fix(cursor): map taskToolCall so subagent dispatches are named

[talma]

Problem

Cursor reports a subagent dispatch as:

{"tool_call": {"taskToolCall": {"args": {"description": "...", "prompt": "..."}}}}

taskToolCall is absent from _CURSOR_TOOL_TYPE_MAP, so _extract_cursor_tool() falls through every branch and returns ("", {}) — even though the payload carries both description and prompt. The data was always there; only the name lookup was missing.

Impact

The empty ToolCall.name breaks three things:

1. Deterministic rule scorers cannot observe the dispatch. tools_invoked, tools_invoked_in_order, tool_args_contain and tool_result_contains all key on tc.name in scorer/rules/trajectory.py. A scenario asserting delegation fails as though the agent never delegated — a false negative, the worst direction for a test. tool_name_in_cli() does not rescue it, since the raw stream contains taskToolCall, not the tool name.

Measured against a real recorded payload:

Assertion	Before	After
tools_invoked: [Task]	✗ False	✓ True
tool_args_contain: {Task: …}	✗ "tool Task not invoked"	✓ matches
skill_invoked(implementation-planner)	✗ (False, '')	✓ (True, 'via Task tool args')

2. LLM judges are shown [TOOL CALL: ] with Input: {} (scorer/llm/scorer.py:145) and are left to guess, which makes any delegation criterion nondeterministic. Observed downstream: identical agent behaviour, opposite verdicts across runs — "the tool sequence ends with an empty call consistent with a Task dispatch" (pass) vs "the tool calls are all Read/Glob/Grep/Shell … no Task dispatch is present" (fail). The dispatch had in fact occurred; the planner sub-transcript was present under tool_calls[N].result.success.conversationSteps.

3. Live progress silently drops it. parse_stream_event() does if not name: return None, so the single most significant event in a subagent flow renders as nothing.

The practical cost downstream is that delegation can only be asserted through prose the agent happens to emit, never through what it actually did.

Fix

One map entry. Named Task to match Claude Code's name for the same primitive, so scenarios can assert on it uniformly across agents.

Tests

• test_parses_cursor_task_tool_call — the dispatch surfaces as Task with args intact, mirroring the existing readToolCall / shellToolCall / grepToolCall cases.
• test_unknown_tool_call_subtype_stays_unnamed — regression guard that unmapped subtype keys still return ("", {}).

Verification

• uv run make check — 3212 passed, 45 skipped, lint clean.
• uv run belt agent info cursor — CLI healthy (per CONTRIBUTING's "unit tests can pass while the CLI is broken").
• No lockfile or doc-parity churn; diff is +37 lines across 2 files.

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[talma]

recheck

[talma]

I have read the CLA Document and I hereby sign the CLA

[talma]

I have read the CLA Document and I hereby sign the CLA

[talma]

recheck

[github-actions]

---

🐸 JFrog Frogbot