[objectstack-fleet]

Requested by Irene · project thread

Fixes #31080

Clause-②: no
Metadata-protocol: none

dispatch-gates --tier reads a MANDATORY model tier for skills/** (clause ① of the model-tiering ruling); it is waived by the standing order of the maintainer in this project (2026-11-03), and this card ran on the project's default model. The dispatching seat runs the contract review.

维护者速读

• 改了什么:skills/objectstack-ai/SKILL.md 的 MCP 工具表:动作行补上 resume_run 和确认门(带 ai.requiresConfirmation 的动作不带 confirm: true 即 428 拒绝、不执行);新增一行 delete_records(单对象、where 或 recordIds 二选一、每次至多 200 条、未确认的调用即预览不删除、OAuth 需 data:read + data:write)。
• 为什么改:表里只教了 delete_record 的确认门,读表的 agent 不知道 run_action / resume_run 也会被拒,也不知道已合入的 delete_records(MCP: delete records by filter in one confirmed call, with refusals grouped by reason, instead of one delete_record per row #30891, PR feat(mcp): delete_records deletes many records of one object in one confirmed call #31024)。
• 风险与回滚:只改技能文字,净增 1 行,token 5563 → 5776(上限 6801 不动);发 @objectstack/skills patch;回滚即 revert。
• 先看:skills/objectstack-ai/SKILL.md L202–204,对照下文「Code sites」各行。
• 两分钟验证:node scripts/check-skills-token-ratchet.mjs,预期退出 0,skills/objectstack-ai/SKILL.md is 5776 tokens (ceiling 6801)。
• 需要你做的:skills/** 属 Tier H(人合),需要 os-zhuang 或 hotlong 的一次 approval,之后席位落地。

What changes

skills/objectstack-ai/SKILL.md, MCP tools table. L202 (create_record, update_record, delete_record) is byte-identical to main.

@@ -202,3 +202,4 @@ permissions and row-level security:
 | `create_record`, `update_record`, `delete_record` | writes; `delete_record` is refused with `ACTION_CONFIRMATION_REQUIRED` (428) unless it sends `confirm: true` — ask the human first, then retry |
-| `list_actions`, `run_action` | your AI-exposed business Actions — the same `ai.exposed` opt-in as above |
+| `delete_records` | up to 200 records of one object per call (fewer under a smaller `OS_DATA_MAX_PAGE_SIZE` or `OS_DATA_QUERY_ROW_BUDGET`), by `where` or `recordIds`, not both; a call without `confirm: true` is the preview (match count, sample ids) and deletes nothing — show it to the human, and send `confirm: true` only once they approve; over OAuth it needs `data:read` and `data:write` |
+| `list_actions`, `run_action`, `resume_run` | your AI-exposed business Actions — the same `ai.exposed` opt-in as above; registered only when the bridge implements `listActions` and `runAction` (the stdio bridge does not, so it serves none), `resume_run` only when it also implements `resumeRun`. An action flagged `ai.requiresConfirmation` (`list_actions` reports it) is refused with the same 428 unless the call sends `confirm: true`, and nothing runs — ask the human first, then retry; `resume_run` submits a paused screen flow's form under the same gates |
 

The "ask the human first, then retry" stance stays on L202 and is repeated on both new rows; they add no gate the code lacks.

Plus .changeset/mcp-tools-table-confirmation-gate.md: '@objectstack/skills': patch.

Code sites (each taught fact, measured on main 24bc259)

