fix(maestro-case): preserve same-name output bindings

[song-zhao-25]

Summary

• preserve every explicit output-binding operator and both operands from SDD through tasks.md
• dispatch output generation by operator so equal-name extracts such as greeting -> greeting retain reassign semantics and originalVar
• define equal-name extraction as a controlled alias while still deduplicating independently minted output IDs
• centralize the canonical output-row grammar in the shared I/O-binding planner; task plugins keep only a compact lossless-row placeholder and link to that contract
• retain the inline-create boundary reminder so sibling resource schemas cannot replace Case binding rows
• add a general io_binding e2e matrix using existing deterministic tenant resources; keep the original single_node/api-workflow test unchanged
• make post-run solution cleanup non-interactive with the CLI-required --yes

Regression coverage

The new tests/tasks/uipath-maestro-case/io_binding/ eval covers:

• literal, Case-variable, prior-task, and JavaScript-expression inputs
• bare output auto-minting
• renamed and nested extraction
• literal, variable-copy, computed, and metadata assignments
• the exact estimatedAge -> estimatedAge controlled-alias regression
• global output-ID uniqueness, lossless tasks.md, tenant-resource reuse, solution shape, and Case debug execution telemetry

Verification

• skill structure validation, task YAML parsing, Python grader compilation, runtime-payload casing guard, and git diff --check pass
• all eight output-capable task templates route to the shared output-list contract; no task plugin retains a duplicated binding grammar
• local coder-eval build produced a valid Case artifact; correcting the flattened grader's _shared import root allowed the official coder-eval re-evaluation of that exact preserved workspace to pass 2/2 criteria with score 1.000
• Case validation passed, every structural I/O assertion passed, all five task executions completed in Case debug, and post-run solution cleanup succeeded
• passing local result: tests/runs/2026-07-18_15-03-43
• CI CLI verb gate and task-schema validation pass

CI note

maestro-case checker unit tests currently fails during pytest 9 collection before executing tests because two pre-existing files share the module name test_checkers.py (athena_cm_event and golden_rebuild/cm_golden_expense). The stacked base PR #2117 fails the same job.

Dependency

Stacked on #2117 because the same-name fix was developed against the inline-resource Case changes introduced there.

[github-actions]

Claude finished @song-zhao-25's task in 2m 46s —— View job

---

Coder-eval task lint (advisory)

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

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

Evidence of passing run

• ✅ Author confirms passing run: "local coder-eval build produced a valid Case artifact; correcting the flattened grader's _shared import root allowed the official coder-eval re-evaluation of that exact preserved workspace to pass 2/2 criteria with score 1.000" and "passing local result: tests/runs/2026-07-18_15-03-43"

Per-task lint

tests/tasks/uipath-maestro-case/io_binding/io_binding.yaml — verdict: OK

No issues. The task is well-structured:

• Self-report: criteria grade real CLI validation (uip maestro case validate) and a comprehensive Python grader against the actual caseplan.json + live debug run — no self-written reports.
• Prompt: states the goal and constraints (build from SDD, preserve output rows, no inline API project). The estimatedAge -> estimatedAge mention is a ground-truth anchor for the regression under test, not procedure leakage.
• Meaningful coverage: the Python grader (check_io_binding_case.py) validates resource bindings, input/output row shapes (bare, reassigned, custom, equal-name), task ordering, $xref resolution, global output-ID uniqueness, tasks.md preservation, variable contracts, solution shape, and all five debug executions reaching Completed/Successful.
• Gameability: requires real tenant resource resolution, real uip maestro case validate, and real uip maestro case debug execution — cannot be faked.
• Near-duplicate: materially distinct from single_node/api_workflow (single task, simple I/O) and entry_points_io (entry-points.json back-fill, not task-level bindings). No other sibling covers the full I/O binding matrix or same-name extraction regression.
• CLI verb reachability: unable to run scripts/check-cli-verbs.py (execution not approved), but the only CLI command is uip maestro case validate — a well-established verb present in dozens of sibling tasks.
• Sandbox: no env_packages, no redundant CLI install.
• Run-limits: correctly placed under top-level run_limits:, not under agent:.

