[loopx-agent]

Goal And Delivered Outcome

Tool-selection advice could explain use cases and prerequisites while omitting the ongoing execution/scheduling budget and maintenance effort needed to choose a workflow. The common non-execution Chat prompt now asks for decision-relevant resource tradeoffs and material unknowns, without inventing prices or savings. CQ02 makes those existing cost expectations explicit.

• Outcome basis: reproduced advice completeness gap; ordinary self-contained repair.
• Intended base: main; no issue is closed by this scoped change.
• Observable before → after: missing recurring-cost guidance → explicit default instruction. The new live-model outcome remains unverified; deterministic checks validate the source change only.

Author Declaration

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

Implemented against

• Specification: community golden queries, CQ02 at f9bdae6.

Criterion	Disposition	Symbol / path	Check
Supported paths, costs and limits	implemented at guidance layer	loopx/chat_agent.py:_turn_prompt; CQ02	Diff inspection; existing Chat regression suite
Simpler path allowed; no assumed team/daemon; external services need a reason	implemented, existing behavior retained	Existing CQ02 row and Chat instructions	Diff inspection; existing Chat regression suite
New candidate actually meets CQ02 through the native model and original reply	deferred	Existing community qualification workflow	Requires candidate adoption and controlled native retest; not claimed here

• Self-check: reviewed the owning common prompt, current public criterion and exact diff. Added no case-name routing, state, permission, transport or model-setting changes. No assertion-only prompt-string tests were added.

Scope And Continuation

• Complete within this source-guidance scope: two shared prompt sentences and clarification of one existing golden row. Execution-mode instructions are unaffected.
• Remaining work: exact-head review/CI, then controlled adoption and native CQ02 qualification in the existing community test workflow. The prior cost failure remains open. Source-read availability and other community qualification gaps are outside this patch.

Validation

• Tested revision: c56f7fbccb9b9129b7ce4263b38da23dae245376.
• Run state: finished.
• Input classes: synthetic, public_fixture, authorized_private_read_only.

Check kind	Result	Public-safe evidence / limitation
unit	passed	Existing test_chat_agent, test_chat_ordinary_project, test_chat_response_parsing, test_chat_codex_context: 373 passed; same Python diff before commit
static	passed	Ruff for changed Python; configured mypy scope; changed-file public-boundary scan; diff and bilingual-RFC checks
integration	passed	Standard loopx canary premerge --from-git-diff: six direct checks, eleven selected checks, zero failures/advisories
real_entrypoint	failed	Pre-change native Chat advice omitted recurring costs despite covering prerequisites; original delivery independently read back. Private evidence retained locally
real_backend	not_run	New prompt has not been installed or invoked through the native model; unit/canary success is not a semantic fix claim

• Coverage and gaps: prompt construction, ordinary/project Chat regression paths and selected repository canaries passed. New-candidate repeated model qualification and remote CI are still required before a behavior/adoption claim.

Frontend / Visual Evidence

• UI impact: none. No layout, chrome or visual-state change; screenshots and attention/craft review are N/A.
• Source data: none.

Type of Change

• Bug fix
• Documentation update

LoopX Area

• Host or runtime integration
• Public docs or presentation surface

Technical Direction

• Operator surface and IM integration; existing community CQ02 acceptance.

Shared-authority RFC fixture impact

N/A: no shared-authority protocol, routing, persistence or promotion change.

Boundary Checklist

• Diff and description exclude private state, credentials, raw traces, verifier output, internal links and local machine paths.
• No duplicated maintainer-owned benchmark work.
• Scoped to the reproduced advice gap and existing criterion.
• No paired RFC changed; local bilingual gate passed.
• UI impact marked none.
• Every branch commit has a DCO sign-off.

[songoow]

动机

这条改动要解决的是 LoopX Chat 在“该用哪个工具/工作流”这类咨询上给出不完整取舍的问题。受影响的是向 LoopX Chat 咨询“该用哪个工具/工作流”的普通使用者：他们此前拿到的建议会讲清用途和前置条件，却可能完全不提搭建与维护投入、持续执行或调度成本，导致在自建与现成工具之间做出信息不完整的取舍。改动前，同一段工具选择请求在 Chat 中可能只得到功能与前置条件说明；改动后，非执行态的公共提示词明确要求给出决策相关的搭建/维护投入、持续执行或调度开销与限制，并在缺少证据时说明未知项而不是编造价格。可观测结果是所有复用 _turn_prompt 的 Chat 宿主（chat_agent、chat_dsh、chat_providers、chat_acp）在非执行态都会带上这段取舍要求，而执行态提示词逐字节保持不变。本 PR 不引入价格数据源、不改变任何执行态提示词、不新增状态/权限/传输或模型配置，也不声称新的原生模型回答已经通过 CQ02 验收。新提示词尚未通过原生模型针对 CQ02 做受控复测；该验收仍留在既有社区资格测试流程，由后续受控采用步骤完成。不合并的代价是：用户继续在缺少持续成本信息的情况下挑选工作流，事后才发现维护或调度开销远超预期，而这一缺口只能靠用户自己反复追问才能补齐。

改动思路

做法是在既有的公共提示词构造器 _turn_prompt 的非执行分支里补两句关于取舍的要求，而不是新增一个提示词所有者或一个成本数据服务。控制流很直接：宿主适配层把用户消息与运行标志交给 _turn_prompt，该函数按 execution_mode / project_work / direct_work 组合角色、限制与仓库读取指令，返回一整段系统提示；新加的两句只在 execution_mode 为假时拼接，因此执行态提示词保持原样。这段取舍要求属于既有提示词构造器里的本地措辞，由唯一 _turn_prompt 所有者承载并被各宿主复用；没有新增决策拥有者、状态或对外契约，因此不需要进一步简化或拆分。本 PR 的交付边界是源侧提示词与 CQ02 验收行同步；原生模型复测与采用证据作为后续步骤保留在社区资格流程，理由是该步骤需要受控环境而不是本补丁的源码范围。

