verify-dms-ai
Drive the AntelopeJS dms-ai module (the AI assistant: its sidecar, Overview, Changes, Activity, Skills and Settings pages, the Codex sign-in banner and the owner-only routes) the way the platform owner does, inside a DMS with free ports and its own MongoDB database, and report a verdict with evidence. Use to prove a change to AntelopeJS/dms-ai works before merge, or to verify a PR. A turn with Claude Code or Codex needs credentials a run never holds: that part is NOT VERIFIED, so a verdict is never a plain PASS.
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-14
- Stars
- ★ 2
- Forks
- 0
01 前言 Front Matter
- name
- verify-dms-ai
- description
- Drive the AntelopeJS dms-ai module (the AI assistant: its sidecar, Overview, Changes, Activity, Skills and Settings pages, the Codex sign-in banner and the owner-only routes) the way the platform owner does, inside a DMS with free ports and its own MongoDB database, and report a verdict with evidence. Use to prove a change to AntelopeJS/dms-ai works before merge, or to verify a PR. A turn with Claude Code or Codex needs credentials a run never holds: that part is NOT VERIFIED, so a verdict is never a plain PASS.
02 资源文件 49
- 脚本ai.mjs16.9 KB
- 参考文档features/access.md3.0 KB
- 参考文档features/activity.md2.3 KB
- 参考文档features/changes.md2.3 KB
- 参考文档features/chat.md2.5 KB
- 参考文档features/codex.md1.1 KB
- 参考文档features/lifecycle.md2.9 KB
- 参考文档features/overview.md2.5 KB
- 参考文档features/package.md793 B
- 参考文档features/README.md4.3 KB
- 参考文档features/settings.md2.2 KB
- 参考文档features/skills.md2.4 KB
- 参考文档features/smoke.md876 B
- 脚本guard.mjs4.7 KB
- 脚本journeys/access.mjs14.1 KB
- 脚本journeys/activity.mjs17.0 KB
- 脚本journeys/activity/export.mjs3.6 KB
- 脚本journeys/activity/helpers.mjs1.8 KB
- 脚本journeys/activity/legacy.mjs2.5 KB
- 脚本journeys/changes.mjs9.0 KB
- 脚本journeys/changes/helpers.mjs1.3 KB
- 脚本journeys/changes/storage.mjs3.5 KB
- 脚本journeys/changes/undo.mjs12.4 KB
- 脚本journeys/chat.mjs1.9 KB
- 脚本journeys/codex.mjs5.8 KB
- 脚本journeys/lifecycle.mjs17.4 KB
- 脚本journeys/overview.mjs11.4 KB
- 脚本journeys/overview/cards-and-period.mjs9.7 KB
- 脚本journeys/overview/helpers.mjs995 B
- 脚本journeys/package.mjs3.6 KB
- 脚本journeys/settings.mjs6.9 KB
- 脚本journeys/settings/helpers.mjs751 B
- 脚本journeys/settings/page.mjs13.1 KB
- 脚本journeys/skills.mjs18.9 KB
- 脚本journeys/smoke.mjs1.8 KB
- 参考文档package.json21 B
- 脚本playground/antelope.config.ts2.9 KB
- 参考文档playground/package.json986 B
- 参考文档playground/pnpm-lock.yaml70.3 KB
- 脚本playground/src/fixture.ts2.2 KB
- 脚本playground/src/index.ts182 B
- 脚本playground/src/pages.ts1.1 KB
- 脚本playground/src/skills.ts1.9 KB
- 参考文档playground/tsconfig.build.json451 B
- 脚本sample.mjs10.4 KB
- 脚本seed.mjs8.0 KB
- 参考文档tsconfig.json613 B
- 脚本verify.config.ts3.7 KB
- 脚本windows.mjs1005 B
03 SKILL.md 正文
Verify dms-ai
Every deterministic step is a dms-control command. You choose which journeys to run and judge what they show; you never write a multi-step shell recipe. When a journey you need does not exist, add it to journeys/<feature>.mjs, describe it in its feature file, and run it.
dms-control is the bin of @antelopejs/dms, a devDependency of this repository. Run it with Node 24 after pnpm install:
C="pnpm exec dms-control"
$C # lists every command
Give doors and --skill absolute paths. mongod comes from DMS_CONTROL_MONGOD or PATH. A verifier that runs with a temporary HOME does not see ~/.cache/mongodb-binaries: pass the binary in DMS_CONTROL_MONGOD.
