@dimi-agent/cli 0.5.4 → 0.6.0

package/dist/main.mjs

@@ -119585,22 +119585,6 @@ function wrapSubagentModelError(error, boundModel, callerModelAlias) {
 		}
 	});
 }
-/** Human-readable duration for the subagent timeout message. */
-function formatSubagentTimeoutDescription(ms) {
-	if (ms % (3600 * 1e3) === 0) {
-		const h = ms / (3600 * 1e3);
-		return `${h} hour${h === 1 ? "" : "s"}`;
-	}
-	if (ms % (60 * 1e3) === 0) {
-		const m = ms / (60 * 1e3);
-		return `${m} minute${m === 1 ? "" : "s"}`;
-	}
-	if (ms % 1e3 === 0) {
-		const s = ms / 1e3;
-		return `${s} second${s === 1 ? "" : "s"}`;
-	}
-	return `${ms} ms`;
-}
 var SUBAGENT_SECTION, SubagentConfigSchema, DEFAULT_SUBAGENT_TIMEOUT_MS, SUBAGENT_TIMEOUT_ENV, subagentEnvBindings, stripSubagentEnv;
 var init_configSection$7 = __esmMin((() => {
 	init_zod$1();
@@ -121101,6 +121085,56 @@ var init_toolPolicyService = __esmMin((() => {
 	registerScopedService(2, IAgentToolPolicyService, AgentToolPolicyService, 0, "toolPolicy");
 }));
 //#endregion
+//#region ../../packages/agent-core-v2/src/agent/loop/contextSize.ts
+/**
+* Scale a token count by `percent` of the model's default window, keeping
+* at least `CONTEXT_SIZE_FLOOR_TOKENS` — but never more than the original
+* window, so models already below the floor (and `percent` at or above
+* 100, or unset) pass through unchanged.
+*/
+function scaleContextTokens(tokens, percent) {
+	if (!Number.isFinite(tokens) || tokens <= 0) return tokens;
+	if (percent === void 0 || percent <= 0 || percent >= 100) return tokens;
+	const scaled = Math.floor(tokens * percent / 100);
+	return Math.max(scaled, Math.min(tokens, CONTEXT_SIZE_FLOOR_TOKENS));
+}
+/**
+* Apply the configured context-size percentage to a model capability.
+* Returns the original object unchanged when nothing changes, so
+* `UNKNOWN_CAPABILITY` identity and other callers relying on object
+* equality keep working.
+*/
+function scaleModelCapabilityContext(capability, percent) {
+	if (percent === void 0 || percent <= 0 || percent >= 100) return capability;
+	const maxContextTokens = scaleContextTokens(capability.max_context_tokens, percent);
+	const maxInputTokens = capability.max_input_tokens === void 0 ? void 0 : scaleContextTokens(capability.max_input_tokens, percent);
+	if (maxContextTokens === capability.max_context_tokens && maxInputTokens === capability.max_input_tokens) return capability;
+	return {
+		...capability,
+		max_context_tokens: maxContextTokens,
+		max_input_tokens: maxInputTokens
+	};
+}
+/**
+* Offered percentage levels for a model window: `100`, `95`, … descending
+* in `CONTEXT_SIZE_STEP_PERCENT` steps while the scaled size stays at or
+* above `CONTEXT_SIZE_FLOOR_TOKENS`. Empty when the window is below the
+* floor — the model's context size is not adjustable.
+*/
+function contextSizePercentOptions(contextWindow) {
+	if (!Number.isFinite(contextWindow) || contextWindow <= 0) return [];
+	const options = [];
+	for (let percent = 100; percent >= 5; percent -= 5) {
+		if (Math.floor(contextWindow * percent / 100) < 2e5) break;
+		options.push(percent);
+	}
+	return options;
+}
+var CONTEXT_SIZE_FLOOR_TOKENS;
+var init_contextSize$1 = __esmMin((() => {
+	CONTEXT_SIZE_FLOOR_TOKENS = 2e5;
+}));
+//#endregion
 //#region ../../packages/agent-core-v2/src/agent/loop/configSection.ts
 function parseNonNegativeInt(raw) {
 	const value = raw.trim();
@@ -121114,6 +121148,7 @@ var init_configSection$5 = __esmMin((() => {
 	init_config$5();
 	init_configSectionContributions();
 	init_toml();
+	init_contextSize$1();
 	LOOP_CONTROL_SECTION = "loopControl";
 	LOOP_MAX_STEPS_PER_TURN_ENV = "DIMI_LOOP_MAX_STEPS_PER_TURN";
 	LOOP_MAX_RETRIES_PER_STEP_ENV = "DIMI_LOOP_MAX_RETRIES_PER_STEP";
@@ -121122,7 +121157,13 @@ var init_configSection$5 = __esmMin((() => {
 		maxRetriesPerStep: number$2().int().min(0).optional(),
 		maxRalphIterations: number$2().int().min(-1).optional(),
 		reservedContextSize: number$2().int().min(0).optional(),
-		compactionTriggerRatio: number$2().min(.5).max(.99).optional()
+		compactionTriggerRatio: number$2().min(.5).max(.99).optional(),
+		/**
+		* Effective context window as a percentage of the model's default, in
+		* 5% steps, never below `CONTEXT_SIZE_FLOOR_TOKENS`. See
+		* `agent/loop/contextSize.ts` for the scaling semantics.
