Add an integration test

Integration tests live at the repo root in package remindb_test, not inside a pkg/. They exercise the full pipeline — a real store, real compiler, real MCP transport — against fixtures under testdata/. There are two flavors:

Flavor	File	Use when
Direct API	integration_test.go	Testing compiler / query / store / transformer end-to-end without MCP
MCP	mcp_integration_test.go	Testing the user-visible tool surface as an agent would call it

Pick the flavor by the question you're asking. "Does the compiler emit the right shape from this fixture?" → direct. "Does an agent calling MemorySearch after MemoryCompile see the expected ranking?" → MCP.

Where it lands

File	What changes
integration_test.go or mcp_integration_test.go	Add a Test<Scenario> function
testdata/<scenario-name>/	New fixture directory (only if existing fixtures don't match)

Fixture directories are flat namespaces (testdata/openclaw/, testdata/<your-scenario>/). One fixture set per scenario family; reuse aggressively.

The two helpers

Both files lean on package helpers — never reach for raw store.Open or raw mcp.NewServer in a test.

• internal/testutil.OpenTestDB(t) — opens an in-memory :memory: SQLite store with migrations applied; auto-cleans on t.Cleanup. Use for direct-API tests.
• internal/mcptest.NewEnv(t) — opens the test DB, wires a default-config server + tracker, connects an in-memory MCP transport, and exposes env.CallTool(t, name, args) and env.TextContent(t, result). Use for MCP tests.

Direct-API template

Mirror integration_test.go:TestOpenClawAgentWorkflow. Walk a real workflow stage by stage, log progress, assert at each stage:

package remindb_test

import (
    "context"
    "strings"
    "testing"

    "github.com/radimsem/remindb/internal/testutil"
    "github.com/radimsem/remindb/pkg/compiler"
    "github.com/radimsem/remindb/pkg/query"
)

func TestNewScenario(t *testing.T) {
    st := testutil.OpenTestDB(t)
    ctx := context.Background()

    // Stage 1: compile fixtures.
    result, err := compiler.CompileDir(ctx, st, "testdata/<scenario>", "<source-id>")
    if err != nil {
        t.Fatalf("CompileDir: %v", err)
    }
    if result.Added == 0 {
        t.Fatal("expected nodes from fixtures")
    }
    t.Logf("compiled: +%d ~%d -%d (%d total)", result.Added, result.Modified, result.Removed, result.Total)

    // Stage 2: query.
    engine := query.NewEngine(st)
    res, err := engine.Search(ctx, "<query>", 1000)
    if err != nil {
        t.Fatalf("Search: %v", err)
    }
    if len(res.Nodes) == 0 {
        t.Fatal("expected search hits")
    }
}

MCP template

Mirror mcp_integration_test.go:TestMcp_OpenClawAgent. The numbered-step structure is convention — each step represents one tool call an agent would make:

func TestMcp_NewScenario(t *testing.T) {
    env := mcptest.NewEnv(t)
    dir, _ := filepath.Abs("testdata/<scenario>")

    // 1. Compile.
    compileResult := env.CallTool(t, "MemoryCompile", map[string]any{
        "path":    dir,
        "message": "<scenario>-init",
    })
    text := env.TextContent(t, compileResult)
    if !strings.Contains(text, "compiled") {
        t.Fatalf("unexpected compile result: %s", text)
    }

    // 2. Search.
    searchResult := env.CallTool(t, "MemorySearch", map[string]any{
        "query":  "<keyword list>",
        "budget": 1000,
    })
    searchText := env.TextContent(t, searchResult)
    if searchText == "no results" {
        t.Fatal("expected results")
    }

    // 3. Further stages...
}

env.CallTool already calls t.Fatalf on transport / tool errors and logs both request and response preview to t.Log. You only assert the behavioral expectations.

Cleanup is automatic

testutil.OpenTestDB(t) and mcptest.NewEnv(t) both register t.Cleanup handlers — DB is dropped, MCP session is closed, in-memory transports are torn down. Don't write your own cleanup unless you allocate something outside these helpers (e.g., a t.TempDir() you wrote files into — that's also auto-cleaned by the testing framework, no manual work needed).

Fixtures under testdata/

testdata/ is the Go-recognized name; go test excludes it from build/lint. Place a fixture directory under it per scenario, with the same file extensions the parser recognizes (.md, .yaml, .json, .jsonl, .toon).

Name fixtures by concept (testdata/openclaw/, testdata/multi-format/), not by test (testdata/test_compile_1/). Multiple tests can share a fixture; one-test-per-fixture is anti-pattern.

Quick reference

Direct API:
1. integration_test.go              (Test<Scenario> with testutil.OpenTestDB)
2. testdata/<scenario>/             (fixtures matching parser-supported extensions)
3. go test ./... -run Test<Scenario>

MCP:
1. mcp_integration_test.go          (TestMcp_<Scenario> with mcptest.NewEnv)
2. testdata/<scenario>/             (same fixture rules)
3. go test ./... -run TestMcp_<Scenario>

Common mistakes

• Reaching for store.Open directly. Use testutil.OpenTestDB(t) — it applies migrations, registers cleanup, and matches the assertion style of every other test. Direct store.Open skips migrations and gives you an empty schema.
• Setting up the MCP transport by hand. pkg/mcp/server_test.go does this for unit-testing the server itself; integration tests should use mcptest.NewEnv(t). The handcrafted form has 6 lines of ceremony (NewInMemoryTransports, srv.Connect, mcp.NewClient, client.Connect, t.Cleanup for both, SetLoggingLevel if needed) — NewEnv collapses all of it.
• Asserting on log output. env.CallTool logs request/response previews via t.Log for debugging. They're not part of the test contract; assert on env.TextContent(t, result) instead.
• One fixture per test. Reuse testdata/openclaw/ for any scenario where its files are appropriate. New fixture directories are for genuinely new content shapes, not new test names.
• Hardcoding the absolute path to testdata/. Use filepath.Abs("testdata/<scenario>") like mcp_integration_test.go:14 — MemoryCompile requires absolute paths.
• Forgetting package remindb_test. Integration tests are external (use _test suffix on the package). They exercise the public API of multiple packages — they're not allowed to import unexported names.

Cross-references

• .claude/rules/go-concise.md — early returns, error handling style, named function over closure-as-var
• .claude/skills/add-mcp-tool/SKILL.md — when adding a tool, also extend mcp_integration_test.go with a scenario that exercises it through env.CallTool
• .claude/skills/add-store-query/SKILL.md — when adding a query, the unit test goes in pkg/store/store_test.go; an integration test only makes sense if it changes a multi-package workflow