Files
coder/coderd/x/chatd/forced_mcp_test.go
T
Michael Suchacz 7724ee281a feat: defer MCP tool schemas behind a find_tools search (#28225)
## Summary

When the `mcp-tool-search` experiment is enabled, chatd stops inlining
connected MCP tool schemas into every generation. It instead exposes a
built-in `find_tools` tool whose description carries a compact catalog
of the deferred tools, and only ships full JSON schemas for tools the
model has activated by searching or by calling them directly.

Closes [CODAGT-760](https://linear.app/coder/issue/CODAGT-760).

## Problem

Tool-heavy agent configurations (GitHub, Linear, Notion, and dev-tooling
MCP servers) inline over 100k tokens of tool schema definitions into
every generation. Initial uncached requests reached ~216k tokens with
time-to-first-token close to nine minutes, while the model typically
invokes only a handful of tools per turn.

## How it works

- `decideMCPToolSearch` defers external and workspace `.mcp.json` MCP
tools whenever the experiment is enabled. Native, dynamic, provider,
skill, and transport tools are never deferred.
- `find_tools` embeds a server-grouped catalog in its tool description
(degrading to names-only, then counts-only, then a constant-size summary
past a context-scaled size cap) and scores keyword matches across tool
names, descriptions, parameter schemas, and server metadata. Queries can
scope to one server with a `server:` prefix, and exact `names` arguments
always activate.
- Activation state is ephemeral: it is re-derived each generation from
surviving chat history (`find_tools` results and direct calls to
deferred tools), so activations naturally lapse when compaction
summarizes them away. Aggregate activated schema weight is capped at 10%
of the context window, shedding the least recently activated schemas
first; `find_tools` shares that budget across parallel calls in one
step. No new persistence.
- Deferred tools stay registered for execution, so the model can call a
cataloged tool directly without searching first; the schema is activated
for subsequent steps.
- Fail-open: the experiment being disabled, an empty candidate set, or
an MCP tool named `find_tools` all disable deferral, leaving today's
behavior byte-identical on the wire.
- Prometheus counters/histograms track `find_tools` calls, matches,
activations, and deferred token weight.
- The conversation timeline renders `find_tools` calls with a collapsed
search summary and expandable match list, falling back to the generic
renderer on malformed payloads.

## Validation

- Unit tests for the catalog, matcher, experiment-gated decision, and
activation derivation; end-to-end chatd generation tests covering
search-then-call, direct-call activation, experiment-off wire parity,
compaction lapse, and subagent tool gating.
- Storybook interaction tests for the timeline rendering and
malformed-payload fallback.
- Remote dogfood UAT on dev.coder.com passed: deferral with a real MCP
server and Anthropic model, direct calls without prior search,
activation persistence across turns, experiment-off parity, and clean
UI/console.

> Disclosure: Mux (AI agent) authored this PR on Mike's behalf.
2026-08-18 19:12:47 +02:00

