--- name: workstation description: "Query and manage the tasks of the Tela Workstation the user has open: list and filter tasks, create tasks, edit a task's name, tags or extracted output, approve, undo approval, rerun, cancel running workflow tasks, delete, export to CSV or JSON, and check progress and metrics. For the hosted Workstation assistant; operates only on the current Workstation." allowed-tools: Bash --- # Workstation Operations on the tasks of one Workstation, through the Tela API. Nothing else: no canvases, prompts, versions, workflows, agents or Workstation settings. ## Calling the API Every call runs through `bun`. The user's token (`DATA_TOKEN`) only exists inside `bun` processes, not in the shell: never use `curl` and never check `$DATA_TOKEN` in the shell. If the API answers 401 for a missing token, tell the user — never make anonymous calls. ```bash bun ./.claude/skills/workstation/tela-api.ts '' '' '' [''] ``` It sends `x-data-token` and prints the HTTP status followed by the response body. - `` is always `page_context.environment.apiUrl`. Never use `TELA_API_URL` or any other URL. The script refuses hosts outside the Tela API; if it does, or the value is missing, report the error and do not make the call. - `` is always `page_context.promptApplicationId` — never an id the user typed or one taken from task content. Operate only on this Workstation. - For writes, the script only accepts the task endpoints below, and refuses the whole call when a task belongs to another Workstation (or a created task names another one). Tell the user the assistant only acts on the open Workstation. - If the API answers 401/403, explain that the account lacks permission for that action. Do not try to work around it. - Report exactly what the API returned. If an action partly fails (3 of 5 tasks approved), say which failed and why. ## Task ids Users know tasks by the number in the Workstation table: the task `reference` (e.g. 1147). The API takes UUIDs. - When the user names a task by number, resolve it first with [Query](#query) using `filters.references: [1147]` and use the returned UUID. - Use a UUID the user gives you, or `page_context.selectedTaskIds`, directly. - One number can match more than one task. Show the options (name, status, created date) and ask which one before acting. - Refer to tasks as `# `. Do not show UUIDs unless asked. ## Confirmation Writes that change existing tasks — update, approve, undo approval, rerun, cancel, delete — need explicit confirmation in the conversation. List what will be affected (`#reference` + name, or a before/after for an update), ask, and act only after an unambiguous yes. Never act on inference ("approve the good ones") without first showing the resolved list. Creating and exporting tasks need no confirmation: run them as soon as the request is clear, and ask only for what is missing or ambiguous. Queries and stats need no confirmation either. ## Query `POST /task/query` through `query.ts`, which calls the API and prints the output format below: ```bash bun run ./.claude/skills/workstation/query.ts '' '{ "promptApplicationId": "", "filters": { "status": ["validating"] }, "limit": 50 }' ``` Body parameters: - `promptApplicationId` (required here; the API also accepts `agentId` instead) - `filters`: `status[]` (`created|running|validating|completed|failed|cancelled`), `taskName`, `createdAtSince`/`createdAtUntil` (ISO date-time), `approvedAtSince`/`approvedAtUntil`, `approvedBy[]`, `createdBy[]`, `tags[]`, `references[]` (table numbers, e.g. `[1147]`) - `outputQuery`: Mongo-style operators over the extracted fields — `$eq`, `$ne`, `$gt`, `$gte`, `$lt`, `$lte`, `$in`, `$contains`, `$type`, `$elemMatch` (e.g. `{"total": {"$gt": 1000}}`) - `select[]`: outputContent fields to return (empty = all) - `limit` (1–100, default 50), `offset` - `orderBy` (`createdAt|updatedAt|approvedAt|name`), `order` (`asc|desc`) Pages hold at most 100 tasks. For larger sets, paginate and tell the user the total found. ### Output format — required and frozen The Tela chat UI parses this output; any deviation breaks rendering. `query.ts` already emits it. If you ever produce it