+		*/
+		contextSizePercent: number$2().int().min(5).max(100).multipleOf(5).optional()
 	});
 	loopControlEnvBindings = envBindings(LoopControlSchema, {
 		maxStepsPerTurn: {
@@ -138782,7 +138823,15 @@ var init_subagent = __esmMin((() => {
 async function runAgentTurn(target, request, options) {
 	options.signal.throwIfAborted();
 	const promptService = target.accessor.get(IAgentPromptService);
-	const turn = request.kind === "prompt" ? await (await promptService.enqueue({ message: {
+	const turn = request.kind === "prompt" ? await (await (options.steer === true ? promptService.enqueueOrSteer({ message: {
+		role: "user",
+		content: [{
+			type: "text",
+			text: request.prompt
+		}],
+		toolCalls: [],
+		origin: AGENT_RUN_PROMPT_ORIGIN
+	} }) : promptService.enqueue({ message: {
 		role: "user",
 		content: [{
 			type: "text",
@@ -138790,7 +138839,7 @@ async function runAgentTurn(target, request, options) {
 		}],
 		toolCalls: [],
 		origin: AGENT_RUN_PROMPT_ORIGIN
-	} })).launched : await promptService.retry();
+	} }))).launched : await promptService.retry();
 	if (turn === void 0) throw new Error("Agent turn could not be started");
 	if (options.onReady !== void 0) turn.ready.then(() => options.onReady?.()).catch(() => {});
 	const completion = awaitRun(target, turn, options);
@@ -138964,7 +139013,8 @@ var init_subagentService = __esmMin((() => {
 			return runAgentTurn(handle, request, {
 				summaryPolicy: opts.summaryPolicy ?? this.summaryPolicyFor(handle),
 				signal: opts.signal,
-				onReady: opts.onReady
+				onReady: opts.onReady,
+				steer: opts.steer
 			});
 		}
 		notifyAgentTaskStopped(context) {
@@ -139236,7 +139286,7 @@ var init_mirrorAgentRun = __esmMin((() => {
 	init_eventBus();
 	init_agentLifecycle();
 	init_subagent();
-})), DEFAULT_PROFILE_NAME, SubagentToolInputSchema, BACKGROUND_AGENT_UNAVAILABLE, RESUME_WITH_TYPE_UNAVAILABLE, USER_INTERRUPTED_SUBAGENT_MESSAGE, SUBAGENT_STOPPED_MESSAGE, ISubagentTool;
+})), DEFAULT_PROFILE_NAME, SubagentToolInputSchema, RESUME_WITH_TYPE_UNAVAILABLE, USER_INTERRUPTED_SUBAGENT_MESSAGE, SUBAGENT_STOPPED_MESSAGE, ISubagentTool;
 var init_agent$2 = __esmMin((() => {
 	init_zod$1();
 	init_instantiation();
@@ -139253,8 +139303,7 @@ var init_agent$2 = __esmMin((() => {
 		prompt: string$2().describe("Full task prompt for the subagent"),
 		description: string$2().describe("Short task description (3-5 words) for UI display"),
 		subagent_type: string$2().optional().describe("One of the available agent types (see \"Available agent types\" in this tool description). Defaults to \"coder\" when omitted."),
-		resume: string$2().optional().describe("Optional agent ID to resume instead of creating a new instance. When set, do not also pass subagent_type — the resumed agent keeps its own type, and supplying both is rejected."),
-		run_in_background: boolean$2().optional().describe("If true, return immediately without waiting for completion. Prefer false unless the task can run independently and there is a clear benefit to not waiting."),
+		resume: string$2().optional().describe("Agent ID to message or continue instead of creating a new instance. Messaging an existing subagent works exactly like a human steering the agent: if it is still running, the prompt is injected into its current turn immediately; if it is idle, it starts a normal turn. When set, do not also pass subagent_type — the resumed agent keeps its own type, and supplying both is rejected."),
 		model: _enum(["secondary", "primary"]).optional().describe("Which model to run the subagent on: \"secondary\" = the configured secondary model; \"primary\" = the main model you are running on (for hard, quality-sensitive tasks). This explicit choice overrides the selected agent type's model_preference; without either, secondary is the default when configured. Only effective when a secondary model is configured; otherwise the subagent inherits your model. Ignored when resuming — resumed subagents keep their own model.")
 	}));
 	object({
@@ -139266,7 +139315,6 @@ var init_agent$2 = __esmMin((() => {
 			cache_write: number$2().int().nonnegative().optional()
 		}).describe("Cumulative token usage")
 	});
-	BACKGROUND_AGENT_UNAVAILABLE = "Background agent execution is not available for this agent because TaskList, TaskOutput, and TaskStop are not enabled.";
 	RESUME_WITH_TYPE_UNAVAILABLE = "Cannot set subagent_type when resuming an existing agent. Resume by agent id only.";
 	USER_INTERRUPTED_SUBAGENT_MESSAGE = "The subagent was stopped before it finished by user.";
 	SUBAGENT_STOPPED_MESSAGE = "The subagent was stopped before it finished.";
@@ -139312,22 +139360,10 @@ function firstNonEmpty(...values) {
 }
 var init_subagentMetadata = __esmMin((() => {}));
 //#endregion
-//#region ../../packages/agent-core-v2/src/agent/tools/agent/agent-background-disabled.md?raw
-var agent_background_disabled_default;
-var init_agent_background_disabled = __esmMin((() => {
-	agent_background_disabled_default = "Background agent execution is disabled for this agent. Do not set `run_in_background=true` — any call that sets it is rejected before the subagent launches. Run every subagent in the foreground and wait for its result.";
-}));
-//#endregion
-//#region ../../packages/agent-core-v2/src/agent/tools/agent/agent-background-enabled.md?raw
-var agent_background_enabled_default;
-var init_agent_background_enabled = __esmMin((() => {
-	agent_background_enabled_default = "When `run_in_background=true`, the subagent runs detached from this turn. The completion arrives in a later turn as a synthetic user-role message containing its result — you do not need to poll, sleep, or check on its progress. Continue with other work or respond to the user. Never fabricate or predict what the result will say.\n\nDefault to a foreground subagent (omit `run_in_background`) when your next step needs its result — foreground hands the result straight back. Reach for `run_in_background=true` only when you have other work to do while it runs and do not need its result to proceed. Never launch in the background and then immediately wait on it (by polling `TaskOutput`, sleeping, or otherwise): that just blocks the turn for no benefit — run it in the foreground instead.\n";
-}));
-//#endregion
 //#region ../../packages/agent-core-v2/src/agent/tools/agent/agent.md?raw
 var agent_default;