Within-PR duplicates

• No duplicate clusters detected (only one task YAML added).

Conclusion

• ✅ All changed tasks pass the rubric. Evidence of passing run confirmed.

[github-actions]

Claude finished @song-zhao-25's task in 6m 48s —— View job

---

PR Review: fix(maestro-case): preserve same-name output bindings

• Read review criteria and project conventions
• Analyze full diff against origin/main
• Read existing PR conversation
• Review skill documentation changes
• Review test changes
• Check cross-skill references and link validity
• Check for secrets/tokens/paths
• Post final review

---

Summary

This PR fixes a regression where an equal-name output extraction like greeting -> greeting was incorrectly collapsed to bare greeting during planning, losing the reassignment semantics (originalVar) needed by the frontend. The fix centralizes the canonical output-row grammar in the shared I/O-binding planning plugin, updates all 8 task-type planning templates to point to that contract, and adds a comprehensive e2e test covering the full I/O binding matrix plus the specific equal-name regression. The cleanup script also gains --yes to prevent interactive prompts.

Change-by-Change Review

1. skills/uipath-maestro-case/SKILL.md (Critical Rule 6)

Severity: OK

Adds "and lossless" to the rule title plus a concise inline explanation of why greeting -> greeting differs from bare greeting. Well-scoped addition that strengthens the critical rule without bloating it.

2. skills/uipath-maestro-case/references/bindings-and-expressions.md

Severity: OK

Changes inline comma-separated outputs: decision, comments to list format (- outputs: / - decision / - comments). Consistent with the new one-item-per-row contract established in io-binding/planning.md.

3. skills/uipath-maestro-case/references/planning.md

Severity: OK

Two clean changes: (1) adds parenthetical "always read alongside the matching task plugin" to the I/O binding nav entry; (2) replaces the task-plugin-only instruction with an explicit dual-read directive ("Read both the task-type plugin and the shared I/O-binding plugin before writing the entry"). The new "Outputs are a lossless handoff" blockquote is well-placed after the field list and before the registry handoff. No duplication with the I/O-binding plugin — this is bridge content tying planning-level context to the shared contract.

4–9. references/plugins/tasks/{action,agent,api-workflow,case-management,process,rpa}/planning.md

Severity: OK

All six files receive the same pattern: (a) outputs field row changes from "Discovered via tasks describe" to "sdd.md task Outputs + resolved schema → Follow the shared I/O-binding output-list contract" with a link to io-binding/planning.md#canonical-tasksmd-output-list; (b) template code block changes from - outputs: <out1>, <out2> to the list form - outputs: / - <SDD output row, copied verbatim>. All links verified — the anchor target exists at io-binding/planning.md:60.

10–11. references/plugins/tasks/{connector-activity,connector-trigger}/planning.md

Severity: OK

These two files, which previously had no explicit output guidance, now get a one-line directive ("Populate outputs: using the shared I/O-binding output-list contract") plus the list-form template entries. Consistent treatment across all 8 task types.

12. references/plugins/tasks/create-inline-common.md

Severity: OK

Adds a "Projection is not normalization" blockquote reminding agents that the pinned builder contract (field names/types) does not replace the case binding rows — the original SDD operator and destination must be preserved in tasks.md. Good boundary reminder for the inline-resource creation flow.

13. references/plugins/variables/global-vars/impl-json.md (Uniqueness Rule)

Severity: OK

Three additions: (1) clarifies "independently minted" qualifier on uniqueness rule; (2) new "Reassigned-output collision" paragraph explaining that -> rows have independently minted source-side id that must be deduped; (3) new "Controlled alias" paragraph explaining the greeting -> greeting exception where id == var == originalVar and neither side gets a counter suffix. Both paragraphs include concrete examples. The prose is dense but technically precise and matches the impl-json dispatch rules.

