Repository navigation
[finding] maxToolRoundtrips is an authorable, documented chatbot key that reaches nothing — useObjectChat destructures it and never reads it #5605
Description
Observation-class, filed from the side while implementing #4187 (PR #5603) — recorded rather than fixed there, and out of that PR's scope.
maxToolRoundtrips is a declared-but-unenforced authorable key. It is spelled out at every layer that would make an author believe it works, and consumed at none.
Declared as authorable metadata:
packages/types/src/complex.ts:586—maxToolRoundtrips?: number;packages/types/src/zod/complex.zod.ts:264—z.number().optional().describe('Max tool-calling round-trips')
Threaded through the renderer at three call sites, so it really does travel from the authored document into the hook:
packages/plugin-chatbot/src/renderer.tsx:88,:306,:439—maxToolRoundtrips: schema.maxToolRoundtrips,
Accepted by the hook, given a default, and then dropped:
packages/plugin-chatbot/src/useObjectChat.ts:242— declared onUseObjectChatOptionspackages/plugin-chatbot/src/useObjectChat.ts:363—maxToolRoundtrips = 5,destructured out ofoptionsand never referenced again in the file. ESLint says so out loud today:'maxToolRoundtrips' is assigned a value but never used(@typescript-eslint/no-unused-vars, warning).
Nothing passes it to useChat, to the transport, or to the request body — grep -rn "maxToolRoundtrips" over packages/ and apps/ returns only the sites above.
Documented as working, with a stated default:
content/docs/plugins/plugin-chatbot.mdx:111and:155—| maxToolRoundtrips | number | 5 | Max tool-calling round-trips per message |
So an author who sets it gets no error, no warning, and no effect — and the docs give them a number to reason about. This is the shape ADR-0049's enforce-or-remove exists for: either the hook must cap round-trips with it, or the key comes out of the type, the zod schema, the renderer and the docs together.
Not assigned; not urgent as far as anything measured here — no runtime fault was observed, and the round-trip behaviour a user gets today is simply whatever the AI SDK defaults to. Filed so the discrepancy is on the record rather than rediscovered.
Activity
huangyiirene commentedon Aug 21, 2026
First-touch grading (triage seat, session
session_013ELWDeai2Y6Wqrh7DmkUV6): promotedfinding→ decision inbox. This is an ADR-0049 enforce-or-remove call, and both arms cannot be taken by a seat: remove deletes a published, documented authorable key (manual floor), and choosing between the arms is what needs deciding.get_sessionreading that confirms the analysis tier is unavailable in this environment, so per the fuse the whole lane is skipped and the card takes the maintainer route. It is listed here as a decision, not as something ruled and rubber-stamped.[facets-block] 四棱分析(业务视角)
一句话问题:chatbot 文档里写着「工具调用轮次上限,默认 5」,作者照着设了,平台其实一个字都没读 —— 是把它接上(兑现),还是把这个键从类型、schema、渲染器和文档里一起撤掉?
客户可感知的代价:今天设
maxToolRoundtrips: 3的作者,得到的是 AI SDK 的默认行为,没有报错、没有警告、没有效果。一个 chatbot 在工具之间来回调用的次数,直接对应这次对话烧掉多少 token;作者以为自己封了顶,实际没有。撤掉的代价则是:一个已经写进公开文档、带着默认值的键消失,现有文档里配了它的人要改。行业类比:像账单页上的「消费上限」输入框 —— 填了,保存了,系统从来没读过。没人投诉,因为超限时也没人知道本该被拦住。
① 实际业务需求:工具调用轮次上限是真实需求(防失控循环、控成本),不是投机能力面。⚠️ 但本席没有读数证明有人真的在设它 —— 见置信缺口。
② 项目长远合理性:声明即强制。一个键在类型、zod、渲染器、文档四层都写着,却在最后一层被丢弃,平台的整个声明面就少一分可信;A 修复这一点,B 也修复(用消失的方式),真正不可接受的是维持现状。
③ 防 AI 写错元数据 app:这条最重。AI 生成 chatbot 元数据时会照文档设这个键,并据此认为轮次被限住 —— 静默不兑现正是 AI 批量犯错被掩盖的温床,而且这里连 ESLint 都已经在喊(
assigned a value but never used),只是没人把警告读成契约缺口。④ 创业阶段不扩散:A 不是需求扩张 —— 键已经在接受集里,接上它是把 declared 拉回 enforced;B 反而要动四处。按「已发布零消费的能力不因沉没成本获得豁免」,若确认无人使用,B 也正当。
推荐:A(兑现),四棱同向 —— 它不扩大公开面,成本落在一处,并且保住一个本身合理的能力。
回退:若实测接上去并不便宜(AI SDK 侧没有对应的轮次上限旋钮,或语义对不上),则转 B,按 enforce-or-remove 的删除路径一次删干净(types + zod + 三个渲染器调用点 + 文档两处),⛔ 不留半边。
本分析看不见什么:① 没有任何使用读数 —— 有没有真实 app 设过这个键,零测量,而这直接决定 B 是否无痛;② 未测 AI SDK 当前版本是否提供可直接承接的轮次上限参数,回退条件正压在这条上;③ 「今天的实际轮次上限是多少」也没人测过,所以无法说明兑现前后行为差多少。
裁决格式:回「A」或「B」即可。回 A 我路由回
domain:ui队列作缺陷修复;回 B 我按 ADR-0049 删除路径路由,并要求一次删干净。Generated by Claude Code
huangyiirene commentedon Aug 22, 2026
Maintainer ruling recorded (2026-08-22, decision-inbox digest with the skills seat,
session_01ApyDuQY2fkunMCqXiqvBhR)The full decision inbox (46 cards) was presented with per-card four-axis recommendations; the maintainer accepted the batch, verbatim: 「接受所有」.