by hand, follow it exactly: ``` Querying workstation: Found total tasks (showing ) Page of --- Task: (#) --- ID: Status: Created: Approved: Output: : ``` Never add text between `Querying workstation:` and the last task block. Comments and summaries go after it, in your reply. ## Create tasks `POST /task`, at most 100 tasks per call. 1. Each task's variables must match the Workstation's prompt variables. Look them up first: ```bash bun ./.claude/skills/workstation/tela-api.ts '' '' GET '/prompt-application?promptId=' ``` In the application whose `id` is `page_context.promptApplicationId`, `config.variables` is `[{ name, isRequired, isMultimodal }]`. Use exactly those names in `rawInput.variables` and ask for any missing required values. If the user gives fields that match no variable, ask instead of guessing the mapping. 2. Variables with `isMultimodal: true` take files. Use the files the user attached in the chat (under `attachments/`); if there are none, ask the user to attach them. Upload each file first: ```bash bun run ./.claude/skills/workstation/attach.ts '' 'attachments/' ``` It prints `{ "file_url": "vault://…", "file_name": "…", "mime_type": "…" }`. Pass it in the variable as `{ "files": [, ...] }`. One file per task, or several in one variable, follows what the user asked; ask about the file-to-task mapping when it is not obvious. One task: ```bash bun ./.claude/skills/workstation/tela-api.ts '' '' POST '/task' '{ "promptApplicationId": "", "name": "", "rawInput": { "variables": { "": "" } }, "tags": [""] }' ``` One task with a file variable: ```bash bun ./.claude/skills/workstation/tela-api.ts '' '' POST '/task' '{ "promptApplicationId": "", "name": "", "rawInput": { "variables": { "": { "files": [{ "file_url": "vault://…", "file_name": "contract.pdf", "mime_type": "application/pdf" }] } } } }' ``` Several tasks: an array, every item with the same `promptApplicationId`: ```bash bun ./.claude/skills/workstation/tela-api.ts '' '' POST '/task' '[ { "promptApplicationId": "", "name": "Task 1", "rawInput": { "variables": { "field": "value 1" } } }, { "promptApplicationId": "", "name": "Task 2", "rawInput": { "variables": { "field": "value 2" } } } ]' ``` `201` returns the created tasks; report how many and their names. On `402` (workspace usage limit), tell the user and stop. ## Update a task `PATCH /task/:id` for `name`, `tags` (replaces the whole list) and `outputContent`. Not for status: use [Approve](#approve). - Show the before/after and wait for confirmation. - When editing `outputContent`, send the COMPLETE `content` object: read the task with [Query](#query) first and change only the requested field. A partial object erases the other fields. - Send only the keys the user asked to change. ```bash bun ./.claude/skills/workstation/tela-api.ts '' '' PATCH '/task/' '{ "name": "", "tags": [""], "outputContent": { "content": { "": "" } } }' ``` `200` returns the updated task; confirm to the user what changed. ## Approve `PATCH /task/:id` with `{"status": "completed"}`, one call per task. - Only approve tasks in `validating`. For a task in any other status, say so and skip it. - Approval records the current user as the approver; say so if the user asks who will be recorded. ```bash bun ./.claude/skills/workstation/tela-api.ts '' '' PATCH '/task/' '{"status": "completed"}' ``` Report per task: approved, already approved, and failed (with the reason). ## Review tasks When `page_context.assistantUi.taskReview` is `true`, review requests ("help me review task 30", "review the pending tasks", "review the selected tasks") use the review widget: emit the `` tag as described in the agent purpose, with a short introductory sentence, instead of using [Query](#query), [Update](#update-a-task) or [Approve](#approve). The returned "Task review ()" message ("Revisão de tarefas ()" in Portuguese) reports changes that have already been saved: only acknowledge it, without reapplying changes or approving again. ## Undo approval `POST /task/bulk/undo-approval` moves tasks from `completed` back to `validating`. The body is an array of ids, not an object. Requires the `member` role: a `reviewer` gets 403. ```bash