What this skill cannot verify
A run holds no ANTHROPIC_API_KEY, no Claude Code login and no ChatGPT sign-in, so no model turn runs. These are NOT VERIFIED in every verdict, never PASS: a turn with Claude Code or Codex (the chat, tool calls as they happen, approvals, questions, queued messages, auto-fix turns, token counts a turn reports), block previews, and the ChatGPT consent, token exchange and refresh. The chat panel is also out of reach: it exists in the development dashboard only (import.meta.env.DEV) and dms-control up builds a production one (AntelopeJS/dms#749). The best result of a verdict is therefore PASS+NOTES.
What a turn leaves behind is not out of reach. The Activity rows, the Overview figures, the change sets with their diffs and Undo and Redo, the skills use counts and the token usage are read from the sidecar's stored conversations, change sets and shadow repository. The journeys write those files (seed.mjs, sample.mjs), start the assistant on them and drive the pages and routes that read them, so those features are verified.
Launch
RUN=$($C up --prepare --base origin/main --skill "$PWD/.agents/skills/verify-dms-ai")
--prepare runs the prepare commands of verify.config.ts: the frozen install, the builds of the module and of the sidecar, the install of the playground's own dependencies, playwright-core into this folder and the TypeScript build of the playground's fixture module. up reads the playground (playground/ in this folder) and starts an isolated DMS:
- the checkout is laid out again under
$RUN/root, every sibling linked, and the playground copied, so the run's build and.antelope/are its own; - the API, the dashboard and the mail sink get free ports from the kernel, and
HOMEis$RUN/home; - a
mongodthe run owns servesverify_<run>databases; - the backend runs
ajs project start -e docs, the dashboardajs dms start -e docs -b <api>.
The module under verification is a local source at the checkout root, and @antelopejs/dms comes from the registry, pinned by the playground's lockfile. The fixture module (playground/src) serves a member with a role (POST /verify-ai/member) and registers a home page and a plain page; it writes the skills the Skills journeys read into generated-skills/ when it is constructed (a committed SKILL.md would be read as a skill of this repository), among them a name shared by two folders and three SKILL.md the catalog must skip.
Doctor
$C doctor "$RUN" # prints only failing checks; --verbose prints all
Do not drive a run that fails doctor.
Drive
$C plan --base origin/main # the diff mapped to features; unmapped surfaces
$C journey "$RUN" smoke
$C journey "$RUN" settings [journey-id...]
- Each feature file in features/ lists its source globs (
covers:), its docs pages (docs:) and its journeys; the journeys are injourneys/<feature>.mjs, with the helpers inai.mjs(requests, checks, members, the sidecar's folder and lock) andguard.mjs(environment loss). - A journey starts on an empty database with the administrator onboarded. A journey that needs stored records calls
seedStore(ctx, { conversations, settings, build }): it stops the backend, empties the sidecar's folder, writesstate.json,change-sets.jsonandsettings.json, and starts the backend on them.buildrecords real change sets through the sidecar's shadow repository (recordChangeSet), so Undo and Redo move bytes on disk.sample.mjsis the shared sample: 24 calls in the last seven days of every kind (read-only, Builder, approved, rule, Full auto, blocked, denied, expired, failed, stopped), 3 reads ten days ago (the last of them a Skill call) and 2 reads forty days ago, and three change sets. The sidecar keeps its settings in the playground'snode_modules/.cache/dms-ai, so a journey that depends on them starts withfreshSidecar, which clears that folder and the run's~/.claude. A checkout therefore holds one run at a time. withEnv(ctx, { DMS_AI: "0" }, body)restarts the backend with that environment and restarts it again without.- To look at the product yourself:
$C onboard "$RUN",$C browser "$RUN" login,goto --path,snapshot,click --role --name,$C api "$RUN" <method> <path> [--as anon] [--body <json>],$C reset "$RUN". Each writes its evidence under$RUN/evidence/.