 var init_agent$1 = __esmMin((() => {
-	agent_default = "Launch a subagent to handle a task. The subagent runs as a same-process loop instance with its own context and wire file. Delegating also keeps the bulk of intermediate file contents out of your own context — you get a conclusion back instead of a pile of dumps.\n\nWriting the prompt:\n- The subagent starts with zero context — it has not seen this conversation. Brief it like a colleague who just walked into the room: state the goal, list what you already know, hand over the specifics.\n- Lookups (read this file, run that test): put the exact path or command in the prompt. The subagent should not have to search for things you already know.\n- Investigations (figure out X, find why Y): give the question, not prescribed steps — fixed steps become dead weight when the premise is wrong.\n- Do not delegate understanding. If the task hinges on a file path or line number, find it yourself first and write it into the prompt.\n\nUsage notes:\n- When the task continues earlier work a subagent already did, prefer resuming that agent (pass its `resume` id) over spawning a fresh instance — the resumed agent keeps its prior context.\n- A subagent's result is only visible to you, not to the user. When the user needs to see what a subagent produced, summarize the relevant parts yourself in your own reply.\n- Subagents use a fixed 2-hour timeout. If one times out, resume the same agent instead of starting over.\n\nWhen NOT to use Agent: skip delegation for trivial work you can do directly — reading a file whose path you already know, searching a small known set of files, or any task that takes only a step or two. Delegation has a context-handoff cost; it pays off only when the task is substantial enough to outweigh it.\n\nOnce a subagent is running, leave that scope to it: do not redo its searches or reads in parallel, and do not abandon it midway and finish the job manually. Both undo the context savings the delegation was meant to buy.\n";
+	agent_default = "Launch a subagent to handle a task. The subagent runs as a same-process loop instance with its own context and wire file. Delegating also keeps the bulk of intermediate file contents out of your own context — you get a conclusion back instead of a pile of dumps.\n\nThe subagent is **fully asynchronous**: this tool returns immediately with the subagent's `agent_id` and a `task_id` — it never blocks your turn. Do the rest of your work while it runs; its result arrives on its own as a completion notification with the final summary. You do not need to poll, sleep, or check on it — and never fabricate or predict what the result will say.\n\nWhen you genuinely have nothing else to do and want to see how it is going:\n\n- Call `AgentOutput(agent_id=\"...\")` to read its recent rendered output (assistant text, thinking, tool calls, progress) — the same view a human sees in the TUI.\n- If it is still working, call `WaitFor` with a reasonable `timeout_seconds` instead of polling `AgentOutput` in a loop; the wait wakes you on the completion notification or the timeout, then check again with `AgentOutput`.\n\nWriting the prompt:\n- The subagent starts with zero context — it has not seen this conversation. Brief it like a colleague who just walked into the room: state the goal, list what you already know, hand over the specifics.\n- Lookups (read this file, run that test): put the exact path or command in the prompt. The subagent should not have to search for things you already know.\n- Investigations (figure out X, find why Y): give the question, not prescribed steps — fixed steps become dead weight when the premise is wrong.\n- Do not delegate understanding. If the task hinges on a file path or line number, find it yourself first and write it into the prompt.\n\nUsage notes:\n- When the task continues earlier work a subagent already did, prefer resuming that agent (pass its `resume` id) over spawning a fresh instance — the resumed agent keeps its prior context. `resume` works exactly like a human steering the agent: while the subagent is still running, your prompt is injected into its current turn immediately; when it is idle, it starts a normal turn. Use it to redirect, follow up, or send it new information.\n- A subagent's result is only visible to you, not to the user. When the user needs to see what a subagent produced, summarize the relevant parts yourself in your own reply.\n- Subagents use a fixed 2-hour timeout. If one times out, resume the same agent instead of starting over.\n\nWhen NOT to use Agent: skip delegation for trivial work you can do directly — reading a file whose path you already know, searching a small known set of files, or any task that takes only a step or two. Delegation has a context-handoff cost; it pays off only when the task is substantial enough to outweigh it.\n\nOnce a subagent is running, leave that scope to it: do not redo its searches or reads in parallel, and do not abandon it midway and finish the job manually. Both undo the context savings the delegation was meant to buy.\n";
 }));
 //#endregion
 //#region ../../packages/agent-core-v2/src/agent/tools/agent/agentTool.ts
@@ -139351,7 +139387,7 @@ function buildProfileDescriptions(profiles, tools, isToolActive$1, showModelPref
 		return `${headerLines}\n  Tools: ${activeTools.join(", ")}`;
 	}).join("\n");
 }
-function formatBackgroundAgentResult(taskId, handle, description, allowBackground) {
+function formatAsyncAgentResult(taskId, handle, description) {
 	return [
 		`task_id: ${taskId}`,
 		"status: running",
@@ -139361,31 +139397,11 @@ function formatBackgroundAgentResult(taskId, handle, description, allowBackgroun
 		"",
 		`description: ${description}`,
 		"",
-		allowBackground ? `next_step: The completion arrives automatically in a later turn — do NOT wait, poll, or call TaskOutput on it; continue with other work or hand back to the user. (If you have nothing to do until it finishes, run such tasks in the foreground next time.)` : "next_step: The completion arrives automatically in a later turn.",
-		`resume_hint: To continue or recover this same subagent later, call Agent(resume="${handle.agentId}", prompt="..."). The parameter is agent_id ("${handle.agentId}"), NOT task_id ("${taskId}") or source_id from a later <notification>. Recovery cases: a later <notification type="task.lost" | "task.failed" | "task.killed"> for this subagent — its conversation history is preserved across session restarts and resume will pick it up.`