Taught fact	Code site
The member is spelled confirm	packages/spec/src/contracts/ai-service.ts:242 AI_ACTION_CONFIRMATION_MEMBER = 'confirm'
run_action: an action with ai.requiresConfirmation: true is refused with 428 ACTION_CONFIRMATION_REQUIRED unless the request carries confirm: true	packages/runtime/src/action-execution.ts:2911 actionConfirmationRefusal (code and status at :2893, :2901)
Nothing runs on a refused call	packages/runtime/src/action-execution.ts:4513: the gate is the last pre-dispatch check, before the subject record is read; ask the human first: tool description mcp-http-tools.ts:2255-2256, packages/mcp/src/skill-md.ts L294–317
The MCP tool declares and forwards the member	packages/mcp/src/mcp-http-tools.ts:2309 (schema key), :2362 (forwarded to bridge.runAction)
list_actions reports requiresConfirmation	packages/mcp/src/mcp-http-tools.ts:2255, :2313-2314 (tool descriptions)
The action tools are registered only on a bridge with listActions and runAction; the stdio bridge has neither	packages/mcp/src/mcp-http-tools.ts:1302 (wireBridgeTools); packages/mcp/src/stdio-data-bridge.ts implements no listActions / runAction
resume_run is in the registerActionTools family, registered only when the bridge implements resumeRun	packages/mcp/src/mcp-http-tools.ts:343-344 (interface: optional), :2218-2220, :2382; the runtime bridge implements it at packages/runtime/src/domains/mcp.ts:1027
resume_run carries the same gate	packages/runtime/src/domains/mcp.ts:1465 resumeAdmissionRefusal calls actionConfirmationRefusal; the tool's description (mcp-http-tools.ts:2394-2396) says a refused call leaves the run paused
delete_records is registered by registerObjectTools, under data:write and data:read	packages/mcp/src/mcp-http-tools.ts:1924 (if (canWrite)), :2033 (if (canRead)), :2057
One object, where or recordIds, not both	packages/mcp/src/mcp-http-tools.ts:2120-2121
At most 200 per call, fewer under a smaller OS_DATA_MAX_PAGE_SIZE or OS_DATA_QUERY_ROW_BUDGET	packages/mcp/src/mcp-http-tools.ts:632 DELETE_RECORDS_MAX_PER_CALL = 200, :635-648 deleteRecordsCap
The unconfirmed call is the preview (count, sample ids) and deletes nothing; show it to the human and confirm only once they approve	packages/mcp/src/mcp-http-tools.ts:2062, :916, :925-933; .changeset/30891-mcp-delete-records.md

Tests

None: the diff is skill teaching text with no runtime behaviour to test. The gates below are the checks.

Acceptance notes

Token budget (node scripts/check-skills-token-ratchet.mjs, exit 0)

File	main	head	Ceiling
skills/objectstack-ai/SKILL.md	5563	5776	6801 (unchanged)

Net lines +1 (+2 / −1) across both commits. The second commit (review fixes) rewrites only the two new rows. No ceiling moves and no payment was needed (headroom 1238 before, 1025 after). main moved 4 commits after the branch point; none touches this file or the ratchet.

Eval sweep

• skills/objectstack-ai/evals/skills-tools-knowledge.json: no case names run_action, resume_run, delete_record(s), confirm, requiresConfirmation or the MCP tools list. Nothing to fix.
• Other copies of the MCP tools list in skills/**: none. The delete_record hits in skills/objectstack-automation/SKILL.md L87 and skills/objectstack-query/rules/filters.md L193 are flow node types, not MCP tools. Left.

Verification

Local gates on 6a31227: 45 derived, 45 run with exit 0, 0 NOT MEASURED (`dispatch-gates --ran`)

• node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack derived 45 commands (26 pnpm, 19 direct node) for the 2 changed paths. Each ran under a 900 s timeout with NODE_OPTIONS=--max-old-space-size=8192, its exit code captured before any pipe, after a full pnpm build (exit 0). All 45 exited 0 on the first run.
• They include check-skills-token-ratchet.mjs, check:skill-docs, check:skill-compatibility, check:skill-frame-sync, check:skill-identifier-liveness, check:doc-authoring, check:dts-closure, check:dual-build-cjs-loads, check-changeset-no-major.mjs, check-empty-changeset.mjs and check-comment-mask-corpus.mjs.
• --ran printed: 45 derived famil(ies) accounted for — 45 run, 0 NOT-MEASURED (a DERIVED zero — all 45 recorded an exit code and none of them is 3).
• Also run, outside the derived list: pnpm check:changeset-fixed (exit 0), the one artifact-roster family whose roster sits under .changeset.

• node scripts/pm/check-governed-merges.mjs --branch claude/project-thread-qmr8o9: GOVERNED, Tier H (人合): skills/** ×1; 13 changed lines (+12 / −1), under 3000.
• node scripts/check-commit-card-trailers.mjs --range origin/main..HEAD: exit 0 before every push.

