Change only starter/ticket_triage.py. Standard library only. Never modify tests/, solution/, or scenario/. ReplayClient, build_messages, load_scenario, and main are provided and already correct; the work is in the four functions below. A client is any callable client(messages) -> str; every string it returns is untrusted input.

• parse_triage(raw, categories) -> dict:
  • Strip surrounding whitespace and one surrounding Markdown code fence (``` or ```json) before parsing.
  • The reply must be a JSON object with exactly the fields category, priority, summary, needs_human — a missing field or an extra field raises TriageError naming that field.
  • category is one of categories; priority is an int from 1 to 4 (bool and float are rejected); summary is a non-empty string of at most 120 characters after stripping; needs_human is a bool.
  • Every rejection raises TriageError whose message names the offending field. Non-JSON text and non-object JSON also raise TriageError.
  • Return exactly the four fields, with summary stripped.
• run_tool_call(call, tools) -> (name, result):
  • tools maps a tool name to {"function": callable, "parameters": {name: type}} — it is the allowlist.
  • call must have exactly the keys tool and arguments. An unregistered tool name raises ToolCallError naming the tool.
  • arguments must have exactly the declared parameter names, each an instance of its declared type (bool is not accepted for int). Missing, unexpected, or mistyped arguments raise ToolCallError naming the argument. The tool function is never called on a refused call.
  • Return (name, function(**arguments)).
• triage(ticket_text, client, *, categories, tools=None, max_attempts=2, max_tool_calls=3) -> dict:
  • Start from build_messages(...). After each reply, append {"role": "assistant", "content": raw}.
  • A reply that decodes to an object with a tool key is a tool call. On success, append {"role": "tool", "name": name, "content": json.dumps(result)} and ask the client again; successful tool calls do not count as attempts. tool_calls counts only calls that ran. A tool call that would exceed max_tool_calls is not run and ends in the fallback.
  • Any rejected reply — invalid answer or refused tool call — counts as one attempt and appends {"role": "user", "content": ...} that contains the TriageError message, then asks again.
  • A valid answer returns the four fields plus attempts (rejected replies
    • 1), tool_calls, and fallback: False.
  • After max_attempts rejected replies, return the fallback: {"category": None, "priority": None, "summary": ticket_text.strip()[:120], "needs_human": True} plus attempts (the number of rejected replies), tool_calls, and fallback: True. Never guess a category.
• evaluate(cases, client_for, *, categories, tools=None) -> dict:
  • Run triage for every case with client_for(case). A case is correct when category and needs_human both match case["expected"].
  • Return total (every case), correct, fallbacks, schema_valid (total - fallbacks), accuracy (round(correct / total, 3), 0.0 for no cases), and failures: one {"id", "reason"} per incorrect or fallback case, in case order.
  • Never skip a case: the denominator is every case you were given.

python ticket_triage.py <scenario-dir> [--min-accuracy X] prints accuracy=… correct=N/T fallbacks=F min_accuracy=…, one line per failure, and returns 0 when accuracy >= min_accuracy, else 1; usage errors return 2. The threshold defaults to the scenario's min_accuracy.

Done means python verify.py starter exits 0 with all fifteen tests passing.