Repository navigation
feat: 增加逐轮检索、备选、选择与执行结果展示 - #495
Conversation
Group recorded retrievals, alternatives, declared selections and measured results in both dashboard layouts. Preserve missing evidence, exact experiment navigation and live reader state. Agent-assisted-by: Codex (author) Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com>
Keep the published offline example reproducible and teach the isolated profile-story test harness about the shared round decision view. Agent-assisted-by: Codex (author) Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com>
zengchuang-hw
left a comment
There was a problem hiding this comment.
PR review verdict — DO NOT PASS yet (required interaction/integration evidence pending). Independent initial review; COMMENT rather than a Request changes vote.
Revision: base 76ca44a8b77553b2cb7f08b9f861399a72736fe2 → current head 9588bb85144e858612ab6869cbb0151e1ad2f883. I included the synchronize delta from e57acfd5ddf0dc84ddf0d7e62b00ab13aa8dc3ef: it updates the generated profile example and the profile-check component mock, not the new panel implementation. No earlier review from this account exists for this PR.
Independent source assessment
- The four-stage panel uses explicit round/capture links, keeps absent-round records under an unassigned choice and labels caller-written context separately from verified retrieval snapshots. Unverified captures do not supply returned-item counts or assembled context.
- Per-round trace alternatives and explicitly linked option IDs are separate from the disclosed campaign-wide option space. Selection-to-try is not rendered as adoption after validation. Run-record and caller-decision disagreements remain visible.
- Run navigation prefers an experiment identity; the default-layout fallback requires an exact unique round/candidate/run tuple. Duplicate identities suppress the evidence button rather than choosing a convenient result.
- Text is inserted with DOM text nodes/textContent, not interpolated into HTML. The new browser test definitions include HTML-like input, multi-run identities, conflicts, zero hits versus missing snapshots, live selection/focus retention, language changes and narrow widths. These are useful test definitions, not independently observed passes.
- Both templates wire the shared component; the new profile example includes it, and the lightweight mock explicitly stubs its render operation. The mock does not test real browser interaction.
Before approval
The current revision is a draft. dashboard-browser job 134685408582 is skipped; the workflow's !pull_request.draft condition explains this. The dedicated job now selects tests/test_round_decisions_browser.py, but a selected test in a skipped job is not executed evidence. Obtain a successful applicable Chromium run of these cases on matching content, including both layouts and live/offline paths. No request to change the intended draft workflow is implied.
The Python 3.10/3.11 full-checks and several package/host checks were also unfinished at this inspection. Complete the applicable checks and inspect their actual results. Lint/smoke and declaration records currently report success; those records cannot establish interactive behavior.
Scope and limits
I inspected actual source/diffs, test definitions, the isolated original #491 goals/constraints/completion criteria, AGENTS and review/contribution guidance. The docs correctly state that this is a read-only presentation slice: it neither collects missing real workflow history nor completes #491's real-campaign evidence requirement. Synthetic screenshots/data are not real optimization evidence. I did not inspect the PNG screenshots visually or run the rendered page, so visual quality and browser behavior remain unverified here.
No concrete blocking code defect was established in this static review. No local repository tests, browser execution, hardware/real-campaign run, source edits or merge were performed; no other reviewers' opinions were read or used. Optional robustness coverage: explicitly test duplicate run identities disable navigation, and define/test the profile-layout behavior when experiment IDs are absent (the fallback test currently covers the default layout only).
david6666666
left a comment
There was a problem hiding this comment.
One retrieval-evidence gap in the new panel.
| button(box,tr('Open retrieval evidence','查看检索证据'),'capture:'+id,()=>openCapture(id)); | ||
| } | ||
| for(const search of roundSearches){ | ||
| const box=item(retrieved,search.path);field(box,tr('Coverage / state','覆盖 / 状态'),label(search.coverage||search.state)); |
There was a problem hiding this comment.
A round can have a shared file-search record without a capture. Its projected events already retain the recorded query or request, but this loop only shows coverage and materials, so the new Retrieval stage can hide the query even though the existing knowledge view has it. Could this render the recorded query, or link directly to that search record, and cover a shared-search-only round in the browser test?
There was a problem hiding this comment.
Confirmed and fixed in 95bb240. The new shared-search-only regression failed on the reviewed version: the projected event contained the query, but the round panel did not render it.
The Retrieval stage now displays each recorded search query (including request query/relevance or explicit IDs) and expandable request/event evidence. Missing queries remain “Not recorded”; a missing capture still does not imply zero hits. Query text stays literal DOM text.
Added Chromium coverage for a shared-search-only round in both layouts and live/offline modes, including scope/material evidence, missing queries, HTML-like text, and language switching. All 121 affected browser/projection/profile checks passed locally on the content committed above; lint and diff checks passed. The remote PR head is confirmed at this commit; its CI rerun is pending.
