Uipath API Workflow

UiPath API Workflow Assistant

Build, run, and publish UiPath API Workflows — JSON files conforming to the CNCF Serverless Workflow DSL 1.0.0 with UiPath activity-type extensions. Executed by @uipath/api-workflow-executor via uip api-workflow run. Packaged as Type: "Api" projects via uip solution pack.

TDD gate — read this first (rule 22). On ANY request to create or edit a workflow, Phase 0 first checks whether the project has an evals/ folder — it lives inside the project directory, next to Workflow.json (<project>/evals/), not at the workspace or solution root, so list the project directory itself. No evals/ folder → the Evaluations feature is not enabled for this project: skip everything about tests and loop mode and author normally. evals/ present → end your turn with two questions — (1) tests: the eval set has rows → "should anything change, or use as-is?" / it is empty → "want me to add some?"; (2) loop mode: "run the tests and re-try until every case passes, or author once?" — and NOTHING is authored until the user answers. Tests come first: once answered, write or update the eval set, then author the workflow to make it pass. If you are about to write Workflow.json in a project with evals/ and the user has not answered these two questions in this conversation, stop. Later change requests that alter behavior can break the tests — ask before editing whether to update them (rule 22, references/testing-and-evals.md §3). Only a prompt that explicitly says not to ask skips the gate when evals/ exists — and that skips the questions, not the run consent (rule 21): run rows only when the request itself asks for a run.

When to Use This Skill

• User wants to create or edit an API workflow JSON file

• User wants to run an API workflow locally with uip api-workflow run
• User wants to package an API workflow project into .nupkg / solution .zip
• User wants to publish an API workflow to UiPath Cloud / Orchestrator

• User asks about activity types (Sequence, Assign, JavaScript, If, ForEach, DoWhile, Break, TryCatch, Wait, Response, File to Base64, Base64 to File, HTTP Request, Connector)
• User wants to handle a file in an API workflow — take a file input, send a file as base64 to an API, turn a base64 payload back into a file, or asks about JobAttachment, $helpers.file, serializeData(), --input-file / --output-dir. See references/files-and-base64.md
• User asks about nested control flow — If inside ForEach, TryCatch around a loop, conditional Break, multi-way branching, etc.
• User asks for an Integration Service connector activity (Gmail Send Email, Outlook Get Newest Email, GitHub Search Issues, Slack Send Message, etc.) — follow the discovery flow in references/connector-activity-discovery.md
• User asks for a generic HTTP Request that needs to render in StudioWeb's designer — same discovery flow
• User asks about JavaScript expressions, $context, $input, $workflow, WorkflowStart, or the export.as pattern

• User asks how to debug a failing API workflow run — the local validate → run --no-auth loop, or a post-publish cloud run (job logs/traces). See references/operating-published-workflows.md
• User wants to operate a published workflow — invoke it (HTTP/schedule/Integration Service event trigger), start/list/stop its Orchestrator jobs, or manage the Integration Service connections it uses (uip is connections list/ping/edit). See references/operating-published-workflows.md

• User wants to test / evaluate a workflow — author (inputs, expectedOutput) cases in the project's evals/ folder (the dataset Studio Web's Evaluations panel reads), run them, score them by the project's evaluator, or "loop until the tests pass". Follow references/testing-and-evals.md. An evals/ folder alone is not the signal — it is an API-workflow one only when it sits next to Workflow.json and uses the evals/<scope>/eval-sets/ layout; low-code agents use evals/eval-sets/ (no scope) → uipath-agents, Flow evals → uipath-maestro-flow
• Any authoring request (create or edit) — Phase 0 discovery checks for an evals/ folder; when it exists, the two test questions (change / add tests? loop mode?) are asked up front — the turn ends there — whether or not the user mentioned testing (rule 22). No evals/ folder → the feature is off, author normally. See Phase 0 and references/testing-and-evals.md

Do NOT use for: .flow Maestro flows (→ uipath-maestro-flow), .xaml / coded RPA (→ uipath-rpa), coded agents (→ uipath-agents), Coded Web Apps (→ uipath-coded-apps).