Human Minimum

• 1. Schemas start as Zod; packages/spec holds no runtime logic. (No schema touched.)
• 2. Every name I added uses the casing for its kind. (No names added.)
• 3. No formatter ran. (Markdown only.)
• 4. One issue, one branch off an up-to-date main, one pull request.
• 5. skills/** publishes @objectstack/skills, so the PR carries a '@objectstack/skills': patch changeset. Both declaration lines are filled in above.
• 6. packages/spec not touched; nothing to regenerate.
• 7. No hand edits to content/docs/references/, packages/*/CHANGELOG.md or content/docs/releases/.
• 8. Title, body and comments are in English. (The 维护者速读 section is Chinese, per the lane template.)
• The Before Submitting list in CONTRIBUTING.md is done.

🤖 Generated with Claude Code

https://claude.ai/code/session_01J8VVAkL29APmM7okwgntB6

[objectstack-fleet]

domain:skills#3 · huangyiirene · session_01J8VVAj7CvjkGXvgXzd6RZ7

Contract-review-claim: session_01J8VVAj7CvjkGXvgXzd6RZ7 · head 4669f41 · 2026-11-07T05:11Z

The seat review runs at CONTRACT_REVIEW_TIER in an isolated subagent. No review seat (lanes/review.md) is seated.

[objectstack-fleet]

domain:skills#3 · huangyiirene · session_01J8VVAj7CvjkGXvgXzd6RZ7

Contract review

Served-tier: CONTRACT_REVIEW_TIER
Head-sha: 4669f418e996a09e5de970722430b737b750dc54
Local-runs: none
Scope: skills/objectstack-ai/SKILL.md, .changeset/mcp-tools-table-confirmation-gate.md + packages/runtime/src/action-execution.ts, packages/mcp/src/mcp-http-tools.ts, packages/runtime/src/domains/mcp.ts, packages/mcp/src/stdio-data-bridge.ts, packages/mcp/src/mcp-server-runtime.ts, packages/mcp/src/skill-md.ts, packages/spec/src/contracts/ai-service.ts, .changeset/30891-mcp-delete-records.md, skills/objectstack-ai/evals/skills-tools-knowledge.json, scripts/check-skills-token-ratchet.mjs, scripts/pm/check-governed-merges.mjs, AGENTS.md

Inputs: card #31080 (body, triage, claim), the PR #31088 body and file list, the net diff origin/main...4669f418 (2 files, +12 / −1), and the head's check-runs. Code read on origin/main 719a502b. None of packages/mcp, packages/runtime, packages/spec/src/contracts, the ratchet or check-governed-merges changed between the branch point 24bc2593 and 719a502b, so the PR body's line citations still hold.

① Derived judgments

The facts the new and edited rows teach, checked against the code on main:

