diff --git a/tools/canoninja/README.md b/tools/canoninja/README.md
new file mode 100644
index 0000000..506acf7
--- /dev/null
+++ b/tools/canoninja/README.md
@@ -0,0 +1,151 @@
+# Ninja File Canonicalizer
+
+Suppose we have a tool that generates a Ninja file from some other description (think Kati and makefiles), and during
+the testing we discovered a regression. Furthermore, suppose that the generated Ninja file is large (think millions of
+lines). And, the new Ninja file has build statements and rules in a slightly different order. As the tool generates the
+rule names, the real differences in the output of the `diff` command are drowned in noise. Enter Canoninja.
+
+Canoninja renames each Ninja rule to the hash of its contents. After that, we can just sort the build statements, and a
+simple `comm` command immediately reveal the essential difference between the files.
+
+## Example
+
+Consider the following makefile
+
+```makefile
+second :=
+first: foo
+foo:
+	@echo foo
+second: bar
+bar:
+	@echo bar
+```
+
+Depending on Kati version converting it to Ninja file will yield either:
+
+```
+$ cat /tmp/1.ninja
+# Generated by kati 06f2569b2d16628608c000a76e3d495a5a5528cb
+
+pool local_pool
+ depth = 72
+
+build _kati_always_build_: phony
+
+build first: phony foo
+rule rule0
+ description = build $out
+ command = /bin/sh -c "echo foo"
+build foo: rule0
+build second: phony bar
+rule rule1
+ description = build $out
+ command = /bin/sh -c "echo bar"
+build bar: rule1
+
+default first
+```
+
+or
+
+```
+$ cat 2.ninja
+# Generated by kati 371194da71b3e191fea6f2ccceb7b061bd0de310
+
+pool local_pool
+ depth = 72
+
+build _kati_always_build_: phony
+
+build second: phony bar
+rule rule0
+ description = build $out
+ command = /bin/sh -c "echo bar"
+build bar: rule0
+build first: phony foo
+rule rule1
+ description = build $out
+ command = /bin/sh -c "echo foo"
+build foo: rule1
+
+default first
+```
+
+This is a quirk in Kati, see https://github.com/google/kati/issues/238
+
+Trying to find out the difference between the targets even after sorting them isn't too helpful:
+
+```
+diff <(grep '^build' /tmp/1.ninja|sort) <(grep '^build' /tmp/2.ninja | sort)
+1c1
+< build bar: rule1
+---
+> build bar: rule0
+3c3
+< build foo: rule0
+---
+> build foo: rule1
+```
+
+However, running these files through `canoninja` yields
+
+```
+$ canoninja /tmp/1.ninja
+# Generated by kati 06f2569b2d16628608c000a76e3d495a5a5528cb
+
+pool local_pool
+ depth = 72
+
+build _kati_always_build_: phony
+
+build first: phony foo
+rule R2f9981d3c152fc255370dc67028244f7bed72a03
+ description = build $out
+ command = /bin/sh -c "echo foo"
+build foo: R2f9981d3c152fc255370dc67028244f7bed72a03
+build second: phony bar
+rule R62640f3f9095cf2da5b9d9e2a82f746cc710c94c
+ description = build $out
+ command = /bin/sh -c "echo bar"
+build bar: R62640f3f9095cf2da5b9d9e2a82f746cc710c94c
+
+default first
+```
+
+and
+
+```
+~/go/bin/canoninja /tmp/2.ninja
+# Generated by kati 371194da71b3e191fea6f2ccceb7b061bd0de310
+
+pool local_pool
+ depth = 72
+
+build _kati_always_build_: phony
+
+build second: phony bar
+rule R62640f3f9095cf2da5b9d9e2a82f746cc710c94c
+ description = build $out
