Allow tools/skills to programmatically clear context and inject a continuation prompt

Context

When working on long, multi-step tasks, the context window fills up and performance degrades. The natural response is to clear and continue, but this destroys all accumulated context — decisions made, files identified, progress tracked, corrections given.

Problem / Motivation

There is no programmatic way for a skill or tool to:

1. Summarize the current session context
2. Clear the conversation
3. Inject a continuation prompt into the fresh session

This means the human must manually orchestrate the "save context → clear → restore context" bridge every time. It's error-prone, breaks flow, and is exactly the kind of repetitive task that should be automatable.

Real-world use case: A /cls skill that the user invokes when context is getting heavy. The skill gathers git state, conversation decisions, active tasks, and the user's gist of what's next — then clears the conversation and injects the summary so Claude picks up seamlessly. Today this is impossible because:

• Skills cannot invoke /clear
• There is no PostClear hook event
• SessionStart hooks cannot inject prompt content
• Hooks generally cannot submit text into the conversation

The workaround is for the skill to save a summary to a file, then the user manually runs /clear, then manually types "read the file and continue." This defeats the purpose of having a skill.

Related: #32861

Proposed Solution

Enable one or more of these mechanisms (in order of preference):

Option A: Tool-invocable clear with continuation

A tool like ClearContext that a skill can call programmatically:

ClearContext({ continuation_prompt: "I'm continuing from a previous session..." })

This clears the conversation and injects the continuation as the opening message.

Option B: PostClear hook with prompt injection

Add a PostClear hook event that fires after /clear. Allow hooks of type "prompt" to inject text into the fresh session:

{
  "hooks": {
    "PostClear": [{
      "hooks": [{ "type": "prompt", "promptTemplate": "Read {{file}} and continue" }]
    }]
  }
}

Option C: /clear --run <prompt-or-file>

A built-in flag that clears and then submits a prompt or reads a file:

/clear --run "read local/cls-prompt.md and continue"

This is the simplest to implement but limits automation to what the human types.

Any of these would unblock the use case. Option A is the most powerful (skills can fully orchestrate the flow). Option C is the simplest to ship.

Acceptance Criteria

• [ ] A skill or tool can programmatically trigger a context clear
• [ ] A continuation prompt can be injected into the fresh session without user action
• [ ] The user initiates the flow (not automatic) — this is a human-in-the-loop bridge action
• [ ] Works in both CLI and IDE integrations