Repository navigation
Conversation
zengchuang-hw
left a comment
There was a problem hiding this comment.
Independent review of head 92fd8a1fe83abb6f585b70312ff0492a1fafd853 (current base 61303fa8150472dfbda62ec8d77908bc05c73f3b; actual patch merge-base ba233f7cb93b61e19fd07f6f18b6ae2c4b2a4efe). COMMENT, not approval, for two independently reproduced findings.
P1 — Final can report PASS for changed retained code under the original selected Git revision. In src/inferrsi/harness/optimization.py::_service_load (464–493), retained baseline/selected paths are checked with candidates.members, which checks membership/path safety, not correspondence between bytes and the claimed revision. The final path then copies the selected directory with fingerprint=False at line 1173. Independent CPU reproduction using this PR's synthetic fixture: start → baseline → accept a committed candidate → unlink/replace the retained selected candidate/engine.py → final. Result: final PASS, unchanged selected revision in the record, but all observed service events reported the replacement version changed-retained, not the accepted version. The editable project was not the final input; the retained snapshot itself drifted.
This breaks the high-level requirement to retest the retained selected revision, and a matching echoed revision/configuration receipt does not establish what bytes ran. Reject drift or reconstruct/verify against an immutable authoritative selected revision before execution and final settlement. This finding does not prescribe file hashes or treat the Issue's later no-fingerprint proposal as accepted scope. Add a regression that changes retained candidate bytes and requires rejection/INCONCLUSIVE rather than PASS under the old revision.
P2 — The new no-fingerprint test fixture leaks patched functions through lazy imports and makes compatibility tests order-dependent. tests/test_project_service_reuse.py:22–48 monkeypatches infra.artifacts.sha256 to a raising stub. test_exact_round_policy_and_legacy_argument_validation (848–868) imports harness.project while that patch is active; the module's imported sha256 alias retains the stub after fixture teardown. Reproduction: run that single new test, followed in the same pytest process by tests/test_project_optimization.py::test_project_runs_shared_rounds_and_final_from_retained_multifile_candidate. Result: 1 passed, 2 failed, both parameterizations failing with the no-fingerprint stub assertion. In a fresh process, the full project-optimization and optimization-session files pass (43 passed). Isolate/import affected modules before patching, or restore the captured aliases as well, and retain an ordered cross-file regression.
Verification actually run:
- Five-file focused run:
test_project_service_reuse.py test_project_objectives.py test_project_optimization.py test_optimization_session.py test_target_module_import_gate.py: 75 passed, 25 failed, 12 errors, 4 subtests passed (exit 1). The failures/errors observed in this run hit the no-fingerprint stub; they are not being presented as 37 independent product defects. - Fresh-process compatibility run:
test_project_optimization.py test_optimization_session.py: 43 passed. - Minimal ordering reproduction above: 2 failed, 1 passed.
- Separate local retained-candidate-drift probe: 1 passed, confirming the incorrect PASS/old-revision/new-code behavior. The probe was outside the repository; no tracked source or tests changed.
- Head check metadata currently reports
declarations: failureandknowledge-model-policy: failure. I did not use their narrative output as review evidence.
The clean high-level #467 requirements support explicit mode selection, preserved default AB/BA, traceable service identity, deterministic declared objectives and independent final. Detailed later profile-repair/changed-condition acceptance rules and the no-fingerprint proposal lack authenticated original acceptance provenance in the supplied requirement view. I have not accepted those as new governing scope. Code was inspected through comment/docstring/prose-stripped AST views; documentation semantics and the complete new documentation were not independently reviewed. No full-suite/merge-result verification, Windows qualification, real-model/device experiment, paid work or performance/quality acceptance was performed. These limits plus the reproduced findings prevent approval.
david6666666
left a comment
There was a problem hiding this comment.
One service lifecycle gap when the whole attempt reaches its deadline.
| raise | ||
| finally: | ||
| try: | ||
| cleanup = stage("stop", bound=False) |
There was a problem hiding this comment.
The required stop script inherits the deadline that the preceding stage just exhausted. In the exact-head CPU fixture, setting the evaluation timeout to 3 s and using profile_mode=timeout completed baseline, measurement and quality, then recorded both profile and stop as timed out; the stop script was never launched. The owned process group is still terminated, but that does not establish project cleanup. Could we reserve an explicit bounded stop window within the declared attempt budget (or make the separate cleanup budget explicit), and test the whole-attempt timeout, not just a shorter stage timeout?
david6666666
left a comment
There was a problem hiding this comment.
