refactor(chunks): append-only chunk log with per-step cache-stable wire

Replace the message-as-container model with a flat, append-only chunk log.

- chunks table (id, tab_id, seq, turn_id, step, role, type, data_json): one
  row per chunk; tool_call (assistant) and tool_result (tool) are SEPARATE
  rows linked by callId. Message/turn are derived groupings, not stored.
- chunks/transform.ts: DB-free explode (Chunk[] -> rows) / group (rows ->
  messages), shared by backend and the browser frontend.
- Cache fix: toModelMessages segments each turn at tool-batch boundaries into
  stable [assistant, tool] pairs per step, so earlier steps serialize
  byte-identically across requests (kills the prompt-cache churn).
- agent-manager persists a turn's chunks on seal (once), discarding a failed
  fallback attempt's partial chunks; rebuilds agent history from the log.
- GET /messages windows the log by chunk seq then groups; loadMoreMessages
  merges a turn split across the window boundary by turnId.
- One-shot migration drops the legacy messages table and clears tabs;
  settings/credentials/keys/usage preserved.

Full suite green (317 tests); biome, tsc, and svelte-check clean.

2 files changed, 288 insertions, 32 deletions

diff --git a/packages/core/tests/agent/agent.test.ts b/packages/core/tests/agent/agent.test.ts
index d6daac6..6c2b452 100644
--- a/packages/core/tests/agent/agent.test.ts
+++ b/packages/core/tests/agent/agent.test.ts
@@ -450,10 +450,11 @@ describe("Agent", () => {
 		expect(toolContent[0]).not.toHaveProperty("result");
 	});
 
-	it("Anthropic [tool-call, text] split: mixed-order assistant message gets split into [text]+[tool-call]", async () => {
-		// Pre-seed an assistant message with chunks in [tool-batch, text] order —
-		// which produces [tool-call, text] in the ModelMessage content, a shape
-		// Anthropic rejects. Only applies for anthropic / opencode-anthropic provider.
+	it("per-step segmentation: a [tool-batch, text] turn becomes [assistant(tool-call), tool(result), assistant(text)]", async () => {
+		// `toModelMessages` segments a turn at each tool-batch boundary, so the
+		// tool-batch (step 0) and the trailing text (step 1) land in SEPARATE
+		// assistant messages — never a single invalid [tool_use, text] block.
+		// This is the cache-stability fix and is applied for every provider.
 		const agent = new Agent(makeConfig({ provider: "opencode-anthropic" }));
 		agent.messages.push({
 			role: "user",
@@ -490,34 +491,31 @@ describe("Agent", () => {
 		const callArgs = vi.mocked(streamText).mock.calls.at(-1)?.[0];
 		const messages = callArgs?.messages as Array<{ role: string; content: unknown }>;
 
-		// After Anthropic structural normalisation, we should have TWO assistant messages:
-		// 1st: text-only content
-		// 2nd: tool-call-only content
+		// No assistant message may mix tool-call and non-tool-call parts (the
+		// invalid shape Anthropic rejects); segmentation guarantees this.
 		const assistantMsgs = messages.filter((m) => m.role === "assistant");
-		expect(assistantMsgs.length).toBeGreaterThanOrEqual(2);
-
-		// Find the text-only assistant message and the tool-call-only assistant message
-		const textOnlyMsg = assistantMsgs.find((m) => {
+		for (const m of assistantMsgs) {
 			const c = m.content as Array<Record<string, unknown>>;
-			return Array.isArray(c) && c.every((p) => p.type !== "tool-call");
-		});
-		const toolOnlyMsg = assistantMsgs.find((m) => {
+			if (!Array.isArray(c)) continue;
+			const hasToolCall = c.some((p) => p.type === "tool-call");
+			const hasNonToolCall = c.some((p) => p.type !== "tool-call");
+			expect(hasToolCall && hasNonToolCall).toBe(false);
+		}
+
+		// The seeded turn yields a tool-call assistant message immediately
+		// followed by its tool-result message (valid tool_use → tool_result).
+		const toolOnlyIdx = messages.findIndex((m) => {
 			const c = m.content as Array<Record<string, unknown>>;
-			return Array.isArray(c) && c.every((p) => p.type === "tool-call");
+			return m.role === "assistant" && Array.isArray(c) && c.some((p) => p.type === "tool-call");
 		});
-
-		// Narrow the optionals — toBeDefined() already verified non-null,
-		// but TypeScript needs the explicit assertion via local consts so
-		// we can pass them to indexOf without `!`.
-		if (!textOnlyMsg || !toolOnlyMsg) throw new Error("type guard");
-
-		// Text message comes first (before tool-call message) — Anthropic requires this ordering
-		expect(messages.indexOf(textOnlyMsg)).toBeLessThan(messages.indexOf(toolOnlyMsg));
+		expect(toolOnlyIdx).toBeGreaterThanOrEqual(0);
+		expect(messages[toolOnlyIdx + 1]?.role).toBe("tool");
 	});
 