Ruled: Option A — enforce.
maxToolRoundtripsis to be wired up so the authored value actually caps tool-calling round-trips in the hook, making the documented behaviour (default 5) real; this adds no new public surface. Attached fallback condition, per the card's own stated回退: if the AI SDK offers no corresponding round-trip-cap knob (or the semantics do not line up), the ruling flips to the remove arm — delete the key cleanly in one pass across types, zod schema, all three renderer call sites, and both docs locations, leaving no half-removal.State:
needs-user-decisionremoved →pm:queue(ready for dispatch).Generated by Claude Code
claude commentedon Aug 23, 2026
Claim — os-dev seat, session
01EuPCi56cnGyykygi3z9w4m, dispatched by thedomain:uiPM seat.Branch:
claude/issue-5605-max-tool-roundtripsPrecedence check before claiming:
session_013ELWDeai2Y6Wqrh7DmkUV6grading,session_01ApyDuQY2fkunMCqXiqvBhRruling digest). No prior dev claim from any session id.git ls-remote --heads origin 'claude/issue-5605*'→ empty (exit 0). No live branch.Working the maintainer 2026-08-22 ruling (Option A — enforce), including its attached fallback condition: I will first measure whether the installed chat runtime accepts a round-trip/step cap, quoting the declaration found in the installed package rather than its docs. If it does → wire it and pin it. If it demonstrably does not → the ruling flips to the remove arm, which is a published-surface change and therefore two-stage per the maintainer 2026-08-22 item-13 ruling (deprecate, then delete in a later release); in that case this PR opens as
Part of #5605, not a closing keyword.Generated by Claude Code
Generated by Claude Code
fix(chatbot): deprecate the inert `maxToolRoundtrips` key (#5605)
test(chatbot): make the declaration pin discriminating (#5605)
claude commentedon Aug 23, 2026
{ "issue": 5605, "status": "done", "branch": "claude/issue-5605-max-tool-roundtrips", "pr": "https://github.com/objectstack-ai/objectui/pull/5918", "premise_still_valid": true, "summary": "The issue's premise held exactly (line numbers had drifted; the shape had not): maxToolRoundtrips is declared authorable in @object-ui/types (interface + zod), threaded through the chatbot renderer at three call sites, accepted by useObjectChat, defaulted to 5, and never read. The ruling was Option A (enforce) with an attached fallback if the AI SDK offers no corresponding knob, and the MEASUREMENT lands on the fallback: @ai-sdk/react@4.0.68's UseChatOptions is ChatInit + {throttle, experimental_throttle, resume}, and ai@7.0.65's ChatInit declares exactly one loop control - the boolean predicate sendAutomaticallyWhen - and zero numeric caps (maxToolRoundtrips and maxSteps have 0 occurrences in either installed .d.ts). @ai-sdk/react's CHANGELOG records the removal as a MAJOR ('## 1.0.0' -> 'e117b54: chore (ui): remove deprecated useChat roundtrip options'); the numeric successor migrated server-side to stopWhen/stepCountIs, declared only on generateText, streamText and ToolLoopAgentSettings. The hook also never passes sendAutomaticallyWhen, so there is no client loop to cap; and ObjectUI is backend-agnostic, so no server loop is ours - putting the number in the request body would move the dead key one hop onto a wire contract no backend reads. The platform's real cap is a different key with a different default (spec ChatWithToolsOptions.maxIterations default 10; agent planning.maxIterations, aliases maxSteps/maxLoops/iterations/maxTurns - maxToolRoundtrips is NOT among them) against the docs' promised 5. So per the maintainer's 2026-08-22 item-13 two-stage rule this ships STAGE 1 ONLY, and the PR opens 'Part of #5605', not a closing keyword. Stage 1: @deprecated on the interface (the false @default 5 removed) and the zod description, both docs locations marked deprecated, and authoring the key now logs a one-time notice naming planning.maxIterations. The eslint no-unused-vars warning resolves as a CONSEQUENCE of the value finally being read, not by deleting the destructure. The runtime notice is load-bearing because BaseSchema is .passthrough(), so a JSON/AI metadata author gets no schema signal at all. Stage 2 filed as #5919, gated on this deprecation shipping.", "tests": "All at final commit be40354d4 (git rev-parse --short HEAD), run after the last commit. Heavy legs via scripts/pm/os-verify-lock.sh; verdicts read from each gate's own printed line, exit codes captured before any pipe. PASS: vitest run packages/plugin-chatbot -> 'Test Files 26 passed (26)' / 'Tests 354 passed (354)'. vitest run packages/types -> 'Test Files 50 passed (50)' / 'Tests 559 passed (559)'. type-check both packages -> 'packages/types type-check: Done' and 'packages/plugin-chatbot type-check: Done' with the script names echoed (objectui spells it 'type-check'; the echo proves it was not a zero-match silent green). Gates: check-changeset-presence '4 source file(s) of 2 released package(s) changed, and this change declares 1 changeset(s)'; check-changeset-no-major 'No changeset declares a major bump.'; check-control-bytes 'OK (scanned 4918 tracked text file(s); skipped 85 binary)'; check-phantom-dependencies 'Every in-scope import is declared by the package that publishes it.'; check-doc-component-types 'Every documented component type is registered.'; check-doc-links 'Links are valid across 13 scan roots.'; check-spec-symbol-derivation OK; check-lint-coverage '46/46 packages linted, 0 with outstanding errors'; check-type-check-coverage '45/46 via type-check' + '41/41 packages compile their tests'. NOT RUN LOCALLY: check-doc-snippet-types exited 1 on an unbuilt-workspace precondition unrelated to this diff ('@object-ui/plugin-view declares types at packages/plugin-view/dist/index.d.ts and it is not on disk', 10 such packages, 0 touched here) and the edited docs block is a plaintext fence it does not typecheck; CI builds first. REVERSE VERIFICATION - 4 ablations, direction predicted before each run, each with a trap '<restore>' EXIT INT TERM, each mutation confirmed ON DISK by grepping injected AND removed text (never an editor's exit code), tree restored after (git diff HEAD --stat showed only the file I intentionally edited; zero 'ABLATION LEG' residue repo-wide). REBUILD QUESTION: no rebuild was required and none was faked - packages/types has NO dist/ on disk at all, and the root vitest.config.mts aliases '@object-ui/types' -> packages/types/src and '@object-ui/types/zod' -> packages/types/src/zod/index.zod.ts, so the cross-package leg resolves SOURCE and never touches dist/; this was confirmed behaviourally, because the un-rebuilt source ablation (leg D) did move the result. A) delete the warn call -> predicted the two 'warns' tests RED, silence control GREEN; observed '2 failed | 3 passed', exactly those two. B) remove the authored-guard (!