@symerian/symi 3.5.5 → 3.5.6

package/dist/{subagent-registry-DQ_BPb-T.js → subagent-registry-ChDWDl_p.js}

@@ -49,8 +49,8 @@ import { a as normalizeElevatedLevel, c as normalizeUsageDisplay, d as supportsX
 import { i as resolveApiKeyForProvider, n as getCustomProviderApiKey, o as resolveEnvApiKey, r as requireApiKey, s as resolveModelAuthMode, t as getApiKeyForModel } from "./model-auth-Byr7Gic_.js";
 import { b as OLLAMA_NATIVE_BASE_URL, c as normalizeGoogleModelId, t as ensureSymiModelsJson, x as createOllamaStreamFn } from "./models-config-CiR_RUxw.js";
 import { i as resolveAckReaction, o as resolveEffectiveMessagesConfig, r as resolveResponsePrefixTemplate, s as resolveHumanDelayConfig, t as createReplyPrefixOptions } from "./reply-prefix-DV3KhOhL.js";
-import { n as resolveMemoryBackendConfig, t as getMemorySearchManager } from "./memory-d0YI5Lr0.js";
-import { a as parseGeminiAuth, c as resolveMemorySearchConfig, n as retryAsync, o as collectProviderApiKeysForExecution, s as executeWithApiKeyRotation } from "./manager-DghM7T-b.js";
+import { n as resolveMemoryBackendConfig, t as getMemorySearchManager } from "./memory-CbwmxmxW.js";
+import { a as parseGeminiAuth, c as resolveMemorySearchConfig, n as retryAsync, o as collectProviderApiKeysForExecution, s as executeWithApiKeyRotation } from "./manager-BVPXSK-I.js";
 import { f as runTasksWithConcurrency } from "./internal-D3mEiuyB.js";
 import { a as chunkText, c as resolveChunkMode, d as isSafeFenceBreak, f as parseFenceSpans, i as chunkMarkdownTextWithMode, l as resolveTextChunkLimit, o as chunkTextWithMode, r as chunkMarkdownText, t as chunkByNewline, u as findFenceSpanAt } from "./chunk-B2lBXhR0.js";
 import { n as resolveMarkdownTableMode } from "./markdown-tables-BEdkErjm.js";
@@ -82,7 +82,7 @@ import { i as parseAbsoluteTimeMs, r as resolveDefaultCronStaggerMs, t as normal
 import { n as resolveMessageChannelSelection, t as listConfiguredMessageChannels } from "./channel-selection-BrFl0vVK.js";
 import { n as resolveAgentRoute, t as buildAgentSessionKey } from "./resolve-route-D0NrCHnr.js";
 import { t as parseTimeoutMs } from "./parse-timeout-DUdeaMqF.js";
-import { n as registerMemoryCli } from "./memory-cli-DGheAg5i.js";
+import { n as registerMemoryCli } from "./memory-cli-B1dFPZw-.js";
 import { i as resolveSlackThreadTs, n as deliverReplies, t as createSlackReplyDeliveryPlan } from "./replies-_DV8VSSj.js";
 import { a as resolveSubagentToolPolicy, i as resolveGroupToolPolicy, n as isToolAllowedByPolicies, r as resolveEffectiveToolPolicy, t as filterToolsByPolicy } from "./pi-tools.policy-CU8U7--z.js";
 import { createRequire } from "node:module";
@@ -7647,13 +7647,21 @@ async function applyMediaUnderstanding(params) {
 * Soft warning appended when context usage crosses the threshold.
 * Fires once per session (tracked via `SessionEntry.contextWarningSent`).
 */
-const SESSION_CONTEXT_WARNING_MESSAGE = "📋 Heads up — this session is approaching the context limit. Start a new session for best results. Past sessions stay in history and can be revisited there.";
+const SESSION_CONTEXT_WARNING_MESSAGE = "📋 Heads up — this session is approaching the context limit. Symi will compact older context automatically when usage reaches 90%.";
 /**
-* Hard limit message returned when the model API rejects the turn for
-* context-too-long or role-ordering reasons, and on every subsequent
-* turn until the user starts a new session.
+* Inline status emitted by Symi before running compaction proactively
+* at the 90% threshold. Followed by the agent's actual response in the
+* same turn.
 */
-const SESSION_OVERFLOW_MESSAGE = "📋 This session has reached the context limit. Please start a new session — your full conversation is saved in history and can be revisited there.";
+const SESSION_COMPACTING_MESSAGE = "📋 Context limit reached for this session. Compacting older context and continuing…";
+/**
+* Returned when compaction can't recover enough budget to continue —
+* either because the proactive compaction at 90% failed, or because a
+* single user message overshot the window in one turn and the reactive
+* fallback compaction also failed. Same message also short-circuits
+* subsequent turns until /reset or /new.
+*/
+const SESSION_COMPACTION_FAILED_MESSAGE = "📋 Couldn't compact this session further. Please /reset or /new to continue — your full conversation is saved in history.";
 
