[Dante-dan]

Goal And Delivered Outcome

• Outcome basis / optional anchor: the exact-session/run and age acceptance row in [Task][RFC]: Qualify DSH/Pi observation and managed-runtime selection #5208; existing claim comment.
• Goal/source and gap: a caller currently receives one aggregate stage from a goal ledger without a way to require its expected observation treatment and session.
• Observable before → after: manual status --expected-binding <JSON> --as-of <timestamp> --max-observation-age-ms <integer> returns explicit binding eligibility; wrong/ambiguous identities, non-valid integrity and stale/future observations produce stage: unknown. Unbound historical reads keep their response shape.
• Issue/task and intended base: Related to [Task][RFC]: Qualify DSH/Pi observation and managed-runtime selection #5208; base main. This completes only the manual exact-read slice, leaving the parent task open.

Author Declaration

• Written by: model_agent — OpenAI GPT-6.1, under the contributor's task direction.

Implemented against

• Specification and revision: [Task][RFC]: Qualify DSH/Pi observation and managed-runtime selection #5208 Acceptance row 2 and docs/architecture/rfcs/harness-selection-dsh-pi-v0.md Implemented Readback Increment at 2b72d30fca7f452a71dffb723ebf6d1e2e878484.

Criterion (spec clause)	Disposition	Symbol / path	Test or command
Exact session/run identity and observation age prevent stale-ledger current health	implemented	build_diagnostic_projection and reliability diagnostics CLI	test_exact_binding_*, test_cli_exact_binding_*
Observation cannot influence prompts or scheduling	implemented for this read slice	existing read-only ledger/projection path; no write or provider call added	CLI ledger/state byte-preservation regressions
C0/C1, live lifecycle, retention, overhead and promotion	out_of_scope	existing #5208 / owning RFC acceptance	no live qualification claimed

• Self-check before submission: inspected the current typed envelope, stats, receipt and projection owners; preserved exact fractional UTC instant comparison; reviewed the opt-in API and negative cases. BindingIneligibility is local to diagnostic read refusal, not a second runtime-state vocabulary.

Scope And Continuation

• Completed scope and remaining work: an independently usable manual read contract and bilingual capability/RFC documentation. It rejects a conflicting ledger rather than silently filtering other runs. Counts/signals remain historical evidence when the binding is ineligible; eligibility is not task success.
• Slice boundary / successor: [Task][RFC]: Qualify DSH/Pi observation and managed-runtime selection #5208 retains live C0/C1, overhead, retention, lifecycle and runtime-selection qualification. A future UI/polling consumer still needs the existing owner-reviewed read budget; no panel or polling path is introduced here.

Validation

• Tested revision: 77aabe3297d1168d45651e92dc23a545f78c9a13 (same source tree as the working-tree runs).
• Run state: incomplete; remote failures triaged against the exact base.
• Input classes: synthetic, public_fixture.

Check	Outcome	Coverage / limitation
python -m pytest -q tests/capabilities/test_reliability_diagnostics.py tests/capabilities/test_reliability_diagnostics_readback.py	passed: 148	Real CLI positive/negative readback, matching single run, mixed session/treatment, integrity and unchanged ledger/canonical files; includes fractional stale/future boundaries
Changed-source Ruff	passed	Four changed Python files
Required full Ruff command	failed	Same F401 unused explore_history_rows in unchanged tests/presentation/test_linked_goal_sources.py:23 on candidate and exact main baseline
Required mypy	passed	17 configured source files
Diff-scoped semantic advisory	passed, one local enum candidate	No registered vocabulary reuse suggested
Required public boundary check	passed	0 errors; 45 warnings including missing optional registry and nonliteral credential references
Full pytest	incomplete, not passed	Prior local full-suite attempt reported failures/errors but its final result is unavailable. Remote test shards were skipped after the baseline static-check failure; the pytest aggregate failed without running them.
CLI output budget smoke / diff-selected premerge	pending final serial run	Earlier attempt encountered a local Node shim failure; not claimed passed
git diff origin/main...HEAD --check	passed	Whitespace

• Coverage and gaps: no live DSH/model calls, frontend journey, retention deletion, runtime promotion, Windows run. Local checks use the repository's provided test/QA paths. Existing full-suite failures and incomplete attempts are not represented as successes.

Frontend / Visual Evidence

• UI impact: none. This opt-in CLI contract adds no dashboard, documentation chrome or visual component; screenshots/preview are N/A.

Type of Change

• New feature
• Documentation update
• Test update

LoopX Area

• Capability or extension (providers, adapters, skills)

Technical Direction

• Direction / acceptance reference: [Task][RFC]: Qualify DSH/Pi observation and managed-runtime selection #5208 / S4, S10, S11; exact manual read prerequisite only.

