feat: anthropic provider

[zvdy]

Context

Adds a native Anthropic provider for kubectl-ai using the official github.com/anthropics/anthropic-sdk-go SDK, instead of routing Claude through the OpenAI compatibility shim.

The compat shim hides Claude-exclusive features and has caused real bugs (tool_call history corruption in streaming). This adds a first-class anthropic:// provider.

Why native?

• Prompt caching — cache_control on system prompt and tool definitions; the system prompt is large and repeated every turn, caching cuts cost significantly
• Extended thinking — thinking content blocks for complex k8s debugging queries
• Reliable streaming — native SSE events instead of the compat translation layer
• Native error codes — overloaded_error, rate_limit_error map cleanly to retry logic
• Future-proofing — adopts the SDK early so new Claude features (e.g. interleaved thinking) are one import away

Testing

Trace: Prompt Caching

Both API turns show cache_creation_input_tokens: 0 and cache_read_input_tokens: 2254 — the system prompt + tool definitions (2254 tokens) are served from Anthropic's prompt cache on every turn.

Cache breakpoints are set via "cache_control": {"type": "ephemeral"} on the system prompt block and the last tool definition.

Turn 1 — input_tokens: 339, cache_read: 2254, cache_creation: 0

event: message_start
data: {
  "usage": {
    "input_tokens": 339,
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 2254,
    "output_tokens": 3
  }
}

Turn 2 — history added (input_tokens grew 339 → 696), cache still hits at 2254

event: message_start
data: {
  "usage": {
    "input_tokens": 696,
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 2254,
    "output_tokens": 1
  }
}

Trace: Tool Use (parallel, streamed)

Claude streamed two parallel tool_use blocks in a single response. Tool input JSON is assembled incrementally via input_json_delta events, then executed and returned as tool_result blocks.

event: content_block_start
data: {"index":1,"content_block":{"type":"tool_use","id":"toolu_01Co6xj94B74RC3zmsmi7eHs","name":"kubectl","input":{}}}

event: content_block_delta
data: {"index":1,"delta":{"type":"input_json_delta","partial_json":"{\"command\": \"kubectl get namespaces\", \"modifies_resource\": \"no\"}"}}

event: content_block_stop
data: {"index":1}

event: content_block_start
data: {"index":2,"content_block":{"type":"tool_use","id":"toolu_01WitVoRNAaJ39aUSaide8ui","name":"kubectl","input":{}}}

event: content_block_delta
data: {"index":2,"delta":{"type":"input_json_delta","partial_json":"{\"command\": \"kubectl get pods -n default\", \"modifies_resource\": \"no\"}"}}

event: content_block_stop
data: {"index":2}

event: message_delta
data: {"delta":{"stop_reason":"tool_use"}}

Tool execution:

action: tool-request
  name: kubectl
  arguments:
    command: kubectl get namespaces

action: tool-response
  response:
    stdout: |
      NAME                 STATUS   AGE
      app                  Active   18d
      default              Active   18d
      kube-node-lease      Active   18d
      kube-public          Active   18d
      kube-system          Active   18d
      local-path-storage   Active   18d
      pgbouncer-reader     Active   16d
      pgbouncer-writer     Active   16d
      postgres             Active   16d

action: tool-request
  name: kubectl
  arguments:
    command: kubectl get pods -n default

action: tool-response
  response:
    stderr: No resources found in default namespace.

Trace: Native SSE Streaming

The final answer is delivered as a stream of text_delta events — each chunk yielded to the UI incrementally, no buffering until completion.

event: content_block_start
data: {"index":0,"content_block":{"type":"text","text":""}}

event: content_block_delta
data: {"index":0,"delta":{"type":"text_delta","text":"## 📋 All Namespaces in your cluster:"}}

event: content_block_delta
data: {"index":0,"delta":{"type":"text_delta","text":"\n\n| Namespace | Status | Age |\n|-----------|--------|-----|"}}

event: content_block_delta
data: {"index":0,"delta":{"type":"text_delta","text":"\n| app | Active | 18d |\n| default | Active | 18d |"}}

event: content_block_delta
data: {"index":0,"delta":{"type":"text_delta","text":"\n| kube-system | Active | 18d |\n| postgres | Active | 16d |"}}

event: content_block_delta
data: {"index":0,"delta":{"type":"text_delta","text":"\n\n## 🔍 Pods in the default namespace:\n\n**No resources found in default namespace.** ✨"}}

event: content_block_stop
data: {"index":0}

event: message_delta
data: {"delta":{"stop_reason":"end_turn"},"usage":{"output_tokens":246}}

event: message_stop
data: {"type":"message_stop"}

Trace: Multi-Turn History

input_tokens grew from 339 (turn 1) to 696 (turn 2), reflecting the full conversation history appended to each request. cache_read_input_tokens stays at 2254 — the system prompt cache is unaffected by history growth.

Messages array sent in turn 2:

[
  {
    "role": "user",
    "content": [{"type": "text", "text": "list all namespaces\nwhat pods are running in the default namespace?"}]
  },
  {
    "role": "assistant",
    "content": [
      {"type": "text", "text": "I'll help you with those requests! Let me fetch the information from your Kubernetes cluster."},
      {"type": "tool_use", "id": "toolu_01Co6xj94B74RC3zmsmi7eHs", "name": "kubectl", "input": {"command": "kubectl get namespaces"}},
      {"type": "tool_use", "id": "toolu_01WitVoRNAaJ39aUSaide8ui", "name": "kubectl", "input": {"command": "kubectl get pods -n default"}}
    ]
  },
  {
    "role": "user",
    "content": [
      {"type": "tool_result", "tool_use_id": "toolu_01Co6xj94B74RC3zmsmi7eHs", "content": "...kubectl get namespaces output..."},
      {"type": "tool_result", "tool_use_id": "toolu_01WitVoRNAaJ39aUSaide8ui", "content": "...No resources found..."}
    ]
  }
]

@droot

[droot]

Love it!! Looking forward to it. Will be happy to review it :)

[zvdy]

I'm ok with current state, will iterate over more features like temperature and things like that, I' think most of the users don't tweak those so are minor future improvements

ptal @droot

[droot]

Thank you!