]> git.giacomolli.uk - onechat/blobdiff - compactor_test.go
Log the start and the length of each subagent run
[onechat] / compactor_test.go
index 7813c773369287593f07b246c6620c7b9c4560ea..0cab0d0bf15c8553097205369a19c2cbdf3a4b35 100644 (file)
@@ -14,7 +14,7 @@ import (
 )
 
 func TestCompactorSendsThePromptToolsAndTask(t *testing.T) {
 )
 
 func TestCompactorSendsThePromptToolsAndTask(t *testing.T) {
-       server := newFakeResponses(t)
+       server := newModelFake(t)
        compactor := newTestCompactor(t, server)
 
        line, err := compactor.Summarize(t.Context(), Task{
        compactor := newTestCompactor(t, server)
 
        line, err := compactor.Summarize(t.Context(), Task{
@@ -28,30 +28,114 @@ func TestCompactorSendsThePromptToolsAndTask(t *testing.T) {
        require.Equal(t, 1, server.Calls())
 
        body := server.Call(0).body
        require.Equal(t, 1, server.Calls())
 
        body := server.Call(0).body
-       require.Equal(t, prompt("Ada"), body["instructions"])
+       require.Equal(t, prompt("Ada", ""), body["instructions"])
        require.Equal(t, "gpt-test", body["model"])
        require.Equal(t, map[string]any{"effort": "high"}, body["reasoning"])
        require.Equal(t, "agent-key", body["prompt_cache_key"])
 
        require.Equal(t, "gpt-test", body["model"])
        require.Equal(t, map[string]any{"effort": "high"}, body["reasoning"])
        require.Equal(t, "agent-key", body["prompt_cache_key"])
 
-       tools, err := json.Marshal(agentTools())
+       tools, err := json.Marshal(promptTools())
        require.NoError(t, err)
 
        sent, err := json.Marshal(body["tools"])
        require.NoError(t, err)
        require.JSONEq(t, string(tools), string(sent))
 
        require.NoError(t, err)
 
        sent, err := json.Marshal(body["tools"])
        require.NoError(t, err)
        require.JSONEq(t, string(tools), string(sent))
 
-       parts := inputParts(t, server.Call(0))
+       texts := compactorInputTexts(t, server.Call(0))
        require.Equal(t, []string{
        require.Equal(t, []string{
-               "<chat>\n0+4|user: earlier\n</chat>\n",
-               "Compaction: compress message 7 into one line of at most 512 bytes\n" +
+               "<chat>\n0+4|user: earlier\n</chat>\n" +
+                       "Compaction: compress message 7 into one line of at most 512 bytes\n" +
                        "(about 70 words), the length of this ruler:\n" +
                        strings.Repeat("-", 512) + "\n" +
                        "<input>\nuser: " + strings.Repeat("x", 600) + "\n</input>",
                        "(about 70 words), the length of this ruler:\n" +
                        strings.Repeat("-", 512) + "\n" +
                        "<input>\nuser: " + strings.Repeat("x", 600) + "\n</input>",
-       }, parts)
+       }, texts)
+}
+
+func TestCompactorSendsEachWholeBlockOfTheViewAsAMessage(t *testing.T) {
+       testCases := []struct {
+               name    string
+               context string
+               want    []string
+       }{
+               {
+                       name:    "empty view goes with the task",
+                       context: "<chat>\n</chat>",
+                       want:    []string{"<chat>\n</chat>\nTASK"},
+               },
+               {
+                       name:    "lines short of a block go with the task",
+                       context: "<chat>\n0+1|a\n1+1|b\n2+1|c\n</chat>",
+                       want:    []string{"<chat>\n0+1|a\n1+1|b\n2+1|c\n</chat>\nTASK"},
+               },
+               {
+                       name:    "one whole block and the task",
+                       context: "<chat>\n0+1|a\n1+1|b\n2+1|c\n3+1|d\n</chat>",
+                       want:    []string{"<chat>\n0+1|a\n1+1|b\n2+1|c\n3+1|d\n", "</chat>\nTASK"},
+               },
+               {
+                       name:    "two whole blocks and the rest with the task",
+                       context: "<chat>\n0+2|a\n2+1|b\n3+1|c\n4+1|d\n5+1|e\n6+1|f\n7+1|g\n8+1|h\n9+1|i\n</chat>",
+                       want: []string{
+                               "<chat>\n0+2|a\n2+1|b\n3+1|c\n4+1|d\n",
+                               "5+1|e\n6+1|f\n7+1|g\n8+1|h\n",
+                               "9+1|i\n</chat>\nTASK",
+                       },
+               },
+       }
+
+       for _, tc := range testCases {
+               t.Run(tc.name, func(t *testing.T) {
+                       server := newModelFake(t)
+                       task := compactorLeafTask()
+                       task.Context = tc.context
+
+                       _, err := newTestCompactor(t, server).Summarize(t.Context(), task)
+                       require.NoError(t, err)
+
+                       text, err := compactorTask(task)
+                       require.NoError(t, err)
+
+                       want := make([]string, len(tc.want))
+                       for n, message := range tc.want {
+                               want[n] = strings.Replace(message, "TASK", text, 1)
+                       }
+
+                       texts := compactorInputTexts(t, server.Call(0))
+                       require.Equal(t, want, texts)
+                       require.Equal(t, tc.context+"\n"+text, strings.Join(texts, ""))
+               })
+       }
+}
+
+func TestCompactorSendsTheSameBlocksForViewsWithTheSameFirstLines(t *testing.T) {
+       shared := "<chat>\n0+4|a\n4+1|b\n5+1|c\n6+1|d\n7+1|e\n8+1|f\n9+1|g\n10+1|h\n"
+       server := newModelFake(t)
+       compactor := newTestCompactor(t, server)
+       _, err := compactor.Summarize(t.Context(), Task{
+               Level:   0,
+               Index:   11,
+               Context: shared + "11+1|i\n</chat>",
+               Input:   []string{"user: eleven"},
+       })
+       require.NoError(t, err)
+
+       _, err = compactor.Summarize(t.Context(), Task{
+               Level:   1,
+               Index:   5,
+               Context: shared + "11+1|j\n12+1|k\n</chat>",
+               Input:   []string{"left line", "right line"},
+       })
+       require.NoError(t, err)
+
+       first := compactorInputTexts(t, server.Call(0))
+       second := compactorInputTexts(t, server.Call(1))
+       require.Len(t, first, 3)
+       require.Len(t, second, 3)
+       require.Equal(t, first[:2], second[:2])
+       require.NotEqual(t, first[2], second[2])
 }
 
 func TestCompactorWritesTheMergeTask(t *testing.T) {
 }
 
 func TestCompactorWritesTheMergeTask(t *testing.T) {
-       text, err := compactionTask(Task{
+       text, err := compactorTask(Task{
                Level: 2,
                Index: 3,
                Input: []string{"left line", "right line"},
                Level: 2,
                Index: 3,
                Input: []string{"left line", "right line"},
@@ -67,30 +151,38 @@ func TestCompactorWritesTheMergeTask(t *testing.T) {
 
 func TestCompactorRetriesALongLineWithTheNewestNote(t *testing.T) {
        long := []string{strings.Repeat("a", 600), strings.Repeat("b", 550), strings.Repeat("c", 400)}
 
 func TestCompactorRetriesALongLineWithTheNewestNote(t *testing.T) {
        long := []string{strings.Repeat("a", 600), strings.Repeat("b", 550), strings.Repeat("c", 400)}
-       server := newFakeResponses(t)
-       server.reply = func(w http.ResponseWriter, call fakeCall) {
-               writeEvents(w, completedEvent([]any{messageItem(long[call.number])}, 100, 0, 5))
+       server := newModelFake(t)
+       server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
+               modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem(long[call.number])}, 100, 0, 5))
        }
 
        }
 
-       line, err := newTestCompactor(t, server).Summarize(t.Context(), leafTask())
+       task := compactorLeafTask()
+       task.Context = "<chat>\n0+1|a\n1+1|b\n2+1|c\n3+1|d\n4+1|e\n</chat>"
+
+       compactor := newTestCompactor(t, server)
+       line, err := compactor.Summarize(t.Context(), task)
        require.NoError(t, err)
        require.Equal(t, long[2], line)
        require.Equal(t, 3, server.Calls())
 
        require.NoError(t, err)
        require.Equal(t, long[2], line)
        require.Equal(t, 3, server.Calls())
 
-       first := inputParts(t, server.Call(0))
+       records := modelReadUsage(t, compactor.client.usage.path)
+       require.Len(t, records, 3)
+       for n, record := range records {
+               require.Equal(t, &[2]int{0, 0}, record.Node)
+               require.Equal(t, len(long[n]), record.TextBytes)
+       }
+
+       first := compactorInputTexts(t, server.Call(0))
        require.Len(t, first, 2)
        for n, want := range []string{long[0], long[1]} {
                call := server.Call(n + 1)
        require.Len(t, first, 2)
        for n, want := range []string{long[0], long[1]} {
                call := server.Call(n + 1)
-               parts := inputParts(t, call)
-               require.Equal(t, first, parts[:2])
-               require.Equal(t, []string{tooLong(want)}, parts[2:])
-               require.Contains(t, parts[2], "was "+strconv.Itoa(len(want))+" bytes,")
-               require.Contains(t, parts[2], "\n"+want[:512]+"| ← LIMIT")
-
-               note, err := json.Marshal(map[string]string{
-                       "type": "input_text",
-                       "text": parts[2],
-               })
+               texts := compactorInputTexts(t, call)
+               require.Equal(t, first, texts[:2])
+               require.Equal(t, []string{compactorTooLong(want)}, texts[2:])
+               require.Contains(t, texts[2], "was "+strconv.Itoa(len(want))+" bytes,")
+               require.Contains(t, texts[2], "\n"+want[:512]+"| <- LIMIT")
+
+               note, err := json.Marshal(modelUserMessage(texts[2]))
                require.NoError(t, err)
                require.LessOrEqual(t, len(call.raw), len(server.Call(0).raw)+len(",")+len(note))
        }
                require.NoError(t, err)
                require.LessOrEqual(t, len(call.raw), len(server.Call(0).raw)+len(",")+len(note))
        }
@@ -98,26 +190,26 @@ func TestCompactorRetriesALongLineWithTheNewestNote(t *testing.T) {
 
 func TestCompactorKeepsTheShortestOfFiveLines(t *testing.T) {
        sizes := []int{700, 530, 600, 520, 560, 513}
 
 func TestCompactorKeepsTheShortestOfFiveLines(t *testing.T) {
        sizes := []int{700, 530, 600, 520, 560, 513}
-       server := newFakeResponses(t)
-       server.reply = func(w http.ResponseWriter, call fakeCall) {
-               writeEvents(w, completedEvent([]any{messageItem(strings.Repeat("z", sizes[call.number]))}, 100, 0, 5))
+       server := newModelFake(t)
+       server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
+               modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem(strings.Repeat("z", sizes[call.number]))}, 100, 0, 5))
        }
 
        }
 
-       line, err := newTestCompactor(t, server).Summarize(t.Context(), leafTask())
+       line, err := newTestCompactor(t, server).Summarize(t.Context(), compactorLeafTask())
        require.NoError(t, err)
        require.NoError(t, err)
-       require.Equal(t, compactionTries, server.Calls())
+       require.Equal(t, compactorTries, server.Calls())
        require.Len(t, line, 520)
 }
 
 func TestCompactorFailsWithoutALine(t *testing.T) {
        testCases := []struct {
                name  string
        require.Len(t, line, 520)
 }
 
 func TestCompactorFailsWithoutALine(t *testing.T) {
        testCases := []struct {
                name  string
-               reply func(http.ResponseWriter, fakeCall)
+               reply func(http.ResponseWriter, ModelFakeCall)
                want  string
        }{
                {
                        name: "failed call",
                want  string
        }{
                {
                        name: "failed call",
-                       reply: func(w http.ResponseWriter, call fakeCall) {
+                       reply: func(w http.ResponseWriter, call ModelFakeCall) {
                                w.WriteHeader(http.StatusInternalServerError)
                                _, _ = io.WriteString(w, "down")
                        },
                                w.WriteHeader(http.StatusInternalServerError)
                                _, _ = io.WriteString(w, "down")
                        },
@@ -125,8 +217,8 @@ func TestCompactorFailsWithoutALine(t *testing.T) {
                },
                {
                        name: "tool call and no text",
                },
                {
                        name: "tool call and no text",
-                       reply: func(w http.ResponseWriter, call fakeCall) {
-                               writeEvents(w, completedEvent([]any{map[string]any{"type": "function_call", "name": "zoom", "arguments": "{}"}}, 100, 0, 5))
+                       reply: func(w http.ResponseWriter, call ModelFakeCall) {
+                               modelWriteEvents(w, modelCompletedEvent([]any{map[string]any{"type": "function_call", "name": "zoom", "arguments": "{}"}}, 100, 0, 5))
                        },
                        want: "no line",
                },
                        },
                        want: "no line",
                },
@@ -134,10 +226,10 @@ func TestCompactorFailsWithoutALine(t *testing.T) {
 
        for _, tc := range testCases {
                t.Run(tc.name, func(t *testing.T) {
 
        for _, tc := range testCases {
                t.Run(tc.name, func(t *testing.T) {
-                       server := newFakeResponses(t)
+                       server := newModelFake(t)
                        server.reply = tc.reply
 
                        server.reply = tc.reply
 
-                       _, err := newTestCompactor(t, server).Summarize(t.Context(), leafTask())
+                       _, err := newTestCompactor(t, server).Summarize(t.Context(), compactorLeafTask())
                        require.ErrorContains(t, err, tc.want)
                        require.Equal(t, 1, server.Calls())
                })
                        require.ErrorContains(t, err, tc.want)
                        require.Equal(t, 1, server.Calls())
                })
@@ -146,24 +238,24 @@ func TestCompactorFailsWithoutALine(t *testing.T) {
 
 func TestCutLineKeepsWholeCharacters(t *testing.T) {
        line := strings.Repeat("a", 511) + "é" + "tail"
 
 func TestCutLineKeepsWholeCharacters(t *testing.T) {
        line := strings.Repeat("a", 511) + "é" + "tail"
-       cut := cutLine(line)
+       cut := compactorCutLine(line)
        require.Equal(t, strings.Repeat("a", 511), cut)
        require.True(t, utf8.ValidString(cut))
        require.Equal(t, strings.Repeat("a", 511), cut)
        require.True(t, utf8.ValidString(cut))
-       require.Equal(t, "short", cutLine("short"))
+       require.Equal(t, "short", compactorCutLine("short"))
 }
 
 func TestCompactorBuildsNodesInTheCore(t *testing.T) {
 }
 
 func TestCompactorBuildsNodesInTheCore(t *testing.T) {
-       server := newFakeResponses(t)
-       server.reply = func(w http.ResponseWriter, call fakeCall) {
+       server := newModelFake(t)
+       server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
                if call.number == 0 {
                        w.WriteHeader(http.StatusServiceUnavailable)
                        return
                }
 
                if call.number == 0 {
                        w.WriteHeader(http.StatusServiceUnavailable)
                        return
                }
 
-               writeEvents(w, completedEvent([]any{messageItem("summary of the long output")}, 100, 0, 5))
+               modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem("summary of the long output")}, 100, 0, 5))
        }
 
        }
 
-       memory, err := openMemory(t.Context(), t.TempDir(), newTestClock().Now, newTestCompactor(t, server))
+       memory, err := openMemory(t.Context(), t.TempDir(), newTestClock().Now, newTestCompactor(t, server), discardLog())
        require.NoError(t, err)
 
        defer func() {
        require.NoError(t, err)
 
        defer func() {
@@ -173,40 +265,32 @@ func TestCompactorBuildsNodesInTheCore(t *testing.T) {
        _, err = memory.Append("echo", strings.Repeat("output ", 200))
        require.NoError(t, err)
        require.NoError(t, memory.Wait(t.Context()))
        _, err = memory.Append("echo", strings.Repeat("output ", 200))
        require.NoError(t, err)
        require.NoError(t, memory.Wait(t.Context()))
-       require.False(t, isBuilt(t, memory.tree, 0, 0))
+       require.False(t, treeIsBuilt(t, memory.tree, 0, 0))
 