Render recorded search requests when a round has no capture, and preserve unique run navigation without experiment IDs in the profile layout. Cover both layouts, shared-search-only rounds, and ambiguous identities in Chromium. Agent-assisted-by: Codex (author) Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com>
|
Follow-up to the initial review, with fixes pushed as 95bb240:
Current-head validation: 121 passed across the round-decision (26), experiment (26), knowledge-navigation (12), profile (5), and retrieval-projection (52) suites; The read-only scope remains as reviewed: real-campaign collection/history and the broader #491 acceptance remain outstanding. Maintainer re-review and the new CI results are still needed; no threads were resolved and no merge was performed. |
zengchuang-hw
left a comment
There was a problem hiding this comment.
Independent incremental review of 9588bb85144e858612ab6869cbb0151e1ad2f883 → 95bb2407275ca9c79d40523cba6338abd556a1ac, base 76ca44a8b77553b2cb7f08b9f861399a72736fe2. Previous review: #495 (review) . COMMENT pending current-revision verification.
Delta inspected: all five changed files, including the generated profile example. Shared-search events now expose the recorded query/request and expandable event evidence even when no capture exists. I traced the upstream projection: these event fields come through its public-value projection, and event-state warnings remain visible. The panel uses textContent for both query and JSON evidence; absent capture data is not turned into a fabricated returned count.
Profile navigation now stamps a round/candidate/run identity on each actual round summary and requires exactly one matching target, with experiment identity preferred when available. This addresses the profile fallback gap I previously suggested testing. New test definitions cover both layouts without experiment IDs, duplicate identities with/without IDs, and search-only rounds across live/offline layouts, including HTML-like query text and language switching. These additions address both optional coverage suggestions at the source level; execution on this revision remains separate.
Earlier browser gap is resolved for the previous content: I read job 134686329594's actual output: 58 passed, including every then-present round-decision case. It tested merge b940607… for previous head 9588bb85…. This is no longer merely a skipped browser job. It does not, however, execute the changed query/event rendering, new profile navigation or the added regression cases at the new head.
Before approval: current-head browser job 134692338442, both full-check jobs and other ready checks were still running/queued at inspection. Confirm successful applicable checks and inspect the new browser results before approval. I found no new concrete blocking defect in this bounded static delta; I am not reporting a pending check as failed.
No local tests or browser/page execution, source edits or merge were performed. I did not read other reviewers' opinions. This remains a read-only presentation review, not #491 real-campaign integration acceptance, screenshot visual certification or a claim that synthetic data prove optimization results.
zengchuang-hw
left a comment
There was a problem hiding this comment.
LGTM
Independent evidence-completion follow-up to #495 (review) , for unchanged exact head 95bb2407275ca9c79d40523cba6338abd556a1ac (previous source delta 9588bb85144e858612ab6869cbb0151e1ad2f883 → this head).
The prior current-revision verification gap is now resolved. I read the actual dashboard-browser job 134692338442 output: 67 passed, including all 26 round-decision cases. These execute the new shared-search query/request evidence and language-switching cases, exact round/candidate/run fallback in both layouts, and duplicate identity suppression with/without experiment IDs. Thus the earlier optional navigation/ambiguity suggestions are now covered by execution, not only test definitions.
Both Linux full-check logs (134692338149, 134692338712) show Ruff success and 5160 passed, 37 skipped, 1 warning, 672 subtests passed each. Browser modules skipped in these general jobs are exercised separately by the successful browser job above; other optional external-source/model/platform skips are not passes. All 11 current-head check records are successful, including installed-package and native-host checks.
Execution provenance: these logs checked out merge 6df3675512a89ce3f3403c0caabcdb50450f405d, whose parents are current base 76ca44a8b77553b2cb7f08b9f861399a72736fe2 and this exact PR head. The merge tree is not identical to the head tree: it incorporates base changes. I compared the merge delta against all 13 current PR files and found no overlapping paths, so the reviewed panel, integration, documentation and browser-test bytes are the ones exercised in this integration run.
My independent source review remains applicable: explicit round links are not fabricated, missing capture evidence is not converted to zero hits, warnings and declared alternatives remain visible, ambiguous navigation is disabled, and event/query text uses textContent. I found no remaining blocking issue for this bounded read-only presentation change.
Approval is for this PR slice, not whole-Issue #491 real-campaign acceptance, genuine knowledge consumption or optimization effectiveness. No local tests/browser execution, screenshot visual certification, source edits or merge were performed. I used code, primary CI records and my own prior findings, not other reviewers' opinions.
Dong1017
left a comment
There was a problem hiding this comment.
LGTM
PR review verdict — PASS
Reviewed base 76ca44a8b77553b2cb7f08b9f861399a72736fe2 and head 95bb2407275ca9c79d40523cba6338abd556a1ac. This is a re-review of the fix delta 9588bb85 → 95bb2407, requested by the author's @mention.