 //#endregion
 //#region src/sessions/level-overrides.ts
@@ -28801,6 +28809,86 @@ async function handleInlineActions(params) {
 	};
 }
 
+//#endregion
+//#region src/auto-reply/reply/middleware-context-warning.ts
+/** Threshold (fraction of context window) that triggers the soft warning. */
+const CONTEXT_WARNING_THRESHOLD = .85;
+/**
+* Threshold that triggers proactive compaction in `prepareReplyTurn`.
+* Higher than the soft warning so the user sees the heads-up first;
+* compaction runs the next turn the threshold is observed.
+*/
+const CONTEXT_PROACTIVE_COMPACT_THRESHOLD = .9;
+/**
+* Pure decision for "should `prepareReplyTurn` run proactive compaction
+* before the next turn?" — extracted so the trigger logic is unit-testable
+* without mocking the rest of the prep pipeline.
+*
+* Returns true ONLY when:
+*   - we have a session id (otherwise compaction has no target),
+*   - we have a positive context window,
+*   - we have a positive cumulative usage from the previous turn,
+*   - usage / window ≥ CONTEXT_PROACTIVE_COMPACT_THRESHOLD.
+*/
+function shouldProactivelyCompact(args) {
+	if (!args.sessionId) return false;
+	if (!args.contextTokens || args.contextTokens <= 0) return false;
+	const usage = args.totalTokens ?? 0;
+	if (usage <= 0) return false;
+	return usage / args.contextTokens >= CONTEXT_PROACTIVE_COMPACT_THRESHOLD;
+}
+/**
+* Read the model's context-window size, falling back to the global
+* default when the model isn't in the registry yet.
+*/
+function resolveContextWindow(modelId) {
+	const fromRegistry = lookupContextTokens(modelId);
+	if (fromRegistry && fromRegistry > 0) return fromRegistry;
+	return DEFAULT_CONTEXT_TOKENS;
+}
+/**
+* After-turn hook that surfaces a one-time soft warning when usage
+* crosses the context threshold. Mutates the result text to append
+* the warning when triggered, and persists `contextWarningSent: true`
+* on the session entry so the warning doesn't fire again.
+*/
+function withContextUsageWarning() {
+	return (next) => async (params) => {
+		const result = await next(params);
+		if (result.outbound.action === "drop") return result;
+		if (!params.sessionKey) return result;
+		const usage = result.rawResult.meta.agentMeta?.usage?.total ?? 0;
+		if (usage <= 0) return result;
+		if (usage / resolveContextWindow(params.modelId ?? result.modelId) < CONTEXT_WARNING_THRESHOLD) return result;
+		const agentId = resolveAgentIdFromSessionKey(params.sessionKey);
+		const storePath = resolveStorePath(params.config.session?.store, { agentId });
+		let alreadyWarned = false;
+		try {
+			await updateSessionStoreEntry({
+				storePath,
+				sessionKey: params.sessionKey,
+				update: async (entry) => {
+					if (entry.contextWarningSent === true) {
+						alreadyWarned = true;
+						return null;
+					}
+					return { contextWarningSent: true };
+				}
+			});
+		} catch (err) {
+			logVerbose(`context-warning: failed to persist contextWarningSent flag for ${params.sessionKey}: ${String(err)}`);
+			return result;
+		}
+		if (alreadyWarned) return result;
+		const trimmedText = result.text.trimEnd();
+		const separator = trimmedText ? "\n\n" : "";
+		return {
+			...result,
+			text: `${trimmedText}${separator}${SESSION_CONTEXT_WARNING_MESSAGE}`
+		};
+	};
+}
+
 //#endregion
 //#region src/auto-reply/reply/session-reset-model.ts
 function splitBody(body) {
@@ -29697,7 +29785,7 @@ async function prepareReplyTurn(ctx, opts, configOverride) {
 		typing.cleanup();
 		return {
 			kind: "early-reply",
-			reply: { text: SESSION_OVERFLOW_MESSAGE }
+			reply: { text: SESSION_COMPACTION_FAILED_MESSAGE }
 		};
 	}
 	await applyResetModelOverride({
@@ -29810,6 +29898,82 @@ async function prepareReplyTurn(ctx, opts, configOverride) {
 	};
 	directives = inlineActionResult.directives;