所有权边界没有变化：loopx/chat_dsh.py、loopx/chat_providers.py、loopx/chat_acp.py 都是从 loopx.chat_agent 导入 _turn_prompt，所以这一处修改覆盖了这些宿主，而不需要各自复制一份措辞。替代方案是把成本口径做成结构化配置或接一个价格数据源，但那会引入本问题不需要的持久状态与外部依赖；本 PR 选择让模型在缺少证据时明确说明未知项，符合“不编造价格或节省额”的边界。

具体改动

语义上只有两处改动，合计 +3/-1 行。其一在 loopx/chat_agent.py 的 _turn_prompt（第 539-540 行）非执行分支末尾新增两句：要求在做工具/工作流建议时说明决策相关的搭建与维护投入、持续执行或调度成本与限制，并要求把取舍落在可用证据上、把重要未知项讲清楚而不是编造价格或节省额。其二在 docs/product/use-cases/community/golden-queries.md 的 CQ02 验收行中把“costs and limits”扩写为包含搭建/维护投入与持续执行/调度成本、且不得编造价格或节省额，使既有公开验收描述与新提示词一致。

关键符号映射：loopx/chat_agent.py:539 的 _turn_prompt 是本仓库唯一的 Chat 系统提示构造器，负责把角色、限制、仓库读取与工具建议要求拼成一段提示；它的 execution_mode 参数决定是否进入“已确认任务执行”分支。loopx/chat_dsh.py:36、loopx/chat_providers.py:17、loopx/chat_acp.py:22 从该模块导入，构成实际消费者链路，因此单点修改即可覆盖多个宿主。golden-queries.md:105 的 CQ02 行是公开验收描述，评审者据此判断建议是否完整。

规格依据：docs/product/use-cases/community/golden-queries.md，审查修订 f9bdae67596c185cebc81dc555dd6a3e0242e9c5。验收判据 CQ02 要求把需求映射到受支持路径、成本与限制，并能推荐更简单的路径、不假设每个任务都需要团队或常驻守护进程、提出外部服务时必须给出具体理由；本 PR 在指引层实现该判据，并把“新候选通过原生模型与原始回复满足 CQ02”的复测作为受控后续步骤。一个具体例子：用户问“我主要想跟踪开源项目、定期整理资料，偶尔修代码，哪些部分适合 LoopX”，改动前系统提示只要求讲清用途与前置条件，改动后同一请求要求同时给出维护投入与持续调度开销，并允许回答“具体价格未知”。

对主干的风险

这段改动属于系统提示措辞，风险集中在“模型是否照做”和“是否影响执行态”。我实测了两侧：非执行态普通请求的提示长度从 13419 增至 13690（恰好是这两句 271 个字符），执行态提示在基线与 head 逐字节相同（均为 8935 字符，diff 无差异），说明任务执行回合的指令没有漂移。负向链路同样成立：当 execution_mode=True 时新句子不会出现，因此“已确认任务执行”不受这条取舍要求约束；project_work 与 direct_work 仍走非执行分支，会带上该要求，这与它们都属于“非执行态建议/工作”的定位一致。

兼容性与权限方面没有变化：不新增字段、状态、CLI 参数或协议标识，不改动 protected_action、context_handoff 或任何权限判定；这段文字是建议性引导而非机器强制约束，忽略它不会改变任何门禁结果，只会让建议重新变得不完整，最终由 CQ02 这类社区资格验收暴露。残余风险是真实的语义效果尚未验证：作者明确标注原生模型复测 not_run，我的本地证据也只覆盖提示词构造与既有 Chat 回归（346 passed），因此“模型会稳定给出成本取舍”这一主张仍属未证实，需要在既有社区资格流程里做受控复测。

我的整体评价

交付判断是 justified_increment：本 PR 交付了一个可直接验证的增量（非执行态建议必须覆盖成本与未知项，执行态逐字节不变，公开验收行同步），剩余缺口是新提示词尚未经过原生模型的 CQ02 复测。没有阻塞性发现：改动规模与问题相称，没有引入价格数据源、第二套提示词所有者或新的权限面，且明确要求“不确定就说未知”，避免了编造成本这一常见失败模式。

我实际执行的验证：在受审查 head 的工作树上直接调用 _turn_prompt 对比三种分支（普通非执行态从 13419 增至 13690 且包含新句子；执行态 8935 与基线逐字节相同；project_work 同样带上新句子）；python -m pytest tests/test_chat_agent.py tests/test_chat_ordinary_project.py → 346 passed；grep 确认 chat_dsh、chat_providers、chat_acp 均从 loopx.chat_agent 导入 _turn_prompt，不存在需要同步的第二份提示词。审查 head：c56f7fbccb9b9129b7ce4263b38da23dae245376。

Reviewer: model_agent, model DeepSeek Flash, provider DeepSeek, self_reported.

English verdict: APPROVE - c56f7fb: two-line non-execution prompt clause plus the CQ02 acceptance row add cost/maintenance tradeoff guidance while the execution-mode prompt stays byte-identical; 346 chat tests pass and no second prompt owner exists. Native CQ02 model qualification remains deferred.