uipath-review

name	uipath-review
description	UiPath read-only reviewer — audit structure, quality, best practices for RPA (.xaml/.cs), agents (.py/agent.json), flows (.flow), BPMN (.bpmn), coded apps, solutions (.uipx). Does NOT edit files. For building/editing→domain skills.
allowed-tools	Bash, Read, Glob, Grep, WebFetch, AskUserQuestion
user-invocable	true

UiPath Solution & Artifact Reviewer

Review UiPath solutions and individual artifacts for structural validity, quality, best practices, optimization, and correctness. Produces a structured review report with findings and recommendations.

When to Use This Skill

• User asks to "review", "audit", "check quality of", or "evaluate" a UiPath project or solution
• User asks "is this solution good?" or "what can be improved?"
• User wants a pre-deployment quality gate check
• User wants to understand the business value and architecture of an existing solution
• User asks about best practices for a specific artifact type
• User has inherited a UiPath project and wants to understand its quality

Critical Rules

1. NEVER manually modify any files. This skill is read-only. Exception: The command uip agent refresh is allowed and mandatory for low code agents, because it is not a manual modification, even when the command updates derived files -- do not restore or clean up those CLI-managed changes. If fixes are needed, identify them in the report and tell the user which skill to use (uipath-rpa, uipath-agents, uipath-maestro-flow, uipath-maestro-bpmn, uipath-api-workflow, uipath-coded-apps, uipath-platform, uipath-solution).
2. ALWAYS run validation and Workflow Analyzer before manual review. For RPA projects, run both uip rpa validate on every entry point AND uip rpa build "<PROJECT_DIR>" — per-file validate checks one file deeply (structure, references, analyzer rules, unknown members, invalid enums, expression compilation); build compiles the whole project — every workflow, including files validate was never pointed at, plus project-scope analyzer rules and packaging. For low-code agents, run uip agent refresh and uip agent validate. Run uip maestro flow validate on flows, uip maestro bpmn validate on BPMN processes, uip api-workflow validate on API workflows. Report every command's Error / Warning / Info counts in the validation table, and a detail line for each Error and Warning — never a detail line for a clean result (Step 2d). A review without both validate AND build (for RPA) is incomplete and may ship broken member references.
3. ALWAYS discover and classify before reviewing. For solutions: classify every project before reviewing any individual one. For single projects: identify the project type and find the enclosing project directory before reviewing individual files.
4. Report severity for every finding. Use: Critical (blocks deployment), Warning (should fix), Info (improvement opportunity).
5. Understand business context first. Before evaluating optimization, ask or infer what the solution is trying to accomplish. A queue-based architecture is not "better" if the use case processes 5 items/day.
6. Use --output json on all CLI validation commands for programmatic parsing.
7. Do not duplicate what validation commands catch. Reference the validation output by rule ID and message — do not manually re-describe the same issue, and do not restate what a command checks or that it passed. Every validation result is accounted for by its count in the validation table; Errors and Warnings additionally get a detail line.
8. Cap the review at 30 minutes of analysis. For very large solutions (10+ projects), provide a summary review with deep dives on the 3 highest-risk projects. Offer to review remaining projects if the user wants.
9. Run the review CLI first, then apply the judgment catalog, for every agent encountered. First run uip agent review (low-code) or uip codedagent review (coded) with --output json — it returns the deterministic findings (Step 2.5a). Then load the format-specific judgment catalog (agents-lowcode-rules.md or agents-coded-rules.md). Future phases add catalogs for RPA, flows, coded apps. This holds even when the skill loads mid-task: if review work already started before this skill loaded (e.g., a generic code-review pass produced findings), Step 2.5a and the guardrail Step 0 catalog fetch are still mandatory — run them, then merge the earlier findings into this skill's report format. Prior review output is never a substitute for the review CLI or the live catalog.
10. Rule findings are authoritative as emitted. Carry review-CLI RuleId, Severity, Description, File, and SuggestedFix into the report verbatim. Format the Recommendation as <File>: <Description>. <SuggestedFix>. Write judgment-catalog findings in the same format, with concise wording. Map severity to the report's bands: error → Critical, warning → Warning, info → Info. judgment severity rows default to Warning; the agent may escalate or de-escalate with reasoning logged in the finding's description. Do not re-rank otherwise.
11. Report rules that could not be applied (missing tooling, missing file, review CLI unavailable, status: deferred) in a dedicated "Rules Skipped" subsection of the report — never silently skip. Only report when the rule was intended, but could not be applied for some reason. Non-applicable rules are not skipped.
12. Never invent rule_id values. Every rule_id cited in the report MUST appear verbatim in EITHER a loaded judgment-catalog file (references/agents/agents-*-rules.md) OR the uip agent review / uip codedagent review JSON output. rule_id is a stable contract identifier — consumers grep for it, dashboards aggregate by it, audits trace it. An invented identifier looks authoritative but cannot be looked up, doesn't aggregate, and produces a different name for the same observation on the next run. If you observe a real, critical issue covered by neither source, the finding is still valid — surface it under Critical Findings without a rule_id (no `RULE_ID` backtick token in the line). Only critical issues qualify — drop an unrule'd Warning or Info. Both sources are agent-only, so this governs agent findings. Before emitting the report, scan every cited rule_id and confirm it appears verbatim in a loaded catalog file or the review-CLI output; demote any that don't to rule_id-less findings.
13. Grade every agent project by the rubric — derived, never asserted. For agent projects (phase 1), produce a letter grade (A/B/C/D/F, no +/-) per agent and overall, computed in Step 4.5 as min(G_det, G_jud). G_det is read from the review CLI's Data.Grade (Step 2.5a) — do not recompute it from finding counts. G_jud you compute from the judgment findings (Step 2.5b + Step 3) by severity count. CLI findings already shaped Data.Grade; only judgment findings feed G_jud, so each finding lands in exactly one sub-grade. Show the binding constraint for every grade; a grade with no shown derivation is invalid (low-code reports omit the printed derivation). A security or data-integrity judgment Critical forces F regardless of design quality (hard gate, not a blend). The skill grade is always ≤ Data.Grade (min only lowers) — report both, never overwrite the CLI grade. Do not grade non-agent projects (RPA, flows, coded apps) — that rubric is a future phase. See references/agents/agent-grading-rubric.md.
14. These paths are CLI-managed — uip agent refresh owns them: .agent-builder/, .local/build/, and (low-code only) the root entry-points.json, regenerated from agent.json. Do not open their contents. Exclude them from classification, source-file selection, structural metrics, and manual checks. Raise a finding only when refresh fails to fix them — a pre-refresh mismatch is stale by construction, not a defect. Read low-code schemas from agent.json (.inputSchema / .outputSchema).