Core Principles

0. Tests first — when the project has evals/, and only after asking. Discover <project>/evals/ (next to Workflow.json). Absent → the Evaluations feature is off for this project: author normally, no test or loop-mode questions. Present → ask the two rule-22 questions, end the turn; once answered: declare output.schema, eval set first, workflow second, then run the rows only in loop mode (that "yes" is the rule-21 consent; "author once" means hand over without running). Behavior changes later on re-open the question — the tests may need updating.
1. Know before you write. Read the existing workflow file before editing. Read an example template before creating from scratch.

2. Start minimal, iterate to correct. Add one activity at a time. Run with --no-auth --output json after each addition. Fix what breaks. Repeat.
3. Validate before running. uip api-workflow validate is the offline static pre-flight (autonomous); uip api-workflow run is the runtime validator that catches what static analysis can't (live HTTP, expression evaluation, connection state) and needs user consent. See rules 20–21.

4. Fix errors by category. Triage: Structure > Expression > Activity Config > Logic. Higher-category fixes often resolve lower-category errors automatically.

Critical Rules

Rule 0 — Escalate big design forks before you build (highest priority, read first). When the happy path doesn't work out of the box and the resolution is a judgment call the user would reasonably want to own, STOP and ask before committing to a branch. Present the concrete options with their trade-offs and a recommended default; proceed only on the user's answer. Triggers (non-exhaustive): no valid connection for a required activity (rule 16); no curated activity exists and the choice is generic activity vs. raw Http kind vs. a different connector; the requested operation isn't exposed by any resolvable activity and the fallback is a hand-built HTTP call against an undocumented endpoint; an input the prompt assumed is missing and the alternatives are placeholder, hardcoded value, or new workflow input; the prompt is satisfiable by structurally different workflows (single connector call vs. ForEach over a list). This does NOT cover mechanical choices with an obvious answer (variable names, activity key suffixes, export-pattern selection) — decide those and move on. Reserve escalation for forks where guessing wrong wastes work or ships something the user didn't intend.

Ask the fork BEFORE branch-specific research, not after. Once you spot a structural fork, do only the shared work needed to surface the options (the cheap resolve that proves no curated activity exists, the connections list/ping that proves no connection works), then ask. Do NOT pre-research every branch — stubbing each candidate activity, describing resources, drafting alternative workflow shapes — so the user can "pick from finished work." The user picks one branch; deep work on the others is thrown away. Sequence: detect fork → minimal shared discovery → ask → then research and build only the chosen branch.

1. Workflow file is JSON, not YAML. Top-level keys: document (with dsl: "1.0.0"), evaluate (language: "javascript", mode: "strict"), do (one root sequence — named Sequence_1 in the template skeleton, but the literal key may differ in existing workflows; always read the actual key from the file before editing — containing WorkflowStart + user activities). See references/workflow-file-format.md.
2. WorkflowStart is always the first activity inside the root sequence's do array. It hydrates variable defaults into $context.variables and forwards inputs to $input. Never remove, rename, or modify it. isTransparent: true (only WorkflowStart uses true).
3. Every activity is a single-key object wrapped in the do array: { "<ActivityKey>": { ...activity body... } }. Activity keys must be globally unique across the whole workflow — including #Wrapper, #Then, #Else, #Body suffixes.
4. Every activity should export its output to propagate state. Two patterns:
   • Variables (Assign only): { ...$context, variables: { ...$context.variables, ...$output } }
   • Outputs (everything else): { ...$context, outputs: { ...$context?.outputs, "<ActivityKey>": $output } } See references/expressions-and-context.md.

5. String literals in Assign.set / Response / If when MUST be wrapped as "${'literal'}" — a JS string inside an expression. Plain "literal" runs fine under uip api-workflow run, but StudioWeb's designer normalizes unwrapped values to "${literal}" on save (treating them as expressions you typed into the property panel). At runtime the bare identifier literal has no binding → ReferenceError: literal is not defined. Use single quotes inside the expression to avoid JSON escaping: "set": { "tier": "${'PLATINUM'}" }. Numbers, booleans, and references like ${$context.variables.X} need no extra wrapping. (Response payloads have a related but distinct constraint — see rule 15.) Scope: this rule applies to Assign / Response / If / variable contexts only. It does NOT apply to connector bodyParameters / queryParameters / pathParameters — those take BARE literals; ${'...'} there is read as an expression and the field is cleared on save. See rule 16 and references/connector-activity-discovery.md#field-shape-rules-flat-keys-bare-literals-renamed-export-hub-prefix. See references/troubleshooting.md.