The new Git-backed snapshot can silently omit or rewrite declared inputs.
| if not isinstance(commit, str) or not re.fullmatch(r"[a-f0-9]{40,64}", commit): | ||
| raise ValueError("Candidate requires an existing Git commit") | ||
| archive = subprocess.check_output( | ||
| ["git", "-C", str(source), "archive", "--format=tar", commit, "--", *includes] |
There was a problem hiding this comment.
git archive does not guarantee an exact copy of the declared commit tree. In a CPU repro, two declared tracked files passed revision(), but an export-ignore attribute omitted one and the new nonempty check accepted the partial snapshot. export-subst can also rewrite a file without changing the member list, even at the same commit. Could we enumerate the declared members from the pinned tree and materialize their exact blobs and modes, or reject attribute-driven changes? Add a regression for an omitted member and a rewritten member so final cannot attest to a revision whose inputs differ from its tree.
|
Posted by an agent on behalf of @Dong1017 (incremental review only; the existing review findings and the #467 rule-5 note are not repeated as new findings). Reviewed head New findings1. P2 — An accepted candidate with a declared gain is counted as “no gain” unless it becomes the new best; with
Reproduced with this PR’s CPU fixture (baseline
The docs currently codify this (“Consecutive complete rejections/unselected candidates may trigger”, 2. P3 — The stagnation streak survives The refresh branch updates 3. P3 — With Known blocking dependency (already recorded in #467, not a new finding)The implementation and docs follow the 2026-11-17 proposed amendment, not the frozen 2026-11-16 rule 5: One code-level detail matters for the eventual amendment: Verification and required evidence
Scope: read the PR body and diff from |
david6666666
left a comment
There was a problem hiding this comment.
A separate ordered regression in the legacy packaged-task path.
| "def test_restored_alias():\n" | ||
| " from inferrsi.harness import project\n" | ||
| " from inferrsi.infra import artifacts\n" | ||
| " assert project.sha256 is artifacts.sha256\n" |
There was a problem hiding this comment.
The new fixture only checks project.sha256; a separate legacy alias still leaks. On this head, running test_exact_round_policy_and_legacy_argument_validation immediately before test_started_task_uses_frozen_protocol_without_registry in one pytest process gives one pass and one setup error: task_protocol.sha256 remains bound to the service test's raising guard. Please scope the guard so newly imported consumers cannot capture it, and run the actual packaged-task test after the service test in this regression.
zengchuang-hw
left a comment
There was a problem hiding this comment.
Incremental independent review of 6c468262e802d6c1f6662669a385c83570af5f81, against my previous review of 92fd8a1fe83abb6f585b70312ff0492a1fafd853. Current base remains 61303fa8150472dfbda62ec8d77908bc05c73f3b. COMMENT, not approval.
Previous P1: resolved for the reproduced retained-candidate drift case. The incremental code records the source repository with each selected revision and reconstructs the execution workspace with git archive <recorded-commit>, rather than trusting the mutable retained directory. I reran my own external probe: after accepting original-selected, I replaced the retained candidate/engine.py with changed-retained and invoked final. At this head the final retained the original revision and its observed events showed original-selected, not the replacement. The probe passed. This verifies that specific correction; continued access to the authoritative Git repository/commit is now part of the execution dependency, not an assertion of self-contained offline replay.
Previous P2: partly fixed, but another lazy-import alias still leaks. The original project-helper ordering reproducer now passes (3 passed), consistent with eagerly importing/patching harness.project.sha256. However, harness.optimization_session.session_command imports harness.task_protocol.prepare during start while the service test's no-fingerprint fixture is active. task_protocol.py captures infra.artifacts.sha256 in its module namespace. Fixture teardown restores the source attribute, not this newly captured alias.
Independent minimal reproduction at this head:
pytest -o addopts='' -q \
tests/test_project_service_reuse.py::test_one_service_order_and_independent_final_ignores_workspace \
tests/test_optimization_session.py::test_started_task_uses_frozen_protocol_without_registry
Result: 1 passed, 1 error. The second test fails during setup in harness/task_protocol.py:32 with AssertionError: This test must not calculate or verify a file fingerprint, raised by the first test module's stub. Address this remaining alias in tests/test_project_service_reuse.py:28-55 and retain the ordered cross-file regression. The new isolation test and the project-only ordering check are not sufficient to establish isolation from all affected lazy imports.
The additional cleanup code gives stop its own bounded stage deadline outside the evaluation deadline. I inspected that delta; it should not be confused with full host/device cleanup qualification.
Completed verification
- Same earlier five-file focused run, plus
tests/test_service_fingerprint_isolation.py: 104 passed, 1 failed, 12 errors, 4 subtests passed in 159.31s, exit 1. The observed failure/error assertions are the remaining no-fingerprint stub leakage; this is not a claim of 13 separate product defects. - Original project-alias ordering reproduction: 3 passed in 5.61s.
- Remaining task-protocol ordering reproduction above: 1 passed, 1 error in 3.10s.
- Independent retained-byte-drift probe outside the repository: 1 passed in 3.81s, with actual execution of the original selected version.
- Current head check metadata reports
declarations: failureandknowledge-model-policy: failure; their narrative output was not used as review evidence.
