UiPath Case Management Authoring Assistant

Preview — skill is under active development; surface and behavior may change.

Builds UiPath Case Management definitions from sdd.md. Generates tasks.md plan, then writes caseplan.json directly via per-plugin JSON recipes. CLI is reserved for read-only metadata fetches (registry, validate, debug, tasks describe, is describe) and solution boundary operations (uip solution new / project add / upload).

When sdd.md is absent, Phase 0 interview generates one interactively (listen → sketch → progressive ask-walk → resolve → approve, with optional HTML preview before handing off). Complex / multi-product cases redirect to uipath-solution-design — see references/phase-0-interview.md § Thresholds for caps.

Scope: new case from sdd.md (user-provided or Phase 0-generated). Modifying existing case not supported (no remote fetch tooling).

When to Use This Skill

• User provides sdd.md and wants Case Management project built
• User asks to create new case management project but has no sdd.md (Phase 0 interview generates one)
• User asks to create new case management project or definition
• User asks to generate implementation tasks from sdd.md or convert spec to plan
• User asks about case management JSON schema — nodes, edges, tasks, rules, SLA
• User wants to manage runtime case instances (list, pause, resume, cancel) — see references/case-commands.md

Do not use for: .xaml → uipath-rpa. .flow → uipath-maestro-flow. Standalone agents/APIs/processes outside case context → corresponding UiPath skill.

Critical Rules

1. Phase 0 interview when sdd.md absent. Generate sdd.md via a guided interview (listen → sketch → progressive ask-walk → resolve → approve); output requires explicit user approval (Approve hard-stop) before treating as Rule 2 input. Apply complexity thresholds — soft-redirect to uipath-solution-design on breach. Never overwrite an existing sdd.md. See references/phase-0-interview.md § Thresholds.
2. sdd.md is sole input post-Phase-0. After Phase 0 approval (or when user-provided), trust as written. Skill does not validate or gap-fill. If ambiguous, use AskUserQuestion — never infer silently.
3. Run uip maestro case registry pull before planning. Discovery reads cache files at ~/.uip/case-resources/<type>-index.json directly. registry search has known gaps (esp. action-apps). See references/registry-discovery.md.
4. --output json on every parsed read.
5. Follow plugin per node type. Open matching planning.md during planning + impl-json.md during execution. Never guess JSON shapes from memory.
6. tasks.md declarative only. No shell commands inside. Field names use plain identifiers (e.g., type:, displayName:, lane:), not CLI flag syntax. One T-entry per sdd.md declaration — every stage, edge, task, trigger, condition, SLA rule, variable, and argument gets own T-number, even when value looks like default (current-stage-entered, case-entered, exit-only, is-interrupting: false, runOnlyOnce: true, marks-stage-complete: true). Never group, never silently omit. When an sdd.md row’s format is unrecognized, ambiguous, or cannot be categorized — invoke AskUserQuestion before skipping. Silent omission is forbidden. Always regenerate from scratch. See references/planning.md §4.0.
7. HARD STOP after tasks.md. AskUserQuestion: Approve and proceed / Request changes. Re-read tasks.md before executing.
8. Unresolved resource → placeholder, never fabricate IDs. Keep <UNRESOLVED: ...> markers in tasks.md. Placeholder task: node with type + displayName + structural fields, data: {}; conditions still reference the TaskId. Exception — action placeholder: data.taskTitle is validator-required; populate from sdd.md task-title hint or fall back to displayName; include data.priority / data.recipient if known. Placeholder event trigger: node with render fields + data.uipath: { serviceType: "Intsvc.EventTrigger" } only (no other data.uipath keys); entry-points.json entry appended; trigger-edge to first stage created. See references/placeholder-tasks.md and references/plugins/triggers/event/impl-json.md § Placeholder fallback.
