Repository navigation
[Task] RSI Dashboard: Optimization Workflow Visualization and Trace Capture #314
Description
Current state: Existing owner @david6666666 records requester-confirmed bounded implementation scope: an English, read-only, self-contained offline HTML dashboard, local Markdown review export, explicit campaign/trace associations, referenced evidence hashes and privacy/verification boundaries. Ready #319 is now at 3a900852f7550d076bfeff71e0cf8f0e9bf971b3. The latest five-file delta selectively shows recorded candidate identifiers and bounded GPU ordinals in displayed commands, with regression/documentation changes; it is separate from the earlier live transport work. This synchronization inspected change metadata and hosted results, not a duplicate code/browser review. All six current-head Ready checks and declarations passed. The PR reports optional read-only loopback live observation, independent-round trend restoration, context/optimization-story capture and narrative-versus-recorded-evidence checks. Its demonstration recordings are explicitly synthetic; the author also reports a Chinese live-browser check against an in-progress campaign, with incomplete attempts remaining inconclusive. These do not establish real optimization efficacy or the poor-outcome feedback acceptance case. The earlier translation/verdict collision and round-specific capture consistency findings received scoped repair confirmation at 870e752f; they are not still awaiting that original repair check. The later live-delta review at 806e71dd found no new actionable defect in its inspected paths. Those scoped historical clearances do not resolve the new current-head review concerns. Reviewer @amy-why-3459 finds the candidate/GPU-display delta reasonable but reiterates the existing branch-chart fallback overflow/nonterminating-tick concern and asks for clarification of remote-viewer/loopback binding. The reviewer reports no local test execution; this synchronization does not independently reproduce the hang or turn the binding question into a verified exposure. The relevant chart files are unchanged by this delta. Current-head review disposition and remaining story/trace acceptance are still pending. The PR's language/live/story additions remain distinct from the linked original English/offline scope record; this status does not accept added scope. Caller-authored narrative, unauthenticated capture origin/model consumption, unavailable historical context and local/browser qualification limits remain explicit. All original acceptance items below remain open.
New #344 proposes per-iteration knowledge-hit/adoption attribution, and #345 proposes separate accuracy/performance progress percentages. Both are unassigned Proposed Issues; each explicitly records creation-only authorization and awaits first-version scope, data-contract and owner decisions. They are related proposals, not delivered #319 capabilities or additions accepted into this Issue's original checklist. Attribution must distinguish use from causal benefit, and workload completion from test verdict, as those proposals state.
Decision needed: The existing #314 owner/requester retains the remaining delivery-scope and acceptance decisions. Maintainers separately need to settle #344/#345's initial scope/interfaces and assign their owners; this synchronization does not accept or assign either proposal.
Next action: The existing owner retains integration and the remaining acceptance evidence, including the disclosed real-run/browser verification limits. Performance acceptance remains a separate issue-scoped decision; the original acceptance checklist below remains open.
Parent issue: #316 · InferRSI Roadmap
Goal
Provide an InferRSI dashboard that makes an optimization task observable from diagnosis and candidate search through execution, validation, and review. Correlate each iteration's results with traces to help diagnose why optimization outcomes fall short.
Owner: @david6666666. Corresponds to original roadmap item 3.
Scope
- Workflow visibility: Show task / run status, stages, iterations, candidates, validation results, and available performance metrics.
- Trace capture and correlation: Link optimization steps to their events, logs, tool call records, and existing profiling artifacts so users can trace results back to specific steps and evidence.
- Problem diagnosis: Surface failures, rollbacks, steps with no improvement, and missing evidence. Help identify whether a problem originates in diagnosis, knowledge selection, search, execution, or validation.
- Human review entry point: Support inspecting and referencing traces so findings and recommendations can feed back into the framework, connecting to the shared feedback loop in roadmap item 7.
Build the first version around the existing RSI workflow and evidence artifacts. The owner will refine integration points, trace fields, and the interface design in this issue.
Acceptance Criteria
- Show an optimization task's stages and iteration results as a connected workflow.
