FormLM

Request an email verification code for FormLM login (part of auth_login method 2 — no password needed). Two-step usage: 1. Call WITHOUT `captcha` → returns a captcha image. Show it to the user and ask them to read the 4 digits. (The image is a human-verification gate — you cannot solve it yourself; the user must read it.) 2. Call WITH `captcha` (the 4 digits) → if correct, a 6-digit code is emailed to the user (valid 5 minutes). Then ask the user for that code and call auth_login with { email, code }. Server rate limits (anti-abuse): 1 send per email per 60s, 10 sends per IP per minute. On 429, tell the user to wait a minute. On 403, the digits were wrong — fetch a fresh image and try once more.

Login to FormLM. Three methods, in order of preference: 1. Access Token (recommended): ask the user to copy it from formlm.me → Workspace → Account Settings → Access Token → Copy, then pass it as `token`. 2. Email verification code (no password needed, fully in-chat, works for ALL accounts): a. Call auth_email_code { email } — returns a captcha image (show it to the user) b. Ask the user to read the 4 digits from the image, then call auth_email_code { email, captcha } c. A 6-digit code is emailed to the user (valid 5 minutes) — ask them for it d. Call this tool with { email, code } 3. email + password: only for accounts that HAVE set a password. Accounts registered via email verification code or Google have NO password. If password login fails with 401/403, do NOT retry — switch to method 1 or 2. IMPORTANT: At the START of any FormLM session (before calling formlm_generate/formlm_exec), call auth_status first. If not logged in, authenticate immediately. Do NOT wait for a 401 error.

Check current login status. Tokens expire after 7 days — if this returns "Token invalid (expired)", ask the user to copy a fresh Access Token from formlm.me → Workspace → Account Settings and re-login via auth_login.

Read-only quality inspection of ONE app — the fastest way to verify a generated app before wiring it into a page. Returns {ok, failCount, warnCount, summary, findings[]} where each finding is {name, level(pass|warn|fail), detail, fix?}. Checks: content/scoring coverage, expert enabled-but-empty or referenced-but-missing (dead AI assistant), styling applied, dimensions without report pages, unexpected certificate page, share access semantics (type=all/owner = login required — anonymous respondents get the login page, NOT the form), link expiry, and fill-in URL reachability. Set expectLang (e.g. "en", "zh-hant") to catch text written in the wrong script — the classic batch defect where an English app ends up with Chinese certificate/report boilerplate. Use deep=true to also scan report widget BODY text (extra queries; titles/labels are always scanned). Read-only: this tool never writes. Apply the returned `fix` commands via formlm_exec.

Execute a raw FormLM CLI command directly. Use this for fine-grained control that formlm_generate doesn't cover, or for modifying existing apps. IMPORTANT: All commands must start with "assess". Read formlm_skill first to ensure P0 compliance. SAFETY — destructive operations (remove/delete): Before running ANY "remove" command (app/form/scale/expert), first run the matching "query"/"list"/"find" command to show the user the EXACT name of what will be deleted, then get explicit confirmation ("Delete app 'X' (id: Y)? This cannot be undone."). NEVER call a remove command based on a vague reference ("delete it", "remove that one") without first resolving and confirming the exact target. Deletion is irreversible. Whitelisted commands (3-token granularity; anything else returns 403): App: assess app list/create/use/current/update/remove/urls Form: assess form query/find/types/config/add/update/remove/move/set-property Scale: assess scale query/find/add/update/set/remove/clear/config/keys/data Connect: assess connect query/find/types/config/cover-page/final-page/main-page/style Report: assess report query/find/update/page/widget (conditional rules: assess report widget logic add/list/remove) Expert: assess expert query/find/config/set/avatar/remove/chat Share: assess share set/query/url/api/flavor Smart: assess smart plan/execute Skill: assess skill form/scale/connect/report/expert/share NOT available here: assess snapshot / assess field * (CLI-only wrappers of the module queries) Common patterns: "assess app list --all --with-urls --json" (full inventory incl. shareToken per app) "assess app urls --app <id> --json" (URLs + published/shareType/shareToken) "assess form add --app <id> --id q1 --name \"Name\" --type radio --options \"A:1,B:2,C:3\" --json" "assess scale add --app <id> --id stress --name \"St…