+ command = /bin/sh -c "echo bar"
+build bar: R62640f3f9095cf2da5b9d9e2a82f746cc710c94c
+build first: phony foo
+rule R2f9981d3c152fc255370dc67028244f7bed72a03
+ description = build $out
+ command = /bin/sh -c "echo foo"
+build foo: R2f9981d3c152fc255370dc67028244f7bed72a03
+
+default first
+```
+
+and when we extract only build statements and sort them, we see that both Ninja files define the same graph:
+
+```shell
+$ diff <(~/go/bin/canoninja /tmp/1.ninja | grep '^build' | sort) \
+       <(~/go/bin/canoninja /tmp/2.ninja | grep '^build' | sort)
+```
+
+# Todo
+
+* Optionally output only the build statements, optionally sorted
+* Handle continuation lines correctly
diff --git a/tools/canoninja/canoninja.go b/tools/canoninja/canoninja.go
new file mode 100644
index 0000000..681a694
--- /dev/null
+++ b/tools/canoninja/canoninja.go
@@ -0,0 +1,130 @@
+package canoninja
+
+import (
+	"bytes"
+	"crypto/sha1"
+	"encoding/hex"
+	"fmt"
+	"io"
+)
+
+var (
+	rulePrefix  = []byte("rule ")
+	buildPrefix = []byte("build ")
+	phonyRule   = []byte("phony")
+)
+
+func Generate(path string, buffer []byte, sink io.Writer) error {
+	// Break file into lines
+	from := 0
+	var lines [][]byte
+	for from < len(buffer) {
+		line := getLine(buffer[from:])
+		lines = append(lines, line)
+		from += len(line)
+	}
+
+	// FOr each rule, calculate and remember its digest
+	ruleDigest := make(map[string]string)
+	for i := 0; i < len(lines); {
+		if bytes.HasPrefix(lines[i], rulePrefix) {
+			// Find ruleName
+			rn := ruleName(lines[i])
+			if len(rn) == 0 {
+				return fmt.Errorf("%s:%d: rule name is missing or on the next line", path, i+1)
+			}
+			sRuleName := string(rn)
+			if _, ok := ruleDigest[sRuleName]; ok {
+				return fmt.Errorf("%s:%d: the rule %s has been already defined", path, i+1, sRuleName)
+			}
+			// Calculate rule text digest as a digests of line digests.
+			var digests []byte
+			doDigest := func(b []byte) {
+				h := sha1.New()
+				h.Write(b)
+				digests = h.Sum(digests)
+
+			}
+			// For the first line, digest everything after rule's name
+			doDigest(lines[i][cap(lines[i])+len(rn)-cap(rn):])
+			for i++; i < len(lines) && lines[i][0] == ' '; i++ {
+				doDigest(lines[i])
+			}
+			h := sha1.New()
+			h.Write(digests)
+			ruleDigest[sRuleName] = "R" + hex.EncodeToString(h.Sum(nil))
+
+		} else {
+			i++
+		}
+	}
+
+	// Rewrite rule names.
+	for i, line := range lines {
+		if bytes.HasPrefix(line, buildPrefix) {
+			brn := getBuildRuleName(line)
+			if bytes.Equal(brn, phonyRule) {
+				sink.Write(line)
+				continue
+			}
+			if len(brn) == 0 {
+				return fmt.Errorf("%s:%d: build statement lacks rule name", path, i+1)
+			}
+			sink.Write(line[0 : cap(line)-cap(brn)])
+			if digest, ok := ruleDigest[string(brn)]; ok {
+				sink.Write([]byte(digest))
+			} else {
+				return fmt.Errorf("%s:%d: no rule for this build target", path, i+1)
+			}
+			sink.Write(line[cap(line)+len(brn)-cap(brn):])
+		} else if bytes.HasPrefix(line, rulePrefix) {
+			rn := ruleName(line)
+			// Write everything before it
+			sink.Write(line[0 : cap(line)-cap(rn)])
+			sink.Write([]byte(ruleDigest[string(rn)]))
+			sink.Write(line[cap(line)+len(rn)-cap(rn):])