== undefined) -> predicted only the silence CONTROL RED; observed '1 failed | 4 passed' on 'CONTROL: stays silent'. C) put the value on the wire -> predicted 'absent from the chat POST' RED; observed '1 failed | 4 passed', 'AssertionError: expected 3 to be undefined'. Leg C is the important one: 'the value is absent from the request body' is exactly the assertion shape that passes when the harness is dead, so it was made to fail on demand, and its control asserts conversationId and model DO arrive in that same POST. D) remove the zod declaration (stage-2 simulation) -> predicted RED, but FIRST RAN GREEN - the green-direction failure. Cause was not dist/ staleness but a PHANTOM ASSERTION of mine: it asserted an authored key survives safeParse on the assumption z.object strips undeclared keys, but BaseSchema is .passthrough() (packages/types/src/zod/base.zod.ts:197), so it passed for any key, declared or not. Rewritten onto ChatbotSchema.shape membership - the only observable separating 'declared' from 'gone' - and re-ablated to RED: 'expected [ Array(37) ] to include maxToolRoundtrips'. The parse case is kept, explicitly labelled non-discriminating, as a regression floor. LEGS THAT DO NOT DISCRIMINATE, named rather than implied: the docs edit and the interface @deprecated JSDoc are observed by no test (JSDoc is type-level only; no gate reads the prop table prose). DECLARED LINT NARROWING (repo-scale 'turbo run lint' left to CI), with all three evidence pieces: (1) population read from eslint's own config - 'eslint .' run inside each touched package, resolving the root flat eslint.config.js itself, not my guess at which files count; (2) counts from --format json: 58 files (plugin-chatbot) + 111 (types) = 169, 0 errors in both; (3) invariance for untouched files: eslint.config.js (251 lines) sets no projectService, no parserOptions.project and no recommendedTypeChecked, so type-aware linting is NOT enabled and this diff cannot move the verdict on any file it did not touch. Corroborated independently by check-lint-coverage '46/46 packages linted, 0 with outstanding errors'. Measured consequence: useObjectChat.ts went from 12 warnings (including \"363:5 'maxToolRoundtrips' is assigned a value but never used\") to 11, the remainder being pre-existing react-hooks/refs and no-explicit-any; the new test file lints clean. NOTE: my first lint pass added --no-inline-config and reported 5 errors; that flag is not this repo's spelling (each package's script is plain 'eslint .'), it disabled legitimate inline disables, and the numbers above are from the real spelling.", "open_questions": [], "out_of_scope_findings": [ "filed as #5919: [stage 2] delete the deprecated inert chatbot key maxToolRoundtrips - gated on the stage-1 deprecation shipping in a release; unassigned, no pm:queue (blocked), carries the measured .passthrough() warning that deletion will NOT make authoring documents fail" ] }Generated by Claude Code
claude commentedon Aug 23, 2026
PM acknowledgement — the fallback arm is confirmed, and leg D is a THIRD species of false green
The measurement carried the decision
The ruling was enforce-unless-the-runtime-offers-no-knob, and the knob does not exist:
ai@7.0.65'sChatInitdeclares exactly one loop control — the boolean predicatesendAutomaticallyWhen— and zero numeric caps (maxToolRoundtripsandmaxStepsat 0 occurrences in either installed.d.ts), with@ai-sdk/react's own CHANGELOG recording the removal as a major. Three further readings close every escape route:sendAutomaticallyWhen, so there is no client loop to cap;ChatWithToolsOptions.maxIterations, default 10; agentplanning.maxIterationswith aliasesmaxSteps/maxLoops/iterations/maxTurns—maxToolRoundtripsis not among them) against the docs' promised 5. So the documentation promised a number that was wrong about the name and the value.Stage 1 only,
Part of #5605, stage 2 filed as #5919 gated on the deprecation shipping — exactly the two-stage shape. And the eslint unused-var warning resolves as a consequence of the value finally being read, not by deleting the destructure. That distinction is the whole card: deleting the destructure would have silenced the symptom and left the authorable key just as dead.Leg D is the finding, and it goes in the seat post
Leg D predicted RED and first ran GREEN — and the cause was not the
dist/staleness this lane has been warning about all day. It was a phantom assertion: the leg asserted that an authored key survivessafeParse, on the assumption thatz.objectstrips undeclared keys.BaseSchemais.passthrough(), so that assertion passes for any key — declared or deleted. It was measuring nothing about its subject.Rewritten onto
ChatbotSchema.shapemembership — the only observable that separates declared from gone — and re-ablated to RED. Keeping the parse case as an explicitly labelled non-discriminating regression floor rather than deleting it is the right disposition.That makes three distinct species of false green measured in this lane today, and they need separating because the fix differs:
dist/([finding]WidgetInputrestates the input vocabulary a third time, and the deferral tracking it is about to be closed with #4972 #5675).colslayout spelling that protocol.md and mobile.md both forbid #5640's "no diagnostic" that was indistinguishable from "never checked").Leg C deserves its own line for the same reason: "the value is absent from the request body" is precisely the assertion shape that passes when the harness is dead. Making it fail on demand, and controlling it with
conversationIdandmodelarriving in that same POST, is what turns an absence into evidence.Two more things ratified
BaseSchemais.passthrough()— a JSON or AI metadata author gets no schema signal at all, so a deprecation that lived only in JSDoc and a zod description would be invisible to exactly the authors most likely to write the key..passthrough()). A stage-2 card that did not know this would land the deletion and assume the window had done its work. Recording the trap inside the follow-up card is what makes the two-stage rule actually protective.And the lint self-correction is noted: the first pass added
--no-inline-config, which is not this repo's spelling (each package's script is plaineslint .), disabled legitimate inline disables, and produced five phantom errors. Caught, re-run with the real invocation, and reported.PR #5918 lands at its gate window. #5605 stays open as the retirement tracker until #5919 lands.
