linux success does not establish Windows behavior. Supply a successful Windows rerun of affected cases or concrete same-host baseline/candidate evidence isolating an unchanged environmental failure. I did not execute Windows/macOS checks.

Observed remote evidence: installed-package 3.10/3.11, dashboard-browser, lint-and-smoke and declaration checks are successful. I did not independently repeat the full installed-resource checker or real sdist-to-wheel build; those remote successes and author receipts are distinct from my local tests. Full local suite not run. No non-blocking suggestions. No merge/auto-merge performed.

Copy link
Copy Markdown
Collaborator

Reviewed base f536288710c726a827bfa3e4357c2c911984cbc3 … head 18f189a8873da75b4fbcae2910711b2a37a8bbf3 (new PR; range f536288710c726a827bfa3e4357c2c911984cbc3..18f189a8873da75b4fbcae2910711b2a37a8bbf3). Refs #408 A1. I am not approving this head.

The A1 move itself looks consistent with the accepted scope: inferrsi.apps.cli:main is the canonical entry, root __main__.py only delegates from .apps.cli import main, root cli.py / arguments.py are removed with no forwarders, the R1 temporary leaf-importer exception is gone (REPO_TYPES is taken from inferrsi.knowledge.catalog), and the root inventory / gate pins shrink 10 → 8. I did not rerun the suite locally. Accuracy and performance comparison experiments are not applicable here (entrypoint relocation; command behavior is meant to stay the same).

Outstanding blockers:

  1. full-check is red on this head for both Python 3.10 and 3.11 (1 failed, 4657 passed): https://github.com/InferMatrixLab/InferRSI/actions/runs/43812153335/job/133115992848 and https://github.com/InferMatrixLab/InferRSI/actions/runs/43812153335/job/133115992931

    tests/test_campaign_record_injection.py (test_harness_records_load_no_knowledge_or_apps_module, line 368) still pins len(RETIRED_ROOT_MODULES) == 11. After A1, RETIRED_ROOT_MODULES is RETIRED_BY_R1 | RETIRED_BY_A1 (13 names, including inferrsi.cli and inferrsi.arguments). This file is not in the PR, so the assertion was not updated. Please update the pinned count, and keep the load-isolation check so it also covers the two newly retired names.

  2. pr-references failed on the same run (https://github.com/InferMatrixLab/InferRSI/actions/runs/43812153335/job/133115992805) because the body at check time contained “This does not close task: A-entry — canonical Apps entrypoints and application composition #408 or task: R01 — retire legacy root paths and close core dependencies #412”. Closing syntax is rejected even when negated. The current body uses Refs #408 and no longer has that phrase; please rerun the check on the current text and leave it green.

  3. task: A-entry — canonical Apps entrypoints and application composition #408 acceptance requires preserving command behavior on the existing host matrix. The PR records a Windows run of 63 passed / 8 failed (seven 45s isolated-subprocess timeouts; one DSH empty-output parse), states that Windows is not green, and does not report a Windows installed-wheel run. The seven timeout cases later passed on H200/Linux, and a follow-up DSH comparison versus main returned exit 0, but the original Windows failures are not closed out. Please provide a successful Windows rerun of the affected cases, or same-host baseline vs candidate evidence that isolates an unchanged environmental failure.

Track all13 retired roots without weakening core isolation assertions.
Preserve R1 eleven plus the two exact A1 retired paths and all original subprocess isolation checks. Windows host evidence remains a separate open review item.

Refs #408

Agent-assisted-by: Claude Code / claude-fable-5-2 [effort=max] (author)
Signed-off-by: GUOGUO <xwdong1998@163.com>

Copy link
Copy Markdown
Collaborator Author

@amy-why-3459 @zengchuang-hw Follow-up to Amy's comment and review5915579655, both assessed at 18f189a8873da75b4fbcae2910711b2a37a8bbf3: current head is 7fd5bb1fe388c9046b9053acf61c5d393c0c2cd9, with main still f536288710c726a827bfa3e4357c2c911984cbc3.

