[BUG] Asking broad questions result in prompt too long.

[marcofranssen]

📋 Prerequisites

• I have searched the existing issues to avoid creating a duplicate
• By submitting this issue, you agree to follow our Code of Conduct
• I am using the latest version of the software
• I have tried to clear cache/cookies or used incognito mode (if ui-related)
• I can consistently reproduce this issue

🎯 Affected Service(s)

Not Sure

🚦 Impact/Severity

Minor inconvenience

🐛 Bug Description

Running with Claude Sonnet 4.6 on Bedrock I'm getting following quite easily.

🔄 Steps To Reproduce

E.g.:

1. Ask the k8s-agent to make suggestions on improving the deployments in a namespace.
2. after calling couple of tools it fails with following

Any broad question will make it fail…

🤔 Expected Behavior

Proper compaction happens to ensure the tool outputs and such are compacted to make proper requests to the LLM.

📱 Actual Behavior

LLM error: API_ERROR operation error Bedrock Runtime: Converse, https response error StatusCode: 400, RequestID: 59c80fdf-4b8e-4d50-ac1b-d3b6188c8d58, ValidationException: The model returned the following errors: prompt is too long: 1580999 tokens > 1000000 maximum```

### 💻 Environment

Installed 0.9.4 helm chart.

### 🔧 CLI Bug Report

_No response_

### 🔍 Additional Context

Sonnet 4.6 Bedrock

### 📋 Logs

```shell
{"level":"info","timestamp":"2026-05-15T18:49:39.659Z","caller":"cmd/main.go:53","msg":"Logger initialized","level":"info"}
{"level":"info","timestamp":"2026-05-15T18:49:39.665Z","caller":"cmd/main.go:89","msg":"Loaded agent config","configDir":"/config"}
{"level":"info","timestamp":"2026-05-15T18:49:39.665Z","caller":"cmd/main.go:90","msg":"Agent configuration","model":"bedrock","stream":false,"httpTools":1,"sseTools":0,"remoteAgents":0}
{"level":"info","timestamp":"2026-05-15T18:49:39.665Z","caller":"cmd/main.go:234","msg":"Built app_name from environment variables","KAGENT_NAMESPACE":"kagent","KAGENT_NAME":"k8s-agent","app_name":"kagent__NS__k8s_agent"}
{"level":"info","timestamp":"2026-05-15T18:49:39.680Z","caller":"cmd/main.go:125","msg":"ADK telemetry initialized"}
{"level":"info","timestamp":"2026-05-15T18:49:39.680Z","caller":"cmd/main.go:140","msg":"Token service started"}
{"level":"info","timestamp":"2026-05-15T18:49:39.680Z","caller":"cmd/main.go:151","msg":"Using KAgent session service","url":"http://kagent-controller.kagent:8083"}
{"level":"info","timestamp":"2026-05-15T18:49:39.680Z","caller":"cmd/main.go:173","msg":"Memory service enabled","appName":"kagent__NS__k8s_agent"}
{"level":"info","timestamp":"2026-05-15T18:49:39.681Z","caller":"mcp/registry.go:86","msg":"Processing HTTP MCP tools","httpToolsCount":1}
{"level":"info","timestamp":"2026-05-15T18:49:39.681Z","caller":"mcp/registry.go:143","msg":"Adding HTTP MCP tool","index":1,"url":"http://kagent-tools.kagent:8084/mcp","toolFilterCount":18,"tools":["k8s_check_service_connectivity","k8s_patch_resource","k8s_remove_annotation","k8s_annotate_resource","k8s_remove_label","k8s_label_resource","k8s_create_resource","k8s_create_resource_from_url","k8s_get_events","k8s_get_available_api_resources","k8s_get_cluster_configuration","k8s_describe_resource","k8s_delete_resource","k8s_get_resource_yaml","k8s_execute_command","k8s_apply_manifest","k8s_get_resources","k8s_get_pod_logs"]}
