docs: deepen agent runtime comparison and refine long-running RFCs

[songoow]

Goal And Delivered Outcome

Long-running execution, exploration and correction already have several LoopX owners, but the RFCs did not consistently explain how their current implementations connect or what remains unqualified. This PR traces Grok Build, Codex, OpenCode, Pi and LoopX source, corrects those RFC definitions, and gives reviewers one connected recovery/adoption journey with failure cases.

中文：文档按“入口与准入 → 执行循环 → 纠偏消费 → 探索回流 → 后续工作与产物 → 独立验收 → 中断恢复”展开。此次更新重点补齐当前主干已有的 native Goal、合同验收、显式来源使用和 Explore 分页，修订五份既有 RFC，避免将相同会话、合法引用或局部 accepted 误写成整个目标闭环完成。

• Outcome basis: the documentation request and the existing S3/S4/S6/S7/S10/S11 boundaries; no invented roadmap item or issue closure.
• Before → after: generic continuation/acceptance language and stale implementation gaps become source-linked process diagrams, scoped maturity rows, owning RFC clauses and a connected negative-case matrix.
• Intended base: main, now integrated at 3300dfdd29fe6e53e1aa4a0282ca3952db6e9e8f, preserving published branch history. The detailed source audit remains pinned to 3dd3c550afecfa0c525a34e8744a3cf06774ac94; the current merge additionally preserves live cadence inheritance from main. Initial LoopX comparison was b77f63495f2619237380755178c0780ae10a0f26; external comparisons remain pinned to their inspected commits rather than being presented as latest-branch audits.
• Reader entry: English guide · 中文详解.

Source Comparison And Transfer Decisions

Each comparison follows the caller, state owner, correction-consumption boundary, return path and recovery tests. The guide supplies immutable source links and distinguishes inspected tests from executed checks.

Project	Process traced	What is reusable, and its limit
Grok Build	Work-round evaluation → candidate-only skeptic review → accumulated-gap strategy investigation; cancellation-safe interjections, compaction reseed and paused snapshot restoration	Staged intervention and explicit race cases. Evaluation adds model cost and does not establish independent Goal acceptance; asynchronous snapshots are not a multi-object transaction.
Codex	Tool/pending-input loop → persistent ThreadGoal idle continuation → separate Goal-state and Core admission; captured compaction basis and child-result mailbox	One continuation owner and current admission before work. Response completion, Turn termination and Goal status differ; a mailbox or completion-tool write does not prove useful adoption.
OpenCode	Prepare/admit durable inbox work → coalesced session drain → model/tool attempt; persistent claims and bounded restart	Separate durable work from cheap wake signals; test the final-drain/idle race. Restart is at-least-once under its server assumptions, not a replacement for LoopX effect reconciliation or distributed fencing.
Pi	Inner tool/steering loop, outer follow-up, preparation/compaction refresh, and finishTurn end/explicit-continue branches	Declare actual consumption granularity. Explicit continue adds one context-only turn when no natural request exists; steering does not cancel submitted tools, and saved context is not business acceptance.
LoopX	Trigger/context → admitted packet → Vision/ACK → original settlement → successor; Explore capture/page replay → later review; exact-Turn steering and scoped accepted-result wake	Compose existing owners. Binding, current-use eligibility, actual next work, independent outcome and whole-Goal closeout remain distinct.

The guide now also maps existing implementation PRs #6357, #6338 and #6279 to their distinct request-recovery, material-read and inert-inspection ownership. Candidate work is not described as missing main implementation or as already merged qualification. The two Stride RFC conflicts are reconciled with current main: six effective work Turns is the default, strategy-only configuration inherits the live cadence, and explicit cadence overrides remain pinned.

The refreshed implementation map adds three concrete call chains:

1. Managed ZCode native Goal: bind/start → current Core admission → journaled native set → same-target resume → monitor/status → confirmed pause/cold recovery. Same instance plus changed criteria/source is a qualification case, not proof of a stale-model bug. Connected status can pause work; a two-second post-check delay and per-execution one-hour timer do not establish an end-to-end stop bound or cumulative Goal budget. Unknown usage remains unknown.
2. Acceptance and explicit result use: configured validators → current revision/work basis → provider CAS → scoped acceptance; explicit A → B → C ancestry is rechecked without rewriting historical accepted work. An accepted selected-work contract is not full Goal closeout, and failed/unverified native completion still owes applicable Core settlement.
3. Direction and Explore: Direction M1 is a pure candidate builder without a production receipt writer/consumer; Explore omitted-result pages already replay current sources on writeback. Citation eligibility does not prove model adoption or constant-time source IO.