Evidence
Everything goes under $RUN/evidence/: <feature>/<journey>/NN-<step>/ with action.json, url.txt, aria.yml and screenshot.png for a browser step, request.json and response.json for a request, and checks.json for every step that checks something; <feature>/<journey>/result.json and <feature>/feature.json. Read the evidence before you judge a PASS: a check passing proves only what it asserts.
Cleanup
$C down "$RUN"
Verdict rules
- Run every feature whose
covers:globs the diff touches, plus the smoke journey.planprints that list asrun. Run every feature instead when the diff touches a shared file, and before a release. The shared files arepackage.json,pnpm-lock.yaml,pnpm-workspace.yaml,sidecar/package.json,sidecar/pnpm-lock.yaml,frontend-vue/package.jsonandfrontend-vue/pnpm-lock.yaml(the sidecar has no workspace file),.nvmrc, thetsconfig.jsonof the module and of the sidecar, and this skill's ownverify.config.ts,ai.mjs,guard.mjs,seed.mjs,sample.mjs,windows.mjs(every feature loads it throughai.mjs) andplayground/. A manifest pins the@antelopejs/interface-*contracts and the runtime every feature runs on, soplannaming onlypackageor onlysmokeunderstates the change: for a lockfile,.nvmrcor a tsconfig it names onlysmoke. - Unmapped surface fails the verdict. User-facing code is
src/**,sidecar/src/**,frontend-vue/**,skills/**and the twopackage.json, tests and tool configs excluded (verify.config.ts).mapchecks the whole tree. - Every registered journey of a feature in
runmust have run and passed. Known findings stay as failing journeys (see Gotchas): they fail a verdict only when the diff touches their feature. - Docs. Judge every
docs:page of each touched feature against what the journeys showed. - Formal checks. None.
- One-way doors. The section comes right after the result line and is mandatory.
$C doorsprints theOne-way-door:line for the diff. - Result. FAIL if any of rules 1 to 6 fail. Otherwise PASS+NOTES, with the Not driven list under
notVerified: the verdict is never a plain PASS here.
Write $RUN/evidence/judgement.json (doors, docs, notVerified, attention, verifier), then $C verdict "$RUN" --pr <n>, which writes $RUN/evidence/verdict.md. Report the verdict and what you saw in your own reply.
Gotchas
- Every dashboard page renders on the server and hydrates after
load;guard.mjswaits for the sign-in form to hydrate, andchooseandclickUntilclick again until a radio, a switch or a dialog responds. - A step whose check fails because the backend or the dashboard stopped answering ends the journey with a
NOT VERIFIEDerror and no failing check. - The sample holds one more record than every documented limit it drives (5 change sets for the 4 listed, 9 tools for the 5 listed, 29 log records for the 25 of a page, 10,005 for the 10,000 of an export, 61 skills for the 24 of a page, a usage entry on the first day of the 14 and one the day before), so that N and N+1 look different. Add the same when you add a limit.
- The 24 hour, 7 day and 30 day windows are driven with calls 3 hours, 10 days and 40 days old, not at the edge of each window.
- Text the dashboard uppercases (card and section titles) is matched without regard to case.
- Known findings, each a failing journey that fails a verdict only when the diff touches its feature:
skills/two-skills-of-one-module-with-one-name-shadow-one-copy: two skills of one module with one name are never marked Shadowed (dms-ai#169). A local skill named like a module skill is.
an-idle-sidecar-exits-after-its-idle-timeout-and-the-next-probe-starts-itsetsDMS_AI_IDLE_SHUTDOWN_MSto 20 seconds and restarts the backend with it; it takes about a minute, and nothing else may use the checkout while it runs.