-	it("Anthropic [tool-call, text] split: openai-compatible provider preserves original order (no split)", async () => {
-		// For non-Anthropic providers, the [tool-call, text] split should NOT be applied.
-		// (No provider set → defaults to openai-compatible)
+	it("per-step segmentation also applies to the openai-compatible provider", async () => {
+		// Segmentation is provider-agnostic: a [tool-batch, text] turn is split
+		// into separate assistant messages for openai-compatible too, with the
+		// tool result in its own tool message (the standard OpenAI shape).
 		const agent = new Agent(makeConfig());
 		agent.messages.push({
 			role: "user",
@@ -552,13 +550,24 @@ describe("Agent", () => {
 		const callArgs = vi.mocked(streamText).mock.calls.at(-1)?.[0];
 		const messages = callArgs?.messages as Array<{ role: string; content: unknown }>;
 
-		// For openai-compatible provider, only ONE assistant message with mixed content
+		// The seeded [tool-batch, text] turn is segmented: a tool-call-only
+		// assistant message, its tool message, and a separate text assistant
+		// message (the new turn's "ok" reply adds one more). No assistant
+		// message mixes tool-call and non-tool-call parts.
 		const assistantMsgs = messages.filter((m) => m.role === "assistant");
-		expect(assistantMsgs).toHaveLength(1);
-		const content = assistantMsgs[0]?.content as Array<Record<string, unknown>>;
-		// Both tool-call and text parts should be in the same message
-		expect(content.some((p) => p.type === "tool-call")).toBe(true);
-		expect(content.some((p) => p.type === "text")).toBe(true);
+		for (const m of assistantMsgs) {
+			const c = m.content as Array<Record<string, unknown>>;
+			if (!Array.isArray(c)) continue;
+			expect(c.some((p) => p.type === "tool-call") && c.some((p) => p.type !== "tool-call")).toBe(
+				false,
+			);
+		}
+		expect(messages.some((m) => m.role === "tool")).toBe(true);
+		const toolCallMsg = assistantMsgs.find((m) => {
+			const c = m.content as Array<Record<string, unknown>>;
+			return Array.isArray(c) && c.some((p) => p.type === "tool-call");
+		});
+		expect(toolCallMsg).toBeDefined();
 	});
 
 	it("empty-text-part filter (Anthropic): empty text chunk is not sent", async () => {
@@ -1250,6 +1259,74 @@ describe("Agent", () => {
 		expect(execCount).toBe(2);
 	});
 
+	// ─── Cache stability: per-step wire prefix is immutable ─────────────────────
+
+	it("keeps earlier steps' wire messages byte-identical across requests (cache prefix is stable)", async () => {
+		// A 3-step tool turn. The messages for steps 0 and 1 must serialize
+		// identically in the step-2 request and the step-3 request — that
+		// byte-stability is what lets Anthropic's rolling prompt cache extend
+		// instead of re-writing the whole prefix every step (cache-miss-report.md).
+		// Uses the openai-compatible provider so no cacheControl markers (which
+		// intentionally move each step) obscure the content comparison.
+		let n = 0;
+		// mock.calls accumulates across tests in this file — reset so our
+		// `calls.length` assertions count only this run's requests.
+		vi.mocked(streamText).mockClear();
+		const toolDef = {
+			name: "read_file",
+			description: "reads a file",
+			parameters: z.object({ path: z.string() }),
+			execute: async (args: Record<string, unknown>) => `contents of ${String(args.path)}`,
+		};
+		const toolStep = (id: string, path: string) =>
+			makeMockStreamResult([
+				{ type: "reasoning-delta", id: `r${id}`, text: `thinking ${id}` },
+				{ type: "text-delta", id: `t${id}`, text: `step ${id}` },
+				{ type: "tool-call", toolCallId: id, toolName: "read_file", input: { path } },
+				finishToolCalls,
+			]);
+		vi.mocked(streamText).mockImplementation(() => {
+			n++;
+			if (n === 1) return toolStep("s0", "a.txt");
+			if (n === 2) return toolStep("s1", "b.txt");
+			if (n === 3) return toolStep("s2", "c.txt");
+			return makeMockStreamResult([{ type: "text-delta", id: "tf", text: "done" }, finishStop]);
+		});
+
+		const agent = new Agent(makeConfig({ tools: [toolDef] }));
+		for await (const _ of agent.run("go")) {
+			/* consume */
+		}
+
+		// 4 streamText calls (steps 0..3). Compare the step-2 request (call idx 2)
+		// and step-3 request (call idx 3).
+		const calls = vi.mocked(streamText).mock.calls;
+		expect(calls.length).toBe(4);
+		const req2 = calls[2]?.[0]?.messages as unknown[];
+		const req3 = calls[3]?.[0]?.messages as unknown[];
+
+		// Step-2 request = [system, user, a(s0), tool(s0), a(s1), tool(s1)] (6).
+		// Step-3 request appends a(s2), tool(s2). The shared 6-message prefix
+		// must be byte-identical.
+		expect(req2).toHaveLength(6);
+		expect(req3).toHaveLength(8);
+		expect(JSON.stringify(req3.slice(0, 6))).toBe(JSON.stringify(req2));
+
+		// And each step really is its own [assistant, tool] pair (not one merged
+		// assistant message with all tool calls bunched together).
+		const roles = (req3 as Array<{ role: string }>).map((m) => m.role);
+		expect(roles).toEqual([
+			"system",
+			"user",
+			"assistant",
+			"tool",
+			"assistant",
+			"tool",
+			"assistant",
+			"tool",
+		]);
+	});
+
 	// ─── Usage / cache-rate telemetry ──────────────────────────────────────────
 