+		"next_step: The subagent runs fully asynchronously — continue with other work. Its final result arrives later as a completion notification.",
+		`progress_hint: To check on it, call AgentOutput(agent_id="${handle.agentId}") to read its recent output (assistant text, thinking, tool calls). If you have nothing else to do, call WaitFor with a reasonable timeout_seconds instead of polling, then check again with AgentOutput.`,
+		`resume_hint: To continue, redirect, or send this subagent a message — like a human steering the agent — call Agent(resume="${handle.agentId}", prompt="..."). While it is still running the prompt is injected into its current turn immediately; when idle it starts a normal turn. The parameter is agent_id ("${handle.agentId}"), NOT task_id ("${taskId}") or source_id from a later <notification>. Recovery cases: a later <notification type="task.lost" | "task.failed" | "task.killed"> for this subagent — its conversation history is preserved across session restarts and resume will pick it up.`
 	].join("\n");
 }
-function formatForegroundAgentSuccess(handle, result) {
-	return [
-		`agent_id: ${handle.agentId}`,
-		`actual_subagent_type: ${handle.profileName}`,
-		"status: completed",
-		"",
-		"[summary]",
-		result
-	].join("\n");
-}
-function formatForegroundAgentFailure(handle, message, timedOut) {
-	const lines = [
-		`agent_id: ${handle.agentId}`,
-		`actual_subagent_type: ${handle.profileName}`,
-		"status: failed",
-		"",
-		`subagent error: ${message}`
-	];
-	if (timedOut) lines.push(`resume_hint: Continue with Agent(resume="${handle.agentId}", prompt="continue"). Use agent_id only; do not set subagent_type. The subagent retains its prior context; redo any unfinished tool call if its result was lost.`);
-	return lines.join("\n");
-}
 function launchErrorMessage(error, signal) {
 	if (isUserCancellation(signal.reason)) return USER_INTERRUPTED_SUBAGENT_MESSAGE;
 	if (isAbortError$1(error)) return formatSubagentStoppedMessage(errorMessage$4(signal.reason));
@@ -139413,7 +139429,6 @@ var init_agentTool = __esmMin((() => {
 	init_toolPolicy();
 	init_permissionMode();
 	init_scopeContext();
-	init_loop();
 	init_userTool();
 	init_toolContract();
 	init_toolContribution();
@@ -139437,8 +139452,6 @@ var init_agentTool = __esmMin((() => {
 	init_flag$2();
 	init_agent$2();
 	init_subagent_task();
-	init_agent_background_disabled();
-	init_agent_background_enabled();
 	init_agent$1();
 	init_decorateParam();
 	init_decorate();
@@ -139461,7 +139474,6 @@ var init_agentTool = __esmMin((() => {
 		name = "Agent";
 		parameters = toInputJsonSchema(SubagentToolInputSchema);
 		callerAgentId;
-		canRunInBackground;
 		constructor(lifecycle, subagents, catalog, scopeContext, tasks, profile, toolPolicy, toolRegistry, workspace, processRunner, sessionMetadata, log, permissionMode, config, flags, modelCatalog) {
 			this.lifecycle = lifecycle;
 			this.subagents = subagents;
@@ -139479,10 +139491,9 @@ var init_agentTool = __esmMin((() => {
 			this.flags = flags;
 			this.modelCatalog = modelCatalog;
 			this.callerAgentId = scopeContext.agentId;
-			this.canRunInBackground = () => this.toolPolicy.isToolActive("TaskList") && this.toolPolicy.isToolActive("TaskOutput") && this.toolPolicy.isToolActive("TaskStop");
 		}
 		get description() {
-			let description = `${agent_default}\n\n${this.canRunInBackground() ? agent_background_enabled_default : agent_background_disabled_default}`;
+			let description = agent_default;
 			const allowlist = subagentAllowlistFor(this.catalog, this.profile.data());
 			const typeLines = buildProfileDescriptions(allowlist === void 0 ? this.catalog.list() : this.catalog.list().filter((profile) => allowlist.includes(profile.name)), this.knownToolReferences(), (profile, name, source) => this.toolPolicy.isToolActiveForProfile(profile, name, source), this.flags.enabled(SECONDARY_MODEL_FLAG_ID), this.flags.enabled(BACKGROUND_BASH_STDIN_FLAG_ID));
 			if (typeLines) description += `\n\nAvailable agent types (pass via subagent_type):\n${typeLines}`;
@@ -139508,13 +139519,13 @@ var init_agentTool = __esmMin((() => {
 			};
 			const profileNameForDisplay = resumeAgentId !== void 0 && resumeAgentId.length > 0 ? this.resumeProfileName(resumeAgentId) ?? "subagent" : requestedProfileName ?? "coder";
 			return {
-				description: `${args.run_in_background === true ? "Launching background" : "Launching"} ${profileNameForDisplay} agent: ${args.description}`,
+				description: `Launching ${profileNameForDisplay} agent: ${args.description}`,
 				accesses: ToolAccesses.none(),
 				display: {
 					kind: "agent_call",
 					agent_name: profileNameForDisplay,
 					prompt: args.prompt,
-					background: args.run_in_background
+					background: false
 				},
 				approvalRule: this.name,
 				matchesRule: (ruleArgs) => matchesGlobRuleSubject(ruleArgs, profileNameForDisplay),
@@ -139537,7 +139548,7 @@ var init_agentTool = __esmMin((() => {
 			if (isResume) {
 				const target = this.lifecycle.get(resumeAgentId);
 				if (target === void 0) throw new Error(`Agent instance "${resumeAgentId}" does not exist`);
-				await this.ensureOwnedIdleSubagent(resumeAgentId, target);
+				await this.ensureOwnedSubagent(resumeAgentId);
 				agentId = target.id;
 				profileName = target.accessor.get(IAgentProfileService).data().profileName ?? "subagent";
 			} else {
@@ -139578,17 +139589,19 @@ var init_agentTool = __esmMin((() => {
 					log: this.log
 				});
 			}
-			const runInBackground = args.run_in_background === true;
 			emitAgentRunSpawned(requester, agentId, {
 				profileName,
 				parentToolCallId: toolCallId,
 				description: args.description,
-				runInBackground
+				runInBackground: false
 			});
 			const mirrored = mirrorAgentRun(requester, await this.subagents.run(agentId, {
 				kind: "prompt",
 				prompt: promptText
-			}, { signal: controller.signal }), {
+			}, {
+				signal: controller.signal,
+				steer: isResume
+			}), {
 				profileName,
 				prompt: promptText,
 				signal: controller.signal,
@@ -139605,16 +139618,14 @@ var init_agentTool = __esmMin((() => {
 				}))
 			};
 		}
-		async ensureOwnedIdleSubagent(agentId, target) {
+		async ensureOwnedSubagent(agentId) {