9. Persist every registry resolution to registry-resolved.json — search query, all matches, selected result, rationale.
10. Cross-task refs: "Stage Name"."Task Name".output_name in planning, resolve to =vars.<outputId> at execution by reading source’s id field (the resolver match key; var mirrors id on self-declaring task outputs but .id is symmetric with the runtime resolver per io-binding validator). Discover output names via uip maestro case spec (connector tasks) or uip maestro case tasks describe (non-connector tasks) — never fabricate. See references/bindings-and-expressions.md and plugins/variables/io-binding/impl-json.md.
11. HARD STOP between Phase 2 (Prototyping) and Phase 3 (Implementation) — unconditional, every run. Run validate --skeleton (structural checks only — skips tasks/SLAs/escalations/entry-exit rules), surface counts, present AskUserQuestion: Publish for review / Skip publish and continue / Abort. Do NOT halt on Phase 2 validate errors — advisory only; user inspects via Abort. Never skip prompt for auto mode, non-interactive mode, prior approval. If harness forbids prompts, halt with error. On Publish for review: print DesignerUrl as plain-text output BEFORE invoking the second AskUserQuestion — never embed URL only inside question body. Additional hard stops gate Phase 4 retry exhaustion (Retry with fix / Pause for manual edit / Abort), Phase 5 entry (Run debug session / Skip to Publish), and Phase 6 entry (Publish to Studio Web / Done). Full contract in references/phased-execution.md.
12. Never run uip maestro case debug automatically. Executes case for real — emails, messages, API calls. Explicit user consent only.
13. All skill artifacts: Read + Write/Edit only. Applies to caseplan.json, sdd.md, sdd.draft.md, tasks.md, tasks/registry-resolved.json, tasks/trigger-spec-cache.json, bindings_v2.json, id-map.json, entry-points.json, build-issues.md. No python, node, jq, sed, awk, or scripts that open/parse/modify/save these files. Specifically forbidden (common slip): node -e "...fs.writeFileSync...", node -e "...fs.readFileSync...", node -e "..." > <artifact>, jq '...' <artifact> > <artifact>, python -c "...open(...,'w')...", sed -i, awk -i inplace, or any shell redirection (>, >>, | tee) onto a skill artifact regardless of interpreter. Writing a helper script under /tmp or anywhere else to assemble a skill artifact is also forbidden — the build-assembler pattern (/tmp/build-caseplan.js, /tmp/gen-tasks.py, etc.) is the same Rule 13 violation as inline node -e, regardless of “mechanical copy” or “avoid Read+Write churn” framing. If caseplan.json exceeds ~30KB and a single Write feels too large, split into the Phase-2-skeleton-then-Phase-3-fill cadence (per case-editing-operations.md § Per-section batch write contract) — never via helper script. Bash subprocesses OK ONLY for UUID v4 generation (node -e "console.log(crypto.randomUUID())" for operate.json.projectId and entry-points.json uniqueId — subprocess MUST NOT require('fs') or use redirection), CLI metadata fetches, validate, debug, and solution scaffold/upload. Prefixed IDs (Stage_, t, Rule_, etc.) are picked inline by the agent — no subprocess. See references/case-editing-operations.md § Tool usage.
14. Always run uip solution resource refresh before uip solution upload or uip maestro case debug — syncs resources from bindings_v2.json so Studio Web can resolve connector dependencies.
15. Never auto-invoke uipath-solution. On Phase 0 threshold breach or stuck-round detection, print plain-text suggestion of the skill name. User re-invokes manually. No tool-call cross-skill handoff.

