From 7a69f14277b27a3da7b2651cdc85568df7468559 Mon Sep 17 00:00:00 2001
From: Ollie Agent <ollie@example.com>
Date: Fri, 31 Jul 2026 07:46:32 +0200
Subject: [PATCH] toolsrv: typed RateLimitedError, agent loop: don't count
 rate-limited calls as consecutive errors

---
 agent/loop.go     | 19 +++++++++++++++++--
 toolsrv/server.go | 12 ++++++++++++
 toolsrv/shell.go  |  2 +-
 3 files changed, 30 insertions(+), 3 deletions(-)

diff --git a/agent/loop.go b/agent/loop.go
index 21d647c..54ec5b7 100644
--- a/agent/loop.go
+++ b/agent/loop.go
@@ -339,7 +339,12 @@ func run(ctx context.Context, cfg agentConfig, state state) error {
 						emit(cfg, Event{Role: "tool", Name: tc.Name, Content: result})
 						return toolResult{ToolCallID: tc.ID, Name: tc.Name, Content: result, IsError: true}, true
 					}
-					result = fmt.Sprintf("error: %v", err)
+					var rlErr *toolsrv.RateLimitedError
+					if errors.As(err, &rlErr) {
+						result = fmt.Sprintf("error: shell is rate-limited — blocked for %v. Do not call shell() again until the block expires. Use other tools or wait.", rlErr.Remaining)
+					} else {
+						result = fmt.Sprintf("error: %v", err)
+					}
 				} else {
 					result = out
 					resultBlocks = blocks
@@ -516,14 +521,24 @@ func run(ctx context.Context, cfg agentConfig, state state) error {
 		}
 
 		// Track consecutive rounds where every tool call errored.
+		// Rate-limited errors don't count — they're a temporary block,
+		// not a real failure. Reset consecutiveErrors so the hard abort
+		// doesn't fire.
 		allErrors := len(results) > 0
+		allRateLimited := allErrors
 		for _, r := range results {
 			if !r.IsError {
 				allErrors = false
+				allRateLimited = false
 				break
 			}
+			if !strings.HasPrefix(r.Content, "error: shell is rate-limited") {
+				allRateLimited = false
+			}
 		}
-		if allErrors {
+		if allRateLimited {
+			consecutiveErrors = 0
+		} else if allErrors {
 			consecutiveErrors++
 		} else {
 			consecutiveErrors = 0
diff --git a/toolsrv/server.go b/toolsrv/server.go
index f294e13..a648c69 100644
--- a/toolsrv/server.go
+++ b/toolsrv/server.go
@@ -20,6 +20,18 @@ const (
 	blockDuration = 5 * time.Minute
 )
 
+// RateLimitedError is returned by checkRateLimit when the shell is blocked
+// due to too many validation failures. The agent loop uses this to detect
+// the condition and inject a clear signal rather than counting it as a
+// normal tool error.
+type RateLimitedError struct {
+	Remaining time.Duration
+}
+
+func (e *RateLimitedError) Error() string {
+	return fmt.Sprintf("rate limited: too many validation failures, blocked for %v", e.Remaining)
+}
+
 // Server runs code in a sandboxed environment.
 type Server struct {
 	// cwd is the working directory for sandboxed commands. If empty,
diff --git a/toolsrv/shell.go b/toolsrv/shell.go
index f2de2e1..23dc347 100644
--- a/toolsrv/shell.go
+++ b/toolsrv/shell.go
@@ -383,7 +383,7 @@ func (e *Server) checkRateLimit() error {
 	now := time.Now()
 	if now.Before(e.blockedUntil) {
 		remaining := e.blockedUntil.Sub(now).Round(time.Second)
-		return fmt.Errorf("rate limited: too many validation failures, blocked for %v", remaining)
+		return &RateLimitedError{Remaining: remaining}
 	}
 	return nil
 }