@themoltnet/pi-extension 0.27.2 → 0.28.0

package/dist/index.d.ts

@@ -40,6 +40,16 @@ declare interface BuildAgentSessionArgs {
     piAuthDir: string;
     /** Resolved pi model handle (provider + model id). */
     modelHandle: Model<Api>;
+    /** Optional runtime-profile thinking/reasoning level applied at session start. */
+    thinkingLevel?: PiThinkingLevel | null;
+    /** Optional runtime-profile sampling temperature applied to provider requests. */
+    temperature?: number | null;
+    /** Optional runtime-profile nucleus sampling probability mass. */
+    topP?: number | null;
+    /** Optional runtime-profile top-k sampling cutoff. */
+    topK?: number | null;
+    /** Optional runtime-profile generated output token cap. */
+    maxOutputTokens?: number | null;
     /** Pre-built customTools array. Caller composes Gondolin + MoltNet + submit tools. */
     customTools: ToolDefinition[];
     /** System-prompt fragments appended after pi's defaults. Parent passes the
@@ -157,6 +167,16 @@ export declare interface CreateSubagentToolArgs {
     piAuthDir: string;
     /** Resolved pi model handle — subagents share it. */
     modelHandle: Model<Api>;
+    /** Runtime-profile thinking/reasoning level — subagents inherit it. */
+    thinkingLevel?: PiThinkingLevel | null;
+    /** Runtime-profile sampling temperature — subagents inherit it. */
+    temperature?: number | null;
+    /** Runtime-profile nucleus sampling probability mass — subagents inherit it. */
+    topP?: number | null;
+    /** Runtime-profile top-k sampling cutoff — subagents inherit it. */
+    topK?: number | null;
+    /** Runtime-profile generated output token cap — subagents inherit it. */
+    maxOutputTokens?: number | null;
     /** Agent name for telemetry. */
     agentName: string;
     /**
@@ -256,6 +276,20 @@ export declare interface ExecutePiTaskOptions {
     /** LLM selection. */
     provider: string;
     model: string;
+    /**
+     * Runtime-profile reasoning/thinking level. Null/undefined means use Pi's
+     * configured default; explicit `off` disables provider thinking where
+     * supported.
+     */
+    thinkingLevel?: PiThinkingLevel | null;
+    /** Optional sampling temperature. Null/undefined means provider default. */
+    temperature?: number | null;
+    /** Optional nucleus-sampling probability mass. Null/undefined means provider default. */
+    topP?: number | null;
+    /** Optional top-k sampling cutoff. Null/undefined means provider default. */
+    topK?: number | null;
+    /** Optional cap on generated output tokens. Null/undefined means provider/model default. */
+    maxOutputTokens?: number | null;
     /** Extra hosts to allow in the sandbox egress policy. */
     extraAllowedHosts?: string[];
     /** Sandbox overrides (env, VFS shadows, resources). */
@@ -544,6 +578,8 @@ export declare interface PiTaskExecutionPlan {
 
 export declare type PiTaskExecutionPlanFactory = (claimedTask: ClaimedTask) => Promise<PiTaskExecutionPlan | null> | PiTaskExecutionPlan | null;
 
+declare type PiThinkingLevel = 'off' | 'minimal' | 'low' | 'medium' | 'high' | 'xhigh';
+
 declare interface PiWorkspaceAttachmentPlan {
     mountPath: string;
     cwdPath: string;

package/dist/index.js

@@ -1868,6 +1868,28 @@ var uploadRuntimeSession = (options) => (options.client ?? client).put({
 	}
 });
 /**
+* List recent team-scoped runtime slots for repair/sync.
+*/
+var listRuntimeSlots = (options) => (options.client ?? client).get({
+	security: [
+		{
+			scheme: "bearer",
+			type: "http"
+		},
+		{
+			name: "X-Moltnet-Session-Token",
+			type: "apiKey"
+		},
+		{
+			in: "cookie",
+			name: "ory_kratos_session",
+			type: "apiKey"
+		}
+	],
+	url: "/runtime-slots",
+	...options
+});
+/**
 * Upsert a team-scoped runtime slot for audit and continuation affinity lookup.
