# dispatch-adapter-claude

A [Dispatch](https://github.com/realtradam/dispatch-adapter-interface) adapter
that connects to Anthropic's Claude API using a personal **Pro / Max
subscription** via the same OAuth flow that the Claude Code CLI uses.

---

## ⚠ Status / disclaimer

This gem impersonates the Claude Code CLI HTTP signature
(`User-Agent: claude-cli/2.1.63 (external, cli)` and the matching Stainless
header set). That is the mechanism Anthropic uses to route Pro / Max traffic
through the Claude Code entitlement.

Use this under your **own** Claude Pro / Max subscription and at your own risk
with respect to Anthropic's Terms of Service. It is your responsibility to
ensure your usage complies with those terms.

---

## Installation

Add to your `Gemfile`:

```ruby
gem "dispatch-adapter-claude"
```

Or install directly:

```bash
gem install dispatch-adapter-claude
```

The gem requires Ruby ≥ 3.2 and depends on
`dispatch-adapter-interface ~> 0.2`.

---

## Quick start

```ruby
require "dispatch/adapter/claude"

# Build the adapter (defaults to claude-sonnet-4-5-20250929).
claude = Dispatch::Adapter::Claude.new(
  model: "claude-sonnet-4-5-20250929"
)

# First run: opens a browser for the OAuth PKCE flow and caches the token
# at ~/.config/dispatch/claude_oauth.json (mode 0600).
# Subsequent calls: validates / auto-refreshes the stored token.
claude.authenticate!

# Send a message.
msgs = [
  Dispatch::Adapter::Message.new(
    role: "user",
    content: [Dispatch::Adapter::TextBlock.new(text: "Say hi")]
  )
]
resp = claude.chat(msgs)

puts resp.content                              # => "Hi there! ..."
puts resp.stop_reason                          # => :end_turn
puts resp.usage.cost.total                     # USD-equivalent (computed)

# Check subscription quota (OAuth only).
report = claude.usage_report
entry  = report.limits.find { |e| e.id == "anthropic:5h" }
puts entry.amount.used_fraction                # 0.124  (12.4% of 5-hour window)
puts entry.window.resets_at                    # 2026-04-28 19:00:00 UTC
```

---

## Pricing semantics

Claude Pro / Max returns **no dollar-cost line item** per request — it is a
flat-rate plan. The `usage.cost` field on every `Response` is a
**locally-computed USD-equivalent** derived from a bundled price table that
mirrors what API customers would pay:

```
cost.input       = (price_per_mtok.input  / 1_000_000) × input_tokens
cost.output      = (price_per_mtok.output / 1_000_000) × output_tokens
cost.cache_read  = (price_per_mtok.cache_read  / 1_000_000) × cache_read_tokens
cost.cache_write = (price_per_mtok.cache_write / 1_000_000) × cache_creation_tokens
cost.total       = sum of the above
```

The only **authoritative** consumption signal for Pro / Max is
`usage_report`, which reports what fraction of each rolling window
(5-hour, 7-day, 7-day Opus, 7-day Sonnet) has been used.

---

## Configuration

### Constructor keyword arguments

| Argument | Type | Default | Description |
|---|---|---|---|
| `model` | `String` | `"claude-sonnet-4-5-20250929"` | Anthropic model ID |
| `api_key` | `String, nil` | `nil` | Raw `sk-ant-api…` key; bypasses OAuth when set |
| `token_path` | `String, nil` | `nil` | Override path for the OAuth token file |
| `base_url` | `String` | `"https://api.anthropic.com"` | API base URL |
| `max_tokens` | `Integer, nil` | `nil` | Instance-level default for `max_tokens` |
| `thinking` | `String, Hash, nil` | `nil` | Instance-level thinking config (see below) |
| `cache_retention` | `Symbol, nil` | `nil` | Default cache TTL: `:short` (5 min), `:long` (1 h), `:none` |
| `min_request_interval` | `Float` | `1.0` | Minimum seconds between outbound requests |
| `extra_betas` | `Array<String>` | `[]` | Additional `Anthropic-Beta` header values |
| `is_oauth` | `Boolean, nil` | `nil` | Override OAuth auto-detection |
| `token_store` | `TokenStore, nil` | `nil` | Inject a custom credential store (testing) |

### Environment variables

No required environment variables. The adapter loads credentials from the
token store at `~/.config/dispatch/claude_oauth.json` by default. Override
with `token_path:`.