The single full-check failure is repaired: the test now pins R1's historical eleven plus the exact two A1 retired names, their disjoint union and total thirteen. Its original subprocess isolation script, negative assertions and deadlines are unchanged. The original Fable session authored the fix; independent H200 CPU validation passed the whole affected module (20 passed in 6.26s), changed-file Ruff and the real installed925-resource checker. No application/package/CI-policy changes or timeout increases.

The fresh pr-references check is green on the corrected current body. New-head installed-package checks for Python3.10/3.11, declarations and lint-and-smoke are also green. The two full-check jobs and dashboard-browser are still running in the current CI run; their final results are not yet claimed.

The Windows host-evidence item remains OPEN: the original 63-pass/8-fail run is preserved, and Linux/H200 success is not presented as Windows proof. No new Windows/macOS run or hosted job is claimed. Please recheck the precise test repair and current-head CI; this is not a claim that all review blockers or architecture approval are complete.

@zengchuang-hw zengchuang-hw left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Incremental review: INCONCLUSIVE, not approved.

Range: 18f189a -> 7fd5bb1.
Previous review: #461 (review)

I inspected the entire one-file incremental diff and surrounding isolation subprocess. The update correctly pins the eleven R1 retired modules plus the exact two A1 names, asserts disjointness and union equality, and checks all thirteen names remain unloaded. The subprocess still imports every Harness records module and rejects Knowledge/Apps/Benchmark imports; no isolation check is weakened. No product code changed in this range.

Executed at this exact head in an isolated Linux/Python 3.12 checkout: all 20 tests in tests/test_campaign_record_injection.py passed; Ruff on that file passed. No test/product edits were made for verification. The previous broader test evidence is historical, not a fresh full-suite result for this head.

Resolved: the outdated retired-root count is corrected and the affected module passes locally. The current-head pr-references check is now successful.

Before approval: current-head full-check 3.10/3.11 and dashboard-browser were still in progress at verification; obtain their completed successful results. The previously documented Windows 63-pass/8-fail host-matrix gap remains unclosed in the available author evidence. A successful affected-case Windows rerun, or same-host immutable-baseline/candidate evidence showing an unchanged environmental failure, is still needed. Linux results cannot close this gap.

Observed remote successes: installed-package 3.10/3.11, pr-references, lint-and-smoke, and a declaration check. I did not rerun the entire local suite, Windows/macOS, the installed-resource checker, or a real sdist-to-wheel build this round. No new code defect or non-blocking suggestion found in this incremental diff. No merge or auto-merge.

Copy link
Copy Markdown
Collaborator

PR review verdict — DO NOT PASS

Reviewed range 18f189a8873da75b4fbcae2910711b2a37a8bbf3..7fd5bb1fe388c9046b9053acf61c5d393c0c2cd9 (commit 7fd5bb1fe388c9046b9053acf61c5d393c0c2cd9). Refs #408 A1. I am not approving this head.

The incremental diff is the one-file test repair in tests/test_campaign_record_injection.py (test_harness_records_load_no_knowledge_or_apps_module). No product code changed in this range.

Resolved from my previous comment (#461 (comment)):

  1. The isolation test now pins R1's eleven retired names plus the exact two A1 names (inferrsi.cli, inferrsi.arguments), asserts disjointness and RETIRED_ROOT_MODULES == RETIRED_BY_R1 | RETIRED_BY_A1, and still checks that all thirteen names stay unloaded in the same subprocess. full-check is green on this head for Python 3.10 and 3.11: https://github.com/InferMatrixLab/InferRSI/actions/runs/43817017627/job/133131273042 and https://github.com/InferMatrixLab/InferRSI/actions/runs/43817017627/job/133131272975

  2. pr-references is green on the current body: https://github.com/InferMatrixLab/InferRSI/actions/runs/43817017627/job/133131272669

Installed-package 3.10/3.11, lint-and-smoke, declarations, and dashboard-browser are also green on this head.

