From add1f92bdd2ff96e9df40499e8e83e44cd702343 Mon Sep 17 00:00:00 2001 From: Devon Rifkin Date: Fri, 21 Aug 2026 17:44:44 -0700 Subject: [PATCH] launch: disable claude code token countdown to preserve KV cache (#17918) Claude Code adds a "tokens left" system message after every tool result. Since ollama moves system messages to the front of the prompt, this breaks the KV cache on every request. --- cmd/launch/claude.go | 1 + cmd/launch/claude_test.go | 1 + 2 files changed, 2 insertions(+) diff --git a/cmd/launch/claude.go b/cmd/launch/claude.go index 394feefb5..f2af75f0d 100644 --- a/cmd/launch/claude.go +++ b/cmd/launch/claude.go @@ -70,6 +70,7 @@ func (c *Claude) envVars(model string) []string { "ANTHROPIC_API_KEY=", "ANTHROPIC_AUTH_TOKEN=ollama", "CLAUDE_CODE_ATTRIBUTION_HEADER=0", + "CLAUDE_CODE_TOTAL_TOKENS_REMINDER=off", "DISABLE_ERROR_REPORTING=1", "DISABLE_FEEDBACK_COMMAND=1", "CLAUDE_CODE_DISABLE_FEEDBACK_SURVEY=1", diff --git a/cmd/launch/claude_test.go b/cmd/launch/claude_test.go index e27d8097b..31f0916b9 100644 --- a/cmd/launch/claude_test.go +++ b/cmd/launch/claude_test.go @@ -353,6 +353,7 @@ func TestClaudeEnvVars(t *testing.T) { "ANTHROPIC_API_KEY": "", "ANTHROPIC_AUTH_TOKEN": "ollama", "CLAUDE_CODE_ATTRIBUTION_HEADER": "0", + "CLAUDE_CODE_TOTAL_TOKENS_REMINDER": "off", "DISABLE_ERROR_REPORTING": "1", "DISABLE_FEEDBACK_COMMAND": "1", "CLAUDE_CODE_DISABLE_FEEDBACK_SURVEY": "1",