### Thinking / extended-output

Pass `thinking:` to the constructor (instance default) or to `chat` (per-call):

```ruby
# String shorthand (maps to budget_tokens heuristics internally)
claude = Dispatch::Adapter::Claude.new(thinking: "high")

# Hash — full control
claude = Dispatch::Adapter::Claude.new(
  thinking: { type: :adaptive, display: :summarized }
)

# Per-call override
resp = claude.chat(msgs, thinking: { type: :enabled, budget_tokens: 4096 })
```

For **adaptive thinking** (Opus 4.7+), use `{ type: :adaptive }`.  
For models that don't support thinking, the parameter is silently ignored.

### Prompt caching

Pass `cache_retention:` to enable Anthropic prompt caching:

```ruby
# Short-lived (5 min TTL) — works everywhere including API keys
resp = claude.chat(msgs, cache_retention: :short)

# Long-lived (1 h TTL) — api.anthropic.com only
resp = claude.chat(msgs, cache_retention: :long)

# Disable caching even if the instance default is set
resp = claude.chat(msgs, cache_retention: :none)
```

The adapter automatically places up to 4 cache breakpoints in the order
that Anthropic requires: last tool → last system block → penultimate user
message → last user message. The 4-breakpoint cap and TTL ordering rule
(a 5-min block may not follow a 1-h block) are enforced automatically.

### Tool use

```ruby
add_tool = Dispatch::Adapter::ToolDefinition.new(
  name: "add",
  description: "Return the sum of two integers",
  parameters: {
    type: "object",
    properties: {
      a: { type: "integer" },
      b: { type: "integer" }
    },
    required: %w[a b]
  }
)

resp = claude.chat(msgs, tools: [add_tool])
if resp.stop_reason == :tool_use
  tc = resp.tool_calls.first
  puts "#{tc.name}(#{tc.arguments})"   # => "add({"a"=>2, "b"=>3})"
end
```

When using OAuth (Pro / Max), tool names are automatically prefixed with
`proxy_` on the wire and stripped from the response. Built-in tool names
(`web_search`, `code_execution`, `text_editor`, `computer`) are passed
through unchanged.

### Streaming

```ruby
full_resp = claude.chat(msgs, stream: true) do |delta|
  case delta.type
  when :text_delta     then print delta.text
  when :thinking_delta then print "[thinking] #{delta.text}"
  when :tool_use_start then puts "\n[tool] #{delta.tool_name}"
  end
end

puts full_resp.usage.cost.total
```

---

## Auth lifecycle

```ruby
claude = Dispatch::Adapter::Claude.new

# Interactive OAuth PKCE login (opens browser on first call)
result = claude.authenticate!
# => :logged_in | :cached | :refreshed | :api_key

# Check whether credentials are present
claude.authenticated?    # => true / false

# Remove stored OAuth credentials
claude.logout!
```

Tokens are stored at `~/.config/dispatch/claude_oauth.json` (file mode 0600)
and refreshed automatically 5 minutes before expiry.

---

## Limitations

| Area | Constraint |
|---|---|
| Image modality | JPG, PNG, WEBP, GIF only; no audio or video |
| Tool-name length | Anthropic enforces a maximum on compiled grammar size; a 400 "compiled grammar too large" triggers an automatic retry with `strict: false` |
| Cache breakpoints | At most 4 per request; the adapter enforces the cap and TTL ordering automatically |
| OAuth callback port | Port 54545 is hard-coded by Anthropic's redirect URI; the port must be free during the initial login |
| Thinking on forced tool_choice | When `tool_choice: :any` or `tool_choice: { type: :tool, … }`, thinking and `output_config` are removed (the API rejects them otherwise) |
| Opus 4.7+ sampling params | `top_p` / `top_k` are silently dropped for Opus 4.7+ models |
| `usage_report` | Requires OAuth (Pro / Max); returns `nil` with a raw API key |

---

## Tracking upstream changes

The two constants most likely to drift when Anthropic releases a new
version of Claude Code are:

| Constant | Location | Current value |
|---|---|---|
| `CLAUDE_CODE_VERSION` | `lib/dispatch/adapter/claude/headers.rb` | `"2.1.63"` |
| `STAINLESS_PACKAGE_VERSION` | `lib/dispatch/adapter/claude/headers.rb` | `"0.74.0"` |

When the real Claude Code CLI updates, bump these to match. Anthropic also
rotates the `Anthropic-Beta` header set a few times per year; `DEFAULT_BETAS`
in the same file lists the current set.

Background research, including the original reverse-engineering notes and the
full gap analysis against `dispatch-adapter-interface`, is in
[`.rules/research/research.md`](./.rules/research/research.md).