 */
 var beginRuntimeSlot = (options) => (options.client ?? client).post({
@@ -5171,6 +5193,15 @@ function createRuntimeSlotsNamespace(context) {
 				if (err instanceof MoltNetError && err.statusCode === 404) return null;
 				throw err;
 			}
+		},
+		async list(query, options) {
+			const filteredQuery = Object.fromEntries(Object.entries(query).filter(([, value]) => value !== void 0));
+			return unwrapResult(await listRuntimeSlots({
+				auth,
+				client,
+				headers: requiredTeamHeaders(options),
+				query: filteredQuery
+			})).items;
 		}
 	};
 }
@@ -9520,6 +9551,32 @@ var RuntimeProfileAllowedWorkspaceModes = _Array_(RuntimeProfileWorkspaceMode, {
 	maxItems: 3,
 	uniqueItems: true
 });
+var RuntimeProfileThinkingLevelOptions = [
+	Literal("off"),
+	Literal("minimal"),
+	Literal("low"),
+	Literal("medium"),
+	Literal("high"),
+	Literal("xhigh")
+];
+Union([...RuntimeProfileThinkingLevelOptions]);
+var RuntimeProfileNullableThinkingLevel = Union([...RuntimeProfileThinkingLevelOptions, Null()]);
+var RuntimeProfileNullableTemperature = Union([Null(), Number$1({
+	minimum: 0,
+	maximum: 2
+})]);
+var RuntimeProfileNullableTopP = Union([Null(), Number$1({
+	minimum: 0,
+	maximum: 1
+})]);
+var RuntimeProfileNullableTopK = Union([Integer({
+	minimum: 1,
+	maximum: 1e4
+}), Null()]);
+var RuntimeProfileNullableMaxOutputTokens = Union([Integer({
+	minimum: 1,
+	maximum: 1e6
+}), Null()]);
 var SandboxResumeCommandWhenSchema = _Object_({ workspaceMode: Optional(_Array_(Union([
 	Literal("shared_mount"),
 	Literal("dedicated_worktree"),
@@ -9644,6 +9701,11 @@ _Object_({
 		minLength: 1,
 		maxLength: 200
 	}),
+	thinkingLevel: RuntimeProfileNullableThinkingLevel,
+	temperature: RuntimeProfileNullableTemperature,
+	topP: RuntimeProfileNullableTopP,
+	topK: RuntimeProfileNullableTopK,
+	maxOutputTokens: RuntimeProfileNullableMaxOutputTokens,
 	runtimeKind: Literal("gondolin_pi"),
 	sandbox: RuntimeProfileSandbox,
 	sessionStorageMode: Literal("local"),
@@ -9755,7 +9817,7 @@ var RuntimeWorkspace = _Object_({
 	createdAtMs: Integer({ minimum: 0 }),
 	lastUsedAtMs: Integer({ minimum: 0 })
 }, { $id: "RuntimeWorkspace" });
-_Object_({
+_Object_({ items: _Array_(_Object_({
 	slot: _Object_({
 		id: String$1({ format: "uuid" }),
 		teamId: String$1({ format: "uuid" }),
@@ -9788,7 +9850,7 @@ _Object_({
 		expiresAtMs: Integer({ minimum: 0 })
 	}, { $id: "RuntimeSlot" }),
 	workspace: Union([RuntimeWorkspace, Null()])
-}, { $id: "ResolvedRuntimeSlot" });
+}, { $id: "ResolvedRuntimeSlot" })) }, { $id: "RuntimeSlotListResponse" });
 _Object_({
 	agentName: String$1({
 		minLength: 1,
@@ -9849,6 +9911,21 @@ _Object_({
 	$id: "FindLatestRuntimeSlotForAttemptQuery",
 	additionalProperties: false
 });
+_Object_({
+	agentName: Optional(String$1({
+		minLength: 1,
+		maxLength: 100
+	})),
+	runtimeProfileId: Optional(String$1({ format: "uuid" })),
+	state: Optional(RuntimeSlotState),
+	limit: Optional(Integer({
+		minimum: 1,
+		maximum: 200
+	}))
+}, {
+	$id: "ListRuntimeSlotsQuery",
+	additionalProperties: false
+});
 //#endregion
 //#region ../tasks/src/success-criteria.ts
 /**
@@ -10269,6 +10346,7 @@ var FreeformArtifact = _Object_({
 });
 var FreeformOutput = _Object_({
 	summary: String$1({ minLength: 1 }),
+	branch: Optional(String$1({ minLength: 1 })),
 	artifacts: Optional(_Array_(FreeformArtifact, { maxItems: 20 })),
 	proposedTaskType: Optional(FreeformTaskTypeProposal),
 	diaryEntryIds: Optional(_Array_(String$1({ format: "uuid" }))),
@@ -10281,7 +10359,7 @@ var FreeformOutput = _Object_({
 * Server-side preflight for `freeform` task-create. Runs after the
 * sync TypeBox check passes and only kicks in when
 * `input.continueFrom` is set — i.e. the proposer is asking to
-* resume a prior freeform attempt's warm slot (#1287).