 	abortedLastRun = inlineActionResult.abortedLastRun ?? abortedLastRun;
+	if (shouldProactivelyCompact({
+		totalTokens: sessionEntry.totalTokens,
+		contextTokens,
+		sessionId: sessionEntry.sessionId
+	})) {
+		if (resolvedOpts?.onBlockReply) try {
+			await resolvedOpts.onBlockReply({ text: SESSION_COMPACTING_MESSAGE });
+		} catch (err) {
+			logVerbose(`proactive-compact: inline status delivery failed: ${String(err)}`);
+		}
+		if (isEmbeddedPiRunActive(sessionEntry.sessionId)) {
+			abortEmbeddedPiRun(sessionEntry.sessionId);
+			await waitForEmbeddedPiRunEnd(sessionEntry.sessionId, 15e3);
+		}
+		let compactResult;
+		try {
+			compactResult = await compactEmbeddedPiSession({
+				sessionId: sessionEntry.sessionId,
+				sessionKey,
+				messageChannel: command.channel,
+				groupId: sessionEntry.groupId,
+				groupChannel: sessionEntry.groupChannel,
+				groupSpace: sessionEntry.space,
+				spawnedBy: sessionEntry.spawnedBy,
+				sessionFile: resolveSessionFilePath(sessionEntry.sessionId, sessionEntry, resolveSessionFilePathOptions({
+					agentId,
+					storePath
+				})),
+				workspaceDir,
+				agentDir,
+				config: cfg,
+				skillsSnapshot: sessionEntry.skillsSnapshot,
+				provider,
+				model,
+				thinkLevel: resolvedThinkLevel,
+				reasoningLevel: resolvedReasoningLevel,
+				trigger: "overflow",
+				senderIsOwner: command.senderIsOwner,
+				ownerNumbers: command.ownerList.length > 0 ? command.ownerList : void 0
+			});
+		} catch (err) {
+			logVerbose(`proactive-compact: compactor threw for ${sessionKey}: ${String(err)}`);
+			compactResult = {
+				ok: false,
+				compacted: false,
+				reason: String(err)
+			};
+		}
+		if (compactResult.ok && compactResult.compacted) await incrementCompactionCount({
+			sessionEntry,
+			sessionStore,
+			sessionKey,
+			storePath,
+			tokensAfter: compactResult.result?.tokensAfter
+		});
+		else {
+			logVerbose(`proactive-compact: failed for ${sessionKey} (${compactResult.reason ?? "unknown"})`);
+			if (storePath) try {
+				await updateSessionStoreEntry({
+					storePath,
+					sessionKey,
+					update: async (entry) => {
+						if (entry.endedReason === "context_overflow") return null;
+						return { endedReason: "context_overflow" };
+					}
+				});
+			} catch (err) {
+				logVerbose(`proactive-compact: failed to mark session sealed: ${String(err)}`);
+			}
+			typing.cleanup();
+			return {
+				kind: "early-reply",
+				reply: { text: SESSION_COMPACTION_FAILED_MESSAGE }
+			};
+		}
+	}
 	await stageSandboxMedia({
 		ctx,
 		sessionCtx,
@@ -39123,7 +39287,7 @@ async function runEmbeddedPiAgent(params) {
 				error: err
 			});
 		}
-		const MAX_OVERFLOW_COMPACTION_ATTEMPTS = 3;
+		const MAX_OVERFLOW_COMPACTION_ATTEMPTS = 1;
 		const MAX_RUN_LOOP_ITERATIONS = resolveMaxRunRetryIterations(profileCandidates.length);
 		let overflowCompactionAttempts = 0;
 		let toolResultTruncationAttempted = false;
@@ -40146,8 +40310,8 @@ function extractErrorMessage(err) {
 function buildBoundaryResult(params) {
 	const profile = resolveModelProfile(params.provider && params.modelId ? `${params.provider}/${params.modelId}` : "", params.config.models?.profiles);
 	return {
-		text: SESSION_OVERFLOW_MESSAGE,
-		rawText: SESSION_OVERFLOW_MESSAGE,
+		text: SESSION_COMPACTION_FAILED_MESSAGE,
+		rawText: SESSION_COMPACTION_FAILED_MESSAGE,
 		outbound: {
 			action: "deliver",
 			reason: "deliver"
@@ -40158,7 +40322,7 @@ function buildBoundaryResult(params) {
 		profile,
 		didSendViaMessagingTool: false,
 		rawResult: {
-			payloads: [{ text: SESSION_OVERFLOW_MESSAGE }],
+			payloads: [{ text: SESSION_COMPACTION_FAILED_MESSAGE }],
 			meta: {
 				durationMs: Date.now() - params.startTime,
 				agentMeta: {
@@ -40217,62 +40381,6 @@ function withContextOverflowRecovery() {
 	};
 }
 