 			const meta = (await this.sessionMetadata.read()).agents?.[agentId];
 			if (!isSubagentMeta(meta)) throw new Error(`Agent instance "${agentId}" is not a subagent`);
 			if (subagentParentAgentId(meta) !== this.callerAgentId) throw new Error(`Agent instance "${agentId}" does not belong to this parent agent`);
-			if (target.accessor.get(IAgentLoopService).status().state === "running") throw new Error(`Agent instance "${agentId}" is already running and cannot run concurrently`);
 		}
 		async execution(args, { toolCallId, signal }) {
 			try {
 				signal.throwIfAborted();
-				const runInBackground = args.run_in_background === true;
 				const requestedProfileName = args.subagent_type?.length ? args.subagent_type : void 0;
 				const resumeAgentId = args.resume?.trim();
 				const isResume = resumeAgentId !== void 0 && resumeAgentId.length > 0;
@@ -139622,17 +139633,12 @@ var init_agentTool = __esmMin((() => {
 					output: RESUME_WITH_TYPE_UNAVAILABLE,
 					isError: true
 				};
-				const allowBackground = this.canRunInBackground();
-				if (runInBackground && !allowBackground) return {
-					output: BACKGROUND_AGENT_UNAVAILABLE,
-					isError: true
-				};
 				const timeoutMs = resolveSubagentTimeoutMs(this.config);
 				const controller = new AbortController();
 				const abortBeforeRegister = () => {
 					controller.abort(signal.reason);
 				};
-				if (!runInBackground) signal.addEventListener("abort", abortBeforeRegister, { once: true });
+				signal.addEventListener("abort", abortBeforeRegister, { once: true });
 				let handle;
 				try {
 					handle = await this.launch(args, toolCallId, controller);
@@ -139640,7 +139646,6 @@ var init_agentTool = __esmMin((() => {
 					signal.removeEventListener("abort", abortBeforeRegister);
 					this.log.warn("subagent launch failed", {
 						toolCallId,
-						runInBackground,
 						operation: isResume ? "resume" : "spawn",
 						subagentType: requestedProfileName ?? "coder",
 						resumeAgentId: isResume ? resumeAgentId : void 0,
@@ -139651,9 +139656,9 @@ var init_agentTool = __esmMin((() => {
 				let taskId;
 				try {
 					const registerOptions = {
-						detached: runInBackground,
+						detached: true,
 						timeoutMs,
-						signal: runInBackground ? void 0 : signal
+						signal: void 0
 					};
 					taskId = this.tasks.registerTask(new SubagentTask(handle, args.description, controller), registerOptions);
 					signal.removeEventListener("abort", abortBeforeRegister);
@@ -139661,7 +139666,7 @@ var init_agentTool = __esmMin((() => {
 					controller.abort();
 					handle.completion.catch(() => {});
 					signal.removeEventListener("abort", abortBeforeRegister);
-					this.log?.warn("background agent task registration failed", {
+					this.log?.warn("subagent task registration failed", {
 						toolCallId,
 						agentId: handle.agentId,
 						subagentType: handle.profileName,
@@ -139673,9 +139678,7 @@ var init_agentTool = __esmMin((() => {
 						isError: true
 					};
 				}
-				if (runInBackground) return { output: formatBackgroundAgentResult(taskId, handle, args.description, allowBackground) };
-				if (await this.tasks.waitForForegroundRelease(taskId) === "detached") return { output: formatBackgroundAgentResult(taskId, handle, args.description, allowBackground) };
-				return await this.formatForegroundResult(taskId, handle, timeoutMs);
+				return { output: formatAsyncAgentResult(taskId, handle, args.description) };
 			} catch (error) {
 				return {
 					output: `subagent error: ${launchErrorMessage(error, signal)}`,
@@ -139683,15 +139686,6 @@ var init_agentTool = __esmMin((() => {
 				};
 			}
 		}
-		async formatForegroundResult(taskId, handle, timeoutMs) {
-			const info = this.tasks.getTask(taskId);
-			if (info?.status === "completed") return { output: formatForegroundAgentSuccess(handle, await this.tasks.readOutput(taskId)) };
-			const timedOut = info?.status === "timed_out";
-			return {
-				output: formatForegroundAgentFailure(handle, timedOut ? `Agent timed out after ${formatSubagentTimeoutDescription(timeoutMs)}.` : formatSubagentStoppedMessage(info?.stopReason), timedOut),
-				isError: true
-			};
-		}
 	};
 	SubagentTool = __decorate$1([
 		__decorateParam(0, IAgentLifecycleService),
@@ -211918,6 +211912,8 @@ var init_profileService = __esmMin((() => {
 	init_errors$11();
 	init_bootstrap();
 	init_config$5();
+	init_configSection$5();
+	init_contextSize$1();
 	init_hostEnvironment();
 	init_hostFileSystem();
 	init_sessionContext();
@@ -212005,6 +212001,7 @@ var init_profileService = __esmMin((() => {
 					this.publishToolPatternWarnings();
 					this.refreshSystemPrompt();
 				}
+				if (domain === "loopControl") this.emitStatusUpdated();
 			}));
 			this._register(this.skillCatalog.onDidChange((sourceId) => {
 				if (sourceId === "plugin") this.refreshSystemPrompt();
@@ -212202,7 +212199,7 @@ var init_profileService = __esmMin((() => {
 			return {
 				cwd: this.cwd,
 				modelAlias: this.modelAlias,
-				modelCapabilities: model === void 0 ? UNKNOWN_CAPABILITY : modelCapabilities(model),
+				modelCapabilities: this.capabilitiesFor(model),
 				profileName: this.profileName,
 				thinkingLevel: this.thinkingLevel,
 				systemPrompt: this.systemPrompt,
@@ -212217,10 +212214,10 @@ var init_profileService = __esmMin((() => {
 		resolveModelContext() {
 			const modelAlias = this.model;
 			const model = this.modelCatalog.get(modelAlias);
-			const loopControl = this.config.get("loopControl");
+			const loopControl = this.config.get(LOOP_CONTROL_SECTION);
 			return {
 				modelAlias,
-				modelCapabilities: modelCapabilities(model),
+				modelCapabilities: this.capabilitiesFor(model),
 				maxOutputSize: model.maxTokens,
 				alwaysThinking: modelAlwaysThinking(model) || void 0,
 				thinkingLevel: this.resolveThinkingState(model).effective,
@@ -212245,8 +212242,15 @@ var init_profileService = __esmMin((() => {
 			};
 		}
 		getModelCapabilities() {
-			const model = this.tryResolveRawModel();
-			return model === void 0 ? UNKNOWN_CAPABILITY : modelCapabilities(model);
+			return this.capabilitiesFor(this.tryResolveRawModel());
+		}
+		/**
+		* Model capabilities with the user's `[loop_control] context_size_percent`
+		* applied to the token limits. The percentage is read on every call, so
+		* config changes take effect on the next request without a restart.