14. references/plugins/variables/io-binding/impl-json.md

Severity: OK

The core fix: (1) the -> extract shape now explicitly allocates id from the global uniqueness pool and mirrors the allocated id into originalVar and target; (2) new "Equal-name extract edge case" paragraph after the shape list dispatches greeting -> greeting to the reassign shape with id == var == originalVar == target_var; (3) uniqueness cross-cutting rule clarifies var is never counter-suffixed on a -> row (it points to the existing case variable), while the source-side id is independently minted. Also adds a same-name annotation paragraph in the "Output Binding Shapes" subsection.

15. references/plugins/variables/io-binding/planning.md

Severity: OK

The centralizing change: (1) intro paragraph adds "The output operator is part of the data contract, not formatting: preserve ->, =, or bare form and both operands through tasks.md"; (2) new "Equal-name extract is still an extract" callout under the -> operator section; (3) new "Canonical tasks.md output list" subsection (the anchor target all 8 task plugins now link to) with a concrete example and the rule against comma-separated names; (4) new validation rule row in the table: "Any SDD Outputs row differs from its tasks.md item in operator or operand → ERROR".

16. tests/tasks/uipath-maestro-case/_shared/cleanup_solutions.py

Severity: OK

Three small fixes: (a) adds --yes flag to uip solution delete to prevent interactive prompts in CI; (b) updates the never policy log message to include --yes --output json; (c) falls back to r.stdout when r.stderr is empty for failure diagnostics. All correct and consistent.

17. tests/tasks/uipath-maestro-case/io_binding/check_io_binding_case.py

Severity: OK

424-line Python grader covering: (a) solution shape (no inline API projects, exactly one CaseManagement project); (b) resource bindings (Golden Expense echo and NameToAge each resolve to correct folder/resourceKey); (c) input bindings (literal, case-variable, cross-task, expression); (d) output bindings (bare auto-mint, reassigned ->, nested Error.Message ->, custom = literal/copy/computed/metadata); (e) the specific estimatedAge -> estimatedAge equal-name regression with equal_name=True flag; (f) task ordering via selected-tasks-completed entry conditions; (g) global output-ID uniqueness; (h) tasks.md lossless preservation (regex checks for all 8 output rows); (i) runtime debug execution telemetry. Well-structured with clear helper functions and descriptive failure messages.

18. tests/tasks/uipath-maestro-case/io_binding/fixtures/sdd.md

Severity: OK

212-line SDD fixture designing a single-stage case with 5 sequential api-workflow tasks against two existing tenant resources. Exercises all binding forms systematically: literal input, case-variable input, cross-task input, expression input, bare output, renamed extraction, nested extraction, literal/copy/computed/metadata assignments, and the estimatedAge -> estimatedAge regression. Section numbering jumps from 1.1 to 1.3 (missing 1.2) — this follows the SDD template convention where §1.2 (Stages Summary) is omitted for single-stage cases and is not an error.

19. tests/tasks/uipath-maestro-case/io_binding/io_binding.yaml

Severity: OK

Task YAML is well-structured: task_id: skill-case-io-binding-matrix follows the skill-<domain>-<capability> pattern; tags include uipath-maestro-case as first tag plus e2e, mode:build, and other valid tags; initial_prompt is goal-focused and not hand-holding; two run_command success criteria with appropriate weights (3.0 for validate, 6.0 for the full grader); no env_packages or @uipath/cli pin. post_run cleanup correctly wired.

What's Missing

Nothing identified. The PR is well-scoped: skill doc changes, test coverage, and cleanup fix all align with the bug fix.