The original clean #467 requirement/provenance gaps from the previous review remain: exact later profile-repair/changed-condition rules and the no-fingerprint proposal were not authenticated as accepted original scope. I inspected only the actual incremental code through comment/docstring/prose-stripped views and reused the clean requirement handoff. I did not read author discussion or other reviews. New documentation semantics, full-suite/merge-result behavior, Windows, real-model/device execution and performance/quality acceptance are not established by these CPU checks. No tracked code/test changes or merge were made.
The retained-code issue is specifically reverified as corrected; the test-isolation issue remains reproducible. Approval is withheld.
zengchuang-hw
left a comment
There was a problem hiding this comment.
Incremental independent review at 5398bf5d76ce73d00180fb7cba7c952f5acf132d, relative to my previous checkpoint at 6c468262e802d6c1f6662669a385c83570af5f81. COMMENT, not approval.
P2 remains independently reproducible. The no-fingerprint fixture still leaves the lazily imported harness.task_protocol.sha256 alias bound to its raising stub. At this exact head, running these two tests in order in one fresh pytest process:
pytest -o addopts='' -q \
tests/test_project_service_reuse.py::test_one_service_order_and_independent_final_ignores_workspace \
tests/test_optimization_session.py::test_started_task_uses_frozen_protocol_without_registry
produced 1 passed, 1 error in 3.25s. The second test's setup reaches task_protocol.prepare and raises AssertionError: This test must not calculate or verify a file fingerprint. The incremental patch does not change that fixture/import path. Restore/isolate this captured alias too, and keep the cross-file ordered regression; the successful service-only checks below do not resolve the compatibility failure.
New scope checked
project_service.retain_revisionnow enumerates the pinned Git tree and reads blob bytes directly, replacinggit archive. The inspected code rejects missing declared paths and non-regular modes. The targeted export-ignore/export-subst tests preserve exact binary payload bytes and executable mode through baseline, candidate and final.- Stagnation now resets for a PASS candidate even when deterministic ranking does not select it, while keeping prior events. Condition refresh resets the new streak without deleting history. Refresh is rejected before state mutation when no round capacity remains. These are observable implementation behaviors, not my acceptance of the disputed later policy details.
- The original P1 retained-byte-drift probe still passes: after replacing the retained selected code with
changed-retained, final runs original-selected under the recorded selected revision. This independently confirms the earlier correction remains effective after the snapshot implementation changed.
Checks actually run
- Eight targeted test selections in
test_project_service_reuse.py, including parametrizations for export attributes, invalid tree inputs, refresh/stagnation/capacity, accepted-but-unselected behavior, retained-code restoration, missing source repository and bounded stop after timeout: 13 passed in 38.88s, exit 0. - My separate retained-code-drift probe outside the repository: 1 passed in 3.63s, exit 0.
- The remaining P2 ordering reproduction above: 1 passed, 1 error, exit 1.
- Exact-head check metadata:
declarations: failure;knowledge-model-policy: failure. I did not rely on check narratives.
I reviewed the actual incremental Python code through comment/docstring/prose-stripped views and reused the clean original requirement handoff; author discussion and other reviews were not read. Full documentation semantics, the complete suite/merge result, Windows and real-model/device/performance qualification were not revalidated. No tracked source/test changes or merge were made. The original detailed acceptance-provenance gaps remain. New focused behavior passes, but the reproduced P2 and failing head checks prevent approval.
david6666666
left a comment
There was a problem hiding this comment.
A declared-input cleanliness check still misses file deletions.
| raise ValueError("Candidate requires an existing Git commit") | ||
| selected = members(source, includes) | ||
| tracked = set(git("ls-files", "-z").split("\0")) | ||
| if any(name not in tracked for name in selected) or git( |
There was a problem hiding this comment.
Could we check the entire declared include, not just the files that still exist? With --include inputs and two tracked files in that directory, deleting one without committing still lets revision() accept HEAD: members() leaves the deleted path out of git status. retain_revision() then silently puts it back for evaluation. Please reject uncommitted deletions inside included directories and add a regression test.
zengchuang-hw
left a comment
There was a problem hiding this comment.
Incremental independent review at 31bed72e55b985f93afc357cf8bd27238da1953b, relative to my previous review at 5398bf5d76ce73d00180fb7cba7c952f5acf132d. COMMENT, not approval.
Previous P2 is resolved in the independently reproduced path. This delta changes only four test files. The service guard now patches each relevant module's hashlib binding instead of replacing exported hashing functions or mutating the shared hashlib module. Consequently a lazily imported task_protocol.sha256 function no longer permanently captures the raising stub after fixture teardown. I reran the exact two-test order that failed at the previous head:
pytest -o addopts='' -q \
tests/test_project_service_reuse.py::test_one_service_order_and_independent_final_ignores_workspace \
tests/test_optimization_session.py::test_started_task_uses_frozen_protocol_without_registry
Result: 2 passed in 3.84s, exit 0.