+* continue from a prior freeform attempt (#1287).
 *
 * Failure modes, in evaluation order:
 *  1. `freeform.sourceTaskNotFound` — source task id does not resolve
@@ -10289,23 +10367,17 @@ var FreeformOutput = _Object_({
 *  2. `freeform.sourceTaskTypeNotSupported` — source isn't `freeform`.
 *     v1 only supports freeform → freeform continuation.
 *  3. `freeform.sourceAttemptNotCompleted` — named attempt is missing
-*     or not in `completed` state; warm continuation only makes sense
+*     or not in `completed` state; continuation only makes sense
 *     once the parent has produced a terminal output.
 *  4. `freeform.executionWorkspaceNotInheritable` — caller set
 *     `execution.workspace` together with `continueFrom`. Workspace
-*     mode for a continuation is inherited from the parent slot
-*     (`maybeAttachWarmSlotContext` forces `dedicated_worktree` +
-*     the parent's worktreeBranch), so any caller-supplied override
-*     is silently dropped at the daemon plan stage. Reject explicitly
-*     so misconfiguration surfaces at create time.
-*  5. `freeform.sourceNotResumeEligible` — `daemonState` is null or
-*     `slotResumableUntil` is null. Older completions (pre-#1287) and
-*     daemons that opt out fall here.
-*  6. `freeform.sourceResumeExpired` — `slotResumableUntil` is in the
-*     past; the warm slot's TTL has elapsed and no daemon is
-*     guaranteed to still hold it.
+*     mode for a continuation is derived by the daemon from parent runtime
+*     context (local slot first, durable session + source attempt branch
+*     second), so any caller-supplied override is silently dropped at the
+*     daemon plan stage. Reject explicitly so misconfiguration surfaces at
+*     create time.
 *
-* Returns on the first failure (no "report all six") — the checks
+* Returns on the first failure — the checks
 * are sequential preconditions, later ones presume earlier ones hold.
 */
 async function validateFreeformInputAsync(input, ctx) {
@@ -10324,7 +10396,7 @@ async function validateFreeformInputAsync(input, ctx) {
 	}];
 	if (input.execution?.workspace) return [{
 		field: "input/execution/workspace",
-		message: "execution.workspace is inherited from the parent slot when continueFrom is set; omit it",
+		message: "execution.workspace is derived from parent runtime context when continueFrom is set; omit it",
 		code: "freeform.executionWorkspaceNotInheritable"
 	}];
 	if (ctx.deferReadinessChecks) return [];
@@ -10334,17 +10406,6 @@ async function validateFreeformInputAsync(input, ctx) {
 		message: `Source attempt ${cf.attemptN} on task ${cf.taskId} is not in 'completed' state`,
 		code: "freeform.sourceAttemptNotCompleted"
 	}];
-	if (!attempt.daemonState || attempt.daemonState.slotResumableUntil === null) return [{
-		field: "input/continueFrom",
-		message: "Source attempt did not report continuation eligibility (older completion or daemon opted out)",
-		code: "freeform.sourceNotResumeEligible"
-	}];
-	const expiresAt = new Date(attempt.daemonState.slotResumableUntil).getTime();
-	if (Number.isNaN(expiresAt) || expiresAt <= Date.now()) return [{
-		field: "input/continueFrom",
-		message: `Source attempt's warm slot expired at ${attempt.daemonState.slotResumableUntil} (reported at ${attempt.daemonState.reportedAt})`,
-		code: "freeform.sourceResumeExpired"
-	}];
 	return [];
 }
 //#endregion
@@ -13688,11 +13749,12 @@ var MAX_CLAIM_CONDITION_STATUSES = 8;
 /**
 * Daemon-asserted runtime state stamped onto a `TaskAttemptSummary` at
 * attempt-completion time. The server persists this block verbatim and
-* reads `slotResumableUntil` for `tasks_continue` create-time
-* eligibility; the daemon-side claim-affinity filter is the runtime
-* truth. The block carries its own `reportedAt` so consumers can reason
-* about staleness without reading documentation. All daemon-asserted
-* state lives here — top-level attempt fields stay server-authoritative.
