)

Exposes all Claude Code permission modes as [ACP Session
modes](agentclientprotocol/agent-client-protocol#67).
This allows ACP clients to build mode selectors and to be notified when
the mode changes as a result of `ExitPlanMode`.

See a demo in the [Zed
PR](zed-industries/zed#37632).

Closes agentclientprotocol#31

---------

Co-authored-by: Bennet Bo Fenner <bennetbo@gmx.de>
Co-authored-by: Richard Feldman <oss@rtfeldman.com>
Co-authored-by: Bennet <bennet@zed.dev>

## Summary

The plan text was accidentally dropped from the `ExitPlanMode` content
array during the switch to built-in Claude Code tools (#316, commit
`be618f5`). This meant ACP clients could not display the plan when
asking the user for approval — they only saw the title "Ready to code?"
with no plan body.

We noticed this in the Fabriqa app where our plan approval UI was
rendering an empty card.

## What changed

Restored the original `input.plan` → `content` mapping in
`toolInfoFromToolUse()` so the plan markdown is included in the
`tool_call` and `request_permission` events.

**Before (broken):**
```typescript
case "ExitPlanMode": {
  return {
    title: "Ready to code?",
    kind: "switch_mode",
    content: [],
  };
}
```

**After (fixed):**
```typescript
case "ExitPlanMode": {
  const planInput = toolUse.input as { plan?: string };
  return {
    title: "Ready to code?",
    kind: "switch_mode",
    content:
      planInput?.plan
        ? [{ type: "content" as const, content: { type: "text" as const, text: planInput.plan } }]
        : [],
  };
}
```

## Where does `input.plan` come from?

The Claude Agent SDK passes tool inputs to the `canUseTool` callback as
`Record<string, unknown>` (see `Options.canUseTool` in
`@anthropic-ai/claude-agent-sdk/sdk.d.ts`). When Claude calls
`ExitPlanMode`, the model includes the plan text in the input.

The formal `ExitPlanModeInput` type (in
`@anthropic-ai/claude-agent-sdk/sdk-tools.d.ts`) only declares
`allowedPrompts`, but it has a `[k: string]: unknown` index signature
that allows additional properties. The model consistently sends `plan`
(string) and `planFilePath` (string). SDK v0.2.76 explicitly added
`planFilePath` to the tool input for hooks and SDK consumers, confirming
these fields are intentional.

The original implementation (commit `0289dd9`, "Support plan mode and
all other permission modes" #40) relied on `input.plan` and worked
correctly until the refactor in #316 dropped it.

### SDK history confirms `input.plan` is reliable

The Claude Agent SDK itself had a regression where `ExitPlanMode` input
was empty
([anthropics/claude-code#12288](anthropics/claude-code#12288)):

| Version | Status |
|---------|--------|
| Claude Code 2.0.34 | `input.plan` populated |
| Claude Code 2.0.51 | Regression — `input: {}` empty |
| Agent SDK 0.1.54 | Fixed — `input.plan` restored |
| Agent SDK 0.2.76 (current) | Working — also added `planFilePath` |

The SDK team treated the empty input as a critical bug and fixed it
promptly. The current SDK (0.2.76) reliably sends `input.plan`, so this
adapter fix is safe to depend on.

## Test plan

- [x] Added test: plan text is included in content when `input.plan` is
provided
- [x] Added test: content is empty when `input.plan` is not provided
- [x] All existing tests pass

Fixes #450