+		*/
+		capabilitiesFor(model) {
+			return scaleModelCapabilityContext(model === void 0 ? UNKNOWN_CAPABILITY : modelCapabilities(model), this.config.get(LOOP_CONTROL_SECTION)?.contextSizePercent);
 		}
 		getMaxOutputSize() {
 			return this.tryResolveRawModel()?.maxTokens;
@@ -214622,7 +214626,6 @@ var init_sessionSwarmService = __esmMin((() => {
 	init_abort();
 	init_profile();
 	init_permissionMode();
-	init_loop();
 	init_userTool();
 	init_eventBus();
 	init_sessionAgentProfileCatalog();
@@ -214751,7 +214754,6 @@ var init_sessionSwarmService = __esmMin((() => {
 			await this.requireOwnedSubagent(callerAgentId, agentId);
 			const caller = this.requireHandle(callerAgentId, "Caller agent");
 			const child = this.requireHandle(agentId, "Agent instance");
-			this.requireIdleSubagent(agentId, child);
 			const profileName = child.accessor.get(IAgentProfileService).data().profileName ?? RESUMED_PROFILE_FALLBACK;
 			if (!retryTurn) emitAgentRunSpawned(caller, agentId, {
 				profileName,
@@ -214765,15 +214767,16 @@ var init_sessionSwarmService = __esmMin((() => {
 				kind: "prompt",
 				prompt: options.prompt
 			};
-			return this.observe(caller, child.id, profileName, request, options);
+			return this.observe(caller, child.id, profileName, request, options, !retryTurn);
 		}
-		async observe(caller, agentId, profileName, request, options) {
+		async observe(caller, agentId, profileName, request, options, steer = false) {
 			return {
 				agentId,
 				profileName,
 				completion: mirrorAgentRun(caller, await this.subagents.run(agentId, request, {
 					signal: options.signal,
-					onReady: options.onReady
+					onReady: options.onReady,
+					steer
 				}), {
 					profileName,
 					prompt: request.kind === "prompt" ? request.prompt : void 0,
@@ -214790,9 +214793,6 @@ var init_sessionSwarmService = __esmMin((() => {
 			if (handle === void 0) throw new Error(`${label} "${agentId}" does not exist`);
 			return handle;
 		}
-		requireIdleSubagent(agentId, child) {
-			if (child.accessor.get(IAgentLoopService).status().state === "running") throw new Error(`Agent instance "${agentId}" is already running and cannot run concurrently`);
-		}
 		async requireOwnedSubagent(callerAgentId, agentId) {
 			const meta = await this.agentMeta(agentId);
 			if (!isSubagentMeta(meta)) throw new Error(`Agent instance "${agentId}" is not a subagent`);
@@ -217745,6 +217745,7 @@ var init_src$6 = __esmMin((() => {
 	init_retry();
 	init_configSection$5();
 	init_configSection$5();
+	init_contextSize$1();
 	init_loop();
 	init_loopService();
 	init_loopContinuation();
@@ -309793,6 +309794,134 @@ var init_busy_input_mode_selector = __esmMin((() => {
 	};
 }));
 //#endregion
+//#region src/utils/usage/usage-format.ts
+function usageNumber$1(value) {
+	return typeof value === "number" && Number.isFinite(value) && value > 0 ? value : 0;
+}
+/**
+* Total prompt tokens for a single generation: non-cache input + cache
+* read + cache write/creation. Matches the denominator pi uses for CH%.
+*/
+function promptTokenTotal(usage) {
+	if (usage === null || usage === void 0) return 0;
+	return usageNumber$1(usage.inputOther) + usageNumber$1(usage.inputCacheRead) + usageNumber$1(usage.inputCacheCreation);
+}
+/**
+* Latest-step prompt-cache hit rate as a percentage in [0, 100].
+* Returns `undefined` when there is no prompt traffic, or when the
+* provider reported neither cache reads nor cache writes (no cache
+* signal — same gate as pi's footer `CH` badge).
+*/
+function cacheHitRatePercent(usage) {
+	if (usage === null || usage === void 0) return void 0;
+	const cacheRead = usageNumber$1(usage.inputCacheRead);
+	const cacheWrite = usageNumber$1(usage.inputCacheCreation);
+	if (cacheRead === 0 && cacheWrite === 0) return void 0;
+	const prompt = promptTokenTotal(usage);
+	if (prompt <= 0) return void 0;
+	return Math.min(100, Math.max(0, cacheRead / prompt * 100));
+}
+/** Compact footer/report label, e.g. `CH85.0%`. */
+function formatCacheHitRate(rate) {
+	if (!Number.isFinite(rate)) return "CH?%";
+	return `CH${rate.toFixed(1)}%`;
+}
+/**
+* Format a token count in 1024-based units: context sizes are powers of
+* two, so 262144 reads as "256k", not "262.1k". k values at or above
+* 100 are rounded to whole numbers ("977k").
+*/
+function formatTokenCount(n) {
+	if (!Number.isFinite(n) || n < 0) return "0";
+	if (n >= 1024 * 1024) return `${trimDecimal(n / (1024 * 1024))}M`;
+	if (n >= 1024) {
+		const k = n / 1024;
+		return `${k >= 100 ? Math.round(k) : trimDecimal(k)}k`;
+	}
+	return String(n);
+}
+/** One decimal place, dropping a redundant ".0" ("1.0" → "1", "1.5" stays). */
+function trimDecimal(v) {
+	const s = v.toFixed(1);
+	return s.endsWith(".0") ? s.slice(0, -2) : s;
+}
+/**
+* Format a token count in 1000-based units ("1M", "950k", "200k"). Unlike
+* `formatTokenCount` (1024-based, the footer/usage convention), this keeps
+* exact product numbers readable: the context-size picker's floor is a
+* literal 200k, and 5% steps of a 1M window land on round values.
+*/
+function formatDecimalTokenCount(n) {
+	if (!Number.isFinite(n) || n < 0) return "0";
+	if (n >= 1e6) return `${trimDecimal(n / 1e6)}M`;
+	if (n >= 1e3) return `${trimDecimal(n / 1e3)}k`;
+	return String(n);
+}
+/**
+* Usage as a whole-number percentage of `max`, ceiled so any non-zero
+* usage shows at least 1%, clamped to [0, 100]. A non-positive or
+* non-finite `max` reports 0.