Not driven
List these under notVerified in the judgement of every verdict:
- every model turn (Claude Code, Codex): no credentials in a run. That covers the chat itself, tool calls as they happen, approvals as they are asked and their rules, Full auto, plan mode, questions, queued messages, auto-fix turns, a diff a turn produces, token counts a turn reports, and the Builder and database catalog tools the agent calls. (Their stored results are driven from seeded records.)
- the chat panel, launcher, shortcuts, palette assistant and block-preview dialog: development dashboard only;
dms-control uphas no mode for it (AntelopeJS/dms#749); - block previews and custom sketches (
preview_blocks, the 64 KB and depth limits, the sketch refusals): the agent's tool, reachable only through a turn; - the ChatGPT consent page, the token exchange and refresh,
chatgpt-auth.jsonwith a token, the sign-out of a real session and the pinnedcodexCLI version gate: they need an OpenAI account and the optional peer; - Safe mode and the Builder-backed rows (Builder Connected, Safe mode enabled): the playground does not load
@antelopejs/dms-builder; - the Overview banners "The assistant's last turn failed" with Open activity and "N requests are waiting for you": they read a live turn and live approvals;
- the MCP bridge process (
src/mcp-bridge.ts) that an outside Claude Code session runs from the project's.mcp.json: the backend only resolves its path, and the journey sends the request the bridge would send toPOST /ai/mcp; - the Conversation link of the Changes page, the Open conversation button of the Activity detail, and the reload and toast after Undo: they belong to the chat session;
NODE_ENV=production: the DMS itself changes with it, so the run cannot start the backend that way;- remote access over a LAN, Tailscale, a tunnel or HTTP/2, and the six-connection browser limit (the refusal of a connection to the network address is driven);
- French strings of the pages (the locale files are checked by the repository's own
check-i18n).
04 同仓库更多 Skills
dms-page-design
Designs DMS pages that show data, through the Builder tools. Use when building or changing an analytics, stats or dashboard page or a table page, choosing a chart or a KPI for some data, picking a page's period, naming a page and its icon, or deciding how a table displays (badges, prices, relation labels, status tabs, saved views, kanban, cards, footer totals, KPI cards above the table).
dms-external-database
Connects a DMS project to a client's existing database, runs the Builder's analysis of it, settles what the analysis could not decide from the catalog and aggregates, generates TableViews through the Builder, names them and gives each the table features its data supports, and always builds an analytics dashboard first in the menu. Use when the user wants pages over a database the project does not own, PostgreSQL, MongoDB or RethinkDB ("plug in our client's Postgres", "map their Mongo database", "generate the admin for this database", "finish the mapping the Builder proposed").
dms-builder-safe
Builds or modifies AntelopeJS DMS admin pages through the Builder MCP tools instead of raw code edits. Use when the generation mode is "safe", or when the user wants predictable, typechecked page edits (add/configure/move/remove blocks, create pages and categories).
dms-ai
Sets up and troubleshoots @antelopejs/dms-ai, the AI development assistant of the AntelopeJS DMS (chat panel, sidecar running Claude Code or Codex, change sets with undo, audit log and skills pages under /modules/ai). Use when installing or configuring dms-ai (backendUrl, DMS_AI, providers, Codex), when the assistant is offline, or when a module must ship skills to it through antelopeJs.skills.
dms-ai-code-mode
Runs the dms-ai Code-mode loop for changing a DMS page in code while the user watches it in the dashboard overlay, with the list_pages, navigate_to_page, find_pages_using, typecheck and query_logs host tools. Use when the dms-ai generation mode is Code and a change touches a rendered page, or when the user allows Code mode after a Builder tool answered opaque_target or unsupported.
page-builder
Create and modify DMS pages. Use when the user wants to add or edit a page.