• Code, status, member: TRUE. ACTION_CONFIRMATION_REQUIRED_CODE and _STATUS = 428 are at action-execution.ts L2893 and L2901. The MCP door has ACTION_CONFIRMATION_REQUIRED / 428 at mcp-http-tools.ts L537–538. The member is AI_ACTION_CONFIRMATION_MEMBER = 'confirm' at packages/spec/src/contracts/ai-service.ts L242. Only the boolean true attests (L2921, and isDeleteConfirmed L594).
• Which actions are gated: TRUE. actionConfirmationRefusal (L2911) reads action.ai.requiresConfirmation === true and nothing else. The mode: 'delete' / variant: 'danger' heuristic only widens the listing (actionLooksDestructive, L2829). "(list_actions reports it)" is sound: a flagged action is always listed requiresConfirmation: true. The code's own docblock says a client that confirms whenever the listing says so is always correct.
• Nothing runs: TRUE. The gate is the last pre-dispatch check in invokeBusinessAction (L4513). It runs before loadActionSubjectRecord and before the invocation row and dispatch.
• resume_run carries the gate: TRUE. resumeAdmissionRefusal (domains/mcp.ts L1449) runs run_action's gates in order, ending with actionConfirmationRefusal (L1465). Every refusal is thrown before resume(), so the run stays paused. "Under the same gates" is a lower bound. resume_run adds more gates (caller's own run, screen pause, subject re-read), and the row does not claim the set is exact.
• delete_records facts: TRUE.
  • One object, where or recordIds, not both (L2116–2122). Neither is also refused.
  • Cap: DELETE_RECORDS_MAX_PER_CALL = 200 (L632), lowered by deleteRecordsCap (L647–648).
  • Preview: an unconfirmed call answers the same 428 refusal with a preview member (matched or matchedMoreThan, wouldAttempt, cap, sampleRecordIds). It deletes nothing (L2149–2163).
  • Scopes: registered only inside if (canWrite) and nested if (canRead) (L1924, L2033). grantedScopes exists only for an OAuth envelope (domains/mcp.ts L129), so "over OAuth it needs data:read and data:write" holds.
  • Registration: no bridge condition (aggregate is optional inside it, L733).
• resume_run's registration condition: OMITTED — must-fix 1. registerActionTools registers resume_run only when typeof bridge.resumeRun === 'function' (L2218–2220, L2382). The bridge interface says so outright: "OPTIONAL, like McpDataBridge.aggregate. A host that cannot resume runs omits it, and resume_run is then not registered" (L343–344). The table's intro names registerObjectTools / registerActionTools as its source. Its own aggregate_records row names exactly this kind of condition ("registered only when the data bridge implements aggregate"). The edited row lists resume_run beside list_actions and run_action with no condition. That over-states what a client sees. The code's own rule for run_action's description is to name resume_run "only where resume_run is actually registered … so this description never points an agent at a tool it cannot call" (L2257–2258). The runtime's HTTP bridge implements resumeRun (domains/mcp.ts L1027), so the default /mcp serves it. A host that wires @objectstack/mcp with an action bridge that cannot resume does not.
• Cap wording is incomplete: should-fix 1. "(fewer on a smaller read page size)" leaves out the row budget. deleteRecordsCap takes min(200, OS_DATA_MAX_PAGE_SIZE, OS_DATA_QUERY_ROW_BUDGET) (L635–648), and the PR body's own code-site table says "read page size or row budget". This is not false, but it is a partial teaching of what lowers the cap.
• "Ask the human first" is missing from the new rows: should-fix 2. The delete_record row (L202, unchanged, byte-identical) says "ask the human first, then retry". The run_action row borrows only "the same 428" from it. The delete_records row never says to show the preview to the human and send confirm: true only once they approve. An agent reading L203 alone learns a two-call recipe with no human in it. The code does put the human in it: the tool description says "Show that to the human, and retry with confirm: true only once they approve" (L2065), and the refusal sentence says the same (L933). The server-served mirror packages/mcp/src/skill-md.ts (L294–296, L314–317) carries the stance on both tools. No row states anything false, so this is not a must-fix.
• The action family's own condition is unstated (pre-existing): should-fix 3, optional. wireBridgeTools registers the whole action family only on a bridge carrying listActions and runAction (L1302). The stdio bridge carries neither (stdio-data-bridge.ts), so the stdio transport the section describes just above serves none of L204's tools. This predates the PR. The row is in the claim, and the must-fix 1 edit is the natural place to fix it.
• No tool is named that the code does not register in its family: delete_records is in registerObjectTools, and resume_run is in registerActionTools.

② Semver level

• '@objectstack/skills': patch is right. The diff is skill teaching text: no authorable key, value or meaning moves, so Metadata-protocol: none, and AGENTS.md Post-Task Checklist step 3 says none takes patch.
• Clause-②: no is right, because no published surface changes.
• Both declarations are line-initial in the PR body and in the changeset body. The changeset's prose is true on main.

③ Boundary flags

• The edits stay inside the claim:
  • L202 is byte-identical.
  • L203 is edited, and one row is added (now L203–204).
  • The evals are untouched.
  • One new .changeset/*.md is added.
• The net diff against main is exactly the 2 files in the PR's file list. The other skills/** differences between origin/main and the head come from main moving (docs(skills): a bare field ref in a predicate faults, it does not read null #31028, docs(skills): objectui i18n guide keys the built-in packs by platform locale tag (zh-Hans, zh-Hant, pt-BR) #31043, docs(skills): name the save_notice node in the automation skill #31044), not from this PR.
• The tier comes from scripts/pm/check-governed-merges.mjs: "Published skills/** stays Tier H" makes this PR Tier H (人合). Landing needs an approval from an account in GOVERNED_APPROVERS after this record. The Governed Surface Queue Guard check is green.

