Split Cursor into cursor-cli and cursor-ide, add missing stores

[tony]

Summary

• Replace the single cursor agent with two — cursor-cli (the cursor-agent terminal binary) and cursor-ide (the desktop app) — because they are distinct applications with disjoint data homes and on-disk formats, and one-agent modelling was the only place the catalogue broke its "one agent = one product = one root family" shape. This is breaking; see Breaking change.
• Add cursor-cli.prompt_history — Cursor's typed-prompt recall buffer at ~/.config/cursor/prompt_history.json, giving Cursor parity with the Claude/Codex/Grok prompt-history stores. The catalogue never knew this path because Cursor moved its CLI home to the lowercase ~/.config/cursor/ while only ~/.cursor/ was watched. As with those backends, the store's records are kind="history" (with role="user"), so it surfaces under --type history/--type all rather than the default --type prompts — a bare --agent cursor-cli search still returns typed prompts via the transcripts' user turns. See prompt_history stores parse as kind='history', so the default --type prompts search skips them #37.
• Add cursor-ide.workspace_state — per-workspace workspaceStorage/<hash>/state.vscdb databases, complementing the global store. Both now surface the aiService.prompts history, which prior builds advertised but never actually extracted (those entries carry no role field, so the generic walker skipped them).
• Add cursor-cli.chats — a best-effort, schema-less protobuf reader for the CLI's content-addressed chats/<hash>/<uuid>/store.db blob stores. Inspectable (opt-in), not searched by default, since the extraction overlaps the cleaner JSONL transcripts.
• Drop the now-redundant cli/ide infix from Cursor store and adapter IDs so the first store_id segment matches the agent.

Breaking change

The cursor agent is retired. Anywhere cursor was referenced — the --agent flag, agent: query predicates, the agentgrep://sources/{agent} MCP resource, and record agent fields — choose the surface you mean:

$ agentgrep search foo --agent cursor

$ agentgrep search foo --agent cursor-cli

$ agentgrep search foo --agent cursor-ide

Store and adapter identifiers were renamed to match:

Old	New
cursor.cli.transcripts	cursor-cli.transcripts
cursor.ai_tracking	cursor-cli.ai_tracking
cursor.ide.state_vscdb	cursor-ide.state_vscdb
cursor.cli_jsonl.v1	cursor_cli.transcripts_jsonl.v1
cursor.state_vscdb_modern.v1	cursor_ide.state_vscdb_modern.v1

Changes by area

• src/agentgrep/store_catalog.py: Split _CURSOR_STORES into _CURSOR_CLI_STORES / _CURSOR_IDE_STORES; add the three new descriptors; bump catalog_version.
• src/agentgrep/__init__.py: New parsers parse_cursor_prompt_history, parse_cursor_cli_chats_db, and the schema-less iter_protobuf_text_fields walker; teach parse_cursor_state_db the aiService.prompts shape via iter_cursor_prompt_candidates; split discover_cursor_sources into CLI/IDE discovery with a platform-resolved ide_workspace root; rename dispatch arms and the AgentName literal.
• MCP / CLI / query surface (mcp/_library.py, mcp/models.py, query/registry.py, ui/app.py): Update agent literals, the adapter registries, the query enum, and TUI agent colors.
• docs/: Split the backend page into cursor-cli.md + cursor-ide.md, update the grid/toctree and storage-catalog dev page, and refresh agent-name references across the CLI/library/MCP docs.
• tests/: Update fixtures and assertions to the split names; add parametrized and fixture-based coverage for the new parsers.

Design decisions

• Two agents, not a sub-facet: Each record already carries a store field, so CLI-vs-IDE provenance was distinguishable without splitting. But the two surfaces have disjoint roots and formats (JSONL/protobuf-SQLite vs state.vscdb), and per-surface availability is only honest when they are separate agents (e.g. an IDE that runs on a remote host has no local state.vscdb). Splitting also makes Cursor obey the same "one agent = one product binary" shape as the other four backends.
• Best-effort protobuf over no parser: Cursor publishes no schema for the chat blobs, and they form a content-addressed protobuf graph. Rather than skip them, the walker decodes the wire format generically — text fields begin with a printable byte, nested messages begin with a low tag byte — and surfaces readable runs. The adapter is date-versioned (cursor_cli.chats_protobuf.v1) precisely because the layout is unofficial and may drift.
• chats is opt-in: The protobuf extraction is noisier than and overlaps the JSONL transcripts, so the store is INSPECTABLE to keep default search clean; it still parses fully when explicitly included.
• Drop the infix: With no backward-compat constraint, collapsing cursor.cli.transcripts → cursor-cli.transcripts keeps store_id aligned with the agent — the invariant fixture_path and discovery both rely on.

Verification

No bare cursor agent literal remains in source:

$ rg '"cursor"' src/

The retired agent is rejected by the CLI:

$ agentgrep find --agent cursor

The new adapters are registered and dispatched:

$ rg 'cursor_cli\.(prompt_history|chats)|cursor_ide\.state_vscdb' src/agentgrep/__init__.py src/agentgrep/mcp/_library.py

Test plan

• test_iter_protobuf_text_fields — schema-less walker recovers leaf/nested text and skips varint/truncated/short fields
• test_iter_cursor_prompt_candidates — aiService.prompts shapes surface as user prompts; non-prompt shapes ignored
• test_cursor_cli_prompt_history_surfaces_user_prompts — prompt_history.json becomes deduped user records with mtime fallback
• test_cursor_cli_chats_db_is_opt_in_and_extracts_protobuf_text — chats store is inspectable-only and yields readable blob text
• test_cursor_ide_workspace_state_extracts_aiservice_prompts — per-workspace state.vscdb surfaces its prompt history
• test_actual_cursor_discovery_splits_main_and_subagent_transcripts — discovery splits under the new names
• Full gate: uv run ruff check ., uv run ruff format ., uv run ty check, uv run py.test --reruns 0, just build-docs
• End-to-end on a real $HOME: find cursor-cli surfaces cursor-cli.prompt_history; search --agent cursor-cli --type all returns its prompt-history records (the store is kind="history", so it needs --type history/all, not the default --type prompts — see prompt_history stores parse as kind='history', so the default --type prompts search skips them #37); the chats walker extracts readable text from real store.db files