fix(tests): unstarve three uipath-agents smoke tasks (max_turns -> 60)

[mjnovice]

Why

The uipath-agents smoke suite is flaking below the 95% gate because three tasks keep hitting MAX_TURNS_EXHAUSTED with high partial scores and zero failed commands — pure turn starvation, the same disease #2349 and #2234 fixed for other tasks.

Evidence from PR #1694's three smoke runs (26/29, 27/29, 27/29 — a different failure set each time) plus PR #1801's run:

Task	max_turns	Starvation evidence
skill-agents-coded-is-smoke	30 (below the smoke experiment default of 40)	Starved twice in a row; score 0.84, 0 failed commands, only the doc-read criterion left
skill-agent-context-prompt-reference	30	Starved on #1694's run and #1801's run; mid-flow at score 0.41 after creating agent.json + the context resource
skill-agent-coded-conversational-langgraph	40 (inherited)	Starved on #1694's rerun

What

Raise all three to expected_turns: 40 / max_turns: 60, matching #2349's precedent for eval_custom_evaluator. No prompt or criteria changes.

🤖 Generated with Claude Code

[github-actions]

Claude finished @mjnovice's task in 2m 51s —— View job

---

Coder-eval task lint (advisory)

3 task YAMLs changed; verdict counts: 0 Critical, 0 High, 0 Medium, 0 Low, 3 OK.

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

Evidence of passing run

❌ High — PR body does not claim the changed tasks have been run and passed with the new max_turns values. The body cites starvation evidence from prior PRs (#1694, #1801) showing the tasks fail at their old limits, but doesn't assert a passing run at max_turns: 60. Please edit the PR description to add a line like:

Ran skill-agents-coded-is-smoke, skill-agent-context-prompt-reference, and skill-agent-coded-conversational-langgraph locally and they passed.

(Honor-system — a verbal claim is sufficient.)

Per-task lint

tests/tasks/uipath-agents/coded/conversational_agent_langgraph/conversational_agent_langgraph.yaml — verdict: OK

No issues. Change is run_limits only (expected_turns: 20→40, added max_turns: 60). Task has strong criteria: command_executed for new/init/run + a run_command Python checker (weight 5.0). Not a near-duplicate of langgraph_classifier — that task is e2e with schema-sync coverage and a classifier workflow; this one tests the conversational (isConversational: true) path with two chat turns.

tests/tasks/uipath-agents/coded/is_smoke/is_smoke.yaml — verdict: OK

No issues. Change is run_limits only (expected_turns: 20→40, max_turns: 30→60). Criteria include skill_triggered, command_executed for the doc-read, file_exists, and a run_command checker at weight 6.0. Not a near-duplicate of is_outlook_send_mail — that task is e2e with multipart edge case and tenant verification; this one is a smoke testing artifact shape without a live tenant.

tests/tasks/uipath-agents/lowcode/context_prompt_reference/context_prompt_reference.yaml — verdict: OK

No issues. Change is run_limits only (expected_turns: 20→40, max_turns: 30→60). Criteria cover the full init→refresh→validate lifecycle plus a run_command checker (weight 5.0) verifying @{contexts.Knowledge} expression tokens. Not a near-duplicate of context_index — that task is e2e testing index-type context resources with solution-level commands; this one exercises the @{}→expression contentToken path with an attachments context.

Within-PR duplicates

No duplicate clusters detected — all three tasks target distinct features (conversational LangGraph, coded IS capability, low-code expression tokens).

Conclusion

✅ All changed tasks pass the rubric. The only finding is the missing passing-run claim in the PR body (advisory).

---

[github-actions]

Claude finished @mjnovice's task in 1m 53s —— View job

---

Reviewing PR #2413

• Read review criteria and project conventions
• Run git diff and read full changed files
• Check for precedent consistency and task YAML validity
• Post review

---

Summary

Raises expected_turns / max_turns from 20–30/30–40 to 40/60 on three uipath-agents smoke tasks that are consistently hitting MAX_TURNS_EXHAUSTED with high partial scores and zero failed commands. No prompt or criteria changes — pure turn-budget relief.

Change-by-Change Review

1. tests/tasks/uipath-agents/coded/conversational_agent_langgraph/conversational_agent_langgraph.yaml

Severity: OK

expected_turns raised 20→40, max_turns: 60 added (was absent, so the task previously inherited the experiment default of 200 — the expected_turns: 20 was the effective cap via coder-eval's behavior). The new values are consistent with other tasks in this suite that needed similar relief (bindings_sync, edit_roundtrip).

2. tests/tasks/uipath-agents/coded/is_smoke/is_smoke.yaml

Severity: OK

expected_turns raised 20→40, max_turns raised 30→60. Previously the tightest budget in the suite (max_turns: 30, below even the smoke experiment default). The bump matches the established 40/60 pattern.

3. tests/tasks/uipath-agents/lowcode/context_prompt_reference/context_prompt_reference.yaml

Severity: OK

Same 20→40 / 30→60 change. turn_timeout: 1200 preserved.

What's Missing

Nothing identified. The PR is scoped correctly to the three starving tasks with clear evidence in the description. No prompt or criteria changes needed — the tasks were passing all criteria except they ran out of turns before completion.

Area Ratings

Area	Status	Notes
Frontmatter	N/A	No skill frontmatter changed
E2E Tests	OK	Only run_limits tuned — tags, criteria, prompts untouched
Skill Body	N/A	No skill docs changed
References & Assets	N/A	No references or assets changed
Repo Hygiene	OK	No secrets, no unrelated changes, scoped to the three affected tasks

Issues for Manual Review

None found.

Conclusion

Clean, well-scoped fix. The 40/60 values are consistent with the existing pattern across other uipath-agents tasks that needed similar relief (bindings_sync, edit_roundtrip). The starvation evidence in the PR description is thorough. Approve.