perf: cut startup cost across natives, providers, plugins and MCP (W1-W6)

.gitignore

@@ -58,6 +58,7 @@ packages/ai/test/.temp-images/
 .gjc/rlm/
 .gjc/state/
 .gjc/_session-*/
+.gjc/rss-checkpoints/
 .gjc/metrics.json
 .pi_config/
 .opencode/

biome.json

@@ -32,6 +32,25 @@
 			}
 		}
 	},
+	"overrides": [
+		{
+			"includes": ["packages/coding-agent/src/**"],
+			"linter": {
+				"rules": {
+					"style": {
+						"noRestrictedImports": {
+							"level": "error",
+							"options": {
+								"paths": {
+									"@gajae-code/ai": "Import from @gajae-code/ai/core instead."
+								}
+							}
+						}
+					}
+				}
+			}
+		}
+	],
 	"formatter": {
 		"enabled": true,
 		"indentStyle": "tab",

packages/agent/CHANGELOG.md

@@ -11,6 +11,9 @@
 ### Fixed
 
 - An aborted run whose tool ignores its `AbortSignal` now terminates on its own (#3894). `Promise.allSettled` waited on the unresolved call forever, so the turn only ended when the session's force-abort budget expired; the loop now emits a synthetic aborted result for the outstanding calls and `waitForIdle` settles immediately. Session dispose consequently reaches idle through the cooperative path instead of force-invalidating the run.
+### Changed
+
+- Telemetry configured with `spans: false` now skips span and attribute construction while preserving usage and cost hooks.
 
 ## [0.12.12] - 2026-08-05
 

packages/agent/src/agent-loop.ts

@@ -1521,11 +1521,18 @@ async function runLoopBody(
 					awaitEventDrain: (invocationSignal: AbortSignal) =>
 						stream.waitForConsumerDrain(AbortSignal.any([loopSignal, invocationSignal])),
 				};
-				const maintenanceOutcome = await config.maintainContext(currentContext, lifecycle);
+				const maintenanceResult = await config.maintainContext(currentContext, lifecycle);
+				const maintenance =
+					typeof maintenanceResult === "string" ? { outcome: maintenanceResult } : maintenanceResult;
 				// A callback can settle after its loop has been cancelled. Never let a
 				// stale "not-needed" fall through to streamAssistantResponse, which
 				// invokes the provider before it observes the aborted signal.
-				const outcome = loopSignal.aborted ? "aborted" : maintenanceOutcome;
+				const outcome = loopSignal.aborted ? "aborted" : maintenance.outcome;
+				if (maintenance.releaseCurrentContext) {
+					currentContext.messages.length = 0;
+					newMessages.length = 0;
+					convertedContextCache.delete(config);
+				}
 
 				if (outcome !== "not-needed") {
 					publishAgentEnd(
@@ -2440,6 +2447,7 @@ async function streamAssistantResponse(
 		});
 	} catch (err) {
 		failChatSpan(telemetry, chatSpan, {
+			stepNumber: chatStepNumber,
 			errorObject: err,
 			responseHeaders: capturedHeaders,
 			baseUrl: config.model.baseUrl,

packages/agent/src/agent.ts

@@ -980,7 +980,14 @@ export class Agent {
 		this.#contextRevision++;
 	}
 
-	replaceMessages(ms: AgentMessage[]) {
+	replaceMessages(
+		ms: AgentMessage[],
+		options?: { historyRewrite?: { reason: string; preserveSeededPrefix?: boolean } },
+	) {
+		const rewrite = options?.historyRewrite;
+		if (rewrite && this.#appendOnlyContext) {
+			this.#appendOnlyContext.releaseAfterHistoryRewrite({ preserveSeededPrefix: rewrite.preserveSeededPrefix });
+		}
 		this.#state.messages = ms.slice();
 		this.#contextRevision++;
 	}

packages/agent/src/append-only-context.ts

@@ -297,7 +297,7 @@ export class AppendOnlyContextManager {
 
 		const newMsgs = messagesToSync.slice(this.#lastSyncCount);
 		for (const msg of newMsgs) {
-			this.log.append(msg);
+			this.log.append(cloneJson(msg));
 		}
 
 		this.#lastSyncCount = messagesToSync.length;
@@ -341,6 +341,17 @@ export class AppendOnlyContextManager {
 		this.log.replaceTail(message);
 	}
 
+	/** Release provider-normalized retainers as one history-rewrite transaction. */
+	releaseAfterHistoryRewrite(options: { preserveSeededPrefix?: boolean } = {}): void {
+		const seeded = options.preserveSeededPrefix === true ? this.#seededPrefixCount : 0;
+		const prefix = seeded > 0 ? this.log.entries().slice(0, seeded) : [];
+		this.log.clear();
+		if (prefix.length > 0) this.log.extend(prefix);
+		this.#lastSyncCount = prefix.length;
+		this.#seededPrefixCount = prefix.length;
+		this.#syncedHashes = this.#hashRange(prefix, 0, prefix.length);
+		this.invalidate();
+	}
 	invalidate(): void {
 		this.prefix.invalidate();
 	}
@@ -384,7 +395,7 @@ export class AppendOnlyContextManager {
 	/** F9: reset the log to a new provider-visible baseline after seeded compaction/rebase. */
 	#rebaseToBaseline(messages: readonly unknown[], seededPrefixCount = 0): void {
 		this.log.clear();
-		this.log.extend([...messages]);
+		this.log.extend(messages.map(message => cloneJson(message)));
 		this.#lastSyncCount = messages.length;
 		this.#seededPrefixCount = seededPrefixCount;
 		this.#syncedHashes = this.#hashRange(messages, 0, messages.length);