16. Caseplan task type enum is closed — 9 values, schema-kebab. Any task node written into caseplan.json MUST have type exactly one of: process

    agent

    rpa

    action

    api-workflow

    case-management

    execute-connector-activity

    wait-for-connector

    wait-for-timer. Never write the plugin folder name (connector-activity, connector-trigger) or the CLI --type flag value into the JSON node — those name the planning artifacts, not the schema. Never write external-agent, wait-for-event, or any hallucinated value — there is no plugin to back them. See references/case-schema.md § Task type and the Plugin Index naming-asymmetry table below.

17. Empty registry lookup → AskUserQuestion for force pull BEFORE any placeholder fallback. When a planning-phase lookup returns 0 matches across all relevant cache files, present AskUserQuestion Force pull and re-resolve / Skip and use placeholders BEFORE writing any placeholder T-entries or invoking per-plugin Unresolved Fallback paths. Apply per lookup-batch (one prompt covers all empties in the batch — do not prompt per-task). Do NOT pre-judge based on resource-name heuristics (“looks vendor-specific, won’t match anyway”) — that is the user’s call. Placeholder fallback is only valid AFTER the user explicitly picks Skip. See references/registry-discovery.md § MUST: Confirm Before Placeholder Fallback.
18. Schema version: default v19, opt-in v20, alpha-tenant forced v20. Skill emits v19 ({ root, nodes, edges }) by default. Emits v20 ({ id, version: "20.0.0", name, metadata, bindings, variables, nodes, edges, layout }) when ANY of: (a) Tenant override — Phase 1 Step 1’s uip login status --output json returns Data.BaseUrl == "https://alpha.uipath.com" (exact case-sensitive match). Forces v20 regardless of user prompt; cannot be downgraded by prompt phrasing. (b) Prompt phrase — the user message that activated the skill contains one of these case-insensitive substrings: v20 schema, schema v20, use v20, emit v20, generate v20, unified schema, schema 20.0.0. Resolution order is (a) before (b); first match wins. Detection scope for (b) is the activating message only — never infer from sdd.md content, file paths, or any subsequent message. Detection scope for (a) is the Step 1 login-status response only — do not re-query mid-build. When v20 is selected by either path, Phase 1 Step 2 writes Schema: v20 as the first non-comment line of tasks.md; otherwise writes Schema: v19. Re-entry protocol re-reads this header to recover the choice across hard stops. caseplan.json self-identifies via its top-level version literal. Mid-flight schema switch forbidden — user must re-run from Phase 1 (Rule 6 regenerate-from-scratch applies); changing tenant or prompt mid-build does not re-flip the header. Cached v19 artifacts cannot be in-place migrated to v20 (or vice versa) — a full Phase 1 regenerate is required. v20 mode softens Phase 4 to informational (no retry-cap hard stop) and prints downstream-CLI-may-reject warnings before Phase 5/6 prompts. See references/phased-execution.md § v20 mode and references/case-schema.md § Top-level shape.
19. v20 layout-strip — node/edge layout state lives in top-level layout, not on the node/edge. In v20 mode, do NOT emit node-level position, style, measured, width, height, zIndex. Do NOT compute stage position.x = 100 + count * 500. Do NOT emit edge data.waypoints. Emit top-level layout: {} (empty object) — FE auto-layouts on canvas load. The frontend’s transformCaseInMemoryJsonToDiskJson strips these fields anyway when round-tripping through canvas; emitting them is harmless on read but wastes tokens. v19 mode preserves all current render-field rules (Pre-flight Checklist Items 3, 4 in references/case-editing-operations.md). See references/case-editing-operations.md § v20 layout-strip.

Workflow

Up to seven hard stops (Phase 0 + Phase 2 second prompt + Phase 4 conditional): Phase 0 (interview → sdd.md, only when sdd.md absent) → approve → Phase 1 Planning (sdd.md → tasks.md) → approve → Phase 2 Prototyping (placeholder) → publish-for-review stop → continue-after-publish stop (publish branch only) → Phase 3 Implementation (detail) → Phase 4 Validate (retry-cap stop on 3rd failure) → Phase 5 Debug (Run vs Skip-to-Publish stop) → Phase 6 Publish (Publish vs Done stop).

Phase 0 — Interview (conditional)

Triggered when sdd.md absent at resolved path. Read references/phase-0-interview.md for the interview modes (listen → sketch → progressive ask-walk → resolve → approve), thresholds, soft-redirect contract, resumption, and HTML preview offer. Produces:

Read budget for Phase 0. Read phase-0-interview.md, references/sdd-generation-rules.md (the mental model + task-type reasoning the progressive walk relies on), and assets/templates/sdd-template.md to begin the interview. Do NOT preload plugin impl-json.md files — those are needed only in Phase 2/3 and pulled in just-in-time per T-entry.

• sdd.md — generated against assets/templates/sdd-template.md
• tasks/registry-resolved.json — per-task registry resolutions
• sdd.draft.md — intermediate, deleted at approval
• sdd-viewer.html — optional, rendered from assets/templates/sdd-viewer.html when user accepts the preview offer; Phase 1 ignores it

If sdd.md already exists: skip Phase 0, hand to Phase 1 unchanged.

Phase 1 — Planning

Read references/planning.md. Produces:

• tasks/tasks.md — T-numbered entries (stages → edges → tasks → conditions → SLA)
• tasks/registry-resolved.json — audit trail

HARD STOP: AskUserQuestion approval. Loop on Request changes.