+* exposes `slotResumableUntil` as a legacy/local warm-slot hint; task
+* continuation eligibility is based on the completed source attempt and
+* daemon-side claim-affinity/runtime-session recovery. The block carries
+* its own `reportedAt` so consumers can reason about staleness without
+* reading documentation. All daemon-asserted state lives here —
+* top-level attempt fields stay server-authoritative.
 *
 * Adding new fields requires explicit design review (intentional
 * boundary; see docs/superpowers/specs/2026-06-04-tasks-continue-design.md).
@@ -18770,6 +18832,89 @@ function extractUsage(message) {
 	};
 }
 //#endregion
+//#region src/runtime/model-options-extension.ts
+function hasPiModelOptions(options) {
+	return options.temperature !== void 0 && options.temperature !== null || options.topP !== void 0 && options.topP !== null || options.topK !== void 0 && options.topK !== null || options.maxOutputTokens !== void 0 && options.maxOutputTokens !== null;
+}
+function createPiModelOptionsExtension(options) {
+	return function piModelOptionsExtension(pi) {
+		pi.on("before_provider_request", (event) => {
+			return applyPiModelOptions(event.payload, options);
+		});
+	};
+}
+function applyPiModelOptions(payload, options) {
+	if (!isRecord(payload)) return void 0;
+	if (!hasPiModelOptions(options)) return void 0;
+	if (isGooglePayload(payload)) {
+		const config = isRecord(payload.config) ? payload.config : {};
+		return {
+			...payload,
+			config: applyConfigOptions(config, options)
+		};
+	}
+	if (isBedrockPayload(payload)) {
+		const inferenceConfig = isRecord(payload.inferenceConfig) ? payload.inferenceConfig : {};
+		return {
+			...payload,
+			inferenceConfig: applyBedrockOptions(inferenceConfig, options)
+		};
+	}
+	return applyTopLevelOptions(payload, options);
+}
+function applyConfigOptions(config, options) {
+	return {
+		...config,
+		...options.temperature !== void 0 && options.temperature !== null ? { temperature: options.temperature } : {},
+		...options.topP !== void 0 && options.topP !== null ? { topP: options.topP } : {},
+		...options.topK !== void 0 && options.topK !== null ? { topK: options.topK } : {},
+		...options.maxOutputTokens !== void 0 && options.maxOutputTokens !== null ? { maxOutputTokens: options.maxOutputTokens } : {}
+	};
+}
+function applyBedrockOptions(inferenceConfig, options) {
+	return {
+		...inferenceConfig,
+		...options.temperature !== void 0 && options.temperature !== null ? { temperature: options.temperature } : {},
+		...options.topP !== void 0 && options.topP !== null ? { topP: options.topP } : {},
+		...options.maxOutputTokens !== void 0 && options.maxOutputTokens !== null ? { maxTokens: options.maxOutputTokens } : {}
+	};
+}
+function applyTopLevelOptions(payload, options) {
+	const reasoningEnabled = hasActiveThinking(payload.thinking) || "reasoning" in payload || "reasoning_effort" in payload;
+	const next = { ...payload };
+	if (options.temperature !== void 0 && options.temperature !== null && !reasoningEnabled) next.temperature = options.temperature;
+	if (options.topP !== void 0 && options.topP !== null && !reasoningEnabled) next.top_p = options.topP;
+	if (options.topK !== void 0 && options.topK !== null && !reasoningEnabled && isAnthropicPayload(next)) next.top_k = options.topK;