④ Design judgment

• One home per rule: holds. The runtime 428 gate is taught only in this table. L154 and L318 teach the authoring side (ai.requiresConfirmation on the action, and the removed tool key), not the refusal. No other skills/** copy of the MCP tool list exists.
• Style. The two new rows are long next to the table's terse rows, but they read in its register. The changes asked for in ① (must-fix 1, should-fix 2) cost about 25 tokens.
• Token budget. By ceil(bytes / 4) arithmetic on the two blobs (nothing run), SKILL.md goes from 22272 to 22795 bytes, so +131 tokens. That matches the PR body's 5563 → 5694 once the stamp-line exclusion is applied. The ceiling stays 6801, with 1107 tokens of headroom left, and no ceiling moves.
• Eval sweep. skills-tools-knowledge.json holds 5 cases: skill authoring, provider, RAG, tool vs action exposure, and agent refusal. None names run_action, resume_run, delete_record(s), confirm or the MCP tool list, so none is stale. The claim only lets the PR edit an existing matching case, and there is none, so there is no gap it could fill.
• Check-runs on the head. 38 runs: 20 success, 13 skipped, 5 in progress, 0 failed. In progress: Type Check · source gates, Test Core (1/8), Lint · docs & code guards, Lint · ESLint, Lint · CI, release & engine gates. Only shards 1 and 2 of 8 of Test Core exist on the head so far.

Must-fix

1. L204: name resume_run's registration condition, in the style of the aggregate_records row, e.g. "resume_run (registered only when the bridge implements resumeRun) submits a paused screen flow's form under the same gates".

Should-fix

1. L203: say that the row budget lowers the cap too, e.g. "fewer under a smaller read page size or row budget; the answer's cap says".
2. L203 and L204: carry L202's stance on both rows. For delete_records: show the preview to the human and retry with confirm: true only once they approve. For run_action / resume_run: ask the human first.
3. (Pre-existing, optional) L204: the family is registered only on a bridge with the action seam, and the stdio transport serves none of it.

Implemented-by: claude/project-thread-qmr8o9
Reviewed-by: session_01J8VVAj7CvjkGXvgXzd6RZ7

VERDICT: FAIL

[objectstack-fleet]

domain:skills#3 · huangyiirene · session_01J8VVAj7CvjkGXvgXzd6RZ7

Contract-review-claim: session_01J8VVAj7CvjkGXvgXzd6RZ7 · head 6a31227 · 2026-11-07T05:28Z

Delta review against the FAIL at head 4669f41 (comment 6702719478: must-fix M1, should-fixes S1–S3). The delta is one commit to the tools-table rows answering all four, plus the PR body (rewritten by the seat to match). The seat review runs at CONTRACT_REVIEW_TIER in an isolated subagent.

[objectstack-fleet]

domain:skills#3 · huangyiirene · session_01J8VVAj7CvjkGXvgXzd6RZ7

Contract review

Served-tier: CONTRACT_REVIEW_TIER
Head-sha: 6a3122720afd6a197e82e95ced726f56f5c8c699
Local-runs: none
Scope: skills/objectstack-ai/SKILL.md, .changeset/mcp-tools-table-confirmation-gate.md + packages/mcp/src/mcp-http-tools.ts, packages/mcp/src/stdio-data-bridge.ts, packages/mcp/src/plugin.ts, packages/mcp/src/mcp-server-runtime.ts, packages/types/src/query-read-limits.ts, packages/runtime/src/action-execution.ts, packages/runtime/src/domains/mcp.ts, packages/mcp/src/skill-md.ts, scripts/check-skills-token-ratchet.mjs, scripts/pm/check-governed-merges.mjs

This is a delta review against the FAIL recorded at 4669f418: one must-fix (M1) and three should-fixes (S1–S3). The delta is one commit, 6a312272, which touches only skills/objectstack-ai/SKILL.md (+2 / −2, the two new table rows). It leaves the changeset and the evals unchanged. Code was re-read on origin/main 719a502b, which has not moved since the first review.

