fix(llamacpp,anthropic): SSE headroom, no-choices guard, real context window

- llamacpp: 4MB SSE scanner buffer (the 64KB bufio.Scanner default killed
  streams whose single line exceeded it, e.g. a write tool call carrying a
  whole file) and an empty-choices guard in toResponse instead of a panic;
  request payload now uses bytes.NewReader (drops a full string copy).
- anthropic: Capabilities() reported a 1M-token context window for any
  non-haiku model. Callers use that number to decide when to compact, so
  compaction would have fired far too late and requests overflowed the
  real window. Default is now the standard 200k, configurable via
  Config.ContextWindow for extended-window models/plans.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

pkg/llm/providers/anthropic/client.go

@@ -7,8 +7,8 @@ import (
 	"encoding/json"
 	"fmt"
 	"io"
-	"net/http"
 	"iter"
+	"net/http"
 	"strings"
 
 	"github.com/VictorVargas/rony-llm-agent/pkg/llm"
@@ -19,6 +19,13 @@ const (
 	defaultModel     = "claude-opus-4-8"
 	defaultMaxTokens = 8192
 	anthropicVersion = "2023-06-01"
+	// defaultContextWindow is what Capabilities() reports when
+	// Config.ContextWindow is unset: 200k tokens, the standard window for
+	// Claude models. Callers use this number to decide when to compact
+	// their conversation, so over-reporting it (the old code assumed 1M
+	// for anything that wasn't haiku) meant compaction fired far too late
+	// and requests started overflowing the real window.
+	defaultContextWindow = 200000
 )
 
 // Config holds the settings needed to create an Anthropic client.
@@ -27,6 +34,7 @@ type Config struct {
 	Model         string // defaults to claude-opus-4-8
 	BaseURL       string // defaults to https://api.anthropic.com/v1
 	MaxTokens     int    // default max_tokens sent on every request (Anthropic requires one); 0 = defaultMaxTokens
+	ContextWindow int    // model's context window in tokens (0 = defaultContextWindow); raise it only for models/plans with an extended window
 	Temperature   *float32
 	TopP          *float32
 }
@@ -37,6 +45,7 @@ type Client struct {
 	baseURL       string
 	model         string
 	maxTokens     int
+	contextWindow int
 	temperature   *float32
 	topP          *float32
 	http          *http.Client
@@ -63,11 +72,17 @@ func New(cfg Config) (*Client, error) {
 		maxTokens = defaultMaxTokens
 	}
 
+	contextWindow := cfg.ContextWindow
+	if contextWindow == 0 {
+		contextWindow = defaultContextWindow
+	}
+
 	return &Client{
 		apiKey:        cfg.APIKey,
 		baseURL:       baseURL,
 		model:         model,
 		maxTokens:     maxTokens,
+		contextWindow: contextWindow,
 		temperature:   cfg.Temperature,
 		topP:          cfg.TopP,
 		http:          http.DefaultClient,
@@ -257,15 +272,11 @@ func (c *Client) Name() string {
 }
 
 func (c *Client) Capabilities() llm.ProviderCapabilities {
-	maxContext := 1000000
-	if strings.Contains(c.model, "haiku") {
-		maxContext = 200000
-	}
 	return llm.ProviderCapabilities{
 		SupportsTools:    true,
 		SupportsVision:   true,
 		SupportsJSON:     true,
-		MaxContextWindow: maxContext,
+		MaxContextWindow: c.contextWindow,
 	}
 }
 

pkg/llm/providers/llamacpp/client.go

@@ -2,12 +2,13 @@ package llamacpp
 
 import (
 	"bufio"
+	"bytes"
 	"context"
 	"encoding/json"
 	"fmt"
 	"io"
-	"net/http"
 	"iter"
+	"net/http"
 	"strings"
 	"time"
 
@@ -123,7 +124,7 @@ func (c *Client) Generate(ctx context.Context, req llm.CompletionRequest) (llm.C
 		return llm.CompletionResponse{}, fmt.Errorf("decoding response: %w", err)
 	}
 
-	return c.toResponse(apiResp), nil
+	return c.toResponse(apiResp)
 }
 
 func (c *Client) Stream(ctx context.Context, req llm.CompletionRequest) iter.Seq2[llm.StreamChunk, error] {
@@ -191,6 +192,11 @@ func (c *Client) Stream(ctx context.Context, req llm.CompletionRequest) iter.Seq
 		}
 
 		scanner := bufio.NewScanner(resp.Body)
+		// A single SSE line can exceed bufio.Scanner's 64KB default cap
+		// (e.g. a large tool-call arguments delta or a long reasoning
+		// event), which would kill the stream mid-turn with "token too
+		// long" — same headroom the anthropic client already reserves.
+		scanner.Buffer(make([]byte, 0, 64*1024), 4*1024*1024)
 		for scanner.Scan() {
 			line := scanner.Text()
 			if !strings.HasPrefix(line, "data: ") {
@@ -370,11 +376,16 @@ func (c *Client) buildRequest(req llm.CompletionRequest, stream bool) (io.Reader
 	if err != nil {
 		return nil, fmt.Errorf("marshaling request: %w", err)
 	}
-	return strings.NewReader(string(data)), nil
+	return bytes.NewReader(data), nil
 }
 
 // toResponse converts a llama.cpp API response to our CompletionResponse.
-func (c *Client) toResponse(resp llamaChatResponse) llm.CompletionResponse {
+func (c *Client) toResponse(resp llamaChatResponse) (llm.CompletionResponse, error) {
+	// Guard against a 200 response with no choices (e.g. a misbehaving
+	// server or proxy) — indexing Choices[0] blindly panics the whole app.
+	if len(resp.Choices) == 0 {
+		return llm.CompletionResponse{}, fmt.Errorf("llamacpp: response contained no choices")
+	}
 	choice := resp.Choices[0]
 	result := llm.CompletionResponse{
 		ID:         resp.ID,
@@ -398,7 +409,7 @@ func (c *Client) toResponse(resp llamaChatResponse) llm.CompletionResponse {
 		TotalTokens:  resp.Usage.TotalTokens,
 	}
 
-	return result
+	return result, nil
 }
 
 // llama.cpp API types