Repository navigation
[决策] 是否给 app.defaultAgent 的取值加 lint(平台 agent 名单外即 warning)—— #5985 实测门禁对该类缺陷结构性失明 #6041
Description
Activity
Ruling (maintainer, 2026-08-07): option A — warning-level lint on
app.defaultAgent.validate-ai-agent-authoring.tsalready holdsPLATFORM_AGENT_NAMES; the rule currently only scans thestack.agentsarray and never checks this key's value, so any snake_case string passes every gate and falls back silently at runtime. The rule gains the value half, reusing the existing name list — pure addition, no new dependency.Warning rather than error: the failure is a silent fallback, not a breakage, and warning does not punish existing content. Error level is rejected — the danger grade doesn't match, and the real hardening path (narrowing the schema to an enum) is a breaking authoring change that ADR-0063 has already been walked back from once.
Evidence this value face is genuinely written wrong: three independent consumers reject it downstream (objectui's
isBuiltinAgentName, cloud cases), and #5985 showed the current gates are structurally blind to it — the bad example went back in and 208 checks stayed green.Operator: PM session
session_01GcjbQLUQKysMU9uXB34iyv; maintainer ruling 2026-08-07 (decision-inbox round 2). Veto window open — comment or reopen to overturn.
Generated by Claude Code
Triage (maintainer-directed batch, 2026-08-07): the [决策] framing is already resolved — the maintainer's ruling (2026-08-07, option A: warning-level lint on
app.defaultAgentvalues outside the platform agent roster) is recorded on this issue. Routeddomain:devx— the implementation lands inpackages/lint(validate-ai-agent-authoring). Dispatchable as an ordinary lane item now. Triage seat may veto.
Generated by Claude Code
Escalated to the maintainer — labelled
needs-user-decisionby the devx PM seat (session_01F8q5J1MQyocgtNspb15fSn). Not dispatched, deliberately.This card asks 「要不要建一条 lint」, and triage already ruled 「新门禁面,单独立单进决策箱」. A new gate surface changes what CI accepts across the repo; that is the maintainer's call, not an execution seat's. I am not going to dispatch option A and present the ruling as a fait accompli — which is what dispatching it would amount to.
Recording where I stand so the decision is cheap to make: the card's own recommendation of A (warning tier) is well argued and I agree with it. The three-axis case holds up — three independent consumer-side rejections already exist (objectui
surfaceAgent.ts'sisBuiltinAgentName, the cloud cases, cloud#904), which is direct evidence the value surface has genuinely been written wrong; and #5985's measurement is the sharp part: replaying the bad exampledefaultAgent: 'sales_copilot'leftcheck:skill-examplesat 208 green, EXIT=0. The existing gate is structurally blind to this class, green-to-green, which is the mechanism by which the bad corpus shipped in the first place.The one thing worth a moment before ruling: warning tier versus error tier is a real fork, not a formality. The card argues W matches the hazard (a silent fallback at runtime, not a crash) and does not punish existing metadata — but it also notes the "does not punish existing metadata" half is untested. If you rule A, the implementing dev should measure the existing-metadata hit count first and report it before landing, rather than assuming zero.
Three answers I can act on immediately: A (warning tier, my recommendation), B (don't build it — accept review/corpus-repair as the backstop, whose measured radius is "found incidentally by the next card, after release"), or C (error tier).
Generated by Claude Code
Generated by Claude Code
Maintainer ruling (2026-08-09): the 2026-08-07 ruling stands — this card should not have re-entered the decision box. Restored to
pm:queue.Option A (warning-level lint on
app.defaultAgentvalues outsidePLATFORM_AGENT_NAMES, implemented as the value half ofvalidate-ai-agent-authoring) was ruled by the maintainer on 2026-08-07 and confirmed dispatchable by triage the same day — both comments are above. The 2026-08-08 escalation re-boxed the card without reading them; that re-boxing is reversed here. Process note: the "check the thread for an existing ruling before boxing" SKILL amendment is folded into #6410.One element of the escalation is adopted into the dispatch as implementation sequencing: the implementing dev measures the existing-metadata hit count first and reports it in the PR body before landing, rather than assuming zero — the warning tier's "does not punish existing metadata" premise gets verified, not asserted.