Review Workflow

Step 0 — Discover, Scope, and Locate the PDD

0a. Probe the Filesystem

Run this from the directory the user specified (or the current working directory):

# Discover solution files, project markers, and documentation
find . -maxdepth 3 \( -type d \( -name ".agent-builder" -o -path "*/.local/build" \) \) -prune -o \( -name "*.uipx" -o -name "project.json" -o -name "project.uiproj" -o -name "agent.json" -o -name "*.flow" -o -name "*.bpmn" -o -name "app.config.json" -o -name ".uipath" -o -name "pyproject.toml" -o -name "langgraph.json" -o -name "llama_index.json" -o -name "openai_agents.json" -o -name "uipath.json" -o -name "main.py" \) -print 2>/dev/null

# Search for PDD or design documents
find . -maxdepth 3 \( -type d \( -name ".agent-builder" -o -path "*/.local/build" \) \) -prune -o \( -name "*PDD*" -o -name "*pdd*" -o -name "*Process_Design*" -o -name "*process_design*" -o -name "*Process-Design*" -o -name "*ProcessDesign*" -o -name "*SDD*" -o -name "*Solution_Design*" -o -name "*design_document*" -o -name "*DesignDocument*" -o -name "*requirements*" -o -name "*specification*" \) -print 2>/dev/null

0b. Locate the PDD (Process Design Document)

The PDD is the source of truth for the review. It defines what the automation should do, its business context, expected inputs/outputs, exception handling requirements, and success criteria. The review evaluates whether the implementation matches the PDD.

Search for PDD in this order:

1. Check common locations: ./docs/, ./documentation/, ./Design/, project root
2. Check common names: PDD.docx, PDD.pdf, PDD.md, Process_Design_Document.*, SDD.*, Solution_Design_Document.*, Requirements.*
3. Check AGENTS.md or README.md at project root — may contain or reference the PDD
4. Check project.json description field or any metadata pointing to documentation

If PDD is found:

• Read it (supports .md, .pdf, .docx via appropriate tools)
• Extract the key review criteria: business process description, expected inputs/outputs, exception handling requirements, SLAs, transaction definitions, queue specifications, application list, credential requirements
• Use it as the primary benchmark for all subsequent review steps

If PDD is NOT found:

Use the AskUserQuestion tool to ask interactively:

Question: "I could not find a Process Design Document (PDD) in this project. Do you have one I can use as the source of truth for this review?"
Header: "PDD"
Options:
  1. Label: "Yes, I have a file"
     Description: "I'll provide a file path, URL, or Confluence/SharePoint link to the PDD, SDD, or requirements document"
  2. Label: "I'll paste the content"
     Description: "I'll copy/paste the PDD content (or key sections) directly into the chat"
  3. Label: "No, proceed without"
     Description: "Skip PDD alignment — review will cover technical quality and best practices only, not business logic verification"

• If user selects "Yes, I have a file": they will provide the path in their response. Read the document and proceed with PDD-informed review.
• If user selects "I'll paste the content": they will paste the PDD text (or relevant sections) in their next message. Use that content as the PDD for the review.
• If user selects "No, proceed without": proceed without it — the review will focus on technical quality, best practices, and structural correctness, but cannot verify business logic alignment. Note this limitation in the report.

0c. Determine Review Scope

Workflow labels like "Path A / Path B / Step 3a" are internal to this skill. NEVER use them in the final review report. The report must use user-facing language — see Step 5 for the required Review Scope vocabulary.

Classify the scope internally using these rules:

Scope: Solution or Multi-project — .uipx exists at root, OR 2+ executable project markers exist in different subdirectories.

• Executable project = project.json with outputType of Process/Tests/unspecified, OR a low-code agent.json, OR a coded-agent Python project (pyproject.toml + framework/uipath.json configuration), OR .flow, OR project.uiproj with ProjectType Flow/ProcessOrchestration/Api
• Library projects (outputType: "Library") co-located with consumers do NOT trigger this scope — that is the normal library+consumer pattern
• Windows-Legacy executables do NOT trigger this scope for .uipx purposes: .uipx solutions are not supported for Legacy projects. If any detected executable is Legacy, do not flag missing .uipx — recommend migration to Modern compatibility if solution bundling is desired. Review each Legacy project independently.

Steps for Solution / Multi-project scope:

1. Read the .uipx file (if present) to enumerate all projects
2. Scan subdirectories for project markers not listed in .uipx (orphan executables)
3. Classify each project using the detection table in Step 1
4. Run solution-level checks: missing config.json, version mismatches, cross-project dependencies, circular dependencies
5. Build a solution map: every project with its type, path, and relationship to others
6. Cross-reference with PDD (if available)
7. Read references/solution-review-guide.md for the full procedure
8. Proceed to Step 1 for each project individually

Scope: Single Project — one project.json / agent.json / .flow / coded-app marker, or one Python coded-agent project, at root; no .uipx, no executable siblings.

1. Classify the project using the detection table in Step 1
2. Cross-reference with PDD (if available)
3. Skip solution-level checks; go directly to Step 1

If the user pointed to a specific file (e.g., Main.xaml), walk up to the enclosing project directory and review the full project.

---

Step 1 — Classify the Project Type and Capture Language

For each project discovered (one for single-project scope, multiple for solution/multi-project scope), determine its type AND capture its expression language.

Step 1a — Read expressionLanguage from project.json for every RPA project. This is mandatory. The value (VisualBasic or CSharp) affects everything downstream: expression syntax in If/Switch conditions, null checks, type checks (TypeOf x Is T in VB vs x is T in C#), string operations, LINQ syntax, and naming conventions. All subsequent inspection steps (especially Step 3a Unit of Work grep and expression-dependent checks) MUST adapt patterns to the project's language. Do not assume VB.