- Make key steps traceable to the corresponding traces, logs, and reports, with explicit indicators for missing records.
- Demonstrate “inspect the workflow → diagnose the problem → produce framework improvement feedback” using a case with poor optimization outcomes.
- Provide setup, integration, and usage instructions, including the first version's coverage and limitations.
Activity
[-][Task] RSI Dashboard：优化流程可视化与 Trace 捕捉[/-][+][Task] RSI Dashboard: Optimization Workflow Visualization and Trace Capture[/+]agent-ci-infra-bot commentedon Nov 2, 2026
意图：为 InferRSI 做一个 Dashboard，把一次优化任务从诊断、候选搜索、执行、验证到复盘的阶段与每轮结果可视化，并与 Trace/日志/证据关联，用于定位效果不佳的原因和支撑人工复盘反馈。结论：需信息 + 需人类决策——Issue 自己把接入点、Trace 字段与界面方案留给负责人在本 Issue 内细化，且与 #61「不假设大范围 GUI」冲突，当前没有可派发的有界切片。核对了仓库记录模型与现有报告入口，未执行任何测试（这是 Issue 评审，无代码改动）。
Labels: kind:task, kind:rfc, area:workbench, area:usability（路线图下的交付任务，但当前仍是待细化的设计，涉及 workbench 证据/报告与人工复盘可用性）
承载形态与产品方向未定，且与已生效的路线图约束冲突
pyproject.toml:12dependencies = []），报告是单文件静态 HTML（src/inferrsi/harness/reporting.py:11），只读入口是 CLIstatus/report（src/inferrsi/cli.py:275）。服务化会引入新的依赖、端口、进程清理与安全边界，这不是实现者能自行选择的。Trace 字段与采集点未定义；现有记录里不存在 Issue 描述的五个阶段
src/inferrsi/harness/runner.py:43（events.jsonl）与:46/:51（status.json）只记录source_snapshot、measurement、quality_validation、complete、crashed这几类事件（runner.py:122-205的event(...)调用）；src/inferrsi/runspec.py:10-31的 RunSpec 只有baseline/candidate_repo与repetitions，没有「轮次」「候选集」概念；src/inferrsi/model/anthropic_agent.py:412-435、:556-581另有一套agent.request/agent.response/agent.tool/agent.context事件，与 workbench 的 run 事件没有共同的 run/尝试 ID。缺少仓库要求的 accepted scope 与 scoped-task 记录字段
AGENTS.md:117「Work Issue-first: confirm the governing Issue's accepted scope before changing code…」；docs/agent-collaboration.md:15「Work follows accepted scope and explicit or standing delegation. Record or link that authority.」；.github/ISSUE_TEMPLATE/scoped-task.md要求 Current state / Decision needed / Next action 以及 Ownership and authorization（含 Status、Accepted decision）。范围过大且混合多个交付物，建议拆分
docs/agent-collaboration.md:15之后的 Working rules（每个 PR 应是有界交付切片）。验收演示的数据来源与隐私边界未说明
docs/privacy-security.md:10-24（日志、报告、证据都不得含真实主机、用户名、绝对路径与凭据，需脱敏）。report.html目前会为 run 目录下所有文件建链，src/inferrsi/harness/reporting.py:26-30，这对「证据浏览」也是一个现成的泄露面）。evidence/的案例与允许展示的字段，明确脱敏与只读边界，并把「缺失证据」的展示方式写成可检查的验收条件。Reviewed commit: 2baebf6
Coverage: Snapshot: 2637 files, 42149780 bytes; skipped 0 entries; Initial model context: 28 files; other accepted files available through tools; Git workspace contains a fixed snapshot, not full upstream history
david6666666 commentedon Nov 2, 2026
The request owner confirmed this implementation scope for #314:
This confirms the dashboard implementation scope. A performance acceptance claim remains a separate issue-scoped decision with its own pinned baseline, workload, quality gates, and evidence.
Add campaign optimization story with captured knowledge context