{"level":"info","timestamp":"2026-05-15T18:49:39.681Z","caller":"mcp/registry.go:153","msg":"Successfully added HTTP MCP toolset","url":"http://kagent-tools.kagent:8084/mcp"}
{"level":"info","timestamp":"2026-05-15T18:49:39.681Z","caller":"mcp/registry.go:107","msg":"Processing SSE MCP tools","sseToolsCount":0}
{"level":"info","timestamp":"2026-05-15T18:49:39.681Z","caller":"agent/agent.go:154","msg":"Memory configuration detected, adding memory tools"}
{"level":"info","timestamp":"2026-05-15T18:49:39.716Z","caller":"models/bedrock.go:128","msg":"Initialized Bedrock Converse API model","model":"us.anthropic.claude-sonnet-4-6","region":"us-east-1"}
{"level":"info","timestamp":"2026-05-15T18:49:39.718Z","caller":"agent/agent.go:132","msg":"Creating Google ADK LLM agent","name":"k8s_agent","hasDescription":true,"hasInstruction":true,"toolsCount":4,"toolsetsCount":1}
{"level":"info","timestamp":"2026-05-15T18:49:39.718Z","caller":"agent/agent.go:144","msg":"Successfully created Google ADK LLM agent","toolsCount":4,"toolsetsCount":1}
{"level":"info","timestamp":"2026-05-15T18:49:39.719Z","caller":"app/app.go:114","msg":"Using KAgent session service","url":"http://kagent-controller.kagent:8083"}
{"level":"info","timestamp":"2026-05-15T18:49:39.719Z","caller":"app/app.go:118","msg":"Using KAgent task store","url":"http://kagent-controller.kagent:8083"}
{"level":"info","timestamp":"2026-05-15T18:49:39.722Z","caller":"server/server.go:78","msg":"Starting Go ADK server!","addr":"0.0.0.0:8080"}
{"level":"info","timestamp":"2026-05-15T18:50:11.849Z","logger":"kagent-executor","caller":"a2a/executor.go:123","msg":"Execute","taskID":"019e2cf9-8a09-74ff-bf8b-663e0be2043d","contextID":"ctx-0dd7dd48-f281-4395-9dbc-04138b15cbeb","appName":"kagent__NS__k8s_agent","userID":"372463714525255869"}
{"level":"info","timestamp":"2026-05-15T18:50:20.092Z","caller":"agent/agent.go:362","msg":"Tool execution started","tool":"load_memory","functionCallID":"tooluse_qyHunzcRvwjoEXXvwhpaJs","sessionID":"ctx-0dd7dd48-f281-4395-9dbc-04138b15cbeb","invocationID":"e-f59e5329-3210-47f0-9749-711bf18dcbb0","args":"{\"query\":\"cluster configuration preferences kubernetes best practices\"}"}
{"level":"info","timestamp":"2026-05-15T18:50:20.294Z","caller":"agent/agent.go:384","msg":"Tool execution completed","tool":"load_memory","functionCallID":"tooluse_qyHunzcRvwjoEXXvwhpaJs","sessionID":"ctx-0dd7dd48-f281-4395-9dbc-04138b15cbeb","invocationID":"e-f59e5329-3210-47f0-9749-711bf18dcbb0","resultKeys":["memories"]}
{"level":"info","timestamp":"2026-05-15T18:50:29.753Z","caller":"agent/agent.go:362","msg":"Tool execution started","tool":"k8s_get_resources","functionCallID":"tooluse_4iKxf4190ia92ErFp5D7ls","sessionID":"ctx-0dd7dd48-f281-4395-9dbc-04138b15cbeb","invocationID":"e-f59e5329-3210-47f0-9749-711bf18dcbb0","args":"{\"all_namespaces\":\"true\",\"output\":\"json\",\"resource_type\":\"daemonsets\"}"}