-//#endregion
-//#region src/auto-reply/reply/middleware-context-warning.ts
-/** Threshold (fraction of context window) that triggers the soft warning. */
-const CONTEXT_WARNING_THRESHOLD = .85;
-/**
-* Read the model's context-window size, falling back to the global
-* default when the model isn't in the registry yet.
-*/
-function resolveContextWindow(modelId) {
-	const fromRegistry = lookupContextTokens(modelId);
-	if (fromRegistry && fromRegistry > 0) return fromRegistry;
-	return DEFAULT_CONTEXT_TOKENS;
-}
-/**
-* After-turn hook that surfaces a one-time soft warning when usage
-* crosses the context threshold. Mutates the result text to append
-* the warning when triggered, and persists `contextWarningSent: true`
-* on the session entry so the warning doesn't fire again.
-*/
-function withContextUsageWarning() {
-	return (next) => async (params) => {
-		const result = await next(params);
-		if (result.outbound.action === "drop") return result;
-		if (!params.sessionKey) return result;
-		const usage = result.rawResult.meta.agentMeta?.usage?.total ?? 0;
-		if (usage <= 0) return result;
-		if (usage / resolveContextWindow(params.modelId ?? result.modelId) < CONTEXT_WARNING_THRESHOLD) return result;
-		const agentId = resolveAgentIdFromSessionKey(params.sessionKey);
-		const storePath = resolveStorePath(params.config.session?.store, { agentId });
-		let alreadyWarned = false;
-		try {
-			await updateSessionStoreEntry({
-				storePath,
-				sessionKey: params.sessionKey,
-				update: async (entry) => {
-					if (entry.contextWarningSent === true) {
-						alreadyWarned = true;
-						return null;
-					}
-					return { contextWarningSent: true };
-				}
-			});
-		} catch (err) {
-			logVerbose(`context-warning: failed to persist contextWarningSent flag for ${params.sessionKey}: ${String(err)}`);
-			return result;
-		}
-		if (alreadyWarned) return result;
-		const trimmedText = result.text.trimEnd();
-		const separator = trimmedText ? "\n\n" : "";
-		return {
-			...result,
-			text: `${trimmedText}${separator}${SESSION_CONTEXT_WARNING_MESSAGE}`
-		};
-	};
-}
-
 //#endregion
 //#region src/auto-reply/reply/middleware-diagnostic-logging.ts
 function deriveChannel(ctx) {

package/dist/{synthesis-CSsmBmJc.js → synthesis-CmQvOHg6.js}

@@ -3,7 +3,7 @@ import "./registry-C-JddWwo.js";
 import "./agent-scope-B1Ina6jV.js";
 import { t as createSubsystemLogger } from "./subsystem-DzRUKS9f.js";
 import "./workspace-CsaDUuDn.js";
-import { t as runAgentTurn } from "./unified-runner-DN0asRIR.js";
+import { t as runAgentTurn } from "./unified-runner-Cu3lGa3A.js";
 import "./auth-profiles-BSw0aQND.js";
 import "./model-selection-BqshlIZc.js";
 import "./deliver-BiWlR84Y.js";
@@ -25,7 +25,7 @@ import "./paths-BsT3BvfH.js";
 import "./tokens-D_vCJSzF.js";
 import "./diagnostic-BALvP9wI.js";
 import "./diagnostic-session-state-BDfanaOY.js";
-import "./manager-Ck1YSfr6.js";
+import "./manager-CpBJ1BE6.js";
 import "./github-copilot-token-uhEBNQfj.js";
 import "./pi-auth-json-DCGR3yfh.js";
 import "./reply-prefix-D0NzzC4I.js";