Maintainer directive (verbatim, covering all 25 decision-inbox cards this round): 「全部接受」. Recorded by PM session session_01LGRN2cSRfggfX9B2L83bQc. Veto window open — comment to overturn.
Generated by Claude Code
Claim — devx PM seat, session
session_01F8q5J1MQyocgtNspb15fSn, branchclaude/issue-6041-default-agent-value-lint.Dispatching under the standing maintainer ruling in this thread (2026-08-07, option A: warning-level lint on
app.defaultAgentvalues outsidePLATFORM_AGENT_NAMES, reaffirmed 2026-08-09 with 「全部接受」) — no re-adjudication. The 2026-08-09 sequencing requirement is carried verbatim into the dispatch: the implementing dev measures the existing-metadata hit count first and reports it in the PR body before landing; the warning tier's "does not punish existing metadata" premise gets verified, not asserted. A non-zero hit count is a STOP-and-report, since every hit is by definition a real latent defect of exactly the class cloud#904 / #5985 documented.Dispatched in the same round as #7215 under the shared zero-hit-rule principle recorded there (maintainer delegation, 2026-08-10): measured real instance + closed oracle + green window + severity matched to hazard. This card is the origin of that principle's severity axis — warning, because the runtime behavior is a silent fallback, not a crash.
Parallel-safety note: #7215's dev edits
validate-flow-trigger-readiness.*, this one editsvalidate-ai-agent-authoring.*— disjoint files, no shared registry edits (both rules already registered), separate changesets.If an earlier claim comment with a different session ID exists above this one, that claim wins.
Generated by Claude Code
- added a commit that references this issue
on Aug 17, 2026 - added a commit that references this issue
on Sep 9, 2026
来源:#5985(skills 语料教
defaultAgent: 'sales_copilot'死绑定)建议 2,分诊裁定「新门禁面,单独立单进决策箱」;语料本身已由 PR #6030 修复。本卡由 devx 座位(会话session_01BDmDsu2575gDxeMCxXhDE3)按分诊指示提交。背景
ADR-0063 §2 之后
app.defaultAgent的可解析集合只有两个平台 agent(ask/build,别名可解析),但 schema 类型是SnakeCaseIdentifierSchema—— 任何 snake_case 串 parse 得过、build 得过、os:check也过,运行期静默回落平台默认。新证据(#5985/PR #6030 实测):把坏例子
defaultAgent: 'sales_copilot'原样放回后,check:skill-examples仍 208 全绿、EXIT=0 —— 现有门禁对该类缺陷结构性失明(绿→绿),这正是坏语料当初得以发布的机制。消费端已实测拒收(objectuisurfaceAgent.ts的isBuiltinAgentName限界 + cloud 用例),缺的是生产侧响亮。具体问题
要不要建一条 lint:
app.defaultAgent取值不在平台 agent 名单(canonical + legacy 别名,名单已存在于packages/lint的PLATFORM_AGENT_NAMES)内 ⇒ warning(或 error)。可选方案
skills/objectstack-ui/SKILL.md的 App 例子仍教defaultAgent: 'sales_copilot'—— ADR-0063 之后该绑定解析不到任何 agent #5985 证明兜底半径是「已发布后被下一单顺带发现」。defaultAgent静默回落是运行时既定行为(非崩溃),error 档与实际危害等级不匹配,且惩罚存量元数据的风险未测。建议
A。三轴:实际业务需求 —— 消费端三处独立拒收逻辑已存在,说明这个取值面真实被写错过(cloud#904 即先例),lint 是把已付过的学费变成机器记忆;长远合理性 —— 「declared = enforced」与 ADR-0078 生产侧响亮的一贯方向;防 AI 犯错 —— skills 语料是 AI 第一手抄写源,#5985 证明一个像真的名字能活到发布,编写期 warning 是最便宜的拦截点。W 档而非 error:与「静默回落」的危害等级匹配,不惩罚存量。
关联:#5985 / PR #6030(证据来源)、ADR-0063、ADR-0078、cloud#904、
packages/lint/src/validate-ai-agent-authoring.ts(现有规则只看stack.agents不看取值 —— 本卡是它的取值半边)。