 	it("emits a usage event from the finish-step part with the cache read/write split", async () => {
diff --git a/packages/core/tests/db/chunks.test.ts b/packages/core/tests/db/chunks.test.ts
new file mode 100644
index 0000000..fe54628
--- /dev/null
+++ b/packages/core/tests/db/chunks.test.ts
@@ -0,0 +1,179 @@
+import { describe, expect, it } from "vitest";
+import { explodeTurn, explodeUserText, groupRowsToMessages } from "../../src/chunks/transform.js";
+import type { Chunk, ChunkRow, ChunkRowDraft } from "../../src/types/index.js";
+
+// These tests cover the pure explode/group transforms — the heart of the flat
+// chunk-log storage model. No DB is required.
+
+/** Promote drafts to rows with synthetic seq/id/createdAt (as appendChunks would). */
+function toRows(drafts: ChunkRowDraft[], tabId = "tab-1", startSeq = 0): ChunkRow[] {
+	return drafts.map((d, i) => ({
+		id: `c${i}`,
+		tabId,
+		seq: startSeq + i,
+		turnId: d.turnId,
+		step: d.step,
+		role: d.role,
+		type: d.type,
+		data: d.data,
+		createdAt: 1000 + i,
+	}));
+}
+
+describe("explodeTurn", () => {
+	it("splits a tool-batch into separate tool_call (assistant) and tool_result (tool) rows", () => {
+		const chunks: Chunk[] = [
+			{ type: "thinking", text: "hmm", metadata: { anthropic: { signature: "S" } } },
+			{ type: "text", text: "let me read" },
+			{
+				type: "tool-batch",
+				calls: [
+					{ id: "a1", name: "read_file", arguments: { path: "x" }, result: "X", isError: false },
+					{ id: "a2", name: "read_file", arguments: { path: "y" }, result: "Y", isError: false },
+				],
+			},
+		];
+		const drafts = explodeTurn("turn-1", chunks);
+
+		// thinking, text, tool_call×2 (assistant), tool_result×2 (tool)
+		expect(drafts.map((d) => `${d.role}/${d.type}`)).toEqual([
+			"assistant/thinking",
+			"assistant/text",
+			"assistant/tool_call",
+			"assistant/tool_call",
+			"tool/tool_result",
+			"tool/tool_result",
+		]);
+		// All in the same step (one round-trip).
+		expect(drafts.every((d) => d.step === 0)).toBe(true);
+		expect(drafts.every((d) => d.turnId === "turn-1")).toBe(true);
+	});
+
+	it("increments step after each tool-batch (multi-step turn)", () => {
+		const chunks: Chunk[] = [
+			{ type: "text", text: "s0" },
+			{ type: "tool-batch", calls: [{ id: "a", name: "t", arguments: {}, result: "r" }] },
+			{ type: "text", text: "s1" },
+			{ type: "tool-batch", calls: [{ id: "b", name: "t", arguments: {}, result: "r" }] },
+			{ type: "text", text: "final" },
+		];
+		const drafts = explodeTurn("turn-1", chunks);
+		const byStep = (s: number) => drafts.filter((d) => d.step === s).map((d) => d.type);
+		expect(byStep(0)).toEqual(["text", "tool_call", "tool_result"]);
+		expect(byStep(1)).toEqual(["text", "tool_call", "tool_result"]);
+		expect(byStep(2)).toEqual(["text"]); // trailing final-step text, no tool-batch
+	});
+
+	it("omits tool_result rows for calls without a result", () => {
+		const chunks: Chunk[] = [
+			{ type: "tool-batch", calls: [{ id: "a", name: "t", arguments: {} }] },
+		];
+		const drafts = explodeTurn("turn-1", chunks);
+		expect(drafts.map((d) => d.type)).toEqual(["tool_call"]);
+	});
+});
+
+describe("groupRowsToMessages (round-trip)", () => {
+	it("reconstructs a user message then an assistant message with a per-step tool-batch", () => {
+		const rows = [