package/dist/{synthesis-XbEFEFK1.js → synthesis-DIkOQnhF.js}

@@ -6,7 +6,7 @@ import "./model-selection-BisYvTBb.js";
 import "./reply-prefix-i-FPYcoQ.js";
 import "./plugins-CZ_mYwXq.js";
 import "./replies-0nzkXt6o.js";
-import { n as runAgentTurn } from "./pi-embedded-zSv7VKNQ.js";
+import { n as runAgentTurn } from "./pi-embedded-C9wLrFjj.js";
 import "./deliver-f3cIWxXT.js";
 import "./diagnostic-BdRnGknC.js";
 import "./diagnostic-session-state-DpxCUzoM.js";
@@ -23,7 +23,7 @@ import "./errors-BEU7IHU6.js";
 import "./tokens-yWO1wD7Z.js";
 import "./thinking-DgjGxdSP.js";
 import "./paths-BsT3BvfH.js";
-import "./manager-Bt-1HGg9.js";
+import "./manager-CnBU0aOR.js";
 import "./github-copilot-token-dYUr1mDQ.js";
 import "./sqlite-D8iYC_CU.js";
 import "./markdown-tables-jQzXAsf3.js";

package/dist/{synthesis-CDd1xqI3.js → synthesis-DdP6tbyk.js}

@@ -1,4 +1,4 @@
-import { t as runAgentTurn } from "./unified-runner-Bm3r8UOg.js";
+import { t as runAgentTurn } from "./unified-runner-DL-sfvGT.js";
 import "./registry-Sbac4a4z.js";
 import "./paths-By0XjHBk.js";
 import "./utils-CTPsqyE_.js";
@@ -50,8 +50,8 @@ import "./session-cost-usage-D3mgssM_.js";
 import "./session-utils-Cs1jlD-q.js";
 import "./with-timeout-GbJ1Yzsh.js";
 import "./reply-prefix-BFNeXT65.js";
-import "./memory-CgWOtF4K.js";
-import "./manager-CC13EPO9.js";
+import "./memory-BsGfHO_F.js";
+import "./manager-CvqH_tah.js";
 import "./sqlite-Cod7C5ba.js";
 import "./chunk-BxfKjCUx.js";
 import "./markdown-tables-CS1Tvb3z.js";
@@ -82,7 +82,7 @@ import "./links-CQZxjjCO.js";
 import "./cli-utils-CbnnSB38.js";
 import "./help-format-8IYCRvx1.js";
 import "./progress-DQTvTiEg.js";
-import "./memory-cli-vN3fGs79.js";
+import "./memory-cli-BwQsPEFk.js";
 import "./logging-CLWZ3KQI.js";
 import "./replies-C5CBlnFS.js";
 import "./pi-tools.policy-WaLKhqJQ.js";

package/dist/{synthesis-DIqhHKjV.js → synthesis-m6WCyI41.js}

@@ -17,7 +17,7 @@ import "./client-DhdWcgZP.js";
 import "./call-ogggp9QZ.js";
 import "./message-channel-DXToSaWK.js";
 import "./pairing-token-B-OTYROz.js";
-import { T as runAgentTurn } from "./subagent-registry-DQ_BPb-T.js";
+import { T as runAgentTurn } from "./subagent-registry-ChDWDl_p.js";
 import "./sessions-Dn6VXn4p.js";
 import "./plugins-BsEhWvLn.js";
 import "./send-Dq252-bi.js";
@@ -47,8 +47,8 @@ import "./model-auth-Byr7Gic_.js";
 import "./github-copilot-token-B5y4__YM.js";
 import "./models-config-CiR_RUxw.js";
 import "./reply-prefix-DV3KhOhL.js";
-import "./memory-d0YI5Lr0.js";
-import "./manager-DghM7T-b.js";
+import "./memory-CbwmxmxW.js";
+import "./manager-BVPXSK-I.js";
 import "./sqlite-9jF_eajd.js";
 import "./chunk-B2lBXhR0.js";
 import "./markdown-tables-BEdkErjm.js";
@@ -79,7 +79,7 @@ import "./links-B9CbwY46.js";
 import "./cli-utils-Ca0KE-dW.js";
 import "./help-format-BTNd5kFC.js";
 import "./progress-BOQ0hkeM.js";
-import "./memory-cli-DGheAg5i.js";
+import "./memory-cli-B1dFPZw-.js";
 import "./logging-uD67RPtO.js";
 import "./replies-_DV8VSSj.js";
 import "./pi-tools.policy-CU8U7--z.js";