Earlier findings: resolved
- Shared-search-only round hides the recorded query (@david6666666,
round_decision_view.html:87): fixed. Amy reports the new regression failed on the reviewed version and passes now. The Retrieval stage now shows each recorded query and request, with expandable request/event evidence. A missing query shows "Not recorded", and a missing capture is not read as zero hits. @zengchuang-hw traced the projection and confirmed the text usestextContent. - Profile layout navigation with no experiment ID (@zengchuang-hw's suggested test): fixed. Navigation now needs a unique experiment ID or an exact unique round/candidate/run tuple, in both layouts.
- Duplicate identities: four browser cases cover duplicate experiment IDs and duplicate fallback tuples. Ambiguous records offer no evidence button.
- Validation on the current head: this was the earlier "browser gap" blocker. Per @zengchuang-hw's current-head Approve (
pullrequestreview-5954153069):- The dashboard-browser job passed 67 tests, including all 26 round-decision cases.
- Both Linux full-check jobs passed with Ruff clean and 5160 passed, 37 skipped.
- All 11 current-head check records succeeded, including installed-package and native-host checks.
- Those logs ran on merge
6df36755…, which has no file overlap with the 13 PR files.
Review state: there is no outstanding CHANGES_REQUESTED. @zengchuang-hw's APPROVED review is on this exact head. The read-only scope matches the PR body: it keeps Refs #491 and does not claim to close it.
Not verified here: I did not run tests or open the CI logs myself. The CI and test results above come from the seeded reviews and the PR author's report. I also did not run the screenshots against real campaign data.
Non-blocking follow-ups (tracked in #491, not part of this slice):
- Real-campaign capture, MiniCPM history backfill and end-to-end acceptance.
- The synthetic-data screenshots do not count as optimization-benefit evidence.
This is one Approve from Dong-bot (authorized delegate). The merge ruleset's two-approval and stale-approval rules still apply.
Posted by Dong-bot, an automated agent acting on behalf of @Dong1017 (always-on-bot, model claude-sonnet-5-5).
zengchuang-hw
left a comment
There was a problem hiding this comment.
Independent same-head recheck for the new direct mention in review 5955098143, following my approval.
Range: 95bb2407275ca9c79d40523cba6338abd556a1ac → the same exact head. No code delta or new primary execution evidence was identified. I reopened the panel source: event queries still render through textContent, capture and caller-recorded context remain separate, and run navigation still requires a unique experiment identity or round/candidate/run tuple. The same eleven check records remain successful; this is a recheck of the previously verified executions, not another test run.
The bounded read-only UI approval remains applicable; I am not submitting a duplicate APPROVE. The prior browser evidence remains 67 passed, including 26 round-decision cases, and the full-check evidence and merge-content applicability remain as recorded in that approval. No new blocking defect was established. This still does not accept #491's entire real-campaign/effectiveness scope or certify screenshot visuals.
No repository or browser code was executed, no other reviewers' opinions were read or used, and no edits or merge were performed.
将每轮分散在知识导航、优化日志和候选详情中的信息汇总到「逐轮决策」,直接查看 检索结果 → 备选方案 → 最终选择 → 执行结果。例如第一轮可以同时看到返回的知识条目、明确记录的备选方案、选中尝试的方案及理由,以及该候选实测后是否采纳。
Governing Issue and delivery slice
Refs #491
Changed artifacts and evidence
两种布局共用一个面板;支持中英文、离线 HTML、Live 更新及窄屏。Profile 布局维持先显示性能趋势的入口,顶部「逐轮决策」链接展开新面板。刷新保留所选轮次、展开项与焦点。缺失 round/run 的外部记录使用已有实验身份,不按顺序推断第一轮。两种布局在缺少实验 ID 时均按唯一的 round/candidate/run 组合跳转;身份重复时禁用跳转。
以下是已实现页面的合成数据截图,数字不代表实际优化收益:
Profile 布局 · 移动端
Checks and dependencies
95bb2407275ca9c79d40523cba6338abd556a1ac的本地相关验证:121 passed。运行tests/test_round_decisions_browser.py、tests/test_dashboard_experiments_browser.py、tests/test_knowledge_navigation_browser.py、tests/test_profile_story_checks.py和tests/test_retrieval_projection.py;ruff check src tests tools、git diff --check通过。测试在提交前运行,所测文件内容与该提交一致。9588bb85144e858612ab6869cbb0151e1ad2f883已完成 Ready CI:Chromium 58 passed,Python 3.10/3.11 full-check 以及 package/host 检查通过;这些结果仅属于上一提交。Remaining work and Issue closure
独立 review 已提出验证要求和检索展示缺口;本轮已补齐共享搜索查询、Profile 身份回退及回归覆盖,待维护者复核修复提交和新 CI。#491 的真实 campaign 接入、历史关联证据和目标验收继续保留;本 PR 的合成页面和浏览器测试不替代其验收。