Outstanding blocker (unchanged):

  1. task: A-entry — canonical Apps entrypoints and application composition #408 acceptance still requires preserving command behavior on the existing host matrix. The recorded Windows run remains 63 passed / 8 failed, and no new Windows/macOS rerun or same-host baseline-vs-candidate evidence has been provided. Linux/H200 success does not close this gap. Please provide a successful Windows rerun of the affected cases, or same-host baseline vs candidate evidence that isolates an unchanged environmental failure.

I did not rerun the suite locally. Accuracy and performance comparison experiments are still not applicable (test-pin and entrypoint relocation).

cursor[bot]
cursor Bot previously requested changes Nov 16, 2026

@cursor cursor Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

PR review verdict — DO NOT PASS

Reviewed base f536288710c726a827bfa3e4357c2c911984cbc3 and head 7fd5bb1fe388c9046b9053acf61c5d393c0c2cd9. Refs #408 A1. Independent review via cursor[bot]; not the author account.

A1 code itself matches the accepted slice: inferrsi.apps.cli:main is the console entry, root __main__.py is only from .apps.cli import main, root cli.py/arguments.py are gone with no forwarders or sys.modules aliases, APPLICATION_ENTRIES drops the CLI files and keeps benchmark_operations.py for A2, the temporary leaf-importer exception is removed (REPO_TYPES via inferrsi.knowledge.catalog), and the root inventory pin is 10→8 with thirteen retired roots (R1 eleven ∪ A1 two).

Verified locally on this head (Linux): focused gate/CLI modules 218 passed, 276 subtests; ruff clean on changed A1 leaves; python -m inferrsi --help exits 0; retired inferrsi.cli/inferrsi.arguments raise ModuleNotFoundError; first_wave_report has zero new violations and still reports isolation_complete=False with the expected Harness→Infra / resource-loader blockers (not weakened). Hosted CI on this head is green: full-check 3.10/3.11, installed-package 3.10/3.11, lint-and-smoke, pr-references, declarations, dashboard-browser.

Prior blockers that are resolved on this head:

  1. tests/test_campaign_record_injection.py now pins R1 eleven + A1 two, disjointness, union, and unloaded thirteen.
  2. pr-references and both full-check jobs succeed on 7fd5bb1fe.

Outstanding blocker (required evidence still insufficient):

  1. #408 acceptance requires preserving command behavior on the existing host matrix. Author evidence still records Windows 63 passed / 8 failed (seven 45s isolated-subprocess timeouts; one DSH empty-output parse), states Windows is not green, and provides no successful Windows rerun of the affected cases and no same-host immutable-baseline vs candidate comparison that isolates an unchanged environmental failure for those seven timeouts. Linux/H200 success and hosted Linux CI do not close that gap. I did not execute Windows/macOS checks in this environment.

Before re-review / Approve: supply a successful Windows rerun of the affected cases, or same-host baseline vs candidate evidence isolating an unchanged environmental failure. No merge performed.

…d runners

Add a host-compatibility job to the ready lane. It runs on windows-latest
and macos-latest, one host at a time, and executes only the eight plugin
and DSH nodes from the recorded Windows run (63 passed, 8 failed), then
builds a wheel from the sdist and checks it in a fresh venv outside the
checkout. Existing jobs, triggers and test deadlines are unchanged.

Refs #408

Agent-assisted-by: Claude Code / claude-fable-5-2 [effort=max] (author)
Signed-off-by: GUOGUO <xwdong1998@163.com>
@Dong1017

Copy link
Copy Markdown
Collaborator Author

Windows evidence for the eight affected nodes, current head cb59d3aa135d803d8dfba84d7bbd94a918551bc8.

Same-host comparison on the original Windows machine (Windows 11, Python 3.13.5, pytest 9.0.2, Node 24.14.0). I ran the same eight nodes with the original arguments (-vv -p no:cacheprovider --durations=20 -o faulthandler_timeout=180), one pytest process at a time, on two clean git archive exports. Every exported file matched its Git blob hash.