Execute a single module from the plan generated by formlm_generate. Call this sequentially for each module, following the task list order from formlm_generate. Default order: form → scale → connect → report → expert → share. Takes ~30-120 seconds per module. Do NOT cancel. ## USER FEEDBACK (IMPORTANT for good UX): Before calling, tell the user: "正在生成 [module] 模块..." After calling, tell the user: "✅ [module] 模块完成" After the last module (share), present the 3 URLs to the user. ## PLAN PASSING: Plan is cached server-side (~10 min) after formlm_generate — no need to pass it manually. Just call formlm_execute with appId and module for each module in sequence. If you get "Plan not found in server cache", pass the plan JSON from the formlm_generate output into `plan` — do NOT re-run formlm_generate (that creates a brand-new app). Success is structured: the response carries appId / module / taskStatus / status(success|error) — trust taskStatus, not human wording. A retried `scale` execute appends dimensions unless existing ones are cleared first (formlm_exec: "assess scale clear --app <id>").

Generate an execution plan from natural language description (Phase 1 only — does NOT execute any module). Server runs: Plan AI (assess-plan.md) → generates task list for 6 modules (form/scale/connect/report/expert/share). Returns: appId, planType, name, plan JSON, and task list. Takes ~10-30 seconds. This tool is plan-only — the name "generate" predates the split. There is no single call that produces a live app; that is generate → execute×N (or the CLI `formlm-cli smart generate` wrapper). ## Scene Templates (when user description is vague, present these and ask them to choose): 1. assessment: 评估量表 — 多维度打分 + 分值区间解读报告 (MOST COMMON, for psych/workplace/health) 2. consultation: 咨询评估 — 评估 + AI专家对话解读 (for mental health / coaching) 3. survey: 问卷调查 — 仅收集数据，无打分 (for feedback / research) 4. exam: 考试测验 — 标准答案 + 对错判分 (for education / training) 5. report: 外部评测 — 对外部对象打分评价+报告 (for performance reviews / 360 feedback) 6. learn: 学习卡片 — 知识点 + 自测题 (for micro-learning) ## AFTER SUCCESS — execute modules sequentially via formlm_execute: The plan is cached server-side (~10 min TTL) — just call formlm_execute with appId and module for each module. No need to pass the plan JSON back — it is handled automatically. IMPORTANT: module coverage depends on planType. Only `consultation` plans include the `expert` module; assessment / exam / report / survey / learn plans do NOT. If the app must have an AI expert, create it explicitly afterwards with formlm_exec: "assess expert config --app <id> --name ... --role ... --kbText ..." (otherwise the published page promises an expert the app does not have — silent functional gap). After the ~10 min cache expires, formlm_execute still works if you pass the plan JSON back via `plan` (the full plan is returned in this tool's output) — re-running this tool would create a NEW app instead. ## USER FEEDBACK (IMPORTANT for good UX): After getting the plan, tell the user: "✅ 计划已生成！" and list the modules. Before each formlm_execute call, tell the user whic…

Get the full SKILL.md domain knowledge for a specific skill module. Same documents that AssessAgent loads internally — contains P0/P1/P2 constraints, parameter rules, examples. You can also read these via MCP resources: formlm://skills/<skillId> READ the relevant skill BEFORE constructing formlm_exec commands manually.

Get aggregated snapshot of ALL 6 app modules in a single call. Returns: form fields, scale dimensions+ranges, connect page styles, report pages+widgets, expert config, share status. Call this BEFORE making changes to understand what already exists. Call this AFTER formlm_generate to verify the generated app. Prefer summary=true for batch audits / consistency checks: it returns a compact profile (appName, fieldCount, questionCount, dimCount+dimNames, reportPageCount+reportPages, styled, expertEnabled/expertHasContent, share.type/perm/day/forever/shareToken) instead of the full module payloads, so you do not have to parse widget layoutData or markdown tables. If any module fetch fails the result carries _errors/_degraded — treat it as UNKNOWN, not as "empty module".

Report the logged-in account budget: plan, app-slot usage, and AI credits — check BEFORE batch creating. Free plan caps at 10 apps (in-use AND recycle-bin apps hold a slot; purging frees it): once full, app create / formlm_generate fail with 403 "app-limit". AI credits are billed per completed generation (capped at zero — generation never blocks). When slots are full, tell the user: purge the recycle bin at formlm.me → Workspace or upgrade (formlm.me → #/vip). Requires a server that reports usage fields; on an older server those lines come back "n/a".