Other changed expectations checked against actual code. The new Harness import-closure expectation includes atomic_json; the current local_process.py really imports atomic_write_json from it. Directly running the current first-wave import report yields 14 Harness→Model/Infra edges, zero policy violations, isolation_complete=False, matching the updated compatibility test. This validates the bounded expected inventory, not full module isolation.
No production code changed in this incremental commit, so the earlier retained-code/P1 implementation is unchanged from the last independently passing probe. I do not identify a new product defect in this test-only delta.
Completed checks at this head
- Eight-file focused suite, in order:
test_project_service_reuse.py,test_project_objectives.py,test_project_optimization.py,test_optimization_session.py,test_service_fingerprint_isolation.py,test_harness_execute.py,test_knowledge_compatibility_shells.py,test_target_module_import_gate.py, usingPYTHONPATH=$PWD/src <test-venv>/bin/python -m pytest -o addopts='' -q: 154 passed, 8 skipped, 4 subtests passed in 239.51s, exit 0. Skipped tests are not validated coverage. - Exact minimal P2 reproducer above: 2 passed.
- Direct import-inventory check described above; tracked checkout unchanged.
- Exact-head CI/status metadata still reports
declarations: failureandknowledge-model-policy: failure. I did not use check narratives as technical review evidence.
Both previously reproduced findings now have specifically bounded positive evidence: P1's production fix was independently exercised at the previous head and the retained-code regression remains in this passing suite; P2's formerly failing cross-file order now passes at this head. These findings are no longer being carried as unresolved code defects.
Approval is still withheld: failing head checks remain, the clean #467 handoff does not authenticate the later detailed acceptance/policy decisions as original governing requirements, and full documentation semantics, complete-suite/merge-result coverage, skipped cases, Windows and real-model/device/performance qualification are not established here. This was a test-only incremental inspection through comment/docstring/prose-stripped views, without author discussion or other reviews. No code change, merge or experiment was performed.
zengchuang-hw
left a comment
There was a problem hiding this comment.
Independent incremental review at 1e5fac1, relative to my previous review at 31bed72: #480 (review)
The delta changes project_service.revision's Git dirty check from currently enumerated member paths to the declared input roots (or "." for the whole project). This closes the omitted-deleted-file case: a file removed inside a declared directory is absent from the member enumeration but still makes that declared tree dirty. The unchanged tracked-member check remains in place. The added tests cover staged and unstaged deletion for directory-scoped and whole-project inputs, committed deletion in the retained snapshot, and exclusion of deletions outside explicitly declared inputs. I found no new actionable defect in this bounded delta.
Executed on this exact head: 9 passed in 8.34s, exit 0. This comprised the five new parameterized cases, test_one_service_order_and_independent_final_ignores_workspace, test_started_task_uses_frozen_protocol_without_registry, and tests/test_service_fingerprint_isolation.py. Thus the previously reproduced lazy-hash isolation path remains passing; the earlier retained-revision result is not reopened by this delta. The previous head's 154 passed / 8 skipped / 4 subtests result is historical evidence, not a rerun of that suite on this head.
Still COMMENT, not approval: current-head declarations and knowledge-model-policy report failure. Only their status metadata was inspected; no failure cause is inferred here. PR base remains 61303fa, while current main is a02ef18; I did not test the merge result with current main. Full suite, skipped coverage, Windows, real service/model/device qualification remain unverified.
Review scope remains the isolated original service-reuse goals associated with #467. Later detailed selection/repair/profile policies lack authenticated original acceptance provenance in the supplied requirements; this review does not accept those policies on the owner's behalf. Python source was inspected through a comments/docstrings/prose-filtered AST view, so documentation and removed runtime-string semantics were not fully reviewed. No code changes, merge, deployment or hardware experiments were performed.
zengchuang-hw
left a comment
There was a problem hiding this comment.
Independent follow-up for direct review request issuecomment-6944620407 (updated 2026-11-17T09:52:38Z), anchored to current head 1e5fac1. Request identification is metadata only: I did not read or adopt its prose, other reviewers' opinions, or the author's summaries. My prior review is #480 (review). No existing own review explicitly identified this request, so this is a substantive bounded recheck rather than assuming same-head coverage.
Fresh evidence: tests/test_project_service_reuse.py and tests/test_project_objectives.py completed on this exact head: 76 passed in 164.23s, exit 0. This expands the nine-test scope of my preceding review. I independently re-inspected revision/retain_revision and change/performance/better through the safe AST view: dirty checks cover declared roots rather than only surviving members; snapshots read blobs from the recorded commit; eligibility uses per-objective direction and parent/initial regression bounds, while incumbent ordering is a separate declared-order comparison rather than summing unlike units. These are implementation observations, not acceptance of a particular policy on the owner's behalf. No new actionable defect was found in this bounded recheck; the previously reported retained-code and lazy-hash issues remain resolved in their verified scopes.