The guide groups recovery/fault variants around changed-basis resume, transitive invalidation, scoped completion, stop/wake races, steering during tools/compaction, lost replies, slow checks and cumulative cost, omitted counterevidence, attachment retry, periodic-review/off-mode behavior and actual next-artifact evidence. These are connected qualification targets, not a claim that every combination has passed.

Author Declaration

• Written by: model_agent — GPT-6, OpenAI.

Implemented against

The documentation request and the following existing clauses at source revision 3dd3c550afecfa0c525a34e8744a3cf06774ac94. Dispositions below concern documentation corrections, not completion of the runtime milestones.

Criterion / owning clause	Disposition	Documented correction	Evidence
Hierarchical Stride §§6.4, 7.5–7.7, 11–12	implemented	Preserve all-exit periodic review; separate cadence, admission, stop queueing and drain; distinguish execution deadline from cumulative cost; make M3 an actual result → work → next-review episode	Paired Stride RFC, source map and fault matrix
Research Exploration §§5.1, 11.5	implemented	Reuse ordinary capture and current page replay; observe attributable next artifact and independent outcome, including unavailable/inapplicable/no-change cases	Paired Research RFC; current Explore and explicit result-use callers
Agent Session §§1, 4–5, 10–11	implemented	Record managed native Goal and scoped continuation; qualify current-basis consumption after same-instance resume; retain original effect identity	Paired Session RFC; native controller/bridge sources
Manager A23 and §10	implemented	Preserve exact ingress/Turn and permitted fallback; queue, delivery, next request, adopted work and acceptance each need their own evidence	Paired Manager RFC; existing steering and return owners
Shared Goal Alignment §§3.7–3.8	implemented	Preserve selected-work scope, mandatory Goal coverage and historical receipts; distinguish current eligibility; bound Direction M1 claims	Paired Alignment RFC; acceptance and transitive result-use owners

Self-check included source comparison and independent read-only reviews of the new native, acceptance and Explore/Pi sections. One precision issue was corrected in both languages: Todo-scoped verification is partial only when current and passing without an applicable hold; failed, stale and held outcomes retain their states. This is not a published exact-head approval.

Scope And Continuation

Complete within the documentation scope: bilingual guide, five paired RFC revisions, index/roadmap reconciliation, process/ownership diagrams and qualification matrix. Normative prose is refined; executable files, tests, dependencies, runtime defaults and persisted schemas are unchanged. All changed repository paths are public documentation.

The existing RFC owners retain the next implementation work: qualify one supported native Goal or Goal Chat entry with a real provider result, current-basis recovery, adopted work/artifact, independent outcome and packaged return. Research §11.5 retains its provider/experiment ownership. Live-model behavior, all-Host coverage and matched quality/cost benefits remain unqualified; no new scheduler, adoption ledger, blanket readiness state or runtime promotion is introduced.

Related-boundary pass: replaced stale descriptions and reused existing acceptance, source-use, page replay, ingress/wake and recovery contracts. No new framework or parallel delivery plan. This documentation slice is independently reviewable and reversible.

Validation

• Tested documentation revision: c34e4f1576a69ab0fc98f9253933c1a3ea7929c2 (identical working content checked before commit).
• Detailed source audit and prior implementation tests: 3dd3c550afecfa0c525a34e8744a3cf06774ac94, run for documentation head caf7471b70e95feef861e71f04ec30f91d68558f. Current-main integration: 3300dfdd29fe6e53e1aa4a0282ca3952db6e9e8f. Fresh checks cover documentation governance/index, bilingual references/fences/privacy and conflict resolution; there is no runtime delta in this PR.
• Run state: finished. Input classes: public_fixture, synthetic.

Check kind	Result	Public-safe evidence / limitation
static	passed	python scripts/generate_rfc_status_index.py --check; lifecycle states unchanged
static	passed	python examples/docs-governance-smoke.py; 161 added local links, 27 Markdown fragments and code fences checked; both guides now use the same 51 external URLs (48 comparison references and three implementation PR links)
manual	passed	Fixed external source links checked during comparison; external tests inspected, not executed
static	passed	Prior supplementary loopx check over all 15 changed docs with an isolated empty registry: 0 errors; one expected warning for the nonexistent disposable runtime directory. Separate credential/private-link/local-path scan clean
unit / integration	passed	Prior supplementary run, not rerun after this main merge: six TS suites, 140 passed, 19 skipped, 0 failed, using qualified Node 22.22.3 / SQLite 3.51.3. Skips are unconfigured PostgreSQL cases; controlled native Host tests are not live vendor/model execution
integration / real_backend	passed	Prior supplementary run, not rerun after this main merge: python -m pytest -q tests/test_delegation_result_use.py, 7 passed. Synthetic worker tasks use production local provider/admission/validation with isolated File/SQLite state; only the test-owned runtime was stopped
manual	passed	Read-only cross-review and English/Chinese semantic check; one state-condition wording issue fixed. No public approval is inferred
static	passed	git diff --check, docs-only base/head scope, explicit-path staging and DCO sign-offs
real_entrypoint / comparative benefit	not_run	No live-model, installed multi-Host/product journey, matched-cost study or PostgreSQL qualification. The new combined fault matrix is proposed acceptance, not executed evidence