Area Ratings

Area	Status	Notes
Frontmatter	OK	SKILL.md frontmatter unchanged; body change is scoped to Critical Rule 6
E2E Tests	OK	New io_binding e2e task covers the regression and the full binding matrix
Skill Body	OK	Critical Rule 6 updated with lossless semantics; no other body sections modified
References & Assets	OK	15 reference files updated consistently; all links resolve; no content duplication beyond bridge sentences
Repo Hygiene	OK	CODEOWNERS entry exists; no secrets/paths; no cross-skill dependencies; --yes flag fix prevents interactive CI prompts

Issues for Manual Review

• Domain accuracy of the equal-name controlled-alias rule. The PR defines that greeting -> greeting produces id == var == originalVar == "greeting" and skips the uniqueness counter. This is a semantic contract with the UiPath frontend (mutateRootVariables in VariableMutationUtils.ts). Verifying that this matches the actual FE behavior requires domain expertise beyond what the skill docs can validate.
• Tenant resource availability. The e2e test depends on two specific tenant resources (CM-Golden-Expense-Reporting-106 and NameToAgeFixed2 in Shared/uipath-maestro-case/). If these are removed or renamed in the tenant, the test will fail. The PR body mentions "existing deterministic tenant resources" — confirming their availability is a manual check.
• Stacked PR dependency. This PR is stacked on test(maestro-case): cover inline resource creation #2117. Merging out of order would break the base. Ensure test(maestro-case): cover inline resource creation #2117 merges first.

Conclusion

Clean PR. The documentation changes are internally consistent, well-centralized (single source of truth for the output-list grammar, 8 task plugins link to it), and the new e2e test provides thorough regression coverage for the equal-name binding fix plus the broader I/O binding matrix. The cleanup script fix (--yes flag) is a sensible CI hardening. No blocking issues found — approve.
|

[charlesliu9]

is the test failure due to the issue fixed in another PR?

[song-zhao-25]

@charlesliu9 yes

[github-actions]

Claude finished @song-zhao-25's task in 1m 18s —— View job

---

Coder-eval task lint (advisory)

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

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

Evidence of passing run

• ✅ Author confirms passing run: "local coder-eval build produced a valid Case artifact; correcting the flattened grader's _shared import root allowed the official coder-eval re-evaluation of that exact preserved workspace to pass 2/2 criteria with score 1.000" and "passing local result: tests/runs/2026-07-18_15-03-43"

Per-task lint

tests/tasks/uipath-maestro-case/io_binding/io_binding.yaml — verdict: OK

No issues. The task is well-structured:

• Self-report: criteria grade real CLI validation (uip maestro case validate) and a comprehensive Python grader against the actual caseplan.json + live debug run — no self-written reports.
• Prompt: states the goal and constraints (build from SDD, preserve output rows, no inline API project). The estimatedAge -> estimatedAge mention is a ground-truth anchor for the regression under test, not procedure leakage.
• Meaningful coverage: the Python grader (check_io_binding_case.py) validates resource bindings, input/output row shapes (bare, reassigned, custom, equal-name), task ordering, $xref resolution, global output-ID uniqueness, tasks.md preservation, variable contracts, solution shape, and all five debug executions reaching Completed/Successful.
• Gameability: requires real tenant resource resolution, real uip maestro case validate, and real uip maestro case debug execution — cannot be faked.
• Near-duplicate: materially distinct from single_node/api_workflow (single task, simple I/O) and entry_points_io (entry-points.json back-fill, not task-level bindings). No other sibling covers the full I/O binding matrix or same-name extraction regression.
• CLI verb reachability: unable to run scripts/check-cli-verbs.py (execution not approved), but the only CLI command is uip maestro case validate — a well-established verb present in dozens of sibling tasks.
• Sandbox: no env_packages, no redundant CLI install.
• Run-limits: correctly placed under top-level run_limits:, not under agent:.

Within-PR duplicates

• No duplicate clusters detected (only one task YAML added).

Conclusion

• ✅ All changed tasks pass the rubric. Evidence of passing run confirmed.