291 lines
10 KiB
Go

package chatd_test
// Regression tests for Cure53 CDM-02-010: the Force On MCP server
// availability policy must be enforced on the backend. A client that
// omits force_on entries from mcp_server_ids when creating a chat or
// sending a message must not be able to exclude those servers from
// the conversation.
import (
"net/http/httptest"
"sync"
"testing"
"github.com/google/uuid"
"github.com/stretchr/testify/require"
"github.com/coder/coder/v2/coderd/database"
"github.com/coder/coder/v2/coderd/database/dbgen"
"github.com/coder/coder/v2/coderd/database/dbtestutil"
"github.com/coder/coder/v2/coderd/x/chatd"
"github.com/coder/coder/v2/coderd/x/chatd/chattest"
"github.com/coder/coder/v2/codersdk"
"github.com/coder/coder/v2/testutil"
)
// newEchoMCPTestServer starts an MCP test server exposing an "echo"
// tool and returns its base URL.
func newEchoMCPTestServer(t *testing.T, name string) string {
t.Helper()
srv := newTestMCPServer(name)
addTestMCPTextTool(srv, "echo", "Echoes the input", "echo: ")
ts := httptest.NewServer(testMCPHTTPHandler(srv))
t.Cleanup(ts.Close)
return ts.URL
}
// newToolRecordingOpenAI returns a mock OpenAI URL that answers every
// streamed call with plain text and records the tool names offered on
// each streamed call, plus an accessor for the recorded calls.
func newToolRecordingOpenAI(t *testing.T) (string, func() [][]string) {
t.Helper()
var (
mu sync.Mutex
calls [][]string
)
url := chattest.NewOpenAI(t, func(req *chattest.OpenAIRequest) chattest.OpenAIResponse {
if !req.Stream {
return chattest.OpenAINonStreamingResponse("title")
}
names := make([]string, 0, len(req.Tools))
for _, tool := range req.Tools {
names = append(names, tool.Function.Name)
}
mu.Lock()
calls = append(calls, names)
mu.Unlock()
return chattest.OpenAIStreamingResponse(chattest.OpenAITextChunks("ok")...)
})
recorded := func() [][]string {
mu.Lock()
defer mu.Unlock()
out := make([][]string, len(calls))
copy(out, calls)
return out
}
return url, recorded
}
// TestCreateChat_ForceOnMCPServerEnforced reproduces CDM-02-010 for
// chat creation: stripping mcp_server_ids from the create request must
// not exclude force_on MCP servers.
func TestCreateChat_ForceOnMCPServerEnforced(t *testing.T) {
t.Parallel()
db, ps := dbtestutil.NewDB(t)
ctx := testutil.Context(t, testutil.WaitLong)
forcedURL := newEchoMCPTestServer(t, "forced-mcp")
openAIURL, recordedCalls := newToolRecordingOpenAI(t)
user, org, model := seedChatDependenciesWithProvider(t, db, "openai-compat", openAIURL)
forcedConfig := dbgen.MCPServerConfig(t, db, database.MCPServerConfig{
DisplayName: "Forced MCP",
Slug: "forced-mcp",
Url: forcedURL,
Availability: "force_on",
CreatedBy: uuid.NullUUID{UUID: user.ID, Valid: true},
UpdatedBy: uuid.NullUUID{UUID: user.ID, Valid: true},
})
server := newActiveTestServer(t, db, ps, func(cfg *chatd.Config) {
withoutMCPToolSearch(cfg)
cfg.AIBridgeTransportFactory = chatAIGatewayTransportFactoryPointer(chattest.NewMockAIBridgeTransport(t, openAIURL))
})
// The attacker strips every MCP server ID from the request.
chat, err := server.CreateChat(ctx, chatd.CreateOptions{
OrganizationID: org.ID,
OwnerID: user.ID,
Title: "forced-mcp-create",
ModelConfigID: model.ID,
MCPServerIDs: []uuid.UUID{},
InitialUserContent: []codersdk.ChatMessagePart{
codersdk.ChatMessageText("hello"),
},
})
require.NoError(t, err)
// The force_on server must be persisted despite the empty list.
dbChat, err := db.GetChatByID(ctx, chat.ID)
require.NoError(t, err)
require.ElementsMatch(t, []uuid.UUID{forcedConfig.ID}, dbChat.MCPServerIDs,
"force_on MCP server must be enforced on chat creation")
waitForChatProcessed(ctx, t, db, chat.ID, server)
chatResult, err := db.GetChatByID(ctx, chat.ID)
require.NoError(t, err)
if chatResult.Status == database.ChatStatusError {
require.FailNowf(t, "chat failed", "last_error=%q", chatLastErrorMessage(chatResult.LastError))
}
// The forced server's tool must be offered to the LLM.
calls := recordedCalls()
require.NotEmpty(t, calls)
require.Contains(t, calls[0], "forced-mcp__echo",
"force_on MCP tools must be offered to the LLM despite a stripped mcp_server_ids list")
}
// TestSendMessage_ForceOnMCPServerEnforced reproduces CDM-02-010 for
// message sends: a tampered mcp_server_ids update must not remove
// force_on MCP servers from the chat.
func TestSendMessage_ForceOnMCPServerEnforced(t *testing.T) {
t.Parallel()
db, ps := dbtestutil.NewDB(t)
ctx := testutil.Context(t, testutil.WaitLong)
forcedURL := newEchoMCPTestServer(t, "forced-mcp")
optionalURL := newEchoMCPTestServer(t, "optional-mcp")
openAIURL, _ := newToolRecordingOpenAI(t)
user, org, model := seedChatDependenciesWithProvider(t, db, "openai-compat", openAIURL)
forcedConfig := dbgen.MCPServerConfig(t, db, database.MCPServerConfig{
DisplayName: "Forced MCP",
Slug: "forced-mcp",
Url: forcedURL,
Availability: "force_on",
CreatedBy: uuid.NullUUID{UUID: user.ID, Valid: true},