6. Each Assign activity MUST set exactly ONE variable. Assign.set is a single-key object, NOT a multi-variable update. StudioWeb's designer collapses multi-key set blocks to one key on save, silently dropping the others — the runtime then only updates the surviving key. To update N variables, use N separate Assign activities placed sequentially in the same do array. Example: instead of "set": { "sum": "${$context.variables.sum + 1}", "count": "${$context.variables.count + 1}" } (loses count after StudioWeb save), write two Assigns — Assign_Sum with "set": { "sum": "${...}" } and Assign_Count with "set": { "count": "${...}" }. Each runs in order; each Assign's variables export merges its single key into $context.variables.
7. If activity requires the wrapper pattern. If_N#Wrapper contains If_N (switch), If_N#Then, If_N#Else. Both #Then and #Else MUST end with "then": "exit" to prevent fall-through. Conditions in when MUST be wrapped in ${...}. For deeply-nested If patterns and multi-way branching, see references/control-flow-patterns.md.
8. Loops (ForEach, DoWhile) require a #Body element inside do. ForEach body uses index-aware accumulation (resets on iteration 0); DoWhile body uses simple accumulation. Loop variables (each, at) are plain strings, NOT expressions.
9. DoWhile for.in is always "${ [1] }". The doWhile condition controls repetition. The body MUST update the condition variable, otherwise the loop runs forever.
10. Nested loops MUST use distinct iterator/index names. Outer for.each: "outerItem", inner for.each: "innerItem". Reusing currentItem shadows the outer. "Distinct" just means "not the same string" — semantic (outerItem / innerItem) and incremental (item1 / item2, currentItem / currentItem2) naming both work.
11. Loop iterators and catch error variables are prefixed with $ in expressions. Declare for.each: "currentItem" (plain string, no $); reference it everywhere else (in when conditions, in script bodies, in set expressions, in body export patterns) as $currentItem — the $ is a literal character in the global identifier name. currentItem is not a reserved name — for.each: "customer" binds $customer, for.each: "row" binds $row, etc. Same shape for for.at ($currentItemIndex, $idx, etc.) and catch.as ($error, $err, etc.). Empirically verified: the executor calls setVariables({"$currentItem": item, ...}) — currentItem (no $) is not bound as a global. Forgetting the $ produces <name> is not defined.
12. Break exits only the innermost enclosing loop. To exit nested loops, set a flag variable + check it in the outer loop. Break value MUST be the string "true", with then: "exit" and set: "${$input}". Only valid inside a #Body.
13. Use $workflow.input.<name> to read workflow inputs, never $input.<name>. $input is the task's input — for any non-first task, it's the previous task's output, NOT the workflow arguments.
14. JavaScript scripts read $context/$workflow/$input as globals. Scripts MUST return a value. The task's run.script.arguments field is StudioWeb designer scaffolding — keep it as the standard "${{ \"$context\": $context, \"$workflow\": $workflow, \"$input\": $input }}" block for designer roundtrip; the runtime ignores it.
15. Response activity shape — STRICT for StudioWeb roundtrip:
    • markJobAsFailed is a sibling of response, not nested inside it.
    • Always include "then": "end" — without it, the workflow does not terminate properly. then: "end" is for Response only; then: "exit" is for control-flow branches/loops.
    • Object-valued responses MUST use the single-expression form, NOT the JSON-object-with-${}-fields form. StudioWeb's designer corrupts the latter on save.
      • ✗ Wrong (CLI runs but StudioWeb corrupts): "response": { "tier": "${$context.variables.tier}", "count": "${$context.variables.count}" }
      • ✓ Correct: "response": "${{ tier: $context.variables.tier, count: $context.variables.count }}" Inside the outer ${{ ... }} you are already in expression scope, so reference variables/outputs directly without an inner ${...} wrapper. JS object literal keys can be unquoted identifiers (tier:, count:); literal string values use single quotes (status: 'ok'); numbers/booleans/references are bare. The designer leaves an already-wrapped single expression alone; the JSON-object form gets flattened to a stringified expression where inner ${...} substitutions are inside JS double-quoted strings (which don't interpolate), turning each field into the literal text of its expression.
      • Either "${ { ... } }" (single-brace, expression-of-object-literal) or "${{ ... }}" (double-brace, object-literal-expression form) is valid — both evaluate to the same JS object. Pick one and stay consistent within a workflow.
    • For single-value responses (returning one variable or one expression), the simple form is fine: "response": "${$context.outputs.Javascript_1}" or "response": "${'done'}".