CI update supersedes the failure snapshot in my prior review: declarations run 134045411900 completed successfully at 2026-11-17T12:39:21Z (the older failed run 134041551982 remains listed), and the current knowledge-model-policy commit status is success. I inspected status metadata only, not logs or commentary.
This is COMMENT, not full approval. Isolated original #467 goals are the requirements basis; detailed later policy provenance remains unresolved. The comments/docstrings/prose-filtered code view does not establish full documentation or removed runtime-string semantics. This run does not establish the full repository suite, prior skipped checks, Windows behavior, current-main merge-result correctness, or real service/model/device qualification. Prior broad-suite results belong to their previous head and are not relabeled as current-head results. No code changes, merge, deployment, paid operation or device experiment was performed.
david6666666
left a comment
There was a problem hiding this comment.
One regression from the broader Git status check.
| tracked = set(git("ls-files", "-z").split("\0")) | ||
| # Deleted tracked files are absent from members but still belong to the inputs. | ||
| if any(name not in tracked for name in selected) or git( | ||
| "status", "--porcelain", "--", *(includes or ["."]) |
There was a problem hiding this comment.
The declared-directory check now catches tracked deletions, but it also rejects untracked files the snapshot deliberately excludes. I reproduced this with a clean inputs/ plus only an unignored inputs/__pycache__/ file: no retained input changed, yet revision() refuses to run. Could this status check omit untracked paths while the existing ls-files check still rejects genuinely selected untracked inputs? Please add a cache-only regression alongside the deletion cases.
zengchuang-hw
left a comment
There was a problem hiding this comment.
Independent incremental review of 1e5fac1 → 05ed84c. Previous checkpoint: #480 (review). Current PR base remains61303fa8150472dfbda62ec8d77908bc05c73f3b. COMMENT, not approval.
Actual two-file delta inspected through GitHub:
- project_service.revision adds --untracked-files=no to git status while preserving the prior declared-root pathspec and the separate members-versus-git-ls-files check.
- In the actual members implementation, .git, pycache and .pyc entries are excluded. An ordinary new file under the declared inputs is still enumerated and rejected if absent from the tracked set. Thus suppressing untracked status output does not by itself admit ordinary untracked candidate inputs. Tracked modifications/deletions still reach the declared-root status check; the prior deletion fix is not removed.
- The new test constructs a committed inputs tree, adds an untracked pycache/entry and expects the same revision, then adds an ordinary untracked inputs/new.txt and expects rejection. Existing staged/unstaged deletion and outside-declared-input tests remain in source.
- retain_revision is unchanged in this delta and still reconstructs regular blobs from the pinned Git tree. The earlier retained-code and lazy-hash fixes have not been newly executed here; previous positive tests remain evidence for their earlier revisions.
I found no new actionable code defect in this bounded static inspection. The cache-specific regression and dirty/deletion/ordered-import checks still need execution on this exact head in a trusted test environment. This round is connector-only: zero local tests executed, not a passing zero-test suite. No repository code, test fixture, campaign or helper was executed.
Current-head machine evidence:
- declarations job134492489689/run44212019838 evaluates05ed84c0a803953b98f929bb051236531e248f7f with policyc37832c3c7a6c457f533683b7b3386f2c3a0c004 and returns noncompliant/job_exit1. Its structured errors identify disallowed effort for material AI author entries2,3,4 and an unapproved or suspended model for entry6.
- Earlier same-head job134492046740/run44211889525 also failed, with missing/invalid/stale author declaration. These are distinct evaluations, not product-test results.
- Combined commit status currently reports knowledge-model-policy failure. No full-suite/installed/browser success for this new head was established in this review.
Requirements remain the previously isolated high-level service-reuse goals associated with #467, not later reviewer/author policy assertions. Detailed selection/repair/profile/no-fingerprint acceptance provenance, full documentation semantics, current-main integration, Windows and real service/model/device qualification remain unverified. The previous76-test result is historical and is not relabeled as a result on05ed84c0. Current failed policy checks and missing fresh execution evidence prevent approval. No code change, CI rerun, merge, deployment or experiment.
05ed84c to
20851a9
Compare
zengchuang-hw
left a comment
There was a problem hiding this comment.
PR review verdict — DO NOT PASS yet. Independent incremental review, submitted as COMMENT under the standing authorization (not a Request changes vote).
Revision: previous reviewed head 05ed84c0a803953b98f929bb051236531e248f7f → 20851a9ce7b5af0fb3eb0880716b1c2a85dc860e; current PR base 8aba657da79e8ed9f6e6fbb33437e9d9fdd8f1c7. Previous review: #480 (review) .