TS files under tests/control_plane_ts/: zcode_native_goal.test.ts, direction_baseline.test.ts, stride_explore_pagination.test.ts, goal_acceptance_runtime.test.ts, goal_acceptance_authority.test.ts, delegation_result_use.test.ts. The first attempt could not load pg in two suites; after installing locked dependencies, all six were rerun for the result above. No tracked dependency or implementation change was made to pass these checks. Those implementation runs remain evidence at their named source/head; they are not fresh checks of later main changes.

Documentation governance, links, source review and public/private checks cover this change. The additional existing implementation tests support the bounded source claims; their counts do not certify semantic adoption or the proposed combined journey. Hosted CI and review of this updated exact head remain separate. Earlier approval on f352d929307c9bd51ae2508f5bf11e6a8fbfe563 is not carried forward as approval of this head.

Frontend / Visual Evidence

• UI impact: none. Markdown architecture/RFC content only; no first viewport, opening navigation or documentation-chrome change.
• Before/after screenshots, visual states, attention/craft review: not applicable; proposed product readback is explicitly separate from a shipped UI change.
• Source data: none for UI evidence.

Type of Change / Area

• Documentation update
• Public docs or presentation surface
• Direction: existing S3/S4/S6/S7/S10/S11 and the five RFC owners above.
• Shared-authority fixture impact: no runtime migration, routing, provider or compatibility-projection change. This PR refines documentation and claims no shared-authority promotion; three-arm promotion rehearsal is not applicable.

Boundary Checklist

• Public docs/body contain no private source material, raw traces, secrets, internal links or local machine paths.
• Existing owners/milestones reused; no benchmark jobs or duplicate runtime implementation.
• Scope remains the requested documentation/RFC integration.
• UI impact is none.
• Every PR-owned commit includes a DCO sign-off.

Combined Integration Readback

A disposable local composition of the four coordinated PR heads passed 18 actual HTTP/source-denial tests, control-plane typecheck and docs governance. The generated registry I/O census was rebuilt from that combined source: 310 sites, zero unclassified, with --check passing. Request, Direction and Recovery routes preserve unique keys and their owning handlers. Source/test insertion conflicts were removed; when #6357 and #6338 are combined, their branch-specific generated census must be regenerated from the combined source, not resolved by accepting one side or concatenating JSON.

This is local integration evidence, not a merged branch, installed product, live-model adoption or resolution of unrelated main-branch checks. Exact heads: #6374 c34e4f1576a69ab0fc98f9253933c1a3ea7929c2; #6357 4746c41141b165c34a6de40ce1fc75a838ce2f24; #6338 082ed83c22e23d1d771da879773b33c66f31d953; #6279 f4bfab4a391e8754cb961471237bc57f4c651eed.

[mergify]

This pull request has merge conflicts with main and cannot be merged
until they are resolved. Please rebase or merge the base branch, @songoow.

Choose the remote for the base repository, not an out-of-date fork.
For a fork clone, first inspect git remote -v; upstream must point
to https://github.com/loopx-project/loopx.git. If it is absent, add it
with git remote add upstream https://github.com/loopx-project/loopx.git.
Then run:

git fetch upstream
git rebase upstream/main
# Resolve each conflict, git add the resolved files, then git rebase --continue.
git push --force-with-lease origin HEAD

For a same-repository clone whose origin points to
https://github.com/loopx-project/loopx.git, use origin instead of
upstream for fetch/rebase. If you prefer merging the base, use
git merge <base-remote>/main and push normally.

Keep the DCO Signed-off-by trailer on every commit when you rebase.
https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[huangruiteng]

LGTM

[loopx-agent]

Reviewer: model_agent; gpt-6.1-sol; OpenAI; runtime_reported; reasoning_effort=xhigh.

动机

实现或评审长程 Agent 续行、探索和纠偏的维护者与贡献者。 过去读完各份 RFC，仍可能把已有反馈投影当作待建功能，或把恢复与额外研究流程混在一起；这份比较把已实现链路、正确续行边界和下一项可验证产物连接起来。 当前交付是成对的源码比较与 RFC 澄清：普通执行继续复用既有 owner，下一阶段明确验证真实结果采用和原会话回传。 本 PR 不安装 Host、不启用策略、不修改运行时，也不宣称真实模型采用或跨项目性能排名。

这解决的是下一步做错归属或重复造机制的风险。源码中已有反馈与捕获，不应继续按过期文字重建；但已有回执也不能冒充真实采用。我的判断依据当前完整提案、变更前的路线图与实际 caller，不继承作者声明或维护者简短 LGTM。