- **On-disk is authoritative.** Even with the single-expression workaround, every StudioWeb designer save can re-trigger normalization passes that may corrupt the Response shape. After any designer roundtrip, re-validate with `uip api-workflow run --no-auth` and re-apply the workaround if needed. Until the designer fix ships, treat the file on disk as truth, not what the designer renders.

16. Connector activities (HTTP + Integration Service) come from uip api-workflow registry resolve + stub — never hand-author or guess. The stub computes metadata.configuration, the kind (UiPath.Http vs UiPath.IntSvc), the endpoint (with hub prefix), SlotKey, and ExportBucketKey (which can differ — HTTP slot HttpRequest_1 vs bucket http_request_1). Use all of them verbatim; NEVER invent a uiPathActivityTypeId, hand-author metadata.configuration, or reconstruct a key from objectName. Non-negotiables (full step-by-step, field-shape rules, multipart, and worked examples in references/connector-activity-discovery.md):
    • A keyword resolve miss is NOT proof no curated activity exists — verify connector-first before giving up. resolve AND-matches every token, so a marketing phrase + guessed verb over-narrows (the product "UiPath Data Fabric" carries connectorKey: uipath-uipath-dataservice and activity names like "Create Entity Record" — resolve "data fabric insert" returns 0; fewer/truer tokens, not more). Before concluding none exists or falling back to a hand-built HTTP call (a Rule 0 fork): map the product/vendor → connector key with uip is connectors list --filter "<product>", then enumerate with uip is activities list <connector-key>. Do NOT hardcode/guess the key — look it up. See the reference's Step 1 recovery.
    • IntSvc/vendor activities require a pinged connection. uip is connections ping <uuid> must succeed before authoring — listing-state ≠ runtime-state; an Enabled connection can still 401 in cloud. An empty listing is NOT proof no connection exists — uip is connections list is folder-scoped. On empty/failed listing, walk the fallbacks in order: unfiltered uip is connections list, then uip is connections list --all-folders (catches connections in other folders), re-pinging a different Id for that ConnectorKey each time.
    • No connection pings cleanly → STOP and ask the user — do not decide alone. Offer: (a) continue with a placeholder (stub without --connection-id, leaving the <REPLACE_WITH_VENDOR_CONNECTION_UUID> sentinel — workflow is structurally complete but 401s until replaced; only with explicit user consent), or (b) stop and wait for the user to create/fix the connection, then re-ping. Never silently emit the placeholder, never silently abort. (Instance of Rule 0 — escalate design forks.)
    • NEVER ship a <REPLACE_WITH_*> placeholder in with.connectionId / connectionResourceId / Http bodyParameters.url. StudioWeb renders it as a broken connection and the workflow 401s. The placeholder is a sentinel for "re-stub with the real value," not a fill-in-later field.
    • After every stub, cross-check required fields — the stub drops required: true request fields (e.g. Outlook getNewestEmail needs parentFolderId). Confirm via uip is resources describe ... --operation <op> or the stub's own metadata.configuration inputFields; re-stub with --inputs if missing.
    • Connector params use flat dotted keys and BARE literals. "message.toRecipients": "...", not nested objects; plain "x@y.com", not "${'x@y.com'}" — rule 5's wrap is inverted here (${'...'} clears the field on save). Real references (${$context...}) stay wrapped.
    • NEVER use Http kind with a vendor connection UUID (401 "Invalid Element token"). IntSvc output is wrapped: read $context.outputs.<ExportBucketKey>.content.<field>.

- **(Solutions-mode + IntSvc only)** sync the connection into the catalogue: `uip api-workflow bindings sync --workflow <Workflow.json>` then `uip solution resource refresh --solution-folder <path>`. Skip for Http kind, non-connector activities, and standalone (no `Solution/`) projects.