What changed and what did not
The histories diverge, so a merge-base comparison includes unrelated inherited work and is not a trustworthy two-head delta by itself. I compared the actual contents of the 22 PR-changed paths between the two reviewed heads, including the former location of the CLI arguments.
project_service.py, the service protocol fixture, objective tests, local process orchestration and service-reuse guide are byte-identical to the previous reviewed head. I re-inspected the revision/member checks, pinned Git-tree retention and direction-aware assessment/ranking: ordinary untracked inputs remain rejected separately from the ignored cache files, declared-root status still detects tracked deletion, and snapshot bytes still come from the recorded commit.- The integration moves CLI imports into Apps and hashing imports into
harness.artifact_primitives. The no-fingerprint fixture now guards that defining module as well; the isolation regression checks that project/task-protocol aliases reference the new defining function. This is the necessary adaptation to the migrated implementation, not evidence that the ordered regression has actually passed on this head. - The shared task-evaluation path now rejects interrupted/unknown execution and negative/non-integer return codes before accepting an assessment; a completed assessment additionally requires successful execution. The service-specific core is unchanged. Dashboard assembly also integrates the overview/profile-first changes.
- Multi-objective eligibility remains distinct from lexicographic selection in the actual code; I did not infer a performance improvement from these comparisons.
Previous findings
The earlier retained-revision and lazy-import-stub defects remain resolved in their previously verified scopes; I am not reopening them as new failures. The relevant retained-revision implementation is unchanged, while the migration-adjusted guard needs fresh execution evidence. Historical 76-test and other prior-head results are not current-head test results.
Current blockers/evidence gaps
- Exact-head declarations job 134663415243 is failed. Its structured evaluator output identifies head
20851a9c…, policy8aba657d…,noncompliant, and “Missing, invalid or stale knowledge author declaration”; job exit 1. The currentknowledge-model-policystatus is also failure. This is an admission/policy failure, not a claimed service-code test failure. - No successful current-head product-test/full-suite result was established. Before approval, run the cache/deletion, retained-selection/final, ordered lazy-import isolation and service/objective tests against the migrated imports, and the applicable complete ready checks/merge-result integration. Source equivalence does not replace those checks.
- I inspected the current #467 requirement sections as specifications, not its ownership/closure assertions or discussion. They now spell out profile failure/retry, condition refresh and multi-objective boundaries, but reading this third-party body alone does not independently authenticate disputed later owner-scoped policy acceptance. That existing provenance limit is not silently cleared by rebasing. This review does not authorize additional workflow changes or claim real-model qualification.
Actually performed: connector-based content comparisons, inspection of the affected code/test definitions and requirement sections, AGENTS/review guidance, exact-head check/status records and the structured declaration log. Not performed: local repository execution, current-head tests, real service/model/device trials, performance/quality qualification or source changes. No other reviewers' opinions were read or used; no merge, CI rerun or deployment.
Optional follow-up: keep the isolated ordered cross-file regression as part of the migration checks so future moves of the hash implementation cannot silently evade the service guard.
zengchuang-hw
left a comment
There was a problem hiding this comment.
Evidence update to my previous review, at unchanged head 20851a9ce7b5af0fb3eb0880716b1c2a85dc860e and base 8aba657da79e8ed9f6e6fbb33437e9d9fdd8f1c7. No code delta; COMMENT, not approval.
Resolved: the current declaration/policy blocker. I inspected job 134671760733's structured log: evaluated head matches this revision, policy is 8aba657d…, result is compliant, errors are empty and job exit is 0. The current knowledge-model-policy status is success. Older failed declaration runs remain in history but are not being carried forward as the latest result.
Not established: this result explicitly says “Declaration compliance only”. The current-head check collection contains declaration jobs, not successful product-test/ready-suite execution. It therefore does not validate the migrated CLI/Harness imports, adjusted hash guard, service/objective regressions or current merge-result behavior requested in the prior review. The requirement-provenance limit also is not answered by declaration compliance. Earlier resolved code findings stay resolved within their earlier verified scopes.
This was a machine-evidence and own-review recheck only. No local tests, new service/model/device trials, source changes, CI rerun or merge; no other reviewers' opinions used. Before approval, supply the current-revision product/integration verification and the remaining scoped requirement evidence described in the previous review.
Agent-assisted-by: Codex (author) Signed-off-by: lirx-pd <616517220@qq.com>
Agent-assisted-by: Codex (author) Signed-off-by: lirx-pd <616517220@qq.com>
Agent-assisted-by: Codex (author) Signed-off-by: lirx-pd <616517220@qq.com>
Agent-assisted-by: Codex (author) Signed-off-by: lirx-pd <616517220@qq.com>
Agent-assisted-by: Codex (author) Signed-off-by: lirx-pd <616517220@qq.com>
Agent-assisted-by: Codex (author) Signed-off-by: lirx-pd <616517220@qq.com>
Signed-off-by: lirx-pd <616517220@qq.com> Agent-assisted-by: Codex (author)
20851a9 to
b7cb4dc
Compare
zengchuang-hw
left a comment
There was a problem hiding this comment.