Phase 2 — Prototyping

Read references/implementation.md + references/phased-execution.md. Builds structural shape only:

1. Solution + project + root case (Step 6)
2. Triggers — manual / timer / event, including placeholder event triggers per Rule 8 (Step 6.1)
3. Global variables + arguments (Step 6.2) — including In arguments whose elementId references a TriggerId captured in Step 6.1
4. Stages (Step 7), edges (Step 8)
5. Tasks — shape only (Step 9): non-connector with full data.inputs[] schema + empty values; connector with typeId + connectionId only (no is describe); unresolved as placeholders per Rule 8
6. Informational validate (Step 9.5.1) — do NOT halt on errors/warnings
7. HARD STOP (Step 9.5.2–9.5.5): Publish for review / Skip publish and continue / Abort. On Publish: uip solution resource refresh --solution-folder <SolutionDir> --output json then uip solution upload, print DesignerUrl, AskUserQuestion: Continue to phase 3 / Abort. On Abort: dump build-issues.md, exit (no cleanup).

Phase 3 — Implementation

Re-read tasks.md AND caseplan.json (Step 9.6). Then:

1. Connector schema + defaults (Step 9.7) — is resources/triggers describe
2. I/O binding all task classes (Step 9.8) — per plugins/variables/io-binding/impl-json.md
3. Conditions all 4 scopes (Step 10)
4. SLA + escalation (Step 11)

No hard stop on Phase 3 exit — proceed directly to Phase 4.

Phase 4 — Validate

1. Full validate (Step 12). Retry up to 3×; on 3rd failure HARD STOP AskUserQuestion: Retry with fix / Pause for manual edit / Abort
2. Dump build-issues.md (Step 12.1)

Phase 5 — Debug

Completion report + HARD STOP AskUserQuestion (Step 13): Run debug session / Skip to Publish. On Run: uip solution resource refresh then uip maestro case debug (never auto-run — Rule 12). Loop on completion until Skip to Publish.

Phase 6 — Publish

HARD STOP AskUserQuestion (Step 14): Publish to Studio Web / Done. On Publish: uip solution resource refresh then uip solution upload, print DesignerUrl (Step 15). Exit on either choice.

Reference Navigation

I need to…	Read
Generate sdd.md interactively when none provided	references/phase-0-interview.md
Plan tasks from sdd.md	references/planning.md
Execute tasks.md into a case	references/implementation.md
Phase 2 → 3 → 4 → 5 → 6 split + hard stop contracts	references/phased-execution.md
Schema v19 vs v20 (Rules 18, 19) — mapping, layout-strip, mode behavior	references/case-schema.md § Top-level shape, references/phased-execution.md § v20 mode
Edit caseplan.json directly	references/case-editing-operations.md
Case JSON schema	references/case-schema.md
Surviving CLI commands (registry, validate, debug, runtime)	references/case-commands.md
Troubleshoot a failed case	references/troubleshooting-guide.md
Resolve task types from registry	references/registry-discovery.md
Wire inputs/outputs + cross-task refs + expression prefixes	references/bindings-and-expressions.md
Configure connector activity / trigger / event	references/connector-integration.md
Construct case spec --input-details JSON	references/case-spec-input-details.md
Placeholder tasks for unresolved resources	references/placeholder-tasks.md
Sync bindings_v2.json + connection resources	references/bindings-v2-sync.md

Plugin Index

Structural:

Plugin	Scope
case	Root case (T01)
stages	Regular and exception stages
edges	Edges between Trigger/Stage nodes
sla	Default SLA, conditional rules, escalation
global-vars	Case variables and arguments
io-binding	Task I/O wiring, cross-task refs
logging	Shared issue log

Tasks (references/plugins/tasks/):

Naming asymmetry — read carefully. Three names exist for connector + timer tasks. Pick the right one by column. Schema-kebab is the only value that goes into caseplan.json type (Rule 16).

sdd.md Type: value / caseplan.json type (schema-kebab)	Plugin folder	CLI --type flag (tasks describe)
process	process	process
agent	agent	agent
rpa	rpa	rpa
action	action	action
api-workflow	api-workflow	api-workflow
case-management	case-management	case-management
execute-connector-activity	connector-activity	connector-activity
wait-for-connector	connector-trigger	connector-trigger
wait-for-timer	wait-for-timer	wait-for-timer (no CLI describe needed)

Triggers (references/plugins/triggers/):

Plugin	When
manual	User-initiated start
timer	Scheduled start
event	External connector event

Conditions (references/plugins/conditions/):

Plugin	Scope
stage-entry-conditions	Stage entered
stage-exit-conditions	Stage exits
task-entry-conditions	Task starts
case-exit-conditions	Case completes/exits