17. Pass input as a JSON string. --input-arguments '{"key":"value"}'. Invalid JSON exits 1.
18. Always --output json when parsing CLI output programmatically. Success → { "Result": "Success", "Code": "WorkflowRun", "Data": {...} }. Failure → { "Result": "Failure", "Message": "...", "Instructions": "..." } with exit 1.

19. Scaffold with uip api-workflow init; publish goes through the solution packager. Create every API workflow project with uip api-workflow init <name> (rule 19a) — never hand-assemble the project files. Project-level CLI commands also exist: uip api-workflow build <projectDir> (compile) and uip api-workflow pack <projectDir> <outputDir> (single-project .nupkg, useful to test one project in isolation). Solution-level build/publish go through uip solution pack <solutionDir> <outputDir> + uip solution publish <package.zip>. There is NO uip api-workflow publish command. Project type must be "Api" in the solution .uipx.

19a. Create projects with uip api-workflow init <name> — it produces the correct Studio Web editable shape and wires the solution. Run it from inside the solution directory (the folder containing the .uipx): bash uip api-workflow init <name> --output json # add --skip-solution-registration for a standalone (no .uipx) project It scaffolds project.uiproj + Workflow.json + entry-points.json + bindings_v2.json and, when run inside a solution, auto-registers the project in the surrounding .uipx (correct ProjectRelativePath + a fresh Id). Success → Code: "ApiWorkflowInit". Then edit Workflow.json only.

**Which mode.** Default = `init` inside a solution (Studio Web-editable + deployable — what a shipped automation needs). Use `--skip-solution-registration` ONLY when the user explicitly wants a CLI-only/local workflow that never opens in Studio Web or ships in a solution; it still emits the full project folder (`<name>/Workflow.json` + siblings), just no `.uipx` wiring. Never emit a lone `Workflow.json` with no project files — even a throwaway local workflow gets a project.

**Why it matters:** a legacy `project.json` + `workflows/WF_*.json` layout (no `.uiproj`) passes every runtime gate — `validate`, `run`, `pack`, `publish`, deploy — but Studio Web rejects it as `invalid_project_folder` and never shows it. `init` is the one step that can't produce the wrong shape. Full layout + field rules: [references/workflow-file-format.md](references/workflow-file-format.md#project-structure-studio-web-editable-contract).

To **convert a legacy `project.json` project**, `init` a fresh sibling and move the existing workflow content into its `Workflow.json` (cleanest), or convert in place — see [references/troubleshooting.md](references/troubleshooting.md). Never wire it with `uip solution projects add/remove` (errors on an already-registered name; `remove`+`add` destroys the project `Id`).

20. uip api-workflow validate <Workflow.json> is the autonomous closure step for every authoring or edit cycle. Run it as the LAST command before asking the user anything about runtime. It's offline (no auth, no network, no side effects): JSON Schema + semantic checks on the static file. Output codes:

    • Result: "Success", Code: "ApiwfValidate", Data.Status: "Valid" (exit 0) — possibly with Data.Warnings. Proceed to rule 21 (ask the user whether to run).
    • Result: "Failure" (exit 1) — do NOT bother the user. Read Instructions, locate the offending activity by its JSON path (e.g. /do/0/Sequence_1/do/2/Mystery_1/metadata/activityType), edit Workflow.json to fix it, then re-validate. Loop until pass.

    Reading the error list. AJV schema errors from oneOf branches produce duplicate "Missing required property" noise (each unmatched variant lists all its required fields). Focus on the semantic-tail errors — the ones with prose messages like Unknown activityType 'X', must contain a 'do' with inner 'switch', is missing 'metadata.configuration', Variable must have a non-empty 'type'. Those uniquely identify the root cause. Fix one root cause, re-validate, repeat — don't chase the schema-level fanout one by one.

**What validate catches:** malformed JSON; unknown `activityType` values (see VALID_ACTIVITY_TYPES list in the validate source); per-activity required keys (If → `do` + inner `switch`, Sequence → `do`, Assign → `set`, ForEach → `for` + `do`, DoWhile → `for` + `doWhile`, Connector → `call` + `metadata.configuration` + `essentialConfiguration`, Response → `response`, etc.); missing `metadata.activityType`/`displayName` (warnings); bad `evaluate.language`/`evaluate.mode`; duplicate or empty-named workflow variables; empty task lists. **What it does NOT catch:** wrong `selectedResourceId`, broken connector connection IDs, runtime expression errors (`ReferenceError: x is not defined`), unwrapped string literals (rule 5), multi-key `Assign.set` (rule 6) — those still need runtime validation via `uip api-workflow run` once the user consents.

