fix(uipath-agents): forbid switching an existing coded agent's framework

[andreibalas-uipath]

Why

skill-agent-coded-antipattern-openai-agents-hitl failed in CI (codex / gpt-5.6-terra, score 0.125). The seed is an OpenAI Agents project that misuses LangGraph's interrupt(...). The agent correctly diagnosed the mismatch, then read the OpenAI integration note ("does not support HITL … use LangGraph or LlamaIndex instead") as a remediation and migrated the whole project to LangGraph — deleting openai_agents.json, swapping deps, rewriting main.py as a StateGraph.

The coded quickstart only stated framework selection for greenfield projects; for existing projects the "keep it" intent lived solely in a step-skip gate (Framework — skip if framework != none), not as a rule.

What

• skills/uipath-agents/references/coded/quickstart.md § Critical Rules — new rule:

  Never switch an existing project's framework. When framework != none (a <framework>.json is already present), the framework is fixed: do not migrate to another framework, swap the <framework>.json, or change framework dependencies in pyproject.toml. Work within the existing framework's capabilities; if a request cannot be met within them, tell the user the limitation and let them decide.

  Deliberately generic — no worked example — so the agent pairs the rule with the capability facts already stated in the framework/capability references.

• tests/.../check_antipattern_openai_agents_hitl.py — dropped the failure hint "or, if HITL is essential, migrate the project to LangGraph", which contradicted the task's own criteria.

Verification

• npm run skills:validate OK (default + studioweb; rule sits outside all flavor markers).

• Local coder-eval runs on this state, experiments/default.yaml, 3 replicates each:

  Harness / model	Pass	Duration
  codex / gpt-5.6-terra	3/3	105–131 s
  claude-code / claude-sonnet-4-6	3/3	80–106 s

• CI brownfield sweep (Run Coder Eval, all 8 lifecycle:edit coded tasks × codex + claude): target task passes on both harnesses; 15/16 agent-graded runs pass with 0 framework migrations. The one non-pass (claude guardrails/byog_middleware) is a missing tenant fixture — byog-smoke-agent-pin isn't registered on the alpha codereval tenant and the agent correctly refused to fabricate it (skill Rule 18); separate follow-up.

The greenfield strict-input flake (strict_input_no_messages) seen on this PR's smoke run is addressed separately in #2888.

🤖 Generated with Claude Code

[github-actions]

Claude finished @andreibalas-uipath's task in 1m 37s —— View job

---

PR Review: fix(uipath-agents): forbid switching an existing coded agent's framework

• Read review criteria and project conventions
• Read full diff and changed files
• Check skill-specific constraints
• Post review findings

---

Summary

Adds a new Critical Rule to quickstart.md forbidding framework migration on existing coded agent projects, and removes a contradictory migration hint from the antipattern test's failure message. Motivated by a CI failure where the agent read "migrate the project to LangGraph" as remediation advice and rewrote the entire project.

Change-by-Change Review

1. skills/uipath-agents/references/coded/quickstart.md (line 39)

Severity: OK

New critical rule is well-placed — immediately after the "Select a framework before writing any code" rule, forming a natural pair (greenfield selection → existing-project lock). The rule is:

• Prescriptive and unambiguous ("do not migrate", "do not swap", "do not change").
• Scoped correctly — keyed on the framework != none detection variable already defined in the Project State Detection table (line 12).
• Includes the right fallback: "tell the user the limitation and let them decide" — avoids a blanket refusal.
• Consistent with the existing step 1 gate at line 85 (Skip if framework != none), which already skipped framework selection but lacked a rule forbidding active migration. The new rule closes that gap.

No token-optimization concerns — rule text is concise and follows the § 0 "keep full sentences for rules" boundary.

2. tests/.../check_antipattern_openai_agents_hitl.py (lines 59–61)

Severity: OK

Removed the parenthetical "(or, if HITL is essential, migrate the project to LangGraph)" from the failure message. This was the root cause — the test's own error text suggested migration, which contradicted both the task's success criteria (which assert openai_agents.json must still exist and Agent construction must be preserved) and the new critical rule. The remaining message is clear: "Remove the interrupt call."

The task YAML (antipattern_openai_agents_hitl.yaml) requires no changes — its initial_prompt and success_criteria were already correct (they grade framework preservation, not migration).

What's Missing

Nothing identified. The change is minimal and correctly scoped:

• No other files in the skill or test suite reference "migrate to LangGraph" (verified via grep).
• The task YAML's description and initial_prompt don't mention migration.
• CODEOWNERS is not affected (no new paths).
• No flavor overrides needed — the new rule sits outside all skill-flavor markers.

Area Ratings

Area	Status	Notes
Frontmatter	OK	Not modified
E2E Tests	OK	Existing test updated to align with the new rule; task YAML unchanged
Skill Body	OK	New critical rule is well-positioned and prescriptive
References & Assets	OK	No broken links; rule references existing detection variables
Repo Hygiene	OK	No secrets, no cross-skill deps, changes scoped to one skill

Issues for Manual Review

None found. The PR description includes passing local coder-eval runs (3/3 on two harness/model combos), confirming the fix prevents migration behavior.

Conclusion

Clean, well-scoped fix. The new rule closes a real gap (the "framework is fixed" intent existed only as a step-skip gate, not as an explicit prohibition), and the test change removes the contradictory advice that caused the CI failure. Approve.

---