Anti-patterns

• Do NOT leave stages without an inbound edge. Orphaned and unreachable. Every stage needs ≥1 inbound edge from Trigger or another stage.
• Do NOT validate after each T-entry. Intermediate states expected invalid. Run validate once at end of Phase 2 (informational) and once in Phase 4 (authoritative).
• tasks.md (Phase 1) uses per-section batched Edit-append — NOT per-T-entry, NOT one mega-Write. One Read + N Edit-appends per section (§4.2.1 vars, §4.3 triggers, §4.4 stages, §4.5 edges, §4.6 tasks, §4.7 conditions, §4.8 SLA). No re-Read between sibling Edits. HARD CAP: after §4.0a Step 1 Seed Write (<1KB header), single Write of whole tasks.md is FORBIDDEN regardless of size. Single Edit-append payload >30KB also FORBIDDEN — split per section even if cumulative payload exceeds 30KB. A 96KB tasks.md Write costs ~360s in one turn (20% of session); section-batched Edit-appends spread across ~7 turns of ~50s. TaskUpdate per T-entry preserves audit trail. Recovery on interruption: re-Read tasks.md, resume from next un-applied T-entry. See planning.md § 4.0a.
• caseplan.json (Phase 2 + 3) uses per-section batched writes — NOT per-T-entry. One Read at section entry + one validate at section end. Tool primitive scales with section size: <10 T-entries → N Edits (one per T-entry, no re-Read between siblings); ≥10 T-entries → may use single whole-section Write covering the section’s nodes/edges array at once, AFTER composing complete section state in reasoning. Untouched siblings (other sections, root fields) MUST be preserved verbatim from the Read — drop nothing. TaskUpdate per T-entry preserves audit trail regardless of write granularity. CLI-gated sections (Phase 2 §4.6 non-connector tasks describe, Phase 3 §9.7 connector case spec) use gather-then-write. Recovery on interruption: re-Read both files, resume from next un-applied T-entry. Full contract in case-editing-operations.md § Per-section batch write contract and implementation.md § Per-plugin execution.
• Do NOT emit standalone text-only assistant turns between tool calls. Status/progress text MUST share its turn with the next tool_use (text block + tool_use block in the same assistant content array). Standalone narration turns each pay full inference latency + prompt cache replay (~5s + ~250K cache-read tokens per turn) for no incremental progress. Cap inline status to ≤1 sentence / ~20 tokens. Per-T-entry audit lives in TaskUpdate, NOT in narration.
  • HARD TOKEN CAP on any single text block: 200 tokens, no exceptions outside the allow-list below. Allow-listed text blocks (hard-stop AskUserQuestion preambles, Phase 5/6 completion reports, Publish for review DesignerUrl print, post-validate result summaries) get a higher ceiling of 500 tokens — never higher. A text block >200 tokens outside the allow-list, or >500 tokens inside it, is a planning monologue, regardless of content or framing.
  • Forbidden announcement verbs. Text blocks (bundled or standalone) starting with Building, Composing, Writing, Drafting, Generating, Now I'll, Next:, Next step:, Approach:, Strategy:, Plan:, Caveman push:, Big single Write:, Let me, or any other narration of the imminent tool call are FORBIDDEN regardless of length. The tool_use input shows what is being built — restating it in prose is pure cost. If the agent feels the urge to write Composing Phase 2 caseplan.json — trigger + 64 variables + 10 stages + 6 edges, it must instead invoke the Write directly with that content as the file body.
  • Allow-listed exceptions (may stand alone, capped at 500 tokens): hard-stop AskUserQuestion preambles, final completion reports (Phase 5/6 exit), Phase 2 Publish for review DesignerUrl print (Rule 11), and post-validate result summaries (N errors, M warnings — fixing X is fine; Composing fix for ... is not). Everything else bundles or omits.
• One task per lane — except parallel members of a runs-sequentially group. Default: each task own lane index in stageNode.data.tasks[laneIndex][], lane is FE layout only. Exception: tasks sharing a runs-sequentially task-entry condition that are meant to run in parallel share the same lane (shared lane = parallel siblings inside the sequential group, carries execution semantics). Solo runs-sequentially tasks still get their own lane.
• Do NOT edit content/*.bpmn. Auto-generated, will be overwritten. Edit content/*.json only.
• Do NOT fabricate expression syntax for conditional SLA rules. Describe condition in natural language; execution phase determines exact form.
• Do NOT invoke other skills automatically. If case needs process/agent/action that doesn’t exist, emit placeholder task (Rule 8) and list missing resources in completion report. On-demand resource creation is future milestone.

Trouble? Use /uipath-feedback to send report.