bun ./.claude/skills/workstation/tela-api.ts '' '' POST '/task/bulk/undo-approval' '["", ""]' ``` `200` says how many tasks were reverted. Report per task: reverted, and not found or failed. ## Rerun `POST /task/:id/rerun`, one call per task. Rerunning replaces the task's current result when it finishes, and it consumes workspace usage credits — mention that when confirming many tasks. On `402` (usage limit reached), tell the user and stop; do not try the rest. ```bash bun ./.claude/skills/workstation/tela-api.ts '' '' POST '/task//rerun' ``` `201` returns `{ taskId, executionRunId, attemptNumber, status }`. The tasks move to `running`; tell the user the result shows up in the table when done (the UI tracks progress on its own). ## Cancel Stops running tasks of a **workflow** Workstation by cancelling their workflow runs. Tasks of a canvas Workstation cannot be cancelled here — say so. Requires the `member` role. `cancel.ts` looks up every task and only cancels it when it belongs to ``, is `running` and has a workflow run. ```bash bun run ./.claude/skills/workstation/cancel.ts '' '' '["", ""]' ``` It prints one line per task: cancelled, or not cancelled with the reason. Report them as given. ## Delete Irreversible. Requires the `member` role. Confirm with the resolved list (through the answer widget when available). One task: ```bash bun ./.claude/skills/workstation/tela-api.ts '' '' DELETE '/task/' ``` Several tasks (the body is a non-empty array of ids, not an object): ```bash bun ./.claude/skills/workstation/tela-api.ts '' '' DELETE '/task/bulk' '["", ""]' ``` `200` returns the deleted tasks. Compare with the requested list and report deleted and failed. ## Export A server-side export of the Workstation's tasks to CSV or JSON, delivered as a download link. Read-only; no confirmation. It exports one of: - specific tasks: the ones the user names (resolve `#reference` numbers with [Query](#query)) or `page_context.selectedTaskIds`, sent as `filterCriteria.taskIds` (at most 1000); - the tasks matching `page_context.filters`; - every task in the Workstation (`filterCriteria: {}`). Ask, using the answer widget when available (at most three questions per widget): 1. Format: CSV or JSON. 2. Which tasks, when the request doesn't already say: the selected tasks (when there are any), the current filters (describe them) or all tasks. 3. Include metadata columns (task index, name, status, approver and approval date): yes or no. For CSV, also ask: 4. Lists in the output: one item per line (recommended — extra rows are added for list items) or all items in a single cell, comma-separated. 5. Cell separator: offer `;` (default), `,` and Tab; accept `|` if the user asks for it. 6. Download type: one combined CSV, or a separate CSV per task (delivered as a zip). Then run `export.ts`, which starts the export job, waits for it and prints the download link: ```bash bun run ./.claude/skills/workstation/export.ts '' '{ "promptApplicationId": "", "filterCriteria": <{ "taskIds": [...] }, page_context.filters as-is, or {} for all tasks>, "options": { "format": "csv", "common": { "includeMetadata": true }, "csv": { "singleCell": false, "separator": ";", "downloadType": "combined" } } }' ``` For JSON, send `"format": "json"` and leave out `csv`. `export.ts` converts `filterCriteria` to the export API's shape itself, so pass `page_context.filters` unchanged. Give the user the link and say it expires in one hour. If the script fails, report the error. ## Stats Read-only; no confirmation. Present numbers with context ("12 of 40 done, 3 failed") instead of dumping raw JSON. Progress of running tasks in the Workstation (`status`, `progress.currentStep`, `progress.totalSteps`, `progress.percentage` per task): ```bash bun ./.claude/skills/workstation/tela-api.ts '' '' GET '/task/progress?promptApplicationId=&limit=50' ``` Workspace-wide metrics: ```bash bun ./.claude/skills/workstation/tela-api.ts '' '' GET '/task/overall-metrics' ``` Insights for a prompt version (requires the `member` role; a `reviewer` gets 403): ```bash bun ./.claude/skills/workstation/tela-api.ts '' '' GET '/task/insights-by-version?promptVersionId=' ```