+*/
+function usagePercent(used, max) {
+	if (!Number.isFinite(max) || max <= 0) return 0;
+	return Math.min(100, Math.max(0, Math.ceil(used / max * 100)));
+}
+/** `usagePercent` for callers that only know the ratio (NaN-safe). */
+function usagePercentFromRatio(ratio) {
+	return Math.min(100, Math.max(0, Math.ceil(safeUsageRatio(ratio) * 100)));
+}
+/**
+* Build a `[███░░░░░░░]` style bar. Returns a plain-ASCII string with
+* `filled`/`empty` glyphs — colouring is the caller's responsibility.
+*/
+function renderProgressBar(ratio, width = 20, filled = "█", empty = "░") {
+	const clamped = safeUsageRatio(ratio);
+	const filledCount = Math.round(clamped * width);
+	return filled.repeat(filledCount) + empty.repeat(Math.max(0, width - filledCount));
+}
+function safeUsageRatio(ratio) {
+	return Number.isFinite(ratio) ? Math.max(0, Math.min(ratio, 1)) : 0;
+}
+/**
+* Map a usage ratio to a semantic colour token — the `/usage` renderer
+* translates these into palette hex values.
+*/
+function ratioSeverity(ratio) {
+	if (ratio >= .85) return "danger";
+	if (ratio >= .5) return "warn";
+	return "ok";
+}
+var init_usage_format = __esmMin((() => {}));
+//#endregion
+//#region src/tui/components/dialogs/context-size-selector.ts
+function buildOptions(contextWindow, percentOptions) {
+	return percentOptions.map((percent) => ({
+		value: String(percent),
+		label: percent === 100 ? "100% (default)" : `${percent}%`,
+		description: `${formatDecimalTokenCount(Math.floor(contextWindow * percent / 100))} tokens`
+	}));
+}
+var ContextSizeSelectorComponent;
+var init_context_size_selector = __esmMin((() => {
+	init_src$6();
+	init_usage_format();
+	init_choice_picker();
+	ContextSizeSelectorComponent = class extends ChoicePickerComponent {
+		constructor(opts) {
+			super({
+				title: "Context size",
+				hint: `Model window ${formatDecimalTokenCount(opts.contextWindow)} · floor ${formatDecimalTokenCount(CONTEXT_SIZE_FLOOR_TOKENS)}`,
+				options: buildOptions(opts.contextWindow, opts.percentOptions),
+				currentValue: String(opts.currentPercent),
+				onSelect: (value) => {
+					opts.onSelect(Number(value));
+				},
+				onCancel: opts.onCancel
+			});
+		}
+	};
+}));
+//#endregion
 //#region src/tui/components/dialogs/editor-selector.ts
 var EDITOR_OPTIONS, EditorSelectorComponent;
 var init_editor_selector = __esmMin((() => {
@@ -310253,7 +310382,7 @@ var init_permission_selector = __esmMin((() => {
 //#endregion
 //#region src/tui/components/dialogs/settings-selector.ts
 function isSettingsSelection(value) {
-	return value === "model" || value === "theme" || value === "editor" || value === "permission" || value === "busy-input" || value === "experiments" || value === "upgrade" || value === "usage";
+	return value === "model" || value === "theme" || value === "editor" || value === "permission" || value === "busy-input" || value === "context-size" || value === "experiments" || value === "upgrade" || value === "usage";
 }
 var SETTINGS_OPTIONS, SettingsSelectorComponent;
 var init_settings_selector = __esmMin((() => {
@@ -310284,6 +310413,11 @@ var init_settings_selector = __esmMin((() => {
 			label: "Busy input",
 			description: "Choose whether Enter queues or steers while the agent is working."
 		},
+		{
+			value: "context-size",
+			label: "Context size",
+			description: "Cap the conversation window as a percentage of the model default (min 200k)."
+		},
 		{
 			value: "experiments",
 			label: "Experiments",
@@ -310529,93 +310663,6 @@ var init_mcp_status_panel = __esmMin((() => {
 	];
 }));
 //#endregion
-//#region src/utils/usage/usage-format.ts
-function usageNumber$1(value) {
-	return typeof value === "number" && Number.isFinite(value) && value > 0 ? value : 0;
-}
-/**
-* Total prompt tokens for a single generation: non-cache input + cache
-* read + cache write/creation. Matches the denominator pi uses for CH%.
-*/
-function promptTokenTotal(usage) {
-	if (usage === null || usage === void 0) return 0;
-	return usageNumber$1(usage.inputOther) + usageNumber$1(usage.inputCacheRead) + usageNumber$1(usage.inputCacheCreation);
-}
-/**
-* Latest-step prompt-cache hit rate as a percentage in [0, 100].
-* Returns `undefined` when there is no prompt traffic, or when the
-* provider reported neither cache reads nor cache writes (no cache
-* signal — same gate as pi's footer `CH` badge).
-*/
-function cacheHitRatePercent(usage) {
-	if (usage === null || usage === void 0) return void 0;
-	const cacheRead = usageNumber$1(usage.inputCacheRead);
-	const cacheWrite = usageNumber$1(usage.inputCacheCreation);
-	if (cacheRead === 0 && cacheWrite === 0) return void 0;
-	const prompt = promptTokenTotal(usage);
-	if (prompt <= 0) return void 0;
-	return Math.min(100, Math.max(0, cacheRead / prompt * 100));
-}
-/** Compact footer/report label, e.g. `CH85.0%`. */
-function formatCacheHitRate(rate) {
-	if (!Number.isFinite(rate)) return "CH?%";
-	return `CH${rate.toFixed(1)}%`;
-}
-/**
-* Format a token count in 1024-based units: context sizes are powers of
-* two, so 262144 reads as "256k", not "262.1k". k values at or above
-* 100 are rounded to whole numbers ("977k").