改动思路

按准入、模型/工具执行、纠正、结果返回、恢复逐层比较，复用现有 Host、replan、Explore 与 acceptance owner。保留周期复核和结算义务，普通恢复先沿原操作处理，新工作再检查当前 Goal、criterion 和显式来源。研究状态的公平对照属于另一项资格，不阻断这条有界业务接续。

最强反面是文档较长且涉及多个 RFC；只追加链接更短，却会留下错误的下一步定义。因此这里把澄清放回现有五个 owner，并明确最小有用的 M3 结果→采用→下一次复盘场景，没有另立账本或额外模型流程。

具体改动

完整范围 6374@f352d929307c9bd51ae2508f5bf11e6a8fbfe563，immutable merge-base b77f63495f2619237380755178c0780ae10a0f26：15 个文档、+1414/-57；运行时、App、包、脚本和测试目录的树哈希与 base 全部相同。

关键内容讲解

• 新英/中文指南分别 591/477 行，区分四个公开 runtime 的处理单位、纠正消费时点与恢复假设。独立读取了八个固定版本的一手源码文件，包含 Codex 子结果不自动触发 Turn、Pi 工具批次边界、OpenCode 持久恢复与 wake 合并、Grok 策略触发及恢复状态；源码阅读不算运行测试或性能证据。
• Stride §11/M3：把“再实现反馈投影”改为复用已有链路，下一项验证真实结果、实际采用、产物与后续复盘。对应文档条款已落实；真实模型/完整成本与 M5 推广仍 deferred。
• Manager A23/§10、Session 输入边界、Alignment §3.8：明确原 Turn/ingress 身份、授权 fallback、原操作恢复和下一动作的新依据；对应澄清已落实。研究 §11.5 复用已有附件，不把研究变成普通任务的必经流程。
• 依据为变更前整体路线图及上述现有 owning clauses；当前指南与成对 RFC 是文档产物，未修改既有 wire vocabulary 或默认执行状态。
• 主规格路径 docs/architecture/rfcs/hierarchical-agent-stride-control-v0.md，版本 b77f63495f2619237380755178c0780ae10a0f26；Stride M5 的真实采用与完整成本验收保持 deferred。

对主干的风险

独立运行 docs-governance、RFC 索引检查和 diff hygiene；119 项相关 typed owner 测试通过、零跳过。全部15文档引用中的734个仓库内目标存在，代码围栏成对；两种语言的48个固定外部 URL 集合一致。这里没有把“八个源码文件已读取”写成“48项均运行验证”。作者的27 Python/35 TS/4 storage 等原分组也没有重标签为本轮结果。CI 未查询。

语义与 CI 对齐

比较了周期复核、原身份恢复、当前依据、停止与 opt-in 的关键条款。指南保留已开启周期复核所有合法出口所欠的 stride 决定记录，未把它降为建议；也没有用另一项目的 at-least-once 或候选评审来替代 LoopX 验收。新增的是待实现的资格澄清，不是机器行为已经改变。新 Host、全 Goal 覆盖、Lark、实际模型采用仍保持明确未测。

我的整体评价

APPROVE，仅文档请求的 goal_achieved。长程义务保持，读者体验的预期正向来自消除过期下一步和重复归属，让第一次有用结果与其验收更清楚；没有实测速度、token 或长期质量增益。相关未来整理已应用在既有 owner 的文字与引用中，无需为此增加运行时抽象或新的跟踪任务。运行时结果采用与实运收益继续由原 M3/M5/Host 阶段负责。

English verdict: APPROVE — f352d92. The complete documentation slice corrects stale owning RFC guidance and specifies a useful result/adoption/recovery episode without changing production trees, defaults or authority. Document/index checks and 119 referenced typed tests passed; eight pinned primary sources were independently inspected. Live adoption, installed Host journeys and comparative full-cost benefit remain unqualified.

[mergify]

This pull request has merge conflicts with main and cannot be merged
until they are resolved. Please rebase or merge the base branch, @songoow.

Choose the remote for the base repository, not an out-of-date fork.
For a fork clone, first inspect git remote -v; upstream must point
to https://github.com/loopx-project/loopx.git. If it is absent, add it
with git remote add upstream https://github.com/loopx-project/loopx.git.
Then run:

git fetch upstream
git rebase upstream/main
# Resolve each conflict, git add the resolved files, then git rebase --continue.
git push --force-with-lease origin HEAD

For a same-repository clone whose origin points to
https://github.com/loopx-project/loopx.git, use origin instead of
upstream for fetch/rebase. If you prefer merging the base, use
git merge <base-remote>/main and push normally.

Keep the DCO Signed-off-by trailer on every commit when you rebase.
https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork