esengine--deepseek-reasonix/internal/agent/postllmcall_flow_test.go

package agent

import (
	"context"
	"strings"
	"testing"

	"reasonix/internal/event"
	"reasonix/internal/provider"
	"reasonix/internal/tool"
)

func reasoningTurn() [][]provider.Chunk {
	return [][]provider.Chunk{{
		{Type: provider.ChunkReasoning, Text: "think A "},
		{Type: provider.ChunkReasoning, Text: "think B"},
		{Type: provider.ChunkText, Text: "the answer"},
		{Type: provider.ChunkDone},
	}}
}

func recordReasoning(events *[]string) event.Sink {
	return event.FuncSink(func(e event.Event) {
		if e.Kind == event.Reasoning {
			*events = append(*events, e.Text)
		}
	})
}

func assistantReasoning(msgs []provider.Message) string {
	for _, m := range msgs {
		if m.Role == provider.RoleAssistant {
			return m.ReasoningContent
		}
	}
	return ""
}

// TestPostLLMCallAbsentStreamsReasoningLive is the regression guard: with no
// PostLLMCall hook, reasoning must still stream chunk-by-chunk (one Reasoning
// event per delta) so the live "thinking…" display keeps working.
func TestPostLLMCallAbsentStreamsReasoningLive(t *testing.T) {
	prov := &scriptedProvider{name: "p", turns: reasoningTurn()}
	var reasoningEvents []string
	a := New(prov, tool.NewRegistry(), NewSession(""), Options{}, recordReasoning(&reasoningEvents))

	if err := a.Run(context.Background(), "go"); err != nil {
		t.Fatalf("Run: %v", err)
	}

	if len(reasoningEvents) != 2 {
		t.Fatalf("want 2 live reasoning events (one per chunk), got %d: %v", len(reasoningEvents), reasoningEvents)
	}
	if joined := strings.Join(reasoningEvents, ""); joined != "think A think B" {
		t.Fatalf("streamed reasoning = %q, want the full chain", joined)
	}
	if got := assistantReasoning(a.session.Messages); got != "think A think B" {
		t.Fatalf("stored reasoning = %q, want the untransformed chain", got)
	}
}

// TestPostLLMCallTransformsReasoningOnce proves a configured hook suppresses the
// live stream, sees the full reasoning, and its output replaces both the single
// emitted Reasoning event and the stored reasoning_content.
func TestPostLLMCallTransformsReasoningOnce(t *testing.T) {
	prov := &scriptedProvider{name: "p", turns: reasoningTurn()}
	var reasoningEvents []string
	h := &stubHooks{hasPostLLM: true, postLLMOut: "TRANSLATED"}
	a := New(prov, tool.NewRegistry(), NewSession(""), Options{Hooks: h}, recordReasoning(&reasoningEvents))

	if err := a.Run(context.Background(), "go"); err != nil {
		t.Fatalf("Run: %v", err)
	}

	if len(reasoningEvents) != 1 || reasoningEvents[0] != "TRANSLATED" {
		t.Fatalf("want one transformed reasoning event, got %v", reasoningEvents)
	}
	if len(h.postLLMSeen) != 1 || h.postLLMSeen[0] != "think A think B" {
		t.Fatalf("hook saw %v, want the full original reasoning once", h.postLLMSeen)
	}
	if len(h.postLLMTurns) != 1 || h.postLLMTurns[0] != 1 {
		t.Fatalf("hook turns = %v, want [1]", h.postLLMTurns)
	}
	if got := assistantReasoning(a.session.Messages); got != "TRANSLATED" {
		t.Fatalf("stored reasoning = %q, want the hook's replacement", got)
	}
}

// TestPostLLMCallConfiguredButNoReasoning makes sure a hook with an empty
// reasoning chain neither calls the hook nor emits a stray Reasoning event.
func TestPostLLMCallConfiguredButNoReasoning(t *testing.T) {
	prov := &scriptedProvider{name: "p", turns: [][]provider.Chunk{{
		{Type: provider.ChunkText, Text: "answer, no thinking"},
		{Type: provider.ChunkDone},
	}}}
	var reasoningEvents []string
	h := &stubHooks{hasPostLLM: true, postLLMOut: "should not be used"}
	a := New(prov, tool.NewRegistry(), NewSession(""), Options{Hooks: h}, recordReasoning(&reasoningEvents))

	if err := a.Run(context.Background(), "go"); err != nil {
		t.Fatalf("Run: %v", err)
	}

	if len(reasoningEvents) != 0 {
		t.Fatalf("no reasoning should emit no Reasoning events, got %v", reasoningEvents)
	}
	if len(h.postLLMSeen) != 0 {
		t.Fatalf("hook should not fire on empty reasoning, saw %v", h.postLLMSeen)
	}
}

// TestPostLLMCallKeepsSignedReasoningOriginal proves that when the reasoning is
// pinned by a provider signature (Anthropic extended thinking), a transform hook
// changes only the live display — the stored reasoning_content stays the original
// so the signed thinking block can be replayed verbatim on the next tool-call
// turn. Storing the transformed text under the original signature is a 400.
func TestPostLLMCallKeepsSignedReasoningOriginal(t *testing.T) {
	prov := &scriptedProvider{name: "p", turns: [][]provider.Chunk{{
		{Type: provider.ChunkReasoning, Text: "think A "},
		{Type: provider.ChunkReasoning, Text: "think B", Signature: "sig-xyz"},
		{Type: provider.ChunkText, Text: "answer"},
		{Type: provider.ChunkDone},
	}}}
	var reasoningEvents []string
	h := &stubHooks{hasPostLLM: true, postLLMOut: "TRANSLATED"}
	a := New(prov, tool.NewRegistry(), NewSession(""), Options{Hooks: h}, recordReasoning(&reasoningEvents))

	if err := a.Run(context.Background(), "go"); err != nil {
		t.Fatalf("Run: %v", err)
	}
	if len(reasoningEvents) != 1 || reasoningEvents[0] != "TRANSLATED" {
		t.Fatalf("want the transformed reasoning shown live, got %v", reasoningEvents)
	}
	if got := assistantReasoning(a.session.Messages); got != "think A think B" {
		t.Fatalf("stored reasoning = %q, want the original (signature pins it)", got)
	}
	for _, m := range a.session.Messages {
		if m.Role == provider.RoleAssistant && m.ReasoningSignature != "sig-xyz" {
			t.Fatalf("stored signature = %q, want sig-xyz alongside its original text", m.ReasoningSignature)
		}
	}
}