+		} else {
+			//goland:noinspection GoUnhandledErrorResult
+			sink.Write(line)
+		}
+	}
+	return nil
+}
+
+func getLine(b []byte) []byte {
+	if n := bytes.IndexByte(b, '\n'); n >= 0 {
+		return b[:n+1]
+	}
+	return b
+}
+
+// Returns build statement's rule name
+func getBuildRuleName(line []byte) []byte {
+	n := bytes.IndexByte(line, ':')
+	if n <= 0 {
+		return nil
+	}
+	ruleName := line[n+1:]
+	if ruleName[0] == ' ' {
+		ruleName = bytes.TrimLeft(ruleName, " ")
+	}
+	if n := bytes.IndexAny(ruleName, " \t\r\n"); n >= 0 {
+		ruleName = ruleName[0:n]
+	}
+	return ruleName
+}
+
+// Returns rule statement's rule name
+func ruleName(lineAfterRule []byte) []byte {
+	ruleName := lineAfterRule[len(rulePrefix):]
+	if len(ruleName) == 0 {
+		return ruleName
+	}
+	if ruleName[0] == ' ' {
+		ruleName = bytes.TrimLeft(ruleName, " ")
+	}
+	if n := bytes.IndexAny(ruleName, " \t\r\n"); n >= 0 {
+		ruleName = ruleName[0:n]
+	}
+	return ruleName
+}
diff --git a/tools/canoninja/canoninja_test.go b/tools/canoninja/canoninja_test.go
new file mode 100644
index 0000000..3c45f8c
--- /dev/null
+++ b/tools/canoninja/canoninja_test.go
@@ -0,0 +1,47 @@
+package canoninja
+
+import (
+	"bytes"
+	"testing"
+)
+
+func TestGenerate(t *testing.T) {
+	tests := []struct {
+		name     string
+		in       []byte
+		wantSink string
+		wantErr  bool
+	}{
+		{
+			name: "1",
+			in: []byte(`
+rule rule1
+  abcd
+rule rule2
+  abcd
+build x: rule1
+`),
+			wantSink: `
+rule R9c97aba7f61994be6862f5ea9a62d26130c7f48b
+  abcd
+rule R9c97aba7f61994be6862f5ea9a62d26130c7f48b
+  abcd
+build x: R9c97aba7f61994be6862f5ea9a62d26130c7f48b
+`,
+			wantErr: false,
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			sink := &bytes.Buffer{}
+			err := Generate("<file>", tt.in, sink)
+			if (err != nil) != tt.wantErr {
+				t.Errorf("Generate() error = %v, wantErr %v", err, tt.wantErr)
+				return
+			}
+			if gotSink := sink.String(); gotSink != tt.wantSink {
+				t.Errorf("Generate() gotSink = %v, want %v", gotSink, tt.wantSink)
+			}
+		})
+	}
+}
diff --git a/tools/canoninja/cmd/main.go b/tools/canoninja/cmd/main.go
new file mode 100644
index 0000000..71802ef
--- /dev/null
+++ b/tools/canoninja/cmd/main.go
@@ -0,0 +1,36 @@
+package main
+
+/*
+   Canoninja reads a Ninja file and changes the rule names to be the digest of the rule contents.
+   Feed  it to a filter that extracts only build statements, sort them, and you will have a crude
+   but effective tool to find small differences between two Ninja files.
+*/
+
+import (
+	"canoninja"
+	"flag"
+	"fmt"
+	"os"
+)
+
+func main() {
+	flag.Parse()
+	files := flag.Args()
+	if len(files) == 0 {
+		files = []string{"/dev/stdin"}
+	}
+	rc := 0
+	for _, f := range files {
+		if buffer, err := os.ReadFile(f); err == nil {
+			err = canoninja.Generate(f, buffer, os.Stdout)
+			if err != nil {
+				fmt.Fprintln(os.Stderr, err)
+				rc = 1
+			}
+		} else {
+			fmt.Fprintf(os.Stderr, "%s: %s\n", f, err)
+			rc = 1
+		}
+	}
+	os.Exit(rc)
+}
diff --git a/tools/canoninja/go.mod b/tools/canoninja/go.mod
new file mode 100644
index 0000000..c5a924e
--- /dev/null
+++ b/tools/canoninja/go.mod
@@ -0,0 +1 @@
+module canoninja