Record the language per project alongside the type (see solution table below).

Step 1b — Determine project type using the detection table:

Filesystem Signal	Project Type	Review Checklist
project.json + .cs files with [Workflow] attributes	RPA (Coded)	rpa-review-checklist.md
project.json + .xaml workflow files	RPA (XAML)	rpa-review-checklist.md
project.json with no targetFramework or targetFramework: "Legacy" (any expression language — Legacy C# exists)	RPA (Windows-Legacy)	rpa-review-checklist.md §10. Also recommend the user invoke uipath-rpa (Legacy mode) for Legacy-specific deep validation. Legacy is supported indefinitely in Studio LTS — do NOT flag as Critical.
project.json + both .cs and .xaml	RPA (Hybrid)	rpa-review-checklist.md
project.json + .xaml + DU packages in dependencies (UiPath.IntelligentOCR.Activities, UiPath.DocumentUnderstanding.ML.Activities)	RPA + Document Understanding	rpa-review-checklist.md + du-review-checklist.md
agent.json with "type": "lowCode"	Agent (Low-Code)	Rule catalog (Step 2.5): agents-lowcode-rules.md
Python coded-agent project, including agent.json with "type": "coded" when present	Agent (Coded)	Rule catalog (Step 2.5):

For Solution / Multi-project scope, record all projects in a table:

| # | Project Path | Type | Language | Entry Points |
|---|---|---|---|---|
| 1 | ./InvoiceProcessor/ | RPA (XAML) | VisualBasic | Main.xaml, Helper.xaml |
| 2 | ./Dispatcher/ | RPA (Coded) | CSharp | Main.cs |
| 3 | ./ClassifierAgent/ | Agent (Coded) | Python | main.py |
| 4 | ./Orchestration.flow | Flow | — | — |

Step 1c — Inventory the authored files for every project you will review. Run this instead of writing your own find, to avoid listing runtime artifacts:

find "<PROJECT_DIR>" \( -type d \( -name ".agent-builder" -o -path "*/.local/build" -o -name "node_modules" -o -name ".venv" -o -name "obj" -o -name "bin" \) \) -prune -o -type f -print 2>/dev/null | sort

The result is the authored-file set for Steps 2.5 and 3. Any path absent from it is out of scope: do not read it, cite it, or name it anywhere in your output.

Step 2 — Run Automated Validation and Workflow Analyzer

This step is mandatory and non-negotiable. You MUST run validation commands yourself (via Bash) before doing any manual review.

• Solution / Multi-project scope: Run validation on every project in the solution. For each RPA project, validate every entry point file.
• Single Project scope: Run validation on the single project. For RPA projects, validate every entry point file.

Account for all results in the final review report: Error / Warning / Info counts in the validation table, plus a detail line per Error and Warning (Step 2d).

2a. RPA Projects — Validate Every Entry Point

1. Read project.json → extract the entryPoints array
2. For each entry point file, run validation yourself:

uip rpa validate --file-path "<ENTRY_FILE>" --project-dir "<PROJECT_DIR>" --output json

3. Then run a project-level build — it covers the whole project: every workflow (including files per-entry-point validate never touched), project-scope analyzer rules, and packaging:

uip rpa build "<PROJECT_DIR>" --log-level Warn --output json

4. Collect all results from both commands — Errors, Warnings, and Info-level messages (Info feeds the table's count; it gets no detail line)
5. If any entry point has validate errors or the project fails to build, the project is not deployable

Do NOT validate only Main.xaml — validate every file listed in entryPoints. A project can have multiple entry points and errors in any of them block deployment.

Do NOT report a clean review based on per-file validate alone — it covers only the files it was pointed at, and a project whose validated files are clean can still fail build on a file nobody validated. A "0 errors" validate result with a failing build is a real bug that ships if the reviewer skips build.

2b. RPA Projects — Run Workflow Analyzer

The Workflow Analyzer checks code quality rules (ST-NMG naming, ST-DBP design, ST-MRD maintainability, ST-USG usage, ST-SEC security, ST-REL reliability). Run it explicitly:

uip rpa analyze --project-dir "<PROJECT_DIR>" --output json

If uip rpa analyze is not available, uip rpa validate includes Workflow Analyzer results. Check the output for all rule violations:

• Error-level violations → report as Critical findings (e.g., ST-SEC-007 SecureString, ST-ANA-005 missing project.json)
• Warning-level violations → report as Warning findings (e.g., ST-DBP-003 empty Catch, ST-MRD-011 Write Line usage, ST-NMG-001 naming)
• Info-level violations → report as Info findings (e.g., ST-ANA-003 workflow count, ST-ANA-009 file activity stats)

Every Workflow Analyzer violation must appear in the review report with its rule ID, affected file, and description. Do not silently skip any severity level.

2c. Other Project Types

Project Type	Validation Command	Report All Severities
Agent (Low-Code)	uip agent refresh "<PROJECT_DIR>" --output json, then uip agent validate "<PROJECT_DIR>" --output json	Yes — errors, warnings, info
Flow	uip maestro flow validate "<PROJECT_NAME>.flow" --output json	Yes — schema errors, reference errors, warnings
Maestro BPMN	uip maestro bpmn validate "<FILE>.bpmn" --output json	Yes — model errors, warnings
API Workflow	uip api-workflow validate "<WORKFLOW_JSON>" --output json	Yes — schema + semantic errors, warnings
Coded App	uip codedapp pack dist --dry-run --output json	Yes — build errors, pack warnings
Solution	uip solution pack "<SOLUTION_DIR>" "<OUTPUT_DIR>" --output json	Yes — per-project pack results

uip api-workflow validate is offline (no auth, no network, no side effects). Do NOT run uip api-workflow run — it executes vendor calls with real side effects. If the CLI reports an unknown command for maestro bpmn validate or api-workflow validate (older CLI), record it under "Rules Skipped" and fall back to the manual structural checks in the type's checklist.