① Derived judgments

• M1 is answered and TRUE. The new L204 text "resume_run only when it also implements resumeRun" matches the code. registerActionTools checks typeof bridge.resumeRun === 'function' (mcp-http-tools.ts L2218–2220) and registers resume_run only inside that check (L2382). The bridge interface marks the member optional (L343–344).
• S1 is answered and TRUE. L203 now reads "fewer under a smaller OS_DATA_MAX_PAGE_SIZE or OS_DATA_QUERY_ROW_BUDGET".
  • deleteRecordsCap takes min(200, maxPageSize, rowBudget) from resolveQueryReadLimits (mcp-http-tools.ts L647–648).
  • Those two limits are read from exactly these env names (packages/types/src/query-read-limits.ts L57, L59).
• S2 is answered and TRUE. The new wording matches what the code tells agents to do.
  • L203 adds "show it to the human, and send confirm: true only once they approve". This matches the delete_records tool description (L2065) and its refusal sentence (L933).
  • L204 adds "ask the human first, then retry". This matches the run_action description (L2255–2256), and the resume_run description and schema (L2394–2396, L2436).
  • Both rows now hold the same stance as L202.
• S3 is answered and TRUE. L204 now says the action tools are "registered only when the bridge implements listActions and runAction (the stdio bridge does not, so it serves none)".
  • wireBridgeTools adds the action family only on that condition (L1302).
  • The stdio transport gets its one bridge from createStdioDataBridge (plugin.ts L744, the only assignment, passed to bridgeDataTools at L848).
  • That bridge returns an McpDataBridge with no listActions, runAction or resumeRun (0 matches in stdio-data-bridge.ts).
  • So the stdio transport serves none of L204's tools.
• The delta adds no false or over-broad claim. Every other clause of the two rows is unchanged from the first review and still TRUE there. The actions:execute scope gate is not restated on the row, but the table intro already covers it ("each family gated by the caller's OAuth scopes").

② Semver level

• The changeset is unchanged in the delta: '@objectstack/skills': patch, Clause-②: no, Metadata-protocol: none, all line-initial. The level is still correct, because the delta is skill teaching text only.
• The rewritten PR body still carries Clause-②: no and Metadata-protocol: none line-initial. Its diff block is byte-for-byte the net diff. It has new code-site rows for the bridge conditions and the env names. It states "5563 → 5776" and "+1 net lines", and both hold.

③ Boundary flags

• L202 is byte-identical to main.
• The net diff against main is still 2 files: SKILL.md +2 / −1 (net lines +1) and the changeset +10. That is still the table rows plus one added row, with no eval edit, all inside the claim.
• The tier is still Tier H (人合): published skills/** in check-governed-merges. Landing needs an approval from an account in GOVERNED_APPROVERS.

④ Design judgment

• Token budget. By ceil(bytes / 4) arithmetic on the blobs (nothing run), SKILL.md grows from 22272 to 23123 bytes. That is 5568 → 5781 raw, or 5563 → 5776 after the stamp-line exclusion, matching the PR body. The ceiling stays 6801, with 1025 tokens of headroom left, and no ceiling moves.
• Style and placement. L204 is now the longest row in the table. Its registration condition is worded like the aggregate_records row, and each rule still lives in this one table only.
• Eval sweep. Unchanged: no case in skills-tools-knowledge.json is stale or affected.
• Check-runs on 6a312272. 42 runs: 17 success, 15 skipped, 10 in progress, 0 failed.
  • In progress: Type Check · source gates, Type Check · consumer gates, Test Core (1/8), Test Core (2/8), Live E2E UI, Lint · repo & tooling gates, Lint · docs & code guards, Lint · ESLint, Lint · CI, release & engine gates, and a second Check Changeset.
  • Governed Surface Queue Guard and the PR-body guards are green.

Must-fix: none. M1 is resolved.
Should-fix: none. S1, S2 and S3 are resolved.