Shared-authority RFC fixture impact

N/A: this change does not migrate authority or alter runtime routing.

Boundary Checklist

• No private state, credentials, traces, internal URLs or local paths in the diff or public evidence.
• No maintainer-owned benchmark duplication.
• Scoped to the existing task's exact read-contract slice.
• Both English/Chinese RFC files updated.
• UI impact marked none.
• Every commit includes a DCO sign-off. No personal certification has been added on another person's behalf; the actual Sign-off result remains a merge gate.

Current CI triage

I checked the logs for this PR at 77aabe3 against its exact base 2b72d30:

• kernel-static-checks fails with Ruff F401 for explore_history_rows at tests/presentation/test_linked_goal_sources.py:23 on both this PR and the base. Newer main b952235 has the same failure. In this PR, Stage2c and test shards were skipped after that failure; the stage2c-correctness-e2e, checks, pytest and merge-gate failures are downstream aggregation failures, not completed test results.
• Release Artifacts reaches smoke:chat-model-menu and fails assert.match(goal, /goal-only-model/): PR build. The exact base's chat-bundle-browser (3) runs that same smoke and fails at the same assertion with the same rendered Goal header. These are different jobs sharing the failing smoke, rather than a claim that a base release build passed or failed.

The separate DCO log confirms missing trailers on 5a1c3bac and 77aabe32; it is not the cause of the Python or release failures. Full-suite/premerge qualification remains incomplete, and #5208 remains open beyond this manual read slice.

[mergify]

Hi @Dante-dan, the DCO Sign-off check did not pass. Please inspect
its details first: checkout, fetch, timeout or infrastructure errors
need their own recovery, not a rewrite of otherwise signed commits.

If the log confirms a missing Signed-off-by trailer, amend the
affected commit with git commit --amend -s; for multiple commits,
use an interactive rebase against the current base from the correct
base-repository remote and sign off each affected commit. Push the
rewritten PR branch with git push --force-with-lease origin HEAD.

[songoow]

Reviewer: model_agent — OpenAI GPT-6 Astra Pro.

详细中文评审

审查版本：77aabe3297d1168d45651e92dc23a545f78c9a13；base：2b72d30fca7f452a71dffb723ebf6d1e2e878484。发布前已重新读取远端，head 未变化。本次覆盖全部 8 个变更文件，并结合 ledger、receipt、stats normalizer 和 DSH observer producer 做了三路独立核查。

结论：未发现本 PR 引入的、需要阻塞合并的正确性缺陷。 本次以 COMMENT 记录代码审查、非阻塞建议和验证边界；完整审查工作流及 merge readiness 尚未验证通过。

1. 动机

本 PR 对应 #5208 第二条 Acceptance及作者的认领说明：原有 Goal ledger 的 aggregate stage 无法向调用方证明它属于所期待的 session、observer 和运行配置。新增显式绑定、评价时间及年龄上限，解决的是手动读取时的证据接纳问题。

我按该边界评审：本次可交付 manual exact-read prerequisite；C0/C1、面板、自动轮询、runtime promotion、overhead 和 retention/deletion 仍属于父任务。没有将这些明确排除的后续工作列为当前 PR 的缺陷，也不把本次测试通过视为父任务完成。

2. 改动思路

实现继续沿用 ledger → LedgerReading → integrity receipt / diagnostic projection → CLI 输出，复用既有 normalize_observer_run_identity() 和 receipt 的完整性判断。新增 BindingIneligibility 表达本地读取拒绝原因，没有成为第二套 runtime-state owner。

发现冲突时拒绝整个绑定，而非过滤其他运行后声称当前健康；拒绝后 stage=unknown，counts/signals/integrity 保留历史证据。这与 RFC 新增读取合同一致。没有新增 provider 调用、调度输入、canonical state 写入或轮询路径。

3. 具体改动

• projection 输入校验：要求精确的四个顶层字段；复用八字段 treatment normalizer；绑定必须提供显式 as_of 和非负整数年龄上限；拒绝 bool、缺失参数以及只有年龄上限没有绑定的组合。
• 绑定判定：检查所有 envelope 的 provider/observer/session，要求唯一 stats identity，逐字段比较完整 treatment，并要求 receipt 为 valid。空 ledger、额外 observer、不同 session/treatment 或非 valid 完整性均不可采用。
• 时间比较沿用精确的时间换算，覆盖 UTC offset 和小数秒边界。任一未来 envelope 都拒绝；最新观测年龄超过上限才拒绝，等于上限允许。
• CLI 将 JSON 解析及 normalizer 的 ValueError 转为退出码 2。不提供绑定时，新增判定不执行，旧 projection 形状及条件式文本输出保持兼容。
• 两个测试模块覆盖真实 CLI 正常/异常 readback、时间边界及文件字节保持；四份文档明确手动读取、历史证据和无控制权的边界。