2d. Record All Results

For the review report, create a validation summary:

### Validation Results

| Project | Command | Errors | Warnings | Info |
|---|---|---|---|---|
| InvoiceProcessor | uip rpa validate (Main.xaml) | 0 | 3 | 1 |
| InvoiceProcessor | uip rpa validate (Helper.cs) | 1 | 0 | 0 |
| InvoiceDispatcher | uip maestro flow validate | 0 | 0 | 0 |
| ClassifierAgent | uip agent validate | 0 | 1 | 0 |

#### Validation Details
- [E-001] InvoiceProcessor/Helper.cs: ST-SEC-007 — Password argument uses String instead of SecureString
- [W-001] InvoiceProcessor/Main.xaml: ST-MRD-011 — Write Line activity used (use Log Message instead)
- [W-002] InvoiceProcessor/Main.xaml: ST-DBP-003 — Empty Catch block in TryCatch_1
- [W-003] InvoiceProcessor/Main.xaml: ST-NMG-001 — Variable 'temp_val' does not match naming convention
- [W-004] ClassifierAgent: Missing tool description for 'lookup_customer'

The validation results section is required in every review report. A review without automated validation is incomplete.

Counts in the table; detail lines for Errors and Warnings only. Every command's Error / Warning / Info counts go in the table — that is where Info is reported. Write a detail line only for an Error or a Warning. Never write a detail line that narrates a clean or successful result (Status: Valid, Verdict: PASS, Score: 100, "no drift", "0 issues", "N files regenerated", "already at schema X") — the 0 in the table already says it, and the Info column already carries the count.

Step 2.5 — Run the Review CLI, then Apply the Judgment Catalog

After Step 2 validation and before manual checklist review, produce rule-ID-level findings in two passes: first the uip agent review / uip codedagent review CLI for the deterministic static checks, then the skill's judgment-only catalog for what code cannot decide reliably.

Late invocation: if a review was already performed or started before this skill loaded, do NOT skip 2.5a/2.5b as "already covered" — no other review flow runs the review CLI or fetches the live guardrail catalog. Run both passes, then fold prior findings into Step 5's report.

2.5a — Run the review CLI first (deterministic findings)

Run the review command for the agent type, once, capturing JSON:

Agent type	Command
Low-code	uip agent review "<PROJECT_DIR>" --output json
Coded	uip codedagent review "<PROJECT_DIR>" --output json

The CLI runs the deterministic static checks its registry ships — structural/schema gates, placeholder cross-refs, eval-set structure and schema cross-refs, guardrail configuration validity, tool count, prompt length/platform — and returns them in rule format. Parse Data.Issues[]; each issue is {RuleId, Category, Severity, Description, File, SuggestedFix}. Carry each into the report verbatim — do not re-derive, rename, or re-rank. These rule IDs are authoritative as emitted by the CLI; they are not listed in the skill catalog.

Guardrail configuration is CLI-only — never eyeball it. Whether a guardrail is well-formed (real validator, allowed scope, required/typed/legal parameters, valid custom-rule shape) is decided only by uip agent review — the GUARDRAIL_* and GUARDRAIL_CUSTOM_* rule IDs come from this command, never from reading agent.json by eye and never from the judgment catalog. So whenever the task involves checking / validating / diagnosing / fixing a guardrail, running the review CLI in this step is mandatory (use --checks guardrails if you only need the guardrail pass), and every GUARDRAIL_* finding it returns must appear verbatim in the report's findings tables — do not replace it with a hand-written description of the problem. (The judgment catalog's LC_GUARDRAIL_* rules are the complement: they audit only guardrails the CLI found format-valid and recommend missing ones at Info — see Step 2.5b and references/agents/guardrails/guardrails-review.md.)

2.5b — Apply the judgment catalog (reasoning the CLI cannot do)

1. Identify which catalog files apply for the current project type:

Signals present	Project type	Catalog files
agent.json.type == "lowCode"	Agent (low-code)	references/agents/agents-lowcode-rules.md
Python coded-agent signals or agent.json.type == "coded"	Agent (coded)	references/agents/agents-coded-rules.md
pyproject.toml + main.py + uipath.json[functions] only (no framework config)	Agent (coded — Simple Function)	same as Agent (coded)
project.json + .xaml / .cs	RPA	(phase 2)
*.flow	Flow	(phase 2)
.uipath/ or app.config.json	Coded App	(phase 2)

2. Read each catalog file in full. Every rule is judgment-form.

3. Guardrails — apply the structured guardrail workflow (project-type specific; Step 0 fetches the authored uip agent guardrails catalog — 30-min cache — plus the never-cached tenant-availability uip agent guardrails list → Audit Mode for existing guardrails + Recommend Mode for missing ones):

   • Low-code (agent.json): when guardrails[] is non-empty or the agent matches a guardrail use case, apply references/agents/guardrails/guardrails-review.md. Emits LC_GUARDRAIL_ACTION_INEFFECTIVE / LC_GUARDRAIL_MISAPPLIED (defects, judgment band) and LC_GUARDRAIL_RECOMMENDED (Info, one per missing guardrail).
   • Coded (SDK middleware / @guardrail decorators wired in the entry .py): when the entry source wires guardrails or the agent matches a use case, apply references/agents/guardrails/coded-guardrails-review.md (its Step 0 fetches the public Python SDK docs only when a finding must name Python classes not already visible in the agent source). Emits CODED_GUARDRAIL_ACTION_INEFFECTIVE / CODED_GUARDRAIL_MISAPPLIED (defects, judgment band) and CODED_GUARDRAIL_RECOMMENDED (Info). The CLI's deterministic CODED_GUARDRAIL_WRONG_IMPORT / CODED_GUARDRAIL_TOOL_SCOPE_NO_TOOLS / CODED_GUARDRAIL_INVALID_CONTRACT (Step 2.5a) are carried verbatim and not re-flagged here.
   • Either way, if the guardrail catalog is unavailable, record the Audit-Mode rules under "Rules Skipped" and keep Recommend Mode's source-only detection.

4. Apply each rule's detection_method: read the named source material (system prompt, tool descriptions, eval datapoints, schemas) and reason about it. Emit a finding when the criteria hold; log the reasoning in the finding's description.