  • Baseline: unchanged main a1093578c725eb7dc34a2bfbbd4fefab5abbc098.
  • Candidate: A1 ddbbb2539305e72ccb8250c606a9e68370bf910d, which is that main merged into A1. The test files are identical on both sides.
Node Original run Baseline main A1 candidate
test_comparison_preview_execution_status_and_report[codex-fail-1-FAIL] failed passed passed
test_isolated_runtime_can_export_dashboard[workswarm] failed failed: dashboard subprocess exceeded 45 s failed: same command, same timeout
test_comparison_preview_execution_status_and_report[workswarm-pass-0-PASS] failed passed passed
test_comparison_preview_execution_status_and_report[workswarm-fail-1-FAIL] failed passed passed
test_comparison_preview_execution_status_and_report[workswarm-missing-2-INCONCLUSIVE] failed passed passed
test_release_archive_is_installable_and_all_versions_match[codex] failed failed: collect --limit 1 --json subprocess exceeded 45 s failed: same command, same timeout
test_release_archive_is_installable_and_all_versions_match[workswarm] failed failed: --plugin-info subprocess exceeded 45 s failed: same command, same timeout
test_remote_ssh_connection_and_scoped_operations (DSH) failed passed passed
Result 63 passed / 8 failed of 71 5 passed / 3 failed, 766.30 s 5 passed / 3 failed, 766.70 s

Unchanged main and the A1 candidate give the same result on the same machine. The three remaining failures are the same isolated-plugin subprocess commands hitting the existing 45 s deadline on both sides. I attribute them to local resource limits on that machine: available physical memory ranged from 605 MB to 1,659 MB of 16 GB during the runs, with memory load up to 96%. The earlier DSH empty-output failure did not recur on either side. No test, assertion or deadline was changed for these runs. The raw logs and JUnit XML for both sides are kept locally and can be attached on request.

Hosted runners. This head adds a host-compatibility job to the ready lane. It runs the same eight nodes on windows-latest and macos-latest, then builds a wheel from the sdist, installs it into a fresh venv outside the checkout and runs the unchanged tests/check_installed.py plus the console and module entry points: https://github.com/InferMatrixLab/InferRSI/actions/runs/43837681649. I will report the result when the run finishes. Until then, Windows is not claimed green.

@Dong1017

Copy link
Copy Markdown
Collaborator Author

Hosted CI result for head cb59d3aa135d803d8dfba84d7bbd94a918551bc8: run 43837681649. Every job passed.

Job Result
host-compatibility (windows-latest) The eight affected nodes: 8 passed in 53.20 s; JUnit shows 8 run, 0 errors, 0 failures, 0 skipped. The wheel was built from the sdist and installed with [skills] into a fresh venv; inferrsi.__file__ resolved to that venv's site-packages; tests/check_installed.py passed under -I; --version and benchmarks list matched between inferrsi.exe and python -I -m inferrsi.
host-compatibility (macos-latest) 8 passed in 20.10 s; the same wheel and entry-point checks passed.
full-check (3.10) and full-check (3.11) 4,665 passed, 27 skipped, 616 subtests passed.
installed-package (3.10), installed-package (3.11), dashboard-browser, pr-references, declarations, lint-and-smoke, knowledge-model-policy Passed.

This is the successful Windows rerun of the affected cases requested in this comment and this review. All eight nodes pass on a hosted Windows runner with their existing 45 s deadlines. The original 63 passed / 8 failed record and the local same-host comparison above remain as history. The wheel version is 0.1.dev1+ge50f44173 because GitHub tested the PR merge commit from a shallow checkout, as the existing installed-package job does.

Architecture approval is still pending.

@amy-why-3459 amy-why-3459 left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

PR review verdict — PASS
Reviewed range 7fd5bb1fe388c9046b9053acf61c5d393c0c2cd9..cb59d3aa135d803d8dfba84d7bbd94a918551bc8 (base 016c7c7df023944f0f46d8e777d6d31f1df1764c, head cb59d3aa135d803d8dfba84d7bbd94a918551bc8). Refs #408 A1.

