Extract shared chatrun package to eliminate chat-runner DRY violations

The CLI, LSP server and tmux code-action tool each carried near-identical
chat-running logic: invoking the LLM (streaming-aware), collecting the
response, and accounting sent/received bytes into the stats package.

Introduce internal/chatrun with:
- Invoke: streaming-aware LLM call that collects the full response and
  optionally mirrors chunks to a writer (nil writer = collect only).
- SentBytes / Account: shared byte counting and stats.Update.

Wire all three surfaces to it:
- hexaicli: runChatRequest delegates to chatrun.Invoke; summarizeChatRun
  uses chatrun.Account. Removed the duplicated streaming/simple helpers.
- hexaiaction: runOnce uses chatrun.Invoke + chatrun.Account, keeping the
  tmux status update local.
- lsp: chatWithStats and the completion path use chatrun.SentBytes/Invoke;
  extracted unavailableClientError to keep chatWithStats small.

chatrun has 100% test coverage; full suite passes with -race.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

2 files changed, 249 insertions, 0 deletions

diff --git a/internal/chatrun/chatrun.go b/internal/chatrun/chatrun.go
new file mode 100644
index 0000000..01400cd
--- /dev/null
+++ b/internal/chatrun/chatrun.go
@@ -0,0 +1,114 @@
+// Package chatrun holds the shared chat-running primitives that used to be
+// duplicated across the three Hexai surfaces: the CLI (internal/hexaicli), the
+// LSP server (internal/lsp) and the tmux code-action tool (internal/hexaiaction).
+//
+// Every surface had to do the same two things when talking to an LLM:
+//
+//  1. Invoke the client, preferring streaming (llm.Streamer) when the provider
+//     implements it and falling back to a single Chat call otherwise, while
+//     collecting the full response text.
+//  2. Account for the exchange: count the bytes sent (sum of message contents)
+//     and received (response length) and feed them into the stats package.
+//
+// Both pieces lived in slightly diverging copies (runChatRequest/runStreaming
+// Chat/runSimpleChat in the CLI, chatWithStats in the LSP, runOnce in the
+// action tool). This package centralises them so all three share one
+// implementation and behaviour stays identical.
+package chatrun
+
+import (
+	"context"
+	"fmt"
+	"io"
+	"strings"
+
+	"codeberg.org/snonux/hexai/internal/llm"
+	"codeberg.org/snonux/hexai/internal/stats"
+)
+
+// Chatter is the minimal client capability Invoke needs: a single Chat call.
+// It is deliberately narrower than llm.Client (no Name/DefaultModel) so callers
+// that only have a chat-doer abstraction — such as the code-action surface —
+// can use Invoke without widening their own interface. Clients that also
+// implement llm.Streamer get streaming automatically via a type assertion.
+type Chatter interface {
+	Chat(ctx context.Context, messages []llm.Message, opts ...llm.RequestOption) (string, error)
+}
+
+// Invoke sends msgs to the client and returns the full assistant response.
+//
+// When the client implements llm.Streamer the response is streamed and, if out
+// is non-nil, each chunk is forwarded to out as it arrives (this is how the CLI
+// renders incremental output). Otherwise a single Chat call is made and, when
+// out is non-nil, the whole response is written to it once.
+//
+// Passing a nil out collects the response without writing it anywhere, which is
+// what the LSP and code-action surfaces need (they post-process the text before
+// applying it to a document).
+func Invoke(ctx context.Context, client Chatter, msgs []llm.Message, opts []llm.RequestOption, out io.Writer) (string, error) {
+	if streamer, ok := client.(llm.Streamer); ok {
+		return invokeStreaming(ctx, streamer, msgs, opts, out)
+	}
+	return invokeSimple(ctx, client, msgs, opts, out)
+}
+
+// invokeStreaming drives ChatStream, accumulating the full text while
+// optionally mirroring each chunk to out. A write error to out aborts further
+// writes and is returned once streaming finishes.
+func invokeStreaming(ctx context.Context, client llm.Streamer, msgs []llm.Message, opts []llm.RequestOption, out io.Writer) (string, error) {
+	var output strings.Builder
+	var writeErr error
+	err := client.ChatStream(ctx, msgs, func(chunk string) {
+		output.WriteString(chunk)
+		if out == nil || writeErr != nil {
+			return
+		}
+		if _, werr := io.WriteString(out, chunk); werr != nil {
+			writeErr = werr
+		}
+	}, opts...)
+	if err != nil {
+		return "", err
+	}
+	if writeErr != nil {
+		return "", writeErr
+	}
+	return output.String(), nil
+}
+
+// invokeSimple performs a single Chat call and, when out is non-nil, writes the
+// whole response to it.
+func invokeSimple(ctx context.Context, client Chatter, msgs []llm.Message, opts []llm.RequestOption, out io.Writer) (string, error) {
+	output, err := client.Chat(ctx, msgs, opts...)
+	if err != nil {
+		return "", err
+	}
+	if out != nil {
+		if _, werr := fmt.Fprint(out, output); werr != nil {
+			return "", werr
+		}
+	}
+	return output, nil
+}
+
