executor: fix Claude rate-limit detection and prioritize Gemini when limited

Updated parseStream to detect 'rate_limit_event' and 'assistant' error:rate_limit messages from the Claude CLI. Updated Classifier to strongly prefer non-rate-limited agents. Added logging to Pool to track rate-limit status during classification.

1 files changed, 9 insertions, 1 deletions

diff --git a/internal/executor/executor.go b/internal/executor/executor.go
index df222f8..f404ddf 100644
--- a/internal/executor/executor.go
+++ b/internal/executor/executor.go
@@ -238,6 +238,7 @@ func (p *Pool) executeResume(ctx context.Context, t *task.Task, exec *storage.Ex
 				retryAfter = 1 * time.Minute
 			}
 			p.rateLimited[agentType] = time.Now().Add(retryAfter)
+			p.logger.Info("agent rate limited", "agent", agentType, "retryAfter", retryAfter)
 			p.mu.Unlock()
 		}
 
@@ -295,12 +296,18 @@ func (p *Pool) execute(ctx context.Context, t *task.Task) {
 		now := time.Now()
 		for agent := range p.runners {
 			activeTasks[agent] = p.activePerAgent[agent]
-			rateLimited[agent] = now.Before(p.rateLimited[agent])
+			activeUntil := p.rateLimited[agent]
+			isLimited := now.Before(activeUntil)
+			rateLimited[agent] = isLimited
+			if isLimited {
+				p.logger.Debug("agent rate limited", "agent", agent, "until", activeUntil)
+			}
 		}
 		status := SystemStatus{
 			ActiveTasks: activeTasks,
 			RateLimited: rateLimited,
 		}
+		p.logger.Debug("classifying task", "taskID", t.ID, "status", status)
 		p.mu.Unlock()
 
 		cls, err := p.Classifier.Classify(ctx, t.Name, t.Agent.Instructions, status)
@@ -433,6 +440,7 @@ func (p *Pool) execute(ctx context.Context, t *task.Task) {
 				retryAfter = 1 * time.Minute
 			}
 			p.rateLimited[agentType] = time.Now().Add(retryAfter)
+			p.logger.Info("agent rate limited", "agent", agentType, "retryAfter", retryAfter)
 			p.mu.Unlock()
 		}