+			...toRows(explodeUserText("turn-1", "hello"), "tab-1", 0),
+			...toRows(
+				explodeTurn("turn-1", [
+					{ type: "text", text: "reading" },
+					{
+						type: "tool-batch",
+						calls: [
+							{
+								id: "a1",
+								name: "read_file",
+								arguments: { path: "x" },
+								result: "X",
+								isError: false,
+							},
+						],
+					},
+					{ type: "text", text: "done" },
+				]),
+				"tab-1",
+				1,
+			),
+		];
+
+		const msgs = groupRowsToMessages(rows);
+		expect(msgs.map((m) => m.role)).toEqual(["user", "assistant"]);
+		expect(msgs[0]?.chunks).toEqual([{ type: "text", text: "hello" }]);
+
+		const a = msgs[1];
+		if (!a) throw new Error("no assistant message");
+		// reconstructed: text, tool-batch(step0), text(step1)
+		expect(a.chunks.map((c) => c.type)).toEqual(["text", "tool-batch", "text"]);
+		const batch = a.chunks.find((c) => c.type === "tool-batch");
+		if (batch?.type !== "tool-batch") throw new Error("no batch");
+		expect(batch.calls[0]).toMatchObject({
+			id: "a1",
+			name: "read_file",
+			arguments: { path: "x" },
+			result: "X",
+			isError: false,
+		});
+	});
+
+	it("keeps each step's tool calls in its own tool-batch chunk", () => {
+		const rows = toRows(
+			explodeTurn("turn-1", [
+				{ type: "tool-batch", calls: [{ id: "a", name: "t", arguments: {}, result: "ra" }] },
+				{ type: "tool-batch", calls: [{ id: "b", name: "t", arguments: {}, result: "rb" }] },
+			]),
+		);
+		const msgs = groupRowsToMessages(rows);
+		expect(msgs).toHaveLength(1);
+		const batches = msgs[0]?.chunks.filter((c) => c.type === "tool-batch") ?? [];
+		expect(batches).toHaveLength(2);
+	});
+
+	it("round-trips a multi-step assistant turn back to its original chunk shape", () => {
+		const original: Chunk[] = [
+			{ type: "thinking", text: "plan", metadata: { anthropic: { signature: "S" } } },
+			{ type: "text", text: "step0" },
+			{
+				type: "tool-batch",
+				calls: [
+					{ id: "a", name: "read_file", arguments: { path: "p" }, result: "R", isError: false },
+				],
+			},
+			{ type: "text", text: "final" },
+		];
+		const rows = toRows(explodeTurn("turn-1", original));
+		const msgs = groupRowsToMessages(rows);
+		expect(msgs).toHaveLength(1);
+		expect(msgs[0]?.chunks).toEqual(original);
+	});
+
+	it("tolerates an orphan tool_result whose tool_call was paged out", () => {
+		const rows = toRows([
+			{
+				turnId: "turn-1",
+				step: 0,
+				role: "tool",
+				type: "tool_result",
+				data: { callId: "z", name: "t", result: "R", isError: false },
+			},
+		]);
+		const msgs = groupRowsToMessages(rows);
+		expect(msgs).toHaveLength(1);
+		const batch = msgs[0]?.chunks[0];
+		if (batch?.type !== "tool-batch") throw new Error("no batch");
+		expect(batch.calls[0]).toMatchObject({ id: "z", result: "R" });
+	});
+
+	it("breaks the assistant grouping on a user or system row", () => {
+		const rows = [
+			...toRows(explodeUserText("t1", "q1"), "tab", 0),
+			...toRows(explodeTurn("t1", [{ type: "text", text: "a1" }]), "tab", 1),
+			...toRows(explodeUserText("t2", "q2"), "tab", 2),
+			...toRows(explodeTurn("t2", [{ type: "system", kind: "notice", text: "n" }]), "tab", 3),
+		];
+		const msgs = groupRowsToMessages(rows);
+		expect(msgs.map((m) => m.role)).toEqual(["user", "assistant", "user", "system"]);
+	});
+});