Split Conversation storage from API view

[shellicar]

Why

Conversation.push() used to clear the entire item array when it saw a compaction block, and ConversationStore did the same on disk load via trimToLastCompaction. That was destructive on two axes:

1. No rollback. Once compaction fired, pre-compaction messages were gone. If the compaction itself errored, the session died with it.
2. No visibility. The CLI, audit log, and any future TUI replay could never see history older than the last compaction.

RequestBuilder also carried a stack of defensive spread-copies ([...messages], { ...block, cache_control }) whose only job was to keep send-time cache_control and system-reminder blocks from leaking back into stored state, because the same array served both purposes.

What

Storage and API view are now separate concerns.

• Conversation stores the full message history forever. Compaction messages are appended like any other message; no clearing.
• New Conversation.cloneForRequest() returns a deep clone (structuredClone) of the slice from the last compaction forward. The returned array is caller-owned and may be mutated freely.
• RequestBuilder's helpers (addCacheControlToLastBlock, cacheLastUserMessage, systemReminder injection) become void functions that mutate in place. The defensive spread-copy dance is gone.
• AgentRun swaps this.#history.messages for this.#history.cloneForRequest() at the send site.
• ConversationStore adds a thin cloneForRequest() passthrough and stops trimming on load.

Payoffs

• Compaction rollback now works for free: remove(id) the compaction block and the next cloneForRequest() slices further back.
• Full history is visible to the CLI, audit log, and future TUI replay.
• On-disk format stays pristine (no cache_control leakage).

Scope

This is PR 1 of a larger SDK reshape. Deferred to a later PR: extracting AnthropicClient from AnthropicAgent, the configure(partial) / run() split, removing ConversationStore / historyFile from the SDK so persistence lives fully in the CLI, and collapsing runAgent.ts. Full design notes in .claude/sessions/2026-04-09.md under "Session 2026-04-09 (SDK refactor planning)".

Tests

• 4 old compaction-clear tests rewritten to assert retention.
• 6 new cloneForRequest tests cover the empty case, no-compaction case, post-compaction slice, deep-clone guarantee (mutating the clone leaves the source untouched at array, message, and content-block levels).
• 1 does not mutate input test on RequestBuilder removed — it encoded the old contract.
• 78/78 claude-sdk tests green.

Related: #232

[bananabot9000]

Exactly the right refactor. Storage and API view separated cleanly. The spread-copy dance is dead, structuredClone handles deep isolation, cloneForRequest() is the single point where "what the API sees" diverges from "what we store."

Tests are thorough — deep-clone proven at array, message, and content-block levels. Compaction-clear tests correctly rewritten as retention tests.

One minor observation: trimToLastCompaction does an O(n) reverse scan which is fine. The structuredClone only runs on the post-compaction slice, so even with large pre-compaction histories the clone cost stays bounded by the active conversation size.

Clean, narrow, well-scoped. The session log's design discussion is excellent context for PR 2.

🍌 Approved.