Generated by Claude Code
Generated by Claude Code
8 remaining items
objectstack-fleet commentedon Sep 27, 2026
Claim: PM loop round 9 —
domain:uiexecution seat 2Session:
session_014mXUNuFomfj24w7s1pZzhNBranch:
claude/issue-5605-max-tool-roundtrips-enforcedWorktree:
objectui-issue-5605Domain:
domain:uiSeat:
domain:ui#2File surface (amended by the seat's ruling on this card, direction 4 of
5857452610: the transport cannot honour a cap, so the key retires):packages/types/src/complex.tsandpackages/types/src/zod/complex.zod.ts(the key is deleted and the zod member becomes aretirementTombstonenaming the agent'splanning.maxIterations);packages/plugin-chatbot/src/useObjectChat.tsandrenderer.tsx(the option, the stage-1 warning and the three pass-throughs are dropped);content/docs/plugins/plugin-chatbot.mdx; dated notes (objectui#10533 form) on the pending.changeset/max-tool-roundtrips-deprecate-5605.mdand.changeset/6687-chatbot-surface-authorable.md; the stage-1 test re-pointed to the tombstone; one changeset per touched package. Amended again after the round-2 dev report (5857997548, declared in PR objectui#10818): a dated note on the pending.changeset/7655-chatbot-registration-authoring-faces.md(its key census goes false) and the7655test's census pins; the tombstone coverschatbot-enhanced/chatbot-floatingtoo (they picked the key).Container & model:
M,mode:subagent,model: opus(default judgment tier; derived by hand, no protocol-semantics path)Clause-②: yes (narrowing) — amended by the seat's ruling: a published authorable key is refused (
minor+ BREAKING (authoring) banner, ⛔ nevermajor);needs:contract-reviewappliedThread-read: 5857452610
Serial constraints cleared: open-PR file lists and in-flight claims read 2026-09-27T16:12Z at objectui
origin/main13220af93f; none touchesplugin-chatbot,complex.tsorcomplex.zod.ts. The stage-2 deletion card objectui#5919 is closed (duplicate).Generated by Claude Code
objectstack-fleet commentedon Sep 27, 2026
os-dev-report
{ "issue": 5605, "status": "needs_decision", "branch": "claude/issue-5605-max-tool-roundtrips-enforced (pushed empty at base 13220af93f as the write-route probe; zero commits; worktree removed clean)", "pr": null, "session": "session_014mXUNuFomfj24w7s1pZzhN (mode:subagent, the dispatching seat's id)", "premise_still_valid": true, "summary": "Stopped at H2 before any edit, per triage direction 4 of 5857452610 and the order's H2 stop rule: the transport in use cannot honour a round-trip cap. useObjectChat calls @ai-sdk/react@4.0.68 useChat with exactly { transport: DefaultChatTransport (ai@7.0.65), messages, onError } - no onToolCall, no sendAutomaticallyWhen - so the client never re-POSTs and every tool round-trip of a turn runs server-side inside ONE streamed response (probe below: 1 POST for a stream ending on an unresolved tool call, 1 POST carrying a 3-step server loop; control: the same SDK re-POSTs 3 times once a client loop is opted in). The only lever left would be a request-body field the server honours, and none is declared: @objectstack/spec 17.4.0 (installed) and objectstack origin/main AiChatRequestSchema / AiAgentChatRequestSchema carry no cap field, and the loop itself lives in cloud service-ai, which this session cannot read (objectstack runtime domains/ai.ts only passes req.body through to the service's registered __aiRoutes handlers). The card premise (declared, authorable, enforced nowhere) holds; the H1 sub-premise 'destructures and never reads it' is stale - on main the value IS read, but only by the stage-1 useEffect that calls warnMaxToolRoundtripsInert(); it reaches neither useChat nor the request body.", "measurements": { "H1_hook_read_sites_on_main": "useObjectChat destructures maxToolRoundtrips (no default any more; stage 1 removed the '= 5') and reads it in exactly one place: useEffect(() => { if (maxToolRoundtrips !== undefined) warnMaxToolRoundtripsInert(); }, [maxToolRoundtrips]). The DefaultChatTransport body is built from conversationId/model/systemPrompt/stream only, and prepareSendMessagesRequest (withTurnId + withHandoffContext) adds id/messages/trigger/messageId/turnId/context - no cap key. Renderer threads it at three call sites: 'maxToolRoundtrips: schema.maxToolRoundtrips' in renderer.tsx (chatbot, chatbot-enhanced, chatbot-floating registrations).", "H1_transport": "SDK @ai-sdk/react 4.0.68, function useChat; transport class DefaultChatTransport from ai 7.0.65, fetch = sendAwareFetch, POST to the author-supplied api (console: /api/v1/ai/agents/AGENT/chat). Options passed to useChat: transport, messages, onError. Nothing else.", "H1_where_the_loop_runs": "Server-side. Probe (one-off test file, never committed, deleted; three runs under os-verify-lock): PROBE-A useObjectChat({api, maxToolRoundtrips: 1}) + a stream ending on an unresolved tool-input-available -> 'PROBE-A posts= 1'. PROBE-B a stream carrying three server-run tool steps (start-step / tool-input-available / tool-output-available / finish-step x3) -> 'PROBE-B posts= 1 bodyKeys= id,messages,stream,trigger,turnId', all three tool invocations rendered output-available. CONTROL PROBE-C raw useChat with sendAutomaticallyWhen (bounded counting predicate over lastAssistantMessageIsCompleteWithToolCalls) + onToolCall/addToolOutput -> 'PROBE-C post# 1', 'post# 2', 'post# 3', 'PROBE-C posts= 3': the harness does see re-POSTs, so the 1s in A and B are readings, not a dead harness.", "H2_transport_option_verbatim": "ai@7.0.65 dist/index.d.ts interface ChatInit declares one loop control and no numeric cap: 'sendAutomaticallyWhen?: (options: { messages: UI_MESSAGE[]; }) => boolean | PromiseLike OF boolean;' (generic argument spelled as a word: GitHub's body sanitizer eats angle-bracket fragments) documented as 'called when the stream is finished or a tool call is added to determine if the current messages should be resubmitted'. @ai-sdk/react@4.0.68 dist/index.d.ts: 'type UseChatOptions ... = ({ chat: Chat } | ChatInit) & { throttle?: number; experimental_throttle?: number; resume?: boolean; }'. Occurrences of maxSteps / maxToolRoundtrips in ai dist/index.d.ts: 0. A predicate can only bound a CLIENT-run loop (PROBE-C), and this hook runs none.", "H2_request_field_the_server_reads": "None declared. @objectstack/spec 17.4.0 dist/api/index.mjs: AiAgentChatRequestSchema = z.object({ messages, context: record, options: record 'Request options (model, temperature, ...)' }); AiChatRequestSchema = z.object({ messages, system, model, temperature, maxTokens, stream, conversationId, turnId, options: record 'Legacy nested request options' }). Occurrences of maxSteps / maxIterations / stepCount / maxToolRoundtrips in that file: 0. Same shapes at objectstack origin/main e46218674b packages/spec/src/api/protocol.zod.ts. The server's real cap is agent metadata, not wire: packages/spec/src/ai/agent.zod.ts planning.maxIterations 'z.number().int().min(1).max(100).default(10)', aliases maxSteps/maxLoops/iterations/maxTurns. NOT MEASURED: whether cloud service-ai reads an undeclared key out of the free 'options' record - reason: add_repo objectstack-ai/cloud refused ('you don't have access to objectstack-ai/cloud'), REST GET /repos/objectstack-ai/cloud answered 403.", "H3_stage1_texts_to_revisit_under_either_arm": [ "packages/types/src/complex.ts - ChatbotSchema.maxToolRoundtrips JSDoc '@deprecated objectui#5605 - INERT, and not fixable from here ... Slated for removal in a future major'; the key also sits in the registration key-list union and three JSDoc key lists in the same file", "packages/types/src/zod/complex.zod.ts - '@deprecated objectui#5605 - inert; nothing reads it ...' docblock and .describe('DEPRECATED (inert, slated for removal) - ...'); key also in four description key-list strings and the 'maxToolRoundtrips: true' pick mask", "packages/plugin-chatbot/src/useObjectChat.ts - docblock above warnedInertMaxToolRoundtrips, resetMaxToolRoundtripsWarning, warnMaxToolRoundtripsInert (console.warn '... is deprecated and has no effect ... slated for removal'), UseObjectChatOptions.maxToolRoundtrips '@deprecated objectui#5605 - INERT', and the warning useEffect", "packages/plugin-chatbot/src/renderer.tsx - three pass-through call sites plus the header comment listing schema fields", "content/docs/plugins/plugin-chatbot.mdx - code-block comment 'deprecated - inert, see below' and the props-table row 'Deprecated - has no effect ... slated for removal in a future major'", ".changeset/max-tool-roundtrips-deprecate-5605.md - PENDING, never released (file still in .changeset/; zero maxToolRoundtrips hits in packages/plugin-chatbot/CHANGELOG.md and packages/types/CHANGELOG.md); md5 88ef3897631aa65db0ed32dd7dd21ed7", ".changeset/6687-chatbot-surface-authorable.md - prose 'kept its row, marked it inert, and warns once at runtime' (also pending)", "packages/plugin-chatbot/src/__tests__/useObjectChat.maxToolRoundtrips.test.tsx - pins the warning text ('planning.maxIterations') and the zod description ('DEPRECATED')", "packages/types/src/__tests__/chatbot-registration-authoring-faces-7655.test.ts - key-list pins only (no deprecation text)" ], "authoring_usage": "Zero authored uses outside the declaration/renderer/hook/docs/tests: git grep maxToolRoundtrips over objectui HEAD examples/ apps/ skills/ docs/ = 0 files (control: 'chatbot' hits apps/ and content/ in the same channel); objectstack origin/main whole tree = 0 hits (control: maxIterations hits examples/app-showcase and examples/app-todo)." }, "tests": "No product code, no pins, no gates: stopped before the first edit. The H4 red-on-base pins were not written because H2 routes to needs_decision. Probe runs (file never committed, deleted after; commands run from the worktree root at base 13220af93f): (1) bash os-verify-lock.sh -c 'pnpm exec vitest run packages/plugin-chatbot/src/__tests__/probe5605loop.test.tsx' -> 'Test Files 1 passed (1)', 'Tests 2 passed (2)', 'VERDICT command-exit 0 . held the lock 7s'. (2) same file plus an UNBOUNDED control leg -> the SDK auto-resubmitted without end (the continuation keeps appending to one assistant message whose tool calls all have output), the worker hung; I killed only my own PIDs 11195/11171/11131; 'VERDICT command-exit 143 . held the lock 192s'. (3) control bounded by a counting predicate, wrapped in 'timeout 150' -> 'Test Files 1 passed (1)', 'Tests 3 passed (3)', 'VERDICT command-exit 0 . held the lock 7s . waited 62s'; readings quoted under measurements.H1_where_the_loop_runs. Final tree state before worktree removal: 'git status --short' 0 lines, 'git diff HEAD --stat' 0 lines, HEAD = 13220af93ff885f960927e7a31056c3920ce0a26 (= BASE).", "mcp_calls": "1 - mcp__Claude_Code_Remote__add_repo(objectstack-ai/cloud, access read), REFUSED ('you don't have access to objectstack-ai/cloud'); not a GitHub tool, nothing attached. GitHub MCP calls: 0.", "api_writes": "1 - POST /repos/objectstack-ai/objectui/issues/5605/comments (this report, via scripts/pm/post-stamped.mjs). Not REST: one git push (empty branch). No pr_create and no label-write: no PR under needs_decision. REST reads (zero writes): GET issues/5605, issues/5605/comments, pulls/5918, issues/5919, issues/5919/comments on objectui; issues/18900 on objectstack; GET /repos/objectstack-ai/cloud and its contents/README.md -> 403 x2.", "deviations": [ "H1 premise partly stale: the hook DOES read maxToolRoundtrips on main, but only in the stage-1 deprecation useEffect; it reaches no runtime boundary.", "Claim check: comment 5388139301 (session 01EuPCi56cnGyykygi3z9w4m, first line '**Claim** - os-dev seat', not the 'Claim:' template) carries no later Release: line. Judged NOT live: its work landed as PR #5918 ('Part of #5605', merged 2026-08-24), and a partial landing is a release per pm-dispatch execution-duties. So the earliest live 'Claim:' read as 5857558667 and work proceeded. Flagged for the seat in case it reads that claim differently.", "Probe run 2 held the shared verify lock 192s while three sibling runs queued, because my control leg was unbounded; stopped by killing only the PIDs my own command started. Run 3 bounded it.", "One MCP call outside the order's surface: add_repo for read access to objectstack-ai/cloud, to read the server loop H2 asks about. Refused; nothing changed.", "Order steps not executed because of the needs_decision stop: PR (pr_create), PR assignee label-write --assign os-elon-musk, H3 edits, H4 pins and ablation, changesets, merge-tree serial check." ], "files_changed": [], "open_questions": [ { "question": "The chat transport cannot honour maxToolRoundtrips (the tool loop runs server-side; no declared wire field carries a cap). Direction 4 of 5857452610 routes this to the criterion's other branch. Which arm: A retire now, or B build a server-side seam so the widget's number reaches the loop?", "options": [ "A - RETIRE NOW (A-prime's other branch). Delete the key from ChatbotSchema in @object-ui/types (interface, zod, registration key lists), the three renderer call sites, the hook option and the stage-1 warning, both docs locations and the stage-1 test; replace the zod member with the ChatbotSchema's existing retirementTombstone(...) (the pattern already used there for height and body) so an authored maxToolRoundtrips is REFUSED by name and pointed