5. Track skipped rules with their reason (, missing optional file, review CLI unavailable). Never silently skip.

See references/rule-catalog-workflow.md for the full procedure including the CLI contract and determinism rules.

Step 3 — Manual Quality Review

For each project (one for single-project, all for solution/multi-project), load the relevant checklist from references/ based on the type classified in Step 1. Read project files, check patterns, evaluate design.

3a. Unit of Work Discovery (mandatory, generic)

Every project has two units of work: what the contract declares one invocation represents, and what the execution body actually does. A mismatch is a Critical-to-Warning finding regardless of project type. Do not ask the user — derive both mechanically from the project.

Step 3a.1 — Discover the declared unit of work (per project type):

Project type	Where the declared unit lives
RPA + queue	Queue item schema (Data/*.json, JSON Schema/, or the SpecificContent fields used by Add Queue Item / Get Transaction Item)
RPA without queue	Main.xaml input arguments
Flow	.flow file → variables.globals → entries with direction: "in" or "inout"
Maestro BPMN	Process start-event payload / process input variables
Agent (low-code)	agent.json → inputSchema
Agent (coded)	Input class in main.py (Pydantic BaseModel)
API workflow	Request input schema in Workflow.json
Coded app	Entry point input schema in operate.json / entry-points.json

Step 3a.2 — Discover the actual unit of work (core execution body):

Identify the core execution file (ProcessTransaction.xaml, Process.xaml, Main.xaml, main.py, flow body, API handler) then run these mechanical checks:

# Detect iteration inside the execution body
grep -n 'ForEach\|While' <EXECUTION_FILE>

# Detect external-effect activities (writes, API calls, queue pushes, workflow invocations)
grep -n 'HttpRequest\|Add Queue Item\|InvokeWorkflowFile\|Write Range\|Write Line\|SqlCommand' <EXECUTION_FILE>

For coded projects, look for for / foreach / while statements and external I/O calls.

Step 3a.3 — Classify using this matrix:

Classify the Transaction Shape using this matrix. Shape is a neutral description of the relationship between input and external effects — it is NOT a pass/fail verdict.

Actual execution pattern	Transaction Shape
One invocation → one atomic external state change (one write, one submission, one workflow call)	One-to-one
Execution iterates over an array/collection field of the declared input, and the loop body contains external effects (see list below)	One-to-many
Iteration only over retry counters, UI element enumeration, or pure in-memory transformations (no external effects in loop body)	One-to-one (in-memory iteration is intra-unit; not a sub-unit of work)
No iteration at all	One-to-one
Contract or execution cannot be deterministically mapped (schema missing/unclear, dynamic dispatch)	Unclear

External effects inside a loop body that make it one-to-many (none of these are defeated by session scope, shared credentials, single portal, or business-model arguments):

• InvokeWorkflowFile / Invoke Method to workflows with external side effects
• HTTP activities (HTTP Request, connector activities, REST calls)
• Queue operations (Add Queue Item, Set Transaction Progress, Set Transaction Status)
• Database writes (Execute Non Query, Insert Data Table, Bulk Insert)
• File writes outside Temp/ directories (Write Range, Write CSV, Append to File)
• UI activities that modify target-system state (Click on submit/save, Type Into fields that persist, SAP Call Transaction)
• Email send activities

Classification is mechanical. It does not change based on:

• "The portal models this as one transaction" (UX framing ≠ atomicity)
• "One browser session" (session ≠ transaction)
• "Idempotency guards exist so it's fine" (guards are a remediation signal, not a reclassifier)
• "The PDD calls it one transaction" (declared intent ≠ execution reality)
• "The queue only has one item" (queue is the declared unit; actual unit is what gets written)

Step 3a.4 — Record shape, then separately assess remediation.

The shape itself is reported neutrally. Whether it becomes a finding — and at what severity — depends on remediation posture:

For One-to-one: No finding. Report the shape observation in Summary, move on.

For One-to-many: Assess two separate questions.

Question A — Can the sub-units be independently queued / split?

• Yes: the proper fix is dispatcher/performer — split the queue so each sub-unit is an atomic transaction. Use this when sub-units are independent (one invoice, one employee record, one order, one file).
• No: the domain forces a sequential session-bound submission (SAP new-plan enrollment, carrier portal group application, bank multi-step wire). Queue splitting is infeasible. The fix is not architectural — it is operational: verify atomicity, error handling, crash recovery, and progress tracking using the 10-point hardening checklist in rpa-common-issues.md → "When it cannot be split — hardening checklist." Each missing safeguard is a separate finding.

Question B — What partial-failure recovery exists today?

Look for any of these patterns (semantically, not by filename):

Pattern	Detection
Read-check-before-write before each sub-unit write	Inspect activity sequence in the loop body
Conditional skip based on "already exists/processed" state	Inspect If/Switch branches wrapping writes
Orchestrator queue dedup via UniqueReference	Check Add Queue Item properties
SQL idempotent writes (MERGE, ON CONFLICT, UPSERT, WHERE NOT EXISTS)	Grep SQL statements
HTTP idempotency (Idempotency-Key header, ETag If-Match / If-None-Match)	Check HTTP Request headers
Status-column filters (WHERE Status != 'Processed')	Grep queries
Pre-check workflow invocation (names often contain check/verify/exists/processed/already/skip/idempoten — one of many forms, not the only signal)	Inspect invoked workflow names and bodies
Per-sub-item progress written to queue Output / Data Service / external state	Inspect what's persisted during the loop

Severity and finding framing:

Scenario	Severity	Finding framing
One-to-many + sub-units splittable + no idempotency guards + MaxRetryNumber < 2	Critical	"Transaction granularity: split into dispatcher/performer. Current architecture risks partial-state corruption on transient failure."
One-to-many + sub-units splittable + idempotency guards exist but progress/output fidelity weak	Warning	"Transaction granularity: consider dispatcher/performer split for better analytics and retry isolation."
One-to-many + sub-units NOT splittable (domain constraint) + missing safeguards	Warning–Critical	"Cannot be split — run the 10-point hardening checklist in rpa-common-issues.md → 'When it cannot be split.' Report each missing safeguard as a separate finding."
One-to-many + splittable + guards + retry + per-sub-item output	Info (tech debt)	"Transaction granularity: working with compensation; consider dispatcher/performer if volume grows."
Unclear	Info	"Unit of work ambiguous — schema/code documentation gap."