额外核查了两个可能的漏洞假设：

1. 同一 observer 的不同 treatment 不会被后写 stats 静默覆盖。既有 _store_stats_record() 会将冲突记录计为无效，新绑定又要求 valid receipt。两种冲突记录顺序均验证为不可采用；正常同运行累计 stats 仍可采用。
2. 八个 treatment 字段本身描述运行配置，但本次还固定 observer/session。DSH producer 默认为实例生成 UUID observer ID。没有发现当前调用链存在必须新增另一套 run ID 才能修复的混淆问题。

4. 对主干的风险

当前变更为 opt-in、只读且局限于既有 owner；审查和聚焦测试未发现默认行为回归。

非阻塞建议 A：统一既有 identity token 校验。

新增顶层身份使用 fullmatch()，而 intake._identity_token() 仍使用 IDENTITY_TOKEN_PATTERN.match(value)。正则结束锚点可匹配最终换行符之前的位置。我复现了 stats 与 expected binding 都包含 model_id="model-fixture\n" 时，binding.eligible=true 且 integrity.status=valid。

这是共享 normalizer 的既有问题，不是本 PR 引入的身份比较绕过；没有证据表明它能把不同身份当成相同身份。建议在共享 owner 中统一为 fullmatch() 并补尾部换行回归，以符合新增文档中非法 token 应被拒绝的约定。

非阻塞建议 B：补两个当前已正确工作的回归场景。

• 第二个 observer 只有 stats、accepted events 为 0：仍应拒绝歧义。
• 同一 observer/treatment 的多条累计 stats：应继续允许，避免以后把正常快照误判为多运行。

两者均已通过额外探针验证，建议固化为测试；它们不是当前缺陷。

5. 我的整体评价

实现方向、owner 边界和拒绝逻辑合理。当前证据支持“该手动读取功能未发现阻塞级新增缺陷”，建议保留现有方向，处理上述非阻塞改进并补齐最终验证记录。

本次独立执行：

验证	结果
python -m pytest -q tests/capabilities/test_reliability_diagnostics.py tests/capabilities/test_reliability_diagnostics_readback.py	148 passed，43.38 秒
四个变更 Python 文件的 Ruff	全部通过
三个顶层身份、八个 treatment 字段逐项 mismatch	均拒绝
冲突 stats 的两种顺序、额外 stats-only observer	均拒绝
同运行累计 stats	保持可采用
真实 CLI readback、异常输入和 ledger 文件字节保持	相关测试通过

未完成/未验证的边界：

• 本次没有独立执行全仓 pytest、完整 premerge、live DSH/model、Windows 或发布资格验证。
• 发布前 PR 描述仍记录全量 pytest 未结束且已有 failures/errors，CLI output budget smoke / diff-selected premerge 待最终运行；这些记录没有被视为通过，也没有由本次审查独立完成失败归因。
• 最新机器人评论 报告 Sign-off check 未通过。具体原因仍需检查该 check 的日志，不能仅凭状态认定一定缺少 Signed-off-by trailer。
• 仓库 loopx-pr-review 编排在审查环境因本地 TypeScript Effect runtime 权限限制未完成，未生成完整 packet / check-result 证明。该限制与上述源码审查和聚焦测试分别记录；本 COMMENT 不授予正式 approval 或 merge readiness。

English verdict

COMMENT — No blocking regression found in the bounded manual exact-read change at 77aabe3297d1168d45651e92dc23a545f78c9a13. The 148 focused tests and changed-source Ruff passed independently; additional identity and ambiguity probes behaved as intended. The trailing-newline token issue is pre-existing and non-blocking; two additional regression cases are recommended. Formal review-workflow validation, full-suite/premerge qualification and merge readiness remain unverified.

[Dante-dan]

@songoow, could you address these two baseline failures? I checked the logs for this PR at 77aabe3 against its exact base 2b72d30:

• kernel-static-checks fails with Ruff F401 for explore_history_rows at tests/presentation/test_linked_goal_sources.py:23 on both this PR and the base. Newer main b952235 has the same failure. In this PR, Stage2c and test shards were skipped after that failure; the stage2c-correctness-e2e, checks, pytest and merge-gate failures are downstream aggregation failures, not completed test results.
• Release Artifacts reaches smoke:chat-model-menu and fails assert.match(goal, /goal-only-model/): PR build. The exact base's chat-bundle-browser (3) runs that same smoke and fails at the same assertion with the same rendered Goal header. These are different jobs sharing the failing smoke, rather than a claim that a base release build passed or failed.

The separate DCO log confirms missing trailers on 5a1c3bac and 77aabe32; it is not the cause of the Python or release failures. Full-suite/premerge qualification remains incomplete, and #5208 remains open beyond this manual read slice.