feat(uipath-rpa): execution maps + pattern card — cut greenfield builds from ~36-40 to ≤10 turns

[AlvinStanescu]

Summary

Implements the turn-count reduction design for uipath-rpa (spec: docs/superpowers/specs/2026-07-02-uipath-rpa-execution-map-design.md, plan: docs/superpowers/plans/2026-07-02-uipath-rpa-execution-maps.md). A greenfield XAML build collapses from per-activity discovery + validate-after-every-activity loops (~36–40 turn budgets on comparable tasks) to a 4-turn happy path with a single validate+build gate.

Four pieces:

1. references/execution-maps-guide.md — turn-by-turn maps for four journeys (greenfield/brownfield × XAML/coded): T1 scaffold+context chain, T2 batch-author ∥ install, T3 validate && build, T4 report. Includes activity source precedence (card → agent memory → Rule 21 triple), never-batch gates, failure exits, and a harness-conditional cross-session memory protocol.
2. references/common-pattern-card.md — 8 multi-activity snippets, every one CLI-verified (uip rpa validate + build clean) in a scratch Portable/VB project against UiPath.System.Activities 26.6.1, UiPath.Excel.Activities 3.6.0-preview (CSV), UiPath.WebAPI.Activities 2.5.1. Card entries supersede the Rule 21 discovery triple for covered activities. Regeneration procedure: .maintenance/pattern-card-maintenance.md.
3. SKILL.md integration (net −10 lines): Rule 18 rewritten (batch-author, single gate), § Call Batching → § Execution Maps, Rule 21 precedence, greenfield skips the discovery subagent, completion saves memory patterns. workflow-guide.md + environment-setup.md aligned.
4. tests/tasks/uipath-rpa/execution_map/execution_map_greenfield.yaml — smoke task enforcing the budget (run_limits.max_turns: 10), with an independent run_command build re-run so invocation-only or garbage-workflow runs fail, and a command_not_executed guard on get-default-xaml proving the card superseded discovery.

Critical Rules change justification

XAML Rule 18 ("Start minimal, iterate to correct — one activity at a time, validate after each") is rewritten, not removed: authoring now batches, but the per-file validate-to-clean requirement, project build gate, 5-attempt caps, and fix-by-category flow (Rule 19) are unchanged. Card/memory-sourced activities carry low hallucination risk, which is what makes batch authoring safe; failures bisect.

Build-verified traps encoded as card Notes

The validation gauntlet caught real doc/platform drift that validate alone misses — now teaching material on the card:

• BuildDataTable absent from cross-platform assemblies (TypeLoadException) → GenerateDataTable
• ParsingMethod is designer-only; runtime member is CSVParsing
• x:TypeArguments="x:Object[]" fails at build → s:Object[]
• HTTP types live in assembly UiPath.Web.Activities, NOT package id UiPath.WebAPI.Activities; HttpResponseSummary is in ...Http.Models
• Excel X family + SMTP SendMail are Windows-only → deliberately off the card with a routing note until a Windows-target validation pass runs

Verification

• Every card snippet: per-file uip rpa validate clean + project uip rpa build clean (uip 1.0.0-alpha, Helm macOS)
• hooks/validate-skill-descriptions.sh, scripts/check-skill-status.py, scripts/check-cli-verbs.py (all uipath-rpa tasks): clean
• All relative links in new/edited files resolve; no UIA CLI leakage (per skills/uipath-rpa/CLAUDE.md boundary)

Pending before merge (why this is a draft)

• coder-eval passing-run claim — skill-rpa-execution-map-greenfield not yet run (no tests/.env agent credentials in the authoring environment). Run: cd tests && make install && SKILLS_REPO_PATH=$(cd .. && pwd) .venv/bin/coder-eval run tasks/uipath-rpa/execution_map/execution_map_greenfield.yaml -e experiments/smoke.yaml, then record the claim here.
• Reviewer sign-off on two agent-defaulted decisions (user AFK at design review): author-all/validate-once over tiered batching; turn budgets ≤5 greenfield / ≤4 brownfield.
• Follow-up (separate PR): Windows-target validation pass to add Excel X + SMTP entries to the card.

🤖 Generated with Claude Code