diff --git a/src/late/mcp/server.py b/src/late/mcp/server.py index 7f0927a0..40497179 100644 --- a/src/late/mcp/server.py +++ b/src/late/mcp/server.py @@ -109,13 +109,10 @@ def _api_transport() -> httpx.BaseTransport: # Always-visible core: the 20 hand-written ergonomic tools (guardrails + # LLM-shaped args, preferred over their raw generated twins) plus the generated -# tools central to working on a profile — posting, the scheduling queue, -# pre-publish validation, account health, cross-platform analytics basics, and -# engagement on published posts. The remaining ~430 generated tools (messaging -# suite, ads, connect, platform-specific analytics, ...) sit behind the BM25 -# tool-search transform (search_tools / call_tool), so clients see ~51 tools up -# front instead of 481 — small enough that clients which cap or truncate large -# tool lists show them all. +# tools most sessions need, chosen by measured use. Every pinned definition is +# sent on every turn of every session, so a tool few sessions call stays behind +# the BM25 tool-search transform (search_tools / call_tool) with the rest of the +# generated tools. _PINNED_TOOLS = [ # Hand-written ergonomic tools "accounts_list", "accounts_get", @@ -127,26 +124,21 @@ def _api_transport() -> httpx.BaseTransport: # Paid diagnosis and its poll fallback "support_ask", "support_assistant_get_support_answer", # Posting operations not covered by the ergonomic set - "posts_bulk_upload_posts", "posts_unpublish_post", "posts_edit_post", - # Scheduling queue - "queue_list_queue_slots", "queue_create_queue_slot", "queue_update_queue_slot", - "queue_delete_queue_slot", "queue_preview_queue", "queue_get_next_queue_slot", + "posts_unpublish_post", "posts_edit_post", "media_get_media_presigned_url", # Pre-publish validation - "validate_post", "validate_post_length", "validate_media", - # Account health & organisation + "validate_post", "validate_media", + # Account health "accounts_get_all_accounts_health", "accounts_get_account_health", - "accounts_get_follower_stats", "accounts_move_account_to_profile", - "account_groups_list_account_groups", + "accounts_get_follower_stats", + # Inbox conversations + "messages_list_inbox_conversations", "messages_get_inbox_conversation_messages", # Cross-platform analytics basics "analytics_get_analytics", "analytics_get_best_time_to_post", "analytics_get_daily_metrics", "analytics_get_post_timeline", # Engagement on published posts "comments_list_inbox_comments", "comments_get_inbox_post_comments", "comments_reply_to_inbox_post", - "mentions_list_inbox_mentions", "mentions_reply_to_mention", - # Campaign tags & usage - "tracking_tags_list_tracking_tags", "tracking_tags_get_tracking_tag_stats", - "usage_get_usage", + "mentions_list_inbox_mentions", # Report bugs and missing features to Zernio FEEDBACK_TOOL, ] @@ -236,15 +228,32 @@ async def _search(self, tools: Sequence[Tool], query: str) -> Sequence[Tool]: async def closest_tools(self, name: str, ctx: Context | None) -> list[str]: """Real tool names nearest a mistyped or guessed one, pinned tools included.""" hidden = await self._get_visible_tools(ctx) - names = [t.name for t in hidden] + sorted(self._always_visible) + pinned = sorted(self._always_visible) + names = [t.name for t in hidden] + pinned by_words = [t.name for t in await self._search(hidden, name)] + by_pinned_words = _names_sharing_words(name, pinned) by_spelling = difflib.get_close_matches(name, names, n=_MAX_TOOL_SUGGESTIONS) # Word matches lead: they named the intended tool first for every # guessed name seen in production, spelling only for some. - interleaved = [n for pair in zip_longest(by_words, by_spelling) for n in pair if n] + interleaved = [ + n + for group in zip_longest(by_words, by_pinned_words, by_spelling) + for n in group + if n + ] return list(dict.fromkeys(interleaved))[:_MAX_TOOL_SUGGESTIONS] +def _names_sharing_words(guess: str, names: Sequence[str]) -> list[str]: + split_camel_case = re.sub(r"([a-z0-9])([A-Z])", r"\1_\2", guess) + guess_words = set(re.findall(r"[a-z0-9]+", split_camel_case.lower())) + shared = {name: len(guess_words & set(name.split("_"))) for name in names} + ranked = sorted( + (n for n, count in shared.items() if count >= 2), key=lambda n: -shared[n] + ) + return ranked[:_MAX_TOOL_SUGGESTIONS] + + _search_transform = _AnnotatedBM25SearchTransform( always_visible=_PINNED_TOOLS, max_results=8, diff --git a/tests/test_mcp_pinned_tools.py b/tests/test_mcp_pinned_tools.py new file mode 100644 index 00000000..9c79dc54 --- /dev/null +++ b/tests/test_mcp_pinned_tools.py @@ -0,0 +1,44 @@ +from __future__ import annotations + +from fastmcp import Client + +from late.mcp import server + + +async def test_every_pinned_name_is_a_registered_tool() -> None: + registered = {t.name for t in await server.mcp.list_tools(run_middleware=False)} + + assert [name for name in server._PINNED_TOOLS if name not in registered] == [] + + +async def test_the_inbox_and_media_url_reads_are_listed_without_a_search() -> None: + listed = {t.name for t in await server.mcp.list_tools()} + + assert { + "messages_list_inbox_conversations", + "messages_get_inbox_conversation_messages", + "media_get_media_presigned_url", + } <= listed + + +async def test_an_unpinned_tool_is_found_by_search_and_not_listed_up_front() -> None: + listed = {t.name for t in await server.mcp.list_tools()} + async with Client(server.mcp) as client: + result = await client.call_tool("search_tools", {"query": "list queue slots"}) + + assert "queue_list_queue_slots" not in listed + assert "queue_list_queue_slots" in result.content[0].text + + +async def test_agents_that_call_a_newly_pinned_tool_through_call_tool_still_reach_it() -> ( + None +): + async with Client(server.mcp) as client: + result = await client.call_tool( + "call_tool", + {"name": "messages_get_inbox_conversation_messages", "arguments": {}}, + raise_on_error=False, + ) + + assert "Unknown tool" not in result.content[0].text + assert "conversation_id" in result.content[0].text diff --git a/tests/test_mcp_tool_guidance.py b/tests/test_mcp_tool_guidance.py index b308a0d2..13cf756c 100644 --- a/tests/test_mcp_tool_guidance.py +++ b/tests/test_mcp_tool_guidance.py @@ -14,6 +14,8 @@ async def _call(name: str, arguments: dict) -> tuple[bool, str]: ("guess", "real"), [ ("inbox_list_conversations", "messages_list_inbox_conversations"), + ("inboxListConversations", "messages_list_inbox_conversations"), + ("inbox-get-conversation-messages", "messages_get_inbox_conversation_messages"), ("ads_list_ads", "ad_campaigns_list_ads"), ("comment_automations_list", "comment_automations_list_comment_automations"), ], @@ -38,11 +40,11 @@ async def test_a_guessed_name_through_call_tool_suggests_the_real_tool(): async def test_call_tool_runs_a_proxied_search_instead_of_refusing_it(): is_error, text = await _call( "call_tool", - {"name": "search_tools", "arguments": {"query": "list inbox conversations"}}, + {"name": "search_tools", "arguments": {"query": "search inbox conversations"}}, ) assert not is_error - assert "messages_list_inbox_conversations" in text + assert "messages_search_inbox_conversations" in text async def test_call_tool_still_refuses_to_call_itself():