+// SentBytes returns the total number of content bytes across msgs. This is the
+// "sent" figure every surface reports and feeds into the stats package.
+func SentBytes(msgs []llm.Message) int {
+	sent := 0
+	for _, m := range msgs {
+		sent += len(m.Content)
+	}
+	return sent
+}
+
+// Account records a completed exchange in the stats package and returns the
+// sent/received byte counts so callers can build their own summaries. The
+// stats.Update error is intentionally swallowed because none of the surfaces
+// treat a stats failure as fatal; callers that want to log it can call
+// stats.Update directly instead.
+func Account(ctx context.Context, provider, model string, msgs []llm.Message, output string) (sent, recv int) {
+	sent = SentBytes(msgs)
+	recv = len(output)
+	_ = stats.Update(ctx, provider, model, sent, recv)
+	return sent, recv
+}
diff --git a/internal/chatrun/chatrun_test.go b/internal/chatrun/chatrun_test.go
new file mode 100644
index 0000000..f2f5806
--- /dev/null
+++ b/internal/chatrun/chatrun_test.go
@@ -0,0 +1,135 @@
+package chatrun
+
+import (
+	"bytes"
+	"context"
+	"errors"
+	"strings"
+	"testing"
+
+	"codeberg.org/snonux/hexai/internal/llm"
+)
+
+// simpleClient implements only Chatter (no streaming).
+type simpleClient struct {
+	resp string
+	err  error
+}
+
+func (c simpleClient) Chat(_ context.Context, _ []llm.Message, _ ...llm.RequestOption) (string, error) {
+	return c.resp, c.err
+}
+
+// streamClient implements Chatter and llm.Streamer.
+type streamClient struct {
+	chunks []string
+	err    error
+}
+
+func (c streamClient) Chat(_ context.Context, _ []llm.Message, _ ...llm.RequestOption) (string, error) {
+	return strings.Join(c.chunks, ""), nil
+}
+
+func (c streamClient) ChatStream(_ context.Context, _ []llm.Message, onDelta func(string), _ ...llm.RequestOption) error {
+	if c.err != nil {
+		return c.err
+	}
+	for _, chunk := range c.chunks {
+		onDelta(chunk)
+	}
+	return nil
+}
+
+// errWriter always fails, to exercise the write-error paths.
+type errWriter struct{ err error }
+
+func (w errWriter) Write([]byte) (int, error) { return 0, w.err }
+
+func TestInvoke_SimpleCollectsAndWrites(t *testing.T) {
+	var out bytes.Buffer
+	got, err := Invoke(context.Background(), simpleClient{resp: "hello"}, nil, nil, &out)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if got != "hello" || out.String() != "hello" {
+		t.Fatalf("got=%q out=%q, want hello/hello", got, out.String())
+	}
+}
+
+func TestInvoke_SimpleNilWriter(t *testing.T) {
+	got, err := Invoke(context.Background(), simpleClient{resp: "x"}, nil, nil, nil)
+	if err != nil || got != "x" {
+		t.Fatalf("got=%q err=%v, want x/nil", got, err)
+	}
+}
+
+func TestInvoke_SimpleChatError(t *testing.T) {
+	_, err := Invoke(context.Background(), simpleClient{err: errors.New("boom")}, nil, nil, nil)
+	if err == nil || !strings.Contains(err.Error(), "boom") {
+		t.Fatalf("expected chat error, got %v", err)
+	}
+}
+
+func TestInvoke_SimpleWriteError(t *testing.T) {
+	_, err := Invoke(context.Background(), simpleClient{resp: "x"}, nil, nil, errWriter{err: errors.New("wfail")})
+	if err == nil || !strings.Contains(err.Error(), "wfail") {
+		t.Fatalf("expected write error, got %v", err)
+	}
+}
+
+func TestInvoke_StreamingCollectsAndWrites(t *testing.T) {
+	var out bytes.Buffer
+	got, err := Invoke(context.Background(), streamClient{chunks: []string{"a", "b", "c"}}, nil, nil, &out)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if got != "abc" || out.String() != "abc" {
+		t.Fatalf("got=%q out=%q, want abc/abc", got, out.String())
+	}
+}
+
+func TestInvoke_StreamingNilWriter(t *testing.T) {
+	got, err := Invoke(context.Background(), streamClient{chunks: []string{"a", "b"}}, nil, nil, nil)
+	if err != nil || got != "ab" {
+		t.Fatalf("got=%q err=%v, want ab/nil", got, err)
+	}
+}
+
+func TestInvoke_StreamingError(t *testing.T) {
+	_, err := Invoke(context.Background(), streamClient{err: errors.New("sfail")}, nil, nil, nil)
+	if err == nil || !strings.Contains(err.Error(), "sfail") {
+		t.Fatalf("expected stream error, got %v", err)
+	}
+}
+
+func TestInvoke_StreamingWriteError(t *testing.T) {
+	// Full text must still be collected even though the writer fails.
+	_, err := Invoke(context.Background(), streamClient{chunks: []string{"a", "b"}}, nil, nil, errWriter{err: errors.New("wfail")})
+	if err == nil || !strings.Contains(err.Error(), "wfail") {
+		t.Fatalf("expected write error, got %v", err)
+	}
+}
+
+func TestSentBytes(t *testing.T) {
+	msgs := []llm.Message{
+		{Role: "system", Content: "abc"},
+		{Role: "user", Content: "de"},
+	}
+	if got := SentBytes(msgs); got != 5 {
+		t.Fatalf("SentBytes = %d, want 5", got)
+	}
+	if got := SentBytes(nil); got != 0 {
+		t.Fatalf("SentBytes(nil) = %d, want 0", got)
+	}
+}
+
+func TestAccount(t *testing.T) {
+	msgs := []llm.Message{{Role: "user", Content: "1234"}}
+	sent, recv := Account(context.Background(), "prov", "model", msgs, "response")
+	if sent != 4 {
+		t.Fatalf("sent = %d, want 4", sent)
+	}
+	if recv != len("response") {
+		t.Fatalf("recv = %d, want %d", recv, len("response"))
+	}
+}