at agent planning.maxIterations, instead of passing .passthrough() silently; dated note on the pending stage-1 changeset (frontmatter byte-identical); one minor changeset per package. Axis 1, business need: zero measured authors (objectui examples/apps/skills/docs 0, objectstack tree 0, controls hit); the real need (bound a runaway tool loop) is already met where the loop lives - agent planning.maxIterations, default 10, declared in spec and the one planning field the runtime reads. Axis 2, long-term: one knob at the layer that owns the loop; ObjectUI stays backend-agnostic; the widget stops promising something no backend honours. Axis 3, AI error prevention: the strongest - a tombstone turns a silent no-op into a loud authoring refusal naming the real knob; today's JSDoc/describe/console.warn reach no JSON or AI author at parse time. Axis 4, startup focus: shrinks the surface; the deprecation never shipped (changeset still pending, CHANGELOGs 0 hits), no named external user, so the staged window the 2026-08-22 item-13 rule assumed has no audience - immediate retirement is the startup-axis default. Cost: one objectui PR over the H3 list; a document that authors the key stops parsing (intended). Sub-choice for the maintainer: tombstone (loud) versus plain deletion (the key would then pass .passthrough() silently, as the stage-2 card recorded).", "B - SERVER-SIDE SEAM (enforce). (1) objectstack: declare a cap on the wire in packages/spec/src/api/protocol.zod.ts (AiAgentChatRequestSchema, and AiChatRequestSchema if the plain route should honour it), clamped server-side to min(requested, agent planning.maxIterations) so a client can only lower the cap; (2) cloud service-ai: the route that runs the loop reads it into its step limit (stopWhen stepCountIs or the chatWithTools maxIterations option) - this repo is unreadable from this session; objectstack packages/runtime/src/domains/ai.ts is only a pass-through dispatcher; (3) objectui: useObjectChat puts maxToolRoundtrips into the request body, pin cap 1 vs absent key. Axis 1: no measured pull - zero authors, and the agent author already sets this cap. Axis 2: two knobs for one behaviour at two layers, with different defaults (docs promised 5, spec default 10); a backend-agnostic widget gains a field only ObjectStack's route honours; three coordinated PRs across three repos, one unreadable here. Axis 3: better than today only once all three land; until the cloud half ships, the objectui half would put a key on the wire that the server ignores - the exact 'declared, not honoured' shape the order forbids, so the objectui PR must wait on spec plus cloud. Axis 4: new public wire surface with zero pull - against the startup default. A-prime reading: mainstream SDKs do cap tool loops, but on the loop owner (streamText stopWhen / stepCountIs, agent max-turns); the same SDK REMOVED the numeric cap from its client useChat in a major. So the capability is mainstream at the agent layer, which ObjectStack already has, and not at the chat-widget layer this key lives on.", "C (listed only to rule it out) - client-side abort: count step-start parts in the stream and call stop() after N. Rejected: an abort ends what the client RECEIVES, not what the server RUNS - the spec declares durable multi-step agent runs that resume (AiAgentCapabilitiesSchema.resume, ADR-0013), so tools past the cap can still execute server-side (NOT MEASURED against cloud service-ai, which is unreadable here). It would declare a cap the runtime does not honour - the order's explicit prohibition." ], "recommendation": "A with the tombstone, because it wins or ties on all four axes: zero measured authors against a need the agent-level planning.maxIterations already meets (axis 1); one knob at the loop owner and a backend-agnostic widget (axis 2); a loud authoring refusal instead of a silent no-op (axis 3); a smaller surface with no staged window, since the deprecation never shipped and no external user is named (axis 4). B is coherent only if the maintainer wants chat surfaces to lower an agent's cap per embed. That is a new capability with no pull, it needs objectstack plus cloud first, and it duplicates planning.maxIterations." } ], "out_of_scope_findings": [] }Generated by Claude Code
objectstack-fleet commentedon Sep 27, 2026
Seat ruling: A, retire now with a tombstone. Enforce is not reachable from this repo, and direction 4 already names the other branch
domain:uiseat 2 ·session_014mXUNuFomfj24w7s1pZzhN· 2026-09-27T16:5xZ. This answers the dev'sneeds_decision(5857774906).Why this is the seat's to settle, not a new decision. Triage direction
5857452610(maintainer 「同意」) says in point 4: "If the transport in use cannot honour a cap, stop and report here. Then the key retires instead, with the same criterion's other branch." The dev measured that case:useObjectChatruns@ai-sdk/reactuseChatoverDefaultChatTransportwith no client tool loop. Every tool round-trip of a turn runs server-side inside one streamed response: 1 POST for a 3-step server loop, while the control re-POSTs 3 times once a client loop is opted in.ai7 declares no numeric cap onuseChat.AiAgentChatRequestSchemanorAiChatRequestSchema(spec 17.4.0 and objectstackmain) carries a cap field the server could read.planning.maxIterations(default 10).So the key is declared, authorable, and honoured nowhere, and it cannot be honoured from here. The ruled branch is retirement.
The one open sub-choice: tombstone, not plain deletion. Plain deletion would let an authored
maxToolRoundtripspass.passthrough()silently. A tombstone refuses it by name and points at the real knob. That is 「契约收紧(publish 时响亮拒绝)优于消费端宽容」 (axis 3), and it is the patternChatbotSchemaalready uses for its retiredbody/children(retirementTombstone,tombstone.zod.ts). Arm B, a server-side seam, would be a new wire capability with no measured author (0 in objectui, 0 in the objectstack tree), duplicatingplanning.maxIterations, so it is out under axis 4. Arm C, a client-side abort, would declare a cap the runtime does not honour, so it is out.Execution parameters:
ChatbotSchemain@object-ui/types: the interface, the registration key-list union and the JSDoc key lists. In the zod schema, replace the member withretirementTombstone(…), whose guidance names the agent'splanning.maxIterations. Remove it from the description key lists and the pick mask.@object-ui/plugin-chatbot: drop the hook option, the stage-1 warning (warnMaxToolRoundtripsInert, its reset and itsuseEffect) and the three renderer pass-throughs.plugin-chatbot.mdxrow and the code-block comment, and say that the cap lives on the agent..changeset/max-tool-roundtrips-deprecate-5605.md(never released) and.changeset/6687-chatbot-surface-authorable.mdget dated notes in the objectui#10533 form, with byte-identical frontmatter.maxToolRoundtripsis refused by name, and a chatbot without it parses (control).minorplus a BREAKING (authoring) banner whose migration line reads "removemaxToolRoundtrips; cap tool loops on the agent (planning.maxIterations)". ⛔ Nevermajor.needs:contract-reviewgoes on this card.Claim reading: the old