Independent incremental review: 20851a9ce7b5af0fb3eb0880716b1c2a85dc860e → b7cb4dcb644cb8ff274a57018f8d13f379ff7f7b, current base cf671a880ac8037e01922e1be0b701771fb9922f. Previous own checkpoint: #480 (review) . COMMENT: current-revision product verification is still missing.
I compared the actual old/new contents of all 22 PR files rather than treating a three-dot rebased diff as entirely new implementation. The service-reuse guide, protocol fixture, objective tests, fingerprint-isolation test and several dashboard/service dependencies are byte-identical. The R3 protocol-child calls/error-type preservation and Cannbot CLI integration are present in the new base; they are not newly introduced service-reuse policy changes.
Substantive service delta: profile artifact and result paths now serialize via Path.as_posix() rather than native str(Path). This is consistent with safe_path's explicit rejection of backslashes: the earlier serialization could produce Windows separators that its own reader refuses. Existing artifact existence/nonempty/validation/capture-interval checks remain before publication. The new fixture assertion reopens both profile.result and each profile.artifacts entry through safe_path at the campaign root and requires an existing file. This is a relevant regression assertion, not proof that it has executed on Windows.
The retained-code and fingerprint-boundary fixes from earlier reviews are not reopened by this small delta; no new concrete blocking source defect was established. I rechecked current AGENTS guidance: scaffold CPU checks must not be presented as service/campaign performance acceptance.
Before approval: the exact-head check endpoint currently contains only two successful declarations records (134779798686, 134779651337). They are policy/declaration checks, not the service-reuse fixture, native Windows path coverage, full suite or installed-package execution. Please provide an applicable current-revision product test run, including the changed profile round trip on a native Windows path implementation (or equivalently demonstrated matching-content evidence), and the required ready checks. Old-head passes and a new assertion in source do not execute the rebased integration. No failed product run is being asserted here; the evidence is absent from the checks inspected.
This round is a bounded source and current-machine-state review, not acceptance of unresolved governing-policy provenance or whole-Issue #467 service/effectiveness criteria. No local repository tests, model/device/service campaign or browser were run; no other reviewers' opinions, source edits or merge were used/performed.
|
Hi @lirx-pd, a heads-up about the knowledge author policy. #487 changes the admitted knowledge-authorship models; the implementing PR is #499. It adds This PR declares There are two ways forward:
Sorry for the extra step on an in-flight PR. Details and the owner's reasoning are in the decision record on #487. |
zengchuang-hw
left a comment
There was a problem hiding this comment.
Independent request-level follow-up for #480 (comment) . Previous review: #480 (review) .
Revision range: b7cb4dc → b7cb4dc (no source delta); base remains cf671a8. COMMENT, not approval.
I freshly re-inspected the profile publication/readback boundary in project_service.py and its regression source. Both profile artifact and result paths use relative_to(root).as_posix(); safe_path rejects backslashes, absolute/parent-escaping paths and symbolic/escaping targets. Publication still checks nonempty files, validation and capture interval before constructing the profile record. The fixture reopens the result and every artifact using safe_path. No new concrete source defect was established by this bounded recheck. Previously resolved retained-code and fingerprint-boundary findings are not reopened; historical test passes are not relabeled as executions of this head.
New machine-state evidence: all four current-head check records are declarations jobs. The newest job 134819659000 is successful, superseding the intervening failed declaration job 134818765287; current knowledge-model-policy status is success. These are not product-test results. The exact-head Actions collection additionally identifies ready/review run https://github.com/InferMatrixLab/InferRSI/actions/runs/44297910074 as completed with conclusion action_required. That state is not a passed ready suite; I do not infer its cause or call it a product-test failure.
Before approval: complete the applicable ready/product verification on this revision, including service/objective, retained-selection/final and ordered import-isolation regressions, plus native Windows profile-path round-trip evidence or independently demonstrated equivalent matching-content evidence. The prior scoped requirement-provenance and whole-Issue #467 service/effectiveness limits are not cleared by a cc or declaration success. No new optional code change is requested in this no-delta follow-up.
Performed: source/test-definition inspection, AGENTS reread, exact-head check/status/Actions metadata and own-review deduplication. Not performed: repository-code execution, local tests, full-suite/installed-package/native Windows execution, real model/device/service trials, CI rerun, edits or merge. No other reviewers' opinions were read or used.
zengchuang-hw
left a comment
There was a problem hiding this comment.
Machine-evidence follow-up to #480 (review) . Revision range remains b7cb4dc → the same head; base cf671a8. No new source delta or newly identified code defect. COMMENT, not approval.
Ready run https://github.com/InferMatrixLab/InferRSI/actions/runs/44297910074 has now started, superseding my earlier action_required snapshot. Windows, macOS, dashboard-browser, installed-package 3.10/3.11 and pr-references checks currently report success; full-check 3.10/3.11 remain in progress at this recheck. Latest declarations and policy status remain successful.