21. Never run uip api-workflow run without an explicit user "yes." Validation (rule 20) is autonomous; running is not. Once validate passes, ask the user: (a) run now or skip, (b) if running, with --no-auth (fast, structure-only — IntSvc kind vendor calls fail) or with auth (real Integration Service calls — vendor side effects WILL happen: emails sent, tickets created, files uploaded). Suggest a default based on workflow content (--no-auth for control-flow-only + Http kind ImplicitConnection; with-auth for any IntSvc kind vendor activity), but wait for the user's answer. Never invoke uip api-workflow run with auth on speculation — once a vendor call goes out, it can't be unsent. This rule gates only uip api-workflow run. If the user asked to package or publish, do not stop on the run question: validate, then pack (rule 19). A local run is a separate request; do it only when the user asks for it.

22. TDD gate — only for projects with an evals/ folder; STOP and ask before authoring; tests are written before the workflow. Every authoring request (create OR edit — including a one-line change) runs Phase-0 discovery of the project's evals/ folder — <project>/evals/, the directory next to Workflow.json (list that directory; the workspace or solution root is the wrong place to look):

    • No evals/ folder → the Evaluations feature is not enabled for this project. Skip the gate entirely: do not offer to create tests, do not mention loop mode, do not create the folder — author normally (Phases 1–3).
    • evals/ present → end the turn with two questions. Do not write or edit Workflow.json, and do not create or change eval files, until the user has answered both:
      1. Tests — the eval set has rows → "I found N test case(s): [one line each]. Should anything change, or use them as-is?"; the set is empty (the panel seeds one with 0 rows) → "There are no test cases yet — want me to add some? I'd suggest: [2–3 proposed cases]."
      2. Loop mode — "Run in loop mode (run the tests and re-try — fix, re-run — until every case passes), or author once and you verify?" A "yes" to loop mode is also the consent to run the rows with --no-auth (rule 21) — do not ask "run now?" separately; connector/authed runs still need their own explicit "yes".

    Order after the answers (TDD): first declare input.schema / output.schema in Workflow.json (init scaffolds output: null, so the row keys have no other source — property names and casing come from here, and the Response must emit exactly those keys), then write or update the eval set, then author the workflow, then run the rows (loop mode) or hand over (author once — do not run). Later change requests that alter behavior (threshold, rounding, branch order, a new default — even with the schema unchanged) can make existing expectations wrong: before editing, list the affected rows and ask whether to update them; after editing, re-run the rows. Ask as two specific questions — never one vague "how do you want to verify it?". With evals/ present, the gate is skipped only when the user already answered, or explicitly asked not to be prompted ("don't ask for confirmation"); then keep existing tests as-is and add tests only if requested. "Don't ask" waives the two questions, not rule 21's run consent: run the rows only if the request itself asks for a run or a loop ("run the evaluations", "until they pass" — that wording is the consent); otherwise author once and hand over the exact command to run them. Protocol and wording: references/testing-and-evals.md §3.

23. Files are references, not bytes — and base64 is a file too. A file input or output is a JobAttachment ({ ID, FullName, MimeType, Metadata? }) pointing at a blob in Orchestrator storage; $workflow.input.<file> is that object, never the bytes.

    • The two activities. Both are run.script tasks; the $helpers.file.* call in the script is what identifies them (it is what validate checks). File to Base64 (metadata.activityType: "FileToBase64", code return { output: await $helpers.file.fileToBase64(<file ref>) }) returns a NEW reference whose blob content IS the base64 text (<name>.base64, text/plain, Metadata.Encoding: "base64"). Base64 to File (metadata.activityType: "Base64ToFile", code return { output: await $helpers.file.base64ToFile({ base64: <ref or string>, fileName?, mimeType? }) }) returns a binary reference. The script is that single return expression and nothing else: Studio Web rebuilds the script from the parsed call on every designer save and silently drops a preceding or trailing statement, a second argument, or an extra option key. validate warns only about the extra statements — an extra argument passes as Valid and still breaks — so put any pre-processing in a JavaScript activity before the conversion.
    • Reading results. Read either activity's result as $context.outputs.<Key>.output.
    • Inlining file content. To put a file's content INLINE in an HTTP request body or a Response field, call <ref>.serializeData() right there — it returns a deferred-read marker the engine fills at send time; never store it in a variable or use it in script logic. Nested in a JSON body it works only for a base64 reference (the File to Base64 output): a binary file's marker, or a bare reference, nested in a body is a send-time error — send a binary file as the whole body (bare reference) or convert it first.
    • Naming. fileName / mimeType apply only to a raw base64 string; a reference keeps its own name and the engine sniffs the type from bytes (a .txt round-trips to an extension-less file).
    • Running. Both helpers need Orchestrator blob storage:

