Split ConversationHistory into Conversation and ConversationStore

[shellicar]

What

Step 1a of the architecture refactor plan (.claude/plans/architecture-refactor.md).

Splits ConversationHistory into two files:

• Conversation.ts — pure in-memory data. Owns the business rules: role-alternation merge, compaction-triggered clear, id tagging/remove. No node:fs.
• ConversationStore.ts — file-backed wrapper. Reads the JSONL file on construction (applying trimToLastCompaction), delegates to Conversation for mutations, and persists after every push/remove. No-file construction works as a plain in-memory store.

AgentRun and AnthropicAgent swap ConversationHistory → ConversationStore. Public interface is identical so call sites are unchanged.

Tests

ConversationHistory.spec.ts → Conversation.spec.ts. Rewritten to project convention:

• One expect per it
• const expected = …; const actual = …; expect(actual).toBe(expected)
• New load() section covering the raw-initialization path that ConversationStore uses

21 tests, all green.

Why AgentRun gets ConversationStore (not bare Conversation)

AgentRun.push() calls must trigger disk saves. Decoupling AgentRun from the persistence layer entirely is a future step; passing it a ConversationStore achieves the structural separation (pure data vs I/O) without introducing a callback or hook mechanism here.

[bananabot9000]

Clean split. Conversation has zero I/O -- node:fs import gone, no #save() calls remain. ConversationStore wraps and delegates correctly: both push() and remove() call #save() after delegation.

Callers (AgentRun, AnthropicAgent) are straight type swaps -- 3+3 lines each, external API unchanged (historyFile option still works).

Tests rewritten to project convention (one expect per it, expected/actual). The load() section correctly verifies that raw initialization bypasses merge logic (two consecutive user messages stay separate after load), which is the key behavioral distinction between the two classes.

One minor note: ConversationStore.remove() calls #save() unconditionally, including when remove() returns false (no-op write). Harmless -- the file content is identical -- but could short-circuit with if (result) this.#save() if you wanted to avoid the disk write on miss. Not blocking.

Exported helpers (HistoryItem, hasCompactionBlock, trimToLastCompaction) are the minimum needed for ConversationStore to do its job. No leaky abstractions.

LGTM. Step 1a done. 🍌