-*/
-function formatTokenCount(n) {
-	if (!Number.isFinite(n) || n < 0) return "0";
-	if (n >= 1024 * 1024) return `${trimDecimal(n / (1024 * 1024))}M`;
-	if (n >= 1024) {
-		const k = n / 1024;
-		return `${k >= 100 ? Math.round(k) : trimDecimal(k)}k`;
-	}
-	return String(n);
-}
-/** One decimal place, dropping a redundant ".0" ("1.0" → "1", "1.5" stays). */
-function trimDecimal(v) {
-	const s = v.toFixed(1);
-	return s.endsWith(".0") ? s.slice(0, -2) : s;
-}
-/**
-* Usage as a whole-number percentage of `max`, ceiled so any non-zero
-* usage shows at least 1%, clamped to [0, 100]. A non-positive or
-* non-finite `max` reports 0.
-*/
-function usagePercent(used, max) {
-	if (!Number.isFinite(max) || max <= 0) return 0;
-	return Math.min(100, Math.max(0, Math.ceil(used / max * 100)));
-}
-/** `usagePercent` for callers that only know the ratio (NaN-safe). */
-function usagePercentFromRatio(ratio) {
-	return Math.min(100, Math.max(0, Math.ceil(safeUsageRatio(ratio) * 100)));
-}
-/**
-* Build a `[███░░░░░░░]` style bar. Returns a plain-ASCII string with
-* `filled`/`empty` glyphs — colouring is the caller's responsibility.
-*/
-function renderProgressBar(ratio, width = 20, filled = "█", empty = "░") {
-	const clamped = safeUsageRatio(ratio);
-	const filledCount = Math.round(clamped * width);
-	return filled.repeat(filledCount) + empty.repeat(Math.max(0, width - filledCount));
-}
-function safeUsageRatio(ratio) {
-	return Number.isFinite(ratio) ? Math.max(0, Math.min(ratio, 1)) : 0;
-}
-/**
-* Map a usage ratio to a semantic colour token — the `/usage` renderer
-* translates these into palette hex values.
-*/
-function ratioSeverity(ratio) {
-	if (ratio >= .85) return "danger";
-	if (ratio >= .5) return "warn";
-	return "ok";
-}
-var init_usage_format = __esmMin((() => {}));
-//#endregion
 //#region src/tui/components/messages/usage-panel.ts
 function usageRowLabel(row) {
 	const window = row.window;
@@ -312380,6 +312427,62 @@ function showBusyInputModePicker(host) {
 		}
 	}));
 }
+/** `[loop_control]` from a `getConfig()` result (untyped domain access). */
+function loopControlFromConfig(config) {
+	return config["loopControl"];
+}
+/** The active model's default context window in tokens, or 0 when unknown. */
+function activeModelContextWindow(host) {
+	const alias = host.state.appState.model;
+	return (alias === void 0 ? void 0 : host.state.appState.availableModels[alias])?.maxContextSize ?? 0;
+}
+/**
+* Open the context-size picker: 5% steps from the model's default window
+* (100%) down while the scaled window stays at or above 200k. Models whose
+* window is already below 200k are not adjustable — a notice explains why
+* instead of opening a one-option picker.
+*/
+function showContextSizePicker(host) {
+	const contextWindow = activeModelContextWindow(host);
+	if (contextWindow <= 0) {
+		host.showNotice("Context size unavailable", "No model is active. Choose a model first.");
+		return;
+	}
+	const percentOptions = contextSizePercentOptions(contextWindow);
+	if (percentOptions.length <= 1) {
+		host.showNotice("Context size cannot be adjusted", `This model's window (${formatDecimalTokenCount(contextWindow)}) is already at or below the ${formatDecimalTokenCount(CONTEXT_SIZE_FLOOR_TOKENS)} floor.`);
+		return;
+	}
+	host.harness.getConfig().then((config) => {
+		const currentPercent = loopControlFromConfig(config)?.contextSizePercent ?? 100;
+		host.mountEditorReplacement(new ContextSizeSelectorComponent({
+			contextWindow,
+			percentOptions,
+			currentPercent,
+			onSelect: (percent) => {
+				host.restoreEditor();
+				applyContextSizeChoice(host, percent);
+			},
+			onCancel: () => {
+				host.restoreEditor();
+			}
+		}));
+	}, (error) => {
+		host.showError(`Failed to load context size: ${formatErrorMessage$2(error)}`);
+	});
+}
+async function applyContextSizeChoice(host, percent) {
+	const contextWindow = activeModelContextWindow(host);
+	try {
+		await host.harness.setConfig({ loopControl: { contextSizePercent: percent } });
+	} catch (error) {
+		host.showError(`Failed to save context size: ${formatErrorMessage$2(error)}`);
+		return;
+	}
+	const maxContextTokens = contextWindow > 0 ? scaleContextTokens(contextWindow, percent) : 0;
+	if (maxContextTokens > 0) host.setAppState({ maxContextTokens });
+	host.showNotice(`Context size: ${percent}%`, `Effective window ${formatDecimalTokenCount(maxContextTokens)} (default ${formatDecimalTokenCount(contextWindow)}).`);
+}
 async function showExperimentsPanel(host) {
 	let features;
 	try {
@@ -312524,6 +312627,9 @@ function handleSettingsSelection(host, value) {
 		case "busy-input":
 			showBusyInputModePicker(host);
 			return;
+		case "context-size":
+			showContextSizePicker(host);
+			return;
 		case "experiments":
 			showExperimentsPanel(host);
 			return;
@@ -312539,6 +312645,7 @@ var MODEL_PICKER_REFRESH_TIMEOUT_MS, MODEL_SWITCH_CACHE_WARNING, EFFORT_SWITCH_C
 var init_config = __esmMin((() => {
 	init_src$4();
 	init_busy_input_mode_selector();
+	init_context_size_selector();
 	init_editor_selector();
 	init_effort_selector();
 	init_experiments_selector();
@@ -312553,6 +312660,8 @@ var init_config = __esmMin((() => {
 	init_dimi_tui$1();
 	init_event_payload();
 	init_thinking_config();
+	init_src$6();
+	init_usage_format();
 	init_info();
 	init_experimental_flags();
 	MODEL_PICKER_REFRESH_TIMEOUT_MS = 2e3;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@dimi-agent/cli",
-  "version": "0.5.4",
+  "version": "0.6.0",
   "description": "The Starting Point for Next-Gen Agents",
   "keywords": [
     "agent",