merge: integrate github/main — LocalRunner, real GeminiRunner, llm clientHEADmain

Merges 12 commits from github/main (formerly master) that were developed
independently. Key additions:
- LocalRunner: OpenAI-compatible local LLM execution (Ollama, LM Studio)
- Real GeminiRunner with full sandbox parity to ClaudeRunner
- llm.Client for enriching CI failures and elaboration via local model
- retry.ParseRetryAfter moved to shared package
- tokens_in/tokens_out columns in executions table

Conflict resolutions:
- Kept local main's VAPID/push, stories, projects, agent events schema
- Merged both sets of Config fields (local + LocalModel from github/main)
- Unified activePerAgent accounting (decActiveAgent helper)
- Removed duplicate helpers from claude.go (now in helpers.go)
- Fixed double-decrement bug in handleRunResult vs decActiveAgent

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

1 files changed, 95 insertions, 0 deletions

diff --git a/internal/executor/summary.go b/internal/executor/summary.go
index a942de0..bcf5cfd 100644
--- a/internal/executor/summary.go
+++ b/internal/executor/summary.go
@@ -2,11 +2,26 @@ package executor
 
 import (
 	"bufio"
+	"context"
 	"encoding/json"
+	"io"
 	"os"
 	"strings"
+	"time"
+
+	"github.com/thepeterstone/claudomator/internal/llm"
 )
 
+// synthesizeSummaryMaxBytes caps how much of the stdout log we send to the
+// LLM. Larger values cost more tokens with diminishing returns for a 2-4
+// sentence summary.
+const synthesizeSummaryMaxBytes = 16 * 1024
+
+// synthesizeSummaryTimeout caps the LLM call so a slow local model can't
+// stall executor finalization. On timeout, we return "" (the existing
+// no-summary path takes over).
+const synthesizeSummaryTimeout = 6 * time.Second
+
 // extractSummary reads a stream-json stdout log and returns the text following
 // the last "## Summary" heading found in any assistant text block.
 // Returns empty string if the file cannot be read or no summary is found.
@@ -28,6 +43,86 @@ func extractSummary(stdoutPath string) string {
 	return last
 }
 
+// synthesizeSummary asks the LLM to summarize the assistant text content in
+// stdoutPath when no "## Summary" heading was present. Returns "" on any
+// error, an empty file, or an empty model response — preserving the
+// existing "no summary" behavior so the new path is purely additive.
+func synthesizeSummary(parent context.Context, c *llm.Client, stdoutPath string) string {
+	if c == nil || stdoutPath == "" {
+		return ""
+	}
+	text := readAssistantTextTail(stdoutPath, synthesizeSummaryMaxBytes)
+	if strings.TrimSpace(text) == "" {
+		return ""
+	}
+
+	cctx, cancel := context.WithTimeout(parent, synthesizeSummaryTimeout)
+	defer cancel()
+	resp, err := c.Chat(cctx, llm.ChatRequest{
+		Messages: []llm.Message{
+			{Role: "system", Content: "You summarize what an automated coding agent did. Reply with 2-4 sentences of plain prose. No bullets, no headings, no preamble."},
+			{Role: "user", Content: "Here is the agent's output. Summarize what it accomplished:\n\n" + text},
+		},
+	})
+	if err != nil {
+		return ""
+	}
+	return strings.TrimSpace(resp.Content)
+}
+
+// readAssistantTextTail returns the concatenated `text` blocks from assistant
+// stream-json events in the last maxBytes of the file. Non-assistant events
+// (system, result, tool_use, etc.) are skipped so the LLM sees just what the
+// agent said. Returns "" on any error.
+func readAssistantTextTail(stdoutPath string, maxBytes int64) string {
+	f, err := os.Open(stdoutPath)
+	if err != nil {
+		return ""
+	}
+	defer f.Close()
+
+	stat, err := f.Stat()
+	if err != nil {
+		return ""
+	}
+	size := stat.Size()
+	if size > maxBytes {
+		if _, err := f.Seek(size-maxBytes, io.SeekStart); err != nil {
+			return ""
+		}
+	}
+
+	var sb strings.Builder
+	scanner := bufio.NewScanner(f)
+	scanner.Buffer(make([]byte, 1024*1024), 1024*1024)
+	first := size > maxBytes // if we seeked, drop the first (likely partial) line
+	for scanner.Scan() {
+		if first {
+			first = false
+			continue
+		}
+		var event struct {
+			Type    string `json:"type"`
+			Message struct {
+				Content []struct {
+					Type string `json:"type"`
+					Text string `json:"text"`
+				} `json:"content"`
+			} `json:"message"`
+		}
+		if err := json.Unmarshal(scanner.Bytes(), &event); err != nil || event.Type != "assistant" {
+			continue
+		}
+		for _, block := range event.Message.Content {
+			if block.Type == "text" && block.Text != "" {
+				sb.WriteString(block.Text)
+				sb.WriteString("\n")
+			}
+		}
+	}
+	return sb.String()
+}
+
 // summaryFromLine parses a single stream-json line and returns the text after
 // "## Summary" if the line is an assistant text block containing that heading.
 func summaryFromLine(line []byte) string {