Implemented-by: claude/project-thread-qmr8o9
Reviewed-by: session_01J8VVAj7CvjkGXvgXzd6RZ7

VERDICT: PASS

[objectstack-fleet]

domain:skills#3 · huangyiirene · session_01J8VVAj7CvjkGXvgXzd6RZ7

维护者速读

• 改了什么: skills/objectstack-ai/SKILL.md 的 MCP 工具表补两处：新增 delete_records 一行(按 where 或 recordIds 批删，单次上限 200,可被 OS_DATA_MAX_PAGE_SIZE / OS_DATA_QUERY_ROW_BUDGET 调低；未确认的调用只返回预览、不删，给人看过并获批后再带 confirm: true);动作行写明标了 ai.requiresConfirmation 的动作没带 confirm: true 会被 428 ACTION_CONFIRMATION_REQUIRED 拒绝，同门也管 resume_run,并注明各工具的注册条件(stdio 不提供动作工具)。另加 '@objectstack/skills': patch changeset。2 个文件。
• 为什么改: 原表只教了 delete_record 的确认，读表的 agent 不知道 run_action / resume_run 也会被拒，也不知道有 delete_records(skills: objectstack-ai MCP tools table omits run_action's confirmation gate and the delete_records tool #31080)。
• 风险与回滚: 只改技能文字，无代码改动;5563 → 5776 token(上限 6801 不动),净 +1 行,L202 原样未动。可整 PR revert。
• 先看: skills/objectstack-ai/SKILL.md L202–204。
• 两分钟验证: grep -n "delete_records\|resume_run" skills/objectstack-ai/SKILL.md 命中工具表 L203–204。
• 需要你做的: skills/** 是受治理 Tier H 面，需 os-zhuang 或 hotlong 在 GitHub 上 Approve 一次。达档复核：全量 6702719478 FAIL(M1 + S1–S3),增量 6702842692 PASS,Head-sha 6a3122720afd6a197e82e95ced726f56f5c8c699。另有一个 CI 子任务需要先重跑：lint 工作流 run 42342799994 的 Type Check · consumer gates 在 runner 上卡死(日志停在 05:29 的构建步骤),45 分钟超时被取消，聚合检查 TypeScript Type Check 因此报红。中继只能重跑 ci.yml,请在该 run 页点 Re-run failed jobs。批准后本席落地。
• 席位意见: ACCEPT。

席位核对 — ACCEPT · domain:skills#3 · head 6a3122720afd6a197e82e95ced726f56f5c8c699

1. 形态: draft → main,Fixes #31080;'@objectstack/skills': patch,Clause-②: no,Metadata-protocol: none;作者 objectstack-fleet[bot];2 个文件，在认领面之内(工具表两行 + 一行新增、changeset)。
2. 前提实测: 复核逐项对照 main 代码:action-execution.ts actionConfirmationRefusal(只看 ai.requiresConfirmation === true,在派发前)、domains/mcp.ts 的 resume_run 门、mcp-http-tools.ts 的 delete_records(二选一选择、上限与 deleteRecordsCap、预览、OAuth 双 scope)和 registerActionTools / wireBridgeTools 的注册条件;确认成员拼写为 confirm。该技能 evals 无相关用例。
3. 门禁(钉 head 6a312272): dev 派生门禁 45 条全部 exit 0;check-runs 44 条:28 success、15 skip;Type Check · consumer gates 被取消(runner 卡死：日志停在 2026-11-07T05:29Z 的构建步骤，同一步在其他 run 约 7 秒;card thread 在该 head 本地跑 check:doc-formula-expressions 7.2 秒 exit 0),聚合 TypeScript Type Check 随之 failure,非本 PR 所致。中继 run_rerun_failed 只收 ci.yml,需维护者重跑该 run 的失败任务。check-governed-merges --pr 31088:Tier H(skills/**)。
4. 接受的偏差: 无。首轮 FAIL 的 M1(resume_run 漏写注册条件)与 S1–S3 已在新提交改正并经增量复核确认;PR 正文由本席按新 diff 重写。
5. 范围外去向: 无。
6. 落地条款: Tier H。保持 draft,挂 pm:awaiting-merge,请 os-zhuang、hotlong 审批;获批后本席过落地前检、清标、ready、auto-merge;维护者直合亦可。