`uip api-workflow run --no-auth` refuses such a workflow up front; run it signed in (`uip login`, no `--no-auth`) — it still needs the rule-21 "yes". Pass local files with `--input-file <name>=<path>` (uploaded, arriving as `$workflow.input.<name>`), collect returned files with `--output-dir <dir>` (each reference in the output gains a `LocalPath`), and `--folder-key <guid>` if the tenant's Attachments API requires a folder. In the printed output the CLI PascalCases keys (`ID` → `Id`).

Shapes, worked example and pitfalls: [references/files-and-base64.md](references/files-and-base64.md).

Workflow Phases

Phase 0: Discovery

Before touching anything, understand what exists — the workflow structure and whether the project has tests. Discovery ALWAYS includes checking for an evals/ folder inside the project directory (ls <project>/ — the folder is a sibling of Workflow.json, never at the workspace/solution root): Studio Web creates it when the Evaluations feature is enabled and the panel is opened, so its absence means the feature is off for this project. If present, read evals/<scope>/eval-sets/*.json (<scope> is default unless the project says otherwise) and the evaluator(s) in evals/<scope>/evaluators/ — the same files the Evaluations panel reads and writes. Contract and semantics: references/testing-and-evals.md §1.

Then apply rule 22 (TDD gate): no evals/ folder → the feature is off, continue to Phase 1 without any test or loop-mode question; evals/ present → end the turn with the two questions — tests (change / add?) and loop mode — and start Phase 1 only after the user has answered, writing the eval set before the workflow.

For edit requests:

1. Read the existing workflow file with Read
2. Identify activity keys already in use (avoid collisions)
3. Identify variables, inputs, outputs already declared
4. Identify export patterns in use (stay consistent)

For create requests:

1. Read assets/templates/api-workflow-template.json for the empty skeleton
2. Read a closer example based on need:
   • Conditional branching with error handling → assets/templates/conditional-workflow-example.json
   • Loops with aggregation → assets/templates/loop-aggregation-example.json
   • Heavily nested control flow (TryCatch around DoWhile around If with Break) → assets/templates/nested-control-flow-example.json
3. For nested patterns specifically, read references/control-flow-patterns.md — pattern catalog for If-in-If, ForEach-with-If, TryCatch-around-loop, conditional Break, etc.

Phase 1: Plan

Decide which activities to use and in what order.

Before generating, determine:

1. Which activities are needed and in what order
2. What unique keys to assign (check existing keys to avoid collision)
3. What variables to declare (in document.metadata.variables.schema.document.properties)
4. What inputs/outputs to declare (in input.schema / output.schema)

Phase 2: Generate or Edit

Precondition (projects with evals/): the two rule-22 questions have been answered in this conversation and the eval set (if wanted) is already written. If not, go back to Phase 0 and ask — do not write Workflow.json yet. Projects without evals/ have no such gate.

For each activity, read its reference section in references/task-types.md, copy the minimal JSON, fill in values.

For CREATE: copy from a template, then add user activities AFTER WorkflowStart inside the root sequence (literally Sequence_1.do in the template skeleton).

For EDIT: read the file first, identify the exact insertion / replacement point, use Edit with sufficient context for unique matching.

If the edit changes behavior and the project has an eval set, rule 22 applies first: ask whether to update the expectations the change would invalidate, then edit, then re-run the rows (see references/testing-and-evals.md §3 step 5).

Workflow skeleton:

{
  "document": { "dsl": "1.0.0", "name": "...", "version": "0.0.1", "namespace": "default", "metadata": { "variab

…(truncated)