        _, err = memory.Append("user", "next")
        require.NoError(t, err)
        require.NoError(t, memory.Wait(t.Context()))
 
 
        _, err = memory.Append("user", "next")
        require.NoError(t, err)
        require.NoError(t, memory.Wait(t.Context()))
 
-       node := readNode(t, memory.tree, 0, 0)
+       node := treeReadNode(t, memory.tree, 0, 0)
        require.Equal(t, "summary of the long output", node.Text)
        require.Equal(t, "<chat>\n0+1|summary of the long output\n1+1|user: next\n</chat>", memory.View())
        require.Equal(t, 2, server.Calls())
 }
 
        require.Equal(t, "summary of the long output", node.Text)
        require.Equal(t, "<chat>\n0+1|summary of the long output\n1+1|user: next\n</chat>", memory.View())
        require.Equal(t, 2, server.Calls())
 }
 
-func newTestCompactor(t *testing.T, server *fakeResponses) *Compactor {
+func newTestCompactor(t *testing.T, server *ModelFake) *Compactor {
        t.Helper()
 
        t.Helper()
 
-       client := newClient(server.URL, apiKey(testKey), "agent-key", filepath.Join(t.TempDir(), "usage.jsonl"))
+       client := newModelClient(server.URL, ModelAPIKey(modelTestKey), "agent-key", filepath.Join(t.TempDir(), "usage.jsonl"))
 
 
-       return newCompactor(client, "Ada", ModelConfig{
+       return newCompactor(client, compactorAdaPrompt, ModelConfig{
                Model:  "gpt-test",
                Effort: "high",
        })
 }
 
                Model:  "gpt-test",
                Effort: "high",
        })
 }
 
-func leafTask() Task {
-       return Task{
-               Level:   0,
-               Index:   0,
-               Context: "<chat>\n</chat>",
-               Input:   []string{"user: " + strings.Repeat("y", 600)},
-       }
-}
-
-// inputParts returns the texts of the one user message of call.
-func inputParts(t *testing.T, call fakeCall) []string {
+// compactorInputTexts returns the text of each user message of call. Each
+// message has one text part.
+func compactorInputTexts(t *testing.T, call ModelFakeCall) []string {
        t.Helper()
 
        var body struct {
        t.Helper()
 
        var body struct {
@@ -220,14 +304,28 @@ func inputParts(t *testing.T, call fakeCall) []string {
        }
 
        require.NoError(t, json.Unmarshal(call.raw, &body))
        }
 
        require.NoError(t, json.Unmarshal(call.raw, &body))
-       require.Len(t, body.Input, 1)
-       require.Equal(t, "user", body.Input[0].Role)
 
 
-       texts := make([]string, len(body.Input[0].Content))
-       for n, part := range body.Input[0].Content {
-               require.Equal(t, "input_text", part.Type)
-               texts[n] = part.Text
+       texts := make([]string, len(body.Input))
+       for n, message := range body.Input {
+               require.Equal(t, "user", message.Role)
+               require.Len(t, message.Content, 1)
+               require.Equal(t, "input_text", message.Content[0].Type)
+               texts[n] = message.Content[0].Text
        }
 
        return texts
 }
        }
 
        return texts
 }
+
+func compactorLeafTask() Task {
+       return Task{
+               Level:   0,
+               Index:   0,
+               Context: "<chat>\n</chat>",
+               Input:   []string{"user: " + strings.Repeat("y", 600)},
+       }
+}
+
+// compactorAdaPrompt returns the prompt of the agent Ada, without instructions.
+func compactorAdaPrompt() (string, error) {
+       return prompt("Ada", ""), nil
+}