UpdatedBy: uuid.NullUUID{UUID: user.ID, Valid: true},
})
optionalConfig := dbgen.MCPServerConfig(t, db, database.MCPServerConfig{
DisplayName: "Optional MCP",
Slug: "optional-mcp",
Url: optionalURL,
CreatedBy: uuid.NullUUID{UUID: user.ID, Valid: true},
UpdatedBy: uuid.NullUUID{UUID: user.ID, Valid: true},
})
server := newActiveTestServer(t, db, ps, func(cfg *chatd.Config) {
withoutMCPToolSearch(cfg)
cfg.AIBridgeTransportFactory = chatAIGatewayTransportFactoryPointer(chattest.NewMockAIBridgeTransport(t, openAIURL))
})
// Creation with a tampered list that omits the forced server.
chat, err := server.CreateChat(ctx, chatd.CreateOptions{
OrganizationID: org.ID,
OwnerID: user.ID,
Title: "forced-mcp-send",
ModelConfigID: model.ID,
MCPServerIDs: []uuid.UUID{optionalConfig.ID},
InitialUserContent: []codersdk.ChatMessagePart{
codersdk.ChatMessageText("hello"),
},
})
require.NoError(t, err)
dbChat, err := db.GetChatByID(ctx, chat.ID)
require.NoError(t, err)
require.ElementsMatch(t, []uuid.UUID{optionalConfig.ID, forcedConfig.ID}, dbChat.MCPServerIDs,
"force_on MCP server must be appended to a tampered create list")
waitForChatProcessed(ctx, t, db, chat.ID, server)
// The attacker clears the MCP server list on a message send.
_, err = server.SendMessage(ctx, chatd.SendMessageOptions{
ChatID: chat.ID,
CreatedBy: user.ID,
Content: []codersdk.ChatMessagePart{codersdk.ChatMessageText("clear the list")},
MCPServerIDs: &[]uuid.UUID{},
})
require.NoError(t, err)
dbChat, err = db.GetChatByID(ctx, chat.ID)
require.NoError(t, err)
require.ElementsMatch(t, []uuid.UUID{forcedConfig.ID}, dbChat.MCPServerIDs,
"force_on MCP server must survive an emptied mcp_server_ids update")
waitForChatProcessed(ctx, t, db, chat.ID, server)
// A legitimate update keeps both the selection and the forced server.
_, err = server.SendMessage(ctx, chatd.SendMessageOptions{
ChatID: chat.ID,
CreatedBy: user.ID,
Content: []codersdk.ChatMessagePart{codersdk.ChatMessageText("select optional")},
MCPServerIDs: &[]uuid.UUID{optionalConfig.ID},
})
require.NoError(t, err)
dbChat, err = db.GetChatByID(ctx, chat.ID)
require.NoError(t, err)
require.ElementsMatch(t, []uuid.UUID{optionalConfig.ID, forcedConfig.ID}, dbChat.MCPServerIDs,
"force_on MCP server must be appended to a tampered update list")
waitForChatProcessed(ctx, t, db, chat.ID, server)
}
// TestGeneration_ForceOnMCPServerEnforcedForExistingChats covers chats
// whose stored mcp_server_ids predates the force_on policy (or was
// tampered before enforcement existed): generation must still include
// force_on servers without relying on the stored list.
func TestGeneration_ForceOnMCPServerEnforcedForExistingChats(t *testing.T) {
t.Parallel()
db, ps := dbtestutil.NewDB(t)
ctx := testutil.Context(t, testutil.WaitLong)
forcedURL := newEchoMCPTestServer(t, "forced-mcp")
openAIURL, recordedCalls := newToolRecordingOpenAI(t)
user, org, model := seedChatDependenciesWithProvider(t, db, "openai-compat", openAIURL)
server := newActiveTestServer(t, db, ps, func(cfg *chatd.Config) {
withoutMCPToolSearch(cfg)
cfg.AIBridgeTransportFactory = chatAIGatewayTransportFactoryPointer(chattest.NewMockAIBridgeTransport(t, openAIURL))
})
// The chat is created before any force_on MCP server exists, so
// its stored mcp_server_ids is empty.
chat, err := server.CreateChat(ctx, chatd.CreateOptions{
OrganizationID: org.ID,
OwnerID: user.ID,
Title: "forced-mcp-existing",
ModelConfigID: model.ID,
MCPServerIDs: []uuid.UUID{},
InitialUserContent: []codersdk.ChatMessagePart{
codersdk.ChatMessageText("hello"),
},
})
require.NoError(t, err)
waitForChatProcessed(ctx, t, db, chat.ID, server)
// An admin marks a server force_on after the chat already exists.
dbgen.MCPServerConfig(t, db, database.MCPServerConfig{
DisplayName: "Forced MCP",
Slug: "forced-mcp",
Url: forcedURL,
Availability: "force_on",
CreatedBy: uuid.NullUUID{UUID: user.ID, Valid: true},
UpdatedBy: uuid.NullUUID{UUID: user.ID, Valid: true},
})
// A send that does not touch mcp_server_ids must still pick up
// the force_on server at generation time.
_, err = server.SendMessage(ctx, chatd.SendMessageOptions{
ChatID: chat.ID,
CreatedBy: user.ID,
Content: []codersdk.ChatMessagePart{codersdk.ChatMessageText("again")},
})
require.NoError(t, err)
waitForChatProcessed(ctx, t, db, chat.ID, server)
chatResult, err := db.GetChatByID(ctx, chat.ID)
require.NoError(t, err)
if chatResult.Status == database.ChatStatusError {
require.FailNowf(t, "chat failed", "last_error=%q", chatLastErrorMessage(chatResult.LastError))
}
// nil MCPServerIDs must keep the stored list untouched.
require.Empty(t, chatResult.MCPServerIDs)
calls := recordedCalls()
require.GreaterOrEqual(t, len(calls), 2)
require.NotContains(t, calls[0], "forced-mcp__echo",
"no force_on server existed during the first turn")
require.Contains(t, calls[len(calls)-1], "forced-mcp__echo",
"force_on MCP tools must reach generation for chats created before the policy")
}