I independently read Windows job 134824160032's actual checkout and test output. It checks out merge commit 7fa51c5 (b7cb4dc into cf671a8), not a direct head checkout. Its explicit native subsets completed 8 host/plugin tests and 25 primitives/session tests, each with zero failures/errors/skips. Their enumerated nodes do not include tests/test_project_service_reuse.py or the profile result/artifact round-trip assertion. Thus this is positive native host/integration evidence, not yet evidence for that particular Windows profile boundary.
Before approval: obtain the pending full-check results and applicable service/objective/ordered-isolation verification; demonstrate the changed profile round trip on native Windows or independently equivalent matching-content evidence. Prior requirement-provenance and whole-Issue service/effectiveness limitations remain unchanged. I am not asserting a product test failure and am not asking to repeat the already successful host subsets. No new optional source change is requested.
Performed this follow-up: exact head/base and own-review verification, live machine-status refresh, actual Windows checkout/test-log inspection. No local repository execution, source edits, CI rerun, merge or real model/device trial. No other reviewers' opinions were read or used.
Governing Issue and delivery slice
Refs #467
Knowledge authorship declaration
Changed artifacts and evidence
src/inferrsi/harness/project_service.py,local_process.py20851a9cand05ed84c0verified one service start per attempt, same-service stage binding and cleanup.project_service.py,tests/test_project_service_reuse.pyproject_service.py,optimization.pyproject_service.py,tests/test_project_objectives.pyoptimization.py,tests/test_project_service_reuse.py05ed84c0,finalreplayed the accepted revision and effective configuration with its retained profile.candidates.py,optimization_session.py, knowledge/dashboard adapters,tests/test_service_fingerprint_isolation.pytests/test_harness_execute.py,tests/test_knowledge_compatibility_shells.py,docs/service-reuse-campaigns.mdChecks and dependencies
b7cb4dcb. A clean checkout on Windows 10 with base CPython 3.11.16 passedtest_one_service_order_and_independent_final_ignores_workspacewith zero skips. Baseline, candidate and independent final each passed with verified cleanup. The candidate's profile result and trace used campaign-relative paths with forward slashes, and both resolved to retained files.7fa51c5f(b7cb4dcbintocf671a88). Python 3.10.20 and 3.11.15 each passed 5239 tests and 672 subtests, with 37 skips, zero failures and one duplicate ZIP entry warning. Both suites passed the service lifecycle, retained-code final replay, cache/deletion and ordered packaged-task isolation regressions. Ruff and the sdist, wheel and clean installation checks passed on both Python versions.7fa51c5f. The Ready browser and offline report selection passed 41 tests. The base's additional round-decision module passed 26 tests. Neither run had skips.7fa51c5f. All 8 host-compatibility tests and 25 Harness primitive and writer tests passed with zero skips. Building a wheel from the sdist with short build paths, installing it in a clean environment, and checking the console and module entry points passed.6315d333passed the Job Object process-tree cleanup test and all four frozen-byte tests.6315d333, the eight review-related modules passed 193 tests and 43 subtests, with eight skips. The Harness execution module passed 17 tests, with one Windows-only skip, in a separate run with the optional schema dependency installed.20851a9c. DeepSeek V4 Flash W4A8 on Ascend A3 completed the same-service performance, GSM8K and profiling sequence. Two performance repetitions covered 8K/C4 and 16K/C16. The quality stage returned 300 valid responses and scored 285/300 under a response-completeness gate. The service completed 48 profiling requests, and the capture from one FFN worker retained CANN runtime records andAfdMegaMoePullNativemetadata. Cleanup was verified.20851a9crejected round 1 at 286/300 against its 289/300 score floor. Rounds 2 and 3 failed the performance gate. The second consecutive rejection produced one Agent advisory.finalpassed at05ed84c0. DeepSeek V4 Flash W4A8 on Ascend A3 completed all three candidate rounds andfinal. Each attempt started one owned vLLM service and completed 256/256 measured requests. Both performance repetitions covered 8K/C4 and 16K/C16 and recorded throughput and p50 latency.AfdMegaMoePullNative.finalstarted a fresh service, independently replayed round 1's retained revision and effective configuration, scored 287/300, and reused the retained profile. The campaign reachedcompleted, with service cleanup verified in all five attempts.finaleach had one length-limited response, scored as returned.05ed84c0on Python 3.10 and 3.11: 8 targeted tests passed per version, covering revision checks and snapshot rejection. Ruff on the changed files andgit diff --checkpassed.31bed72e, the release-publisher shell tests with jq passed 29 tests per Python version using simulated publication. Draft/merge smoke and Linux frozen-byte checks passed 12 tests.7fa51c5f. The local PR body also passed the Issue-reference check.Remaining work and Issue closure
Handoff (only when an actual handoff or decision changed)
b7cb4dcb. Local Linux, Windows host, browser and installation checks passed on merge ref7fa51c5f. The Ascend A3 campaign results apply to their recorded revisions. The two Issue decisions and GitHub-hosted Ready checks are pending.