The shape observation belongs in the Executive Summary of the report as a one-liner (see Step 5). Any finding generated from the shape analysis becomes a normal numbered finding in the Critical/Warning/Info sections — not a separate "Unit of Work Analysis" block.

3b. PDD Alignment Review (if PDD is available)

If a PDD was found or provided in Step 0, use it as the primary benchmark for the manual review. For each project, verify:

PDD Section	What to Check	Severity if Mismatched
Business process description	Does the implementation match the described process flow?	Warning
Expected inputs/outputs	Do workflow arguments match PDD-defined inputs and outputs?	Warning
Exception handling requirements	Are Business Exceptions thrown for the cases the PDD defines? Are retries configured per PDD specs?	Warning
Application list	Are all applications from the PDD automated? Any missing? Any extras not in PDD?	Warning
Transaction definition	Does the transaction item structure match the PDD?	Warning
Queue specifications	Queue names, retry counts, SLAs match PDD?	Warning
Credential requirements	Are all credentials from PDD stored securely (assets/vault)?	Critical if hardcoded
SLAs and performance targets	Does the automation design support PDD-defined throughput/timing?	Info
Happy path + exception scenarios	Are all PDD-documented scenarios handled?	Warning
Out of scope items	Does the automation stay within PDD-defined scope?	Info

Report PDD mismatches as a dedicated section in the review report. A technically sound automation that doesn't match its PDD is still a problem.

If no PDD is available, skip this sub-step and note in the report:

Note: No PDD was available for this review. Business logic alignment could not be verified. This review covers technical quality and best practices only.

3c. Technical Quality Review

For each project, load the type-specific checklist:

For Solution / Multi-project scope, also perform solution-level checks from references/solution-review-guide.md:

• Solution structure validation (.uipx, config.json, orphan projects) — skip .uipx checks if any detected executable is Windows-Legacy; recommend migration instead
• Cross-project dependency checks
• Configuration consistency across projects
• Multi-project architecture pattern assessment

For deep-dive RPA reviews, also consult:

• RPA (advanced): rpa-advanced-checklist.md — project organization, selector robustness, variable hygiene, data patterns, error handling depth, testing maturity, idempotency
• RPA (long-running): long-running-workflow-issues.md — load when project uses persistence activities (Suspend, Wait and Resume, Create Form Task, Orchestration Process type)
• RPA (Modern Studio): modern-studio-issues.md — load for Studio 2024.10+ projects (Modern vs Classic mixing, coded/XAML interop, Object Repository, Data Manager, Healing Agent)
• Document Understanding: du-review-checklist.md — load when DU packages detected in project.json dependencies

For common antipatterns per project type, also consult:

• RPA: rpa-common-issues.md
• Flows: flow-common-issues.md

Step 4 — Evaluate Optimization

Only after validation (Step 2) and manual review (Step 3) are complete, evaluate optimization.

Solution / Multi-project scope — evaluate cross-project concerns:

• Architecture: Is the multi-project design appropriate (dispatcher/performer, main + libraries, flow + resources)?
• Cross-project dependencies: Are library versions pinned? Any circular dependencies?
• Queue usage: Should this solution use queues for work distribution?
• Bulk operations: Are there loops that could use bulk APIs?
• Transaction handling: Is error recovery and retry properly implemented across projects?
• Resource efficiency: Are there redundant API calls, excessive logging, or unnecessarily large files?
• Configuration consistency: Do all projects use the same pattern for configuration (assets, config.json)?

Single Project scope — evaluate within-project optimization:

• Queue usage: If processing >50 independent items, should this use queues?
• Bulk operations: Are there loops with individual API calls that could be batched?
• Transaction handling: Is REFramework or equivalent retry logic needed?
• Resource efficiency: File sizes, logging volume, selector efficiency, data handling patterns

Read references/review-workflow-guide.md for the full optimization evaluation criteria.

Read references/architecture-assessment-guide.md for the architecture-level evaluation framework — process suitability, complexity classification, environment separation, and architecture principles scoring.

Step 4.5 — Compute the Agent Letter Grade (A–F)

Agent projects only (phase 1) — matching the Step 2.5 judgment catalog, which is agent-only today. Grade every agent project, and (for a multi-agent solution) the agent set overall, on an A–F scale. Non-agent projects are not graded yet (RPA, flows, coded apps are future phases) — report their findings without a grade. The grade is derived — never a fresh judgment. Take the worse of two sub-grades: G_det is read from the review CLI, G_jud you compute from judgment:

Final grade = min(G_det, G_jud)        where G_det = <review CLI>.Data.Grade

• G_det (deterministic) — read it from the review CLI; do not recompute. uip agent review / uip codedagent review (Step 2.5a) returns Data.Grade — that letter, collapsed to its base letter (C+ → C), is G_det. (Data.Issues[] are still reported verbatim, but the grade comes from Data.Grade, not from tallying them.)
• G_jud (non-deterministic) — the only sub-grade you compute, from the judgment-catalog (2.5b) + manual review (Step 3) findings.

CLI findings already shaped Data.Grade (G_det); only judgment findings feed G_jud — so each finding lands in exactly one sub-grade.

G_jud score — 100 − (15 × Criticals) − (4 × Warnings) − (1 × Infos) over the judgment findings, floored at 0, looked up in the rubric's grade chart: 85–100→A, 65–84→B, 45–64→C, 25–44→D, 0–24→F. Then cap: any unmitigated judgment Critical → at most D; security/data-integrity judgment Critical → F. Architecture-principle scores do not feed the grade.

Overall Agent Grade: single agent → its grade. Multiple agents → the worst per-agent grade. Never average grades.

Report the binding constraint in one line (e.g. "B — gated by G_det = CLI Data.Grade B; judgment clean (G_jud A)"). Since the skill grade is min(Data.Grade, G_jud), it is always ≤ Data.Grade — report both; never overwrite the CLI grade. This goes in the Summary's Agent Grade line, and the letter alone is restated as **Final grade: <A–F>** on the report's last line (Step 5).

Full rubric, grade chart, low-code section omissions, edge cases (no-PDD / CLI-unavailable / no-eval-set), CLI-grade alignment, and worked examples: references/agents/agent-grading-rubric.md.

Step 5 — Produce the Review Report

Output a structured report in chat (do NOT create a file):

Report rules — do not violate:

1. NEVER use internal workflow labels in the output. Forbidden terms: "Path A", "Path B", "Step 3a", "Step 0c", "Mismatch"/"Aligned" (use "one-to-one" / "one-to-many" / "unclear"), "disqualifying criteria", "verdict". The report is for the user, not a trace of the skill's internal workflow.
2. Do NOT create a separate "Unit of Work Analysis" section. The shape observation is a one-liner in the Summary. If the shape analysis produces a concern, it becomes a normal numbered finding.
3. Size metrics per file type use activity / variable / node counts, not "lines". Lines are meaningless for XAML and misleading for any file. See "Structural Metrics" table below.
4. Low-code agent reviews omit some sections — see agent-grading-rubric.md.
5. Validation Status for Legacy projects says "Use uipath-rpa (Legacy mode) for Legacy-specific validation" — it does NOT say "Could not run" or "Failed". Legacy is supported indefinitely in Studio LTS; the uip rpa CLI targets Modern projects (Legacy mode uses the uip rpa-legacy CLI internally).

Structural metrics to report (never "lines"):

File type	Metrics to use
.xaml	Activity count, max nesting depth, root-scope variable count, argument count, invoke-workflow count
.cs (coded workflow)	Method count, statement count (LOC excluding blank/comment), class count
.flow	Node count, gateway count, longest path depth, subflow count
.py (coded agent)	Function count, statement count, import count
Config (JSON/XLSX)	Entry count, nesting depth

Required report structure:

## Review Report: <Project or Solution Name>

### Summary
- **Overall Quality:** Good / Needs Improvement / Critical Issues
- **Agent Grade:** <A–F> — <verdict label> (<binding constraint, e.g. "gated by G_det = CLI Data.Grade B; judgment clean (G_jud A, 91)">) — *agent projects only; omit this line if the review has no agent projects*
- **Business Value:** <1-2 sentence description of what this automation does>
- **Review Scope:** Single project / Solution (N projects) / Multi-project repo (N executables + M libraries)
- **Project Types Found:** <list with type and language, e.g., "RPA (XAML, VisualBasic)", "Agent (Coded, Python)">
- **Validation Status:** <per project: pass with counts, or "Validation via uipath-rpa (Legacy mode)" for Legacy>
- **PDD Available:** Yes (path) / No — business logic alignment not verified
- **Transaction Shape:** <one line per project, e.g., "Processes 1 invoice per invocation (one-to-one)." or "Processes 1 company per invocation; internally writes N employee enrollments (one-to-many) — see [W-002].">

### PDD Alignment (only if PDD was available)

| PDD Requirement | Implementation Status | Finding |
|---|---|---|
| ... | ... | ... |

> If no PDD: "No PDD was available for this review. Business logic alignment could not be verified."

### Automated Validation Results

| Project | File | Command | Errors | Warnings | Info |
|---|---|---|---|---|---|
| ... | ... | ... | ... | ... | ... |

**Validation Details:** *(Errors and Warnings only — omit the heading entirely when there are none)*
- [V-E-001] <project>/<file>: **<rule-id>** — <message>
- [V-W-001] <project>/<file>: **<rule-id>** — <message>

> For Legacy projects, note: "Validation CLI (`uip rpa validate`, `uip rpa analyze`) targets Modern projects. Legacy validation runs through `uipath-rpa` Legacy mode (using the `uip rpa-legacy` CLI)."

### Rules Skipped

| Rule / Command | Why |
|---|---|
| `uip codedagent review` | CLI not available in environment (deterministic checks not run) |
| `LC_GUARDRAIL_ACTION_INEFFECTIVE`, `LC_GUARDRAIL_MISAPPLIED` | Guardrails catalog unavailable (agent declares 2 guardrails, effectiveness unverified) |

> Only rules that were intended but could not be applied (Critical Rule 11). Group skips sharing one cause into a single row.

### Critical Findings (block deployment)

| ID | Rule | Recommendation |
|---|---|---|
| C-D-001 | `LOWCODE_SYSTEM_MESSAGE_MISSING` | `ClassifierAgent/agent.json`: `messages[0]` (system role) has empty content. Set `messages[0].content` to a non-empty system prompt. |
| C-001 | — | `ProjectA/Helper.cs`: Password argument uses `String`. Change the argument type to `SecureString`. |

### Warnings (should fix before production)

| ID | Rule | Recommendation |
|---|---|---|
| W-D-002 | `LC_PROMPT_ROLE_DEFINITION` | `ClassifierAgent/agent.json`: System prompt starts with task instructions before defining the agent's role. Open with: "You are an X that does Y." |

### Improvement Opportunities

| ID | Rule | Recommendation |
|---|---|---|
| I-D-001 | `LC_GUARDRAIL_RECOMMENDED` | `ClassifierAgent/agent.json`: `inputSchema.properties` contains `customer_email` and `ssn` without a PII guardrail. Add an Agent-scope PII guardrail with a block action. |

> One row per finding. Format each recommendation as `<file>: <issue>. <fix>.` Use the CLI's `File`, `Description`, and `SuggestedFix` verbatim; keep judgment and manual findings concise. Review-CLI, judgment-catalog, and manual-checklist findings all go in these three tables — do not split them into separate sections by source, and never list a finding in more than one table. `Rule` is `—` for a finding with no `rule_id` (Critical Rule 12).

### Per-Project Summary
| Project | Type | Language | Size | Validation | Quality | Grade | Key Findings |
|---|---|---|---|---|---|---|---|
| ClassifierAgent | Agent (Coded) | Python | 14 functions, 220 statements | Pass | Good | B | W-D-002 |
| ProjectA | RPA (Coded) | CSharp | 42 methods, 1,300 statements | 1 error, 2 warnings | Needs Improvement | — | V-E-001, W-001 |
| ProjectB | Flow | — | 18 nodes, 3 gateways, depth 5 | Pass | Good | — | I-001 |
| ProjectC | RPA (XAML) | VisualBasic | 84 activities, 50 vars, depth 12 | Via uipath-rpa (Legacy mode) | Needs Improvement | — | C-002, W-003 |

> The **Grade** column is the per-agent `min(G_det, G_jud)` from Step 4.5 — **agent projects only** (`—` for other types, phase 1). Append the review CLI's `Data.Grade` when it differs, e.g. `B (CLI: A)`. The **Quality** column (Good / Needs Improvement / Critical Issues) applies to every project type.

### Recommended Next Steps

Route each fix to the appropriate skill:

| Fix needed | Use skill |
|---|---|
| Fix RPA workflow / coded workflow / XAML / project.json | `uipath-rpa` |
| Fix RPA Windows-Legacy project | `uipath-rpa` (Legacy mode) |
| Fix agent (coded or low-code) | `uipath-agents` |
| Fix flow (.flow) | `uipath-maestro-flow` |
| Fix Maestro BPMN (.bpmn) | `uipath-maestro-bpmn` |
| Fix API workflow (Workflow.json) | `uipath-api-workflow` |
| Fix coded app | `uipath-coded-apps` |
| Fix Orchestrator resources (assets, queues, folders) | `uipath-platform` |
| Fix `.uipx` solution / pack / publish / deploy lifecycle | `uipath-solution` |

1. Fix [C-001] using `uipath-rpa` — change argument type to SecureString
2. ...

### Optimization Notes
- <queue usage, bulk operations, retry/idempotency observations — e.g., partial-failure handling for one-to-many shapes. Only print section when optimization is relevant and applicable to the project or solution.>

**Final grade: <A–F>**

Final grade: is the report's last line — nothing follows it. No notes, caveats, or commentary, inside the report or after it. It restates the Summary's Agent Grade letter so the grade stays visible at the tail of a long report; the two must always match. Letter only. Only print for agent projects.

Finding severity labels (never "Mismatch"/"Aligned"):

• Overall Quality: Good / Needs Improvement / Critical Issues (all project types)
• Agent Grade: A / B / C / D / F (no +/-) — agent projects only; see Step 4.5 and agent-grading-rubric.md
• Transaction Shape: one-to-one / one-to-many / unclear
• Findings: Critical / Warning / Info

Overall Quality thresholds (all project types):

• Good — 0 Critical, 0–3 Warnings
• Needs Improvement — 0 Critical, 4+ Warnings OR 1 Critical with clear fix
• Critical Issues — 2+ Critical OR 1 Critical with security/data-integrity implications

Agent Grade → verdict label (agent projects only; the line reads "B — Good"):

Grade	Verdict label
A / B	Good
C / D	Needs Improvement
F	Critical Issues

This maps the letter to the verdict word only. The agent grade is min(G_det, G_jud) from Step 4.5, where G_det is the review CLI's Data.Grade and the G_jud band lives in Step 4.5 — do not restate either here.

Task Navigation

I need to...	Read this
Compute the A–F letter grade for an agent (Step 4.5)	agent-grading-rubric.md
Understand the rule row schema	rule-format.md
Run the review CLI + judgment catalog (Step 2.5)	rule-catalog-workflow.md
Apply the low-code agent judgment catalog	agents-lowcode-rules.md
Apply the coded agent judgment catalog	agents-coded-rules.md
Understand the full review workflow in detail	review-workflow-guide.md
Review a solution structure (.uipx)	solution-review-guide.md
Review an RPA project (coded or XAML)	rpa-review-checklist.md
Find common RPA issues	rpa-common-issues.md
Review a flow project	flow-review-checklist.md
Find common flow issues	flow-common-issues.md
Review a Maestro BPMN project (.bpmn)	bpmn-review-checklist.md
Review an API workflow project (Workflow.json)	api-workflow-review-checklist.md

Anti-Patterns — What NOT to Do

1. Do not modify files. This is a review skill, not a builder. Identify issues, recommend fixes, and tell the user which skill to use.
2. Do not review without running automated validation first. Manual review alone misses structural issues that CLI tools catch instantly.
3. Do not skip solution-level discovery. Reviewing a single project without understanding the solution context leads to wrong optimization recommendations (e.g., suggesting queues when the solution already has a dispatcher/performer pattern).
4. Do not report validation errors as manual findings. Reference the validation output — do not re-describe what the CLI already reported.
5. Do not provide a review without severity ratings. Every finding must be Critical, Warning, or Info. An undifferentiated list of issues is not actionable.
6. Do not recommend architecture changes without understanding business context. Ask about volume, frequency, SLA, and error tolerance before suggesting queue-based processing, parallel execution, or other architectural patterns.
7. Do not attempt to fix issues yourself. Report the issue, suggest the fix, name the skill that can apply it. Stop there.
8. Do not flag Windows-Legacy compatibility as Critical. Legacy is supported indefinitely in Studio LTS — 2024.10, 2025.10, 2026.10, and all future LTS releases continue to support creating, opening, editing, running, and deploying Legacy projects. It is NOT a deployment blocker and NOT a mid-term support risk. Deprecation means "no new features added to Legacy," not "Legacy will be removed." Flag as Warning (if the project would benefit from capabilities Legacy lacks — see rpa-review-checklist.md §10 for ranked feature list) or Info (if Studio LTS is the organizational standard or SOAP web services are required). When recommending migration, lead with the 2-3 features most relevant to the project's actual pain (typically Healing Agent, Unified Target / Modern UIA, Object Repository, ScreenPlay, coded test cases, Autopilot, Agents/Maestro). Route Legacy-specific deep validation to uipath-rpa (Legacy mode).
9. Do not recommend removing a dependency without grepping for usages. A package may be the sole supplier of an activity used elsewhere — recommend removal only after confirming no consumers exist.
10. Do not flag package versions. Many UiPath packages currently ship preview-by-default during the public preview phase, and resolution defaults to bringing them in with explicit user confirmation. Surface stability concerns through activity-owner channels, not user-facing review reports.