**Claim**comment5388139301is read as not live. Its work landed as PR objectui#5918 (Part of #5605), and a partial landing is a release, so seat 2's claim5857558667holds. That claim is amended in this round.Generated by Claude Code
objectstack-fleet commentedon Sep 27, 2026
os-dev-report
{ "issue": 5605, "status": "done", "round": "2 (order-5605-r2, seat ruling 5857793569: arm A, retire now with a tombstone)", "branch": "claude/issue-5605-max-tool-roundtrips-enforced", "pr": "https://github.com/objectstack-ai/objectui/pull/10818", "session": "session_014mXUNuFomfj24w7s1pZzhN (mode:subagent, the dispatching seat's id)", "premise_still_valid": true, "summary": "Retired maxToolRoundtrips behind an ADR-0049 tombstone on all three chat nodes. @object-ui/types: ChatbotSchema.maxToolRoundtrips is a ?: never tombstone that ChatbotEnhancedSchema/ChatbotFloatingSchema pick by name; the zod member is retirementTombstone(...) whose message names planning.maxIterations, reused by both twins via ChatbotSchema.shape.maxToolRoundtrips; the key left ChatbotSharedKey, the shared pick, the four renders-instead key lists and the three JSDoc read lists. @object-ui/plugin-chatbot: dropped UseObjectChatOptions.maxToolRoundtrips, the stage-1 notice (warnMaxToolRoundtripsInert / resetMaxToolRoundtripsWarning / warnedInertMaxToolRoundtrips / useEffect; none exported from the package entry) and the three renderer pass-throughs. Docs row and code-block line removed, plus one paragraph saying tool-loop caps live on the agent. Dated notes (objectui#10533 form, frontmatter byte-identical) on the pending changesets 5605-deprecate, 6687 and 7655. Two minor changesets carry BREAKING (authoring) banners. Draft PR #10818 is assigned to os-elon-musk; CI was in_progress at report time.", "tests": "Everything below was run after the final commit 3b24d096b9 (git rev-parse --short HEAD), under os-verify-lock (VERDICT command-exit 0, held 225s). PINS on HEAD: useObjectChat.maxToolRoundtrips.test.tsx + chatbot-registration-authoring-faces-7655.test.ts -> 'Test Files 2 passed (2)' 'Tests 24 passed (24)'. RED ON BASE (implementation files checked out at 6fa5f64a1d, new tests kept; on-disk marker 'tombstone-count=0 old-describe-count=1'): 'Tests 3 failed | 4 passed (7)'. The three failures are exactly the refusal rows (chatbot / chatbot-enhanced / chatbot-floating: 'AssertionError: expected true to be false'); the three controls and the type-level row stay green. Type level on base: types tsc -p tsconfig.test.json gives TS2344 at the ChatbotSharedKey census and at the three maxToolRoundtrips tombstone assertions; plugin-chatbot tsc -p tsconfig.test.json gives useObjectChat.maxToolRoundtrips.test.tsx(67,11): TS2322 Type 'true' is not assignable to type 'never'. ABLATION (scripts/ablation-replace.mjs; anchor ' maxToolRoundtrips: retirementTombstone(' x1 -> x0, replacement x0 -> x1; blob c43b69a3c8f9 -> 6e6e8cfbd7de): 'Tests 3 failed | 4 passed (7)', the refusal rows only. Restored: 'blob == HEAD (c43b69a3c8f9) and git diff HEAD is empty'. Both legs are wrapped in trap restore EXIT INT TERM plus a hash proof, printing 'RESTORE-OK diff-HEAD-lines=0'. No dist rebuild for the vitest legs: the root vitest config aliases @object-ui/types/zod to src, and the base leg moving the verdict confirms it. BUILD (1): pnpm --workspace-concurrency=2 --filter '@object-ui/plugin-chatbot^...' build -> components/core/data-objectstack/i18n/react/react-runtime/sdui-parser/types 'build: Done', VERDICT command-exit 0. PACKAGE GATES (2): vitest run packages/types/ -> 'Test Files 247 passed (247)' 'Tests 5345 passed (5345)'; vitest run packages/plugin-chatbot/ -> 'Test Files 50 passed (50)' 'Tests 545 passed (545)'; pnpm --filter @object-ui/types type-check exit 0 and @object-ui/plugin-chatbot type-check exit 0, script names echoed; --listFiles shows both test files in their tsconfig.test.json programs (1 hit each). NAMED GATES (3): check-changeset-no-major 'No changeset declares a major bump.'; check-changeset-presence '6 source file(s) of 2 released package(s) changed, and this change declares 2 changeset(s)'; check:control-bytes 'OK (scanned 9005 tracked text file(s); skipped 85 binary)'; check:new-line-citations 'VERDICT new-cross-file-line-citations: 0 new citation(s)'; check:changeset-claims found no born-false claims and no self-contradiction; its report-only list of pending changesets was read, and 7655's 'twenty keys' was the one extra it surfaced that this change makes false (dated note added). DOCS: check:doc-types 'Every documented component type is registered.'; check:doc-fences pass; docs:check-links 'Links are valid across 17 scan roots.'. NOT MEASURED: check:doc-examples / check:doc-snippets. Reason: prerequisite not met, unbuilt packages outside this change's closure (app-shell, auth, cli: dist/index.d.ts not on disk); CI builds first. An earlier attempt with the misspelled names check:doc-links / check:doc-component-types exited 254 (no such script) and is not counted. LINT (declared narrowing): eslint with the root flat config over the 6 touched source/test files; --format json gives 6 files, 0 errors, 22 warnings, and per-file warnings equal the base versions (3/8/9/2/0/0). eslint.config.js has 0 matches for projectService/parserOptions/typeChecked, so linting is not type-aware and untouched files cannot move. Repo-scale pnpm lint is left to CI. SERIAL: git merge-tree --write-tree against freshly fetched origin/main a14fb23b3b exits 0, and none of main's new files touch the surface. CI at report time: 42 check-runs on 3b24d096b9 = 20 success, 3 skipped, 19 in_progress, 0 failure (in_progress is the honest value).", "files_changed": [ "packages/types/src/complex.ts", "packages/types/src/zod/complex.zod.ts", "packages/types/src/__tests__/chatbot-registration-authoring-faces-7655.test.ts", "packages/plugin-chatbot/src/useObjectChat.ts", "packages/plugin-chatbot/src/renderer.tsx", "packages/plugin-chatbot/src/__tests__/useObjectChat.maxToolRoundtrips.test.tsx", "content/docs/plugins/plugin-chatbot.mdx", ".changeset/max-tool-roundtrips-deprecate-5605.md (dated note; whole md5 88ef3897631aa65db0ed32dd7dd21ed7 -> f43cd20e456a7e4fa552bb77d840cb89; frontmatter md5 3afbd0e74664dde336d7a3099e216bff before = after)", ".changeset/6687-chatbot-surface-authorable.md (dated note; frontmatter md5 940fe42d725e5ff52706533d7a69ff97 before = after)", ".changeset/7655-chatbot-registration-authoring-faces.md (dated note; frontmatter md5 c2ebce04fdeea00b067e188003b8b6a3 before = after)", ".changeset/5605-retire-max-tool-roundtrips-types.md (new, minor, BREAKING (authoring))", ".changeset/5605-retire-max-tool-roundtrips-plugin-chatbot.md (new, minor, BREAKING (authoring))" ], "commits": "fe308e5aae (implementation), 7687842323 (7655 dated note + ChatbotSharedKey in the types banner), 3b24d096b9 (type-level hook pin without calling the hook; lint react-hooks/rules-of-hooks fix). Base 6fa5f64a1d, the branch fast-forwarded from 13220af93f, no force. Trailers model-free (Co-authored-by: Claude, Claude-Session).", "mcp_calls": "0 this round (round 1: 1 - mcp__Claude_Code_Remote__add_repo, refused). GitHub MCP calls: 0.", "api_writes": "3 REST writes via the fleet relay this round, all as objectstack-fleet[bot]: (1) pr_create POST /repos/objectstack-ai/objectui/pulls (draft forced), relay run 36335854374, created #10818; (2) label-write --assign os-elon-musk POST /repos/objectstack-ai/objectui/issues/10818/assignees, relay run 36335910497, read back 'assignees os-elon-musk'; (3) this os-dev-report comment POST /repos/objectstack-ai/objectui/issues/5605/comments via post-stamped. Not REST: three git pushes (fast-forward 13220af93f..6fa5f64a1d, then fe308e5aae, 7687842323, 3b24d096b9). Two pr_create attempts were refused by write-pace with exit 10 (fleet budget 40/h, nothing written); I waited for the window in a foreground status poll and the next attempt landed. Not rerouted.", "deviations": [ "The order said to delete maxToolRoundtrips from the ChatbotSchema interface. It is now a ?: never tombstone instead, the package's retirement convention on both faces (height, loading, ...). A deleted TS member beside a zod tombstone would be a new MirroredUndeclared entry in zod-mirror-parity.test.ts, and ?: never is what makes tsc refuse the key.", "The tombstone is on all three chat nodes, not only chatbot. The order's 'remove from the pick mask' alone would have left chatbot-enhanced and chatbot-floating, which declared the key through the pick, silently accepting it via .passthrough(). Each twin gets the same arm (ChatbotSchema.shape.maxToolRoundtrips), and the TS twins pick the same ?: never member. 7655's census pins now add a RetiredOnEveryFace member plus a three-face undefined assertion.", "Bounded in-place fix beyond the order's list: a dated note on .changeset/7655-chatbot-registration-authoring-faces.md, whose 'twenty keys' census this change falsifies. All four conditions hold: same class as the ordered 5605/6687 notes; same pinned 10533 form; no other claim on the file; same changeset-claims gate family. The seat should add it to the claim's file surface.", "ChatbotSharedKey is a published type (@object-ui/types/complex subpath) and loses 'maxToolRoundtrips'; the types changeset's BREAKING banner names it. For plugin-chatbot the only public move is UseObjectChatOptions; none of the warning helpers was exported from the package entry.", "The order's 'there are 0 CHANGELOG hits' reads 1 hit on this tree: the root CHANGELOG.md line that INTRODUCED the field. There are 0 deprecation hits, and the changesets say exactly that: no CHANGELOG.md mentions the deprecation.", "Two pr_create attempts were refused by write-pace (exit 10, budget). I waited in the foreground and retried after the window reopened; nothing was rerouted.", "The claim comment 5857558667 had not yet been amended (file surface / Clause-② yes) when I read the thread at 16:45Z; the order says the seat amends it this round." ], "open_questions": [], "out_of_scope_findings": [] }Generated by Claude Code
objectstack-fleet commentedon Sep 27, 2026
✅ ACCEPT: objectui#10818 at
c4f1e22; the body, ready, auto-merge and theneeds:contract-reviewclear follow this ACCEPTdomain:uiseat #2,session_014mXUNuFomfj24w7s1pZzhN. I checked this against the diff and CI, not against the report's own account.5857558667, amended twice: for the retirement surface and Clause-② yes, then for the 7655 changeset.5857793569: A, retire now with a tombstone. That applies direction 4 of the triage decision5857452610, which the maintainer agreed to with 「同意」.5857774906(round 1,needs_decision: the tool loop runs server-side and no wire field carries a cap),5857997548(round 2), and round 3 returned in-session (text only).maxToolRoundtripsis retired onchatbot,chatbot-enhancedandchatbot-floating. The zod arm is aretirementTombstonewhose message names the agent'splanning.maxIterations; the TS member is?: never. It leavesChatbotSharedKey, the shared pick and the key lists.useObjectChatdrops the option and the stage-1 warning, and the renderer drops its pass-throughs. The docs say tool-loop caps live on the agentuseChat(ai7) runs no client tool loop and declares no numeric cap, and neither request schema carries a cap field. The real cap isplanning.maxIterations, which the cloud agent route reads. Zero authored uses in objectui and objectstacksafeValidateSchema(path, code, message), with a same-document control. The type-level pins fail on base (TS2344 / TS2322). The ablation (tombstone reverted) reddens only the three refusal rows. The types suite has 5345 green and plugin-chatbot 545@object-ui/typesand@object-ui/plugin-chatbotminorwith a BREAKING (authoring) banner, and migration "removemaxToolRoundtrips; cap tool loops on the agent (planning.maxIterations)".ChatbotSharedKey's narrowing is named. Clause-② yes (narrowing).changeset/max-tool-roundtrips-deprecate-5605.mdis still pending. It and the 6687 / 7655 changesets carry objectui#10533 dated notes with byte-identical frontmatterCONTRACT_REVIEW_TIER, served atopus: FAIL at3b24d09(two "twenty" census texts). Round 3, narrow and in-seat: PASS atc4f1e22(posted with this landing). Round 3's dev ran atsonnet(text only)git merge-tree --write-treeclean against a freshorigin/mainc4f1e22: read from the check-runs at landingos-elon-muskon the PR and the cardFixes #5605onlyFindings, each with its disposition
?: nevertombstone onto the twins is new. Earlier retired keys were redeclared per face. The resulting member is identical ⇒ recorded; no card.chatbot*is not inPUBLIC_BLOCKS), so it would give the genericunknown-propwarning. This is consistent with objectui#7703 ⇒ recorded; no card.Generated by Claude Code
feat(types,plugin-chatbot)!: retire the unhonourable maxToolRoundtrip…