The outstanding host-matrix blocker from #461 (comment) is resolved:

  • Same-host comparison on the original Windows machine (#461 (comment)): unchanged main a1093578c725eb7dc34a2bfbbd4fefab5abbc098 and A1 ddbbb2539305e72ccb8250c606a9e68370bf910d produced the same 5 passed / 3 failed result; the three remaining failures are the same 45 s isolated-plugin timeouts on both sides.
  • Hosted host-compatibility on this head is green: windows-latest and macos-latest both passed the eight affected nodes plus the sdist-to-wheel install and console/module entry checks. This is the successful Windows rerun of the affected cases.

The unique product/test delta versus current main is still the previously reviewed A1 move and the thirteen-root isolation pin. The new commit 960082badc9e0554d12b1563637fb0755dad0f83 only adds the host-compatibility job in .github/workflows/ci-ready.yml. Ready-lane checks on this head are green, including full-check 3.10/3.11, installed-package 3.10/3.11, lint-and-smoke, pr-references, declarations, and dashboard-browser: https://github.com/InferMatrixLab/InferRSI/actions/runs/43837681649

Non-blocking: the new job is now a standing ready-lane check for later PRs, not only this A1 evidence run. That looks reasonable as ongoing host-matrix coverage; please confirm it is intended as permanent policy.

Accuracy and performance comparison experiments are not applicable (entrypoint relocation and CI evidence; command behavior is meant to stay the same). I did not rerun the suite locally.

@zengchuang-hw zengchuang-hw left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

A1 incremental review: APPROVE at cb59d3a.

Range reviewed: 7fd5bb1 -> cb59d3a. Previous review: #461 (review). This follows my earlier actual-code architecture review and retired-root guard verification.

I inspected the new host-compatibility workflow and merge integration. The job has read-only contents permission, does not persist checkout credentials, serializes the two hosted OS jobs, requires every one of the eight selected nodes to run with zero skips/errors/failures, and uses unchanged test deadlines. The fresh wheel environment removes PYTHONPATH, verifies site-packages origin, runs the installed checker under -I, and compares console/module outputs. It is bounded host evidence, not a full Windows/macOS suite or performance qualification.

The incoming main API/catalog/knowledge changes are byte-identical to main 016c7c7. The Apps assessment test retains the A1 canonical CLI import, and installed inventory includes main's two new resources (927). No old CLI forwarder was reintroduced.

Executed locally at this exact head, isolated Linux/Python 3.12 checkout: 168 tests and 274 subtests passed across campaign record injection, Apps assessment, target-module gate, first-wave isolation, package-root inventory and source snapshot modules; ruff check src tests tools passed. No product or test edits were made.

I independently read the hosted job logs, rather than accepting the author summary:

  • Windows: eight affected nodes passed in 53.20s, JUnit 8 tests / 0 errors / 0 failures / 0 skipped; real sdist-to-wheel build and fresh installed checker succeeded.
  • macOS: eight nodes passed in 20.10s, likewise zero skipped/errors/failures, with the same installed-wheel checks.
  • Both installed inventories report 927 resources with matching digest. Console/module version and benchmark-list checks completed successfully.
  • Full-check 3.10 and 3.11 each report 4,665 passed, 27 skipped and 616 subtests passed. Current-head pr-references, dashboard-browser, installed-package and lint/declaration checks are successful.

Hosted checkout was synthetic merge e50f441. I verified its tree ee1449978dc7abfc2bacc8517581d9ef7d6d784f equals this exact head's tree, so this is relevant current-head evidence, not an unrelated revision. The shallow-checkout wheel version is not a separate product revision claim.

Resolved since my previous review: full CI completed successfully; pr-references remains green; the successful real Windows rerun closes the eight-case host evidence gap. The original 63-pass/8-fail history remains history, not rewritten. I find no remaining blocking issue in the reviewed A1 scope. Architecture approval here is for the bounded CLI/parser retirement and exact guards, not A2/M1/R2/R3 or complete core isolation.

Not executed locally this round: full suite, Windows/macOS jobs, separate wheel build, or performance experiments. Hosted checks above were inspected, not rerun by me. No non-blocking suggestions. No merge or auto-merge performed.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

kind:task Bounded delivery task with a delegated slice

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants