fix(maestro-case-tests): steer loan phase-0 prompt to trigger uipath-maestro-case

[abhiram-vad]

Why

skill-case-phase-0-loan-origination failed on 2026-07-16 (score 0.118) with a complete, coherent 34KB draft on disk. The run never invoked uipath-maestro-case: the closing ask "I want a draft design document" matched uipath-planner's Phase-D triggers ("design/architect/generate an SDD → Phase D"), and planner Rule 6 wrote the draft to loan-origination-sdd.md in one shot — no interview. Every criterion and the LLM judge read only sdd.draft.md/sdd.md, so everything scored 0 (judge received <file not found>).

What

Prompt-only change: the closing paragraph now anchors on maestro-case's own trigger clauses — "We don't have an sdd.md yet — interview me…" (matches "via lightweight interview if sdd.md absent") and "build the caseplan later" (matches "Always invoke for caseplan.json"; planner redirects caseplan → maestro-case) — and drops the planner-bait "design document" wording. Criteria, judge, and simulation are untouched.

Note: the planner/maestro-case description overlap is the underlying product gap (planner's deferral_single_case test only passes because its prompt avoids design-document phrasing) — deferred to a separate skill-description change if this recurs.

Verification

• YAML parses; /lint-task verdict: Medium (pre-existing advisory — no skill_triggered gate; unchanged by this PR)
• Passing coder-eval run on this branch: run 29531235811 — skill-case-phase-0-loan-origination: SUCCESS, score 1.000, 6 interview turns (vs 1 turn / 0.118 on the failing 07-16 run), llm_judge 1.00

🤖 Generated with Claude Code

[github-actions]

Claude finished @abhiram-vad's task in 2m 3s —— View job

---

Coder-eval task lint (advisory)

1 task YAML changed; verdicts: 0 Critical, 0 High, 1 Medium, 0 Low, 0 OK.

Rubric: .claude/commands/lint-task.md. This check is advisory and never blocks merge.

Evidence of passing run

❌ High — PR body says a coder-eval CI run was dispatched but does not claim it has passed. Please update the PR description once the run completes, e.g.: Ran skill-case-phase-0-loan-origination and it passed.

Per-task lint

tests/tasks/uipath-maestro-case/phase_0_to_case/loan_origination/loan_origination.yaml — verdict: Medium

Issues:

• [Medium] Could pass for the wrong reason (lines 37–57, 59–139): no skill_triggered criterion — the prompt supplies all domain knowledge (stages, exception lanes, $5M rule), so an agent could write a passing sdd.draft.md directly from the prompt without invoking uipath-maestro-case. The llm_judge (line 104) grades domain coherence but cannot distinguish skill-produced output from agent-authored prose. Pre-existing; unchanged by this PR.

Suggested fixes:

• Consider adding a skill_triggered criterion for uipath-maestro-case to verify the skill was actually invoked. This is the root cause of the original failure (planner handled the request instead of maestro-case) and would catch that class of regression mechanically.

Within-PR duplicates

• No duplicate clusters detected. (Only 1 task YAML changed.)

Regarding the prompt change itself

The diff is well-targeted. The old closing paragraph — "I want a draft design document" — matched uipath-planner's Phase-D triggers. The new wording anchors on two uipath-maestro-case trigger clauses:

• "We don't have an sdd.md yet — interview me" → matches "via lightweight interview if sdd.md absent"
• "build the caseplan later" → matches "Always invoke for caseplan.json" (planner redirects caseplan → maestro-case)

This is a reasonable prompt-steering fix. The PR description correctly identifies the underlying product gap (planner/maestro-case description overlap) and explicitly defers a skill-description change.

Conclusion

⚠ 1 task has issues, max severity Medium. Advisory only — not blocking merge. The skill_triggered gap is pre-existing and the prompt change is a sound fix for the skill-routing failure. Please update the PR description with a passing-run claim once the dispatched coder-eval run completes.

---

[github-actions]

Claude finished @abhiram-vad's task in 1m 41s —— View job

---

Reviewing PR #2095

• Read review criteria and project conventions
• Read full diff and changed file
• Check existing PR conversation
• Perform review checks
• Post review findings

---

Summary

Prompt-only fix to the loan_origination test task. The closing paragraph of initial_prompt is reworded to anchor on uipath-maestro-case trigger clauses ("We don't have an sdd.md yet — interview me", "build the caseplan later") and drops the "design document" phrasing that was matching uipath-planner's Phase D triggers.

Change-by-Change Review

1. tests/tasks/uipath-maestro-case/phase_0_to_case/loan_origination/loan_origination.yaml (lines 54–57)

Severity: OK

The old prompt said "I want a draft design document" — this matches uipath-planner's trigger: "design / architect / generate an SDD → Phase D" (SKILL.md:26). The new wording hits two uipath-maestro-case trigger phrases:

1. "We don't have an sdd.md yet — interview me" → matches maestro-case description: "via lightweight interview if sdd.md absent"
2. "build the caseplan later" → matches maestro-case description: "Always invoke for caseplan.json files"; planner explicitly redirects caseplan work to maestro-case

The remaining phrase "just capture the design" retains the word "design" but in a subordinate clause without "document"/"SDD"/"architect" — unlikely to re-trigger planner. Simulation persona, constraints, success criteria, and judge prompt are all untouched — correct, since the issue was routing, not grading.

No YAML structural changes, no tag changes, no new files. Task parses cleanly.

What's Missing

Nothing identified. This is a narrowly scoped prompt fix; criteria and simulation are appropriately left unchanged.

Area Ratings

Area	Status	Notes
Frontmatter	N/A	No frontmatter changes
E2E Tests	OK	Existing task modified, tags and criteria unchanged
Skill Body	N/A	No skill changes
References & Assets	N/A	No reference changes
Repo Hygiene	OK	No secrets, no cross-skill deps, scoped change

Issues for Manual Review

• Skill description overlap. The PR description correctly identifies that uipath-planner and uipath-maestro-case have overlapping trigger surface around "design document" phrasing. This prompt fix works around the overlap, but the underlying ambiguity remains. If other test prompts (or real user prompts) use "design document" wording with case-management intent, they'll hit the same planner mismatch. The PR defers a skill-description fix — worth tracking.
• CI verification. The PR mentions a coder-eval run was dispatched. Confirming a passing score on the updated prompt would complete validation.

Conclusion

Clean, well-scoped fix. The prompt change correctly re-anchors on uipath-maestro-case's trigger clauses and removes the planner-bait phrasing. No issues found — approve.