{"level":"info","timestamp":"2026-05-15T18:50:29.754Z","caller":"agent/agent.go:362","msg":"Tool execution started","tool":"k8s_get_resources","functionCallID":"tooluse_2Dgujd9VMLkxNsKk0Jf1MQ","sessionID":"ctx-0dd7dd48-f281-4395-9dbc-04138b15cbeb","invocationID":"e-f59e5329-3210-47f0-9749-711bf18dcbb0","args":"{\"all_namespaces\":\"true\",\"output\":\"json\",\"resource_type\":\"pods\"}"}
{"level":"info","timestamp":"2026-05-15T18:50:29.754Z","caller":"agent/agent.go:362","msg":"Tool execution started","tool":"k8s_get_resources","functionCallID":"tooluse_bJMM5ugkE282XsjiO8sXYJ","sessionID":"ctx-0dd7dd48-f281-4395-9dbc-04138b15cbeb","invocationID":"e-f59e5329-3210-47f0-9749-711bf18dcbb0","args":"{\"all_namespaces\":\"true\",\"output\":\"json\",\"resource_type\":\"deployments\"}"}
{"level":"info","timestamp":"2026-05-15T18:50:30.202Z","caller":"agent/agent.go:384","msg":"Tool execution completed","tool":"k8s_get_resources","functionCallID":"tooluse_4iKxf4190ia92ErFp5D7ls","sessionID":"ctx-0dd7dd48-f281-4395-9dbc-04138b15cbeb","invocationID":"e-f59e5329-3210-47f0-9749-711bf18dcbb0","resultKeys":["output"]}
{"level":"info","timestamp":"2026-05-15T18:50:30.673Z","caller":"agent/agent.go:384","msg":"Tool execution completed","tool":"k8s_get_resources","functionCallID":"tooluse_bJMM5ugkE282XsjiO8sXYJ","sessionID":"ctx-0dd7dd48-f281-4395-9dbc-04138b15cbeb","invocationID":"e-f59e5329-3210-47f0-9749-711bf18dcbb0","resultKeys":["output"]}
{"level":"info","timestamp":"2026-05-15T18:50:31.232Z","caller":"agent/agent.go:384","msg":"Tool execution completed","tool":"k8s_get_resources","functionCallID":"tooluse_2Dgujd9VMLkxNsKk0Jf1MQ","sessionID":"ctx-0dd7dd48-f281-4395-9dbc-04138b15cbeb","invocationID":"e-f59e5329-3210-47f0-9749-711bf18dcbb0","resultKeys":["output"]}
2026/05/15 18:50:32 ERROR failed to write an event error="write tcp 10.6.9.53:8080->10.6.9.61:56982: write: broken pipe"
2026/05/15 18:50:34 traces export: rpc error: code = ResourceExhausted desc = grpc: received message larger than max (9950241 vs. 4194304)
{"level":"info","timestamp":"2026-05-15T18:57:05.652Z","logger":"kagent-executor","caller":"a2a/executor.go:123","msg":"Execute","taskID":"019e2cff-da74-7cd8-8dd8-f54cf68d7319","contextID":"ctx-0dd7dd48-f281-4395-9dbc-04138b15cbeb","appName":"kagent__NS__k8s_agent","userID":"372463714525255869"}

📷 Screenshots

🙋 Are you willing to contribute?

• I am willing to submit a PR to fix this issue

[JBunga]

For this class of failure, it may help to split the fix into two separate gates:

1. preflight accounting before the provider call: estimate input + tool payload + requested output budget against the real Bedrock model window
2. a payload-reduction fallback before retry: reduce the assembled context first, then call the provider once with the smaller prompt

I work on Adola/Rose 1, which is a pre-model compression step for long agent/RAG context. Plain text in, smaller plain text out, with token-savings receipts. There is a no-signup demo here if you want an external baseline for the k8s-agent case: https://adola.app/try?sample=agent

Not suggesting this replaces kagent-side compaction/accounting. It is mainly useful to test whether the final payload can be brought under the provider limit without changing the rest of the agent loop.