+	if (options.maxOutputTokens !== void 0 && options.maxOutputTokens !== null) {
+		const maxOutputTokens = options.maxOutputTokens;
+		if ("max_output_tokens" in next || isResponsesPayload(next)) next.max_output_tokens = maxOutputTokens;
+		else if ("max_completion_tokens" in next) next.max_completion_tokens = maxOutputTokens;
+		else if ("maxTokens" in next) next.maxTokens = maxOutputTokens;
+		else next.max_tokens = maxOutputTokens;
+	}
+	return next;
+}
+function isGooglePayload(payload) {
+	return "contents" in payload && ("config" in payload || "model" in payload);
+}
+function isBedrockPayload(payload) {
+	return "inferenceConfig" in payload || "additionalModelRequestFields" in payload;
+}
+function isResponsesPayload(payload) {
+	return "input" in payload && !("messages" in payload);
+}
+function isAnthropicPayload(payload) {
+	return "anthropic_version" in payload;
+}
+function hasActiveThinking(value) {
+	if (!isRecord(value)) return false;
+	const type = value.type;
+	return type !== "disabled" && type !== "off" && type !== false;
+}
+function isRecord(value) {
+	return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+//#endregion
 //#region src/runtime/agent-session-factory.ts
 var NO_SKILLS = () => ({
 	skills: [],
@@ -18787,10 +18932,17 @@ async function buildAgentSession(args) {
 		agentName: args.agentName,
 		spanAttributes: args.otelSpanAttrs
 	});
+	const modelOptions = {
+		temperature: args.temperature,
+		topP: args.topP,
+		topK: args.topK,
+		maxOutputTokens: args.maxOutputTokens
+	};
+	const extensionFactories = hasPiModelOptions(modelOptions) ? [piOtelExtension, createPiModelOptionsExtension(modelOptions)] : [piOtelExtension];
 	const resourceLoader = new DefaultResourceLoader({
 		cwd: args.cwdPath,
 		agentDir: args.piAuthDir,
-		extensionFactories: [piOtelExtension],
+		extensionFactories,
 		appendSystemPrompt: args.appendSystemPrompt,
 		skillsOverride: args.skillsOverride ?? NO_SKILLS
 	});
@@ -18804,6 +18956,7 @@ async function buildAgentSession(args) {
 		agentDir: args.piAuthDir,
 		cwd: args.cwdPath,
 		model: args.modelHandle,
+		thinkingLevel: args.thinkingLevel ?? void 0,
 		customTools: args.customTools,
 		sessionManager,
 		resourceLoader
@@ -19445,7 +19598,9 @@ function buildFreeformUserPrompt(input, ctx) {
 		"2. Gather enough context to avoid guessing.",
 		"3. Complete the requested work when it is safe and bounded.",
 		"4. If the request reveals a recurring task shape, include a",
-		"   `proposedTaskType` in the final output with a concise rationale."
+		"   `proposedTaskType` in the final output with a concise rationale.",
+		"5. If you changed code on a branch, include that branch in",
+		"   `branch` so future continuations can recover git context."
 	].join("\n");
 	const sections = [
 		{
@@ -19497,6 +19652,7 @@ function buildFreeformUserPrompt(input, ctx) {
 				shapeSketch: [
 					"{",
 					"  \"summary\": \"<2-5 sentence result>\",",
+					"  \"branch\": \"<branch name when code changed; omit for prose-only work>\",",
 					"  \"artifacts\": [{ \"kind\": \"...\", \"title\": \"...\", \"description\": \"...\", \"body\": \"<inline content up to 64 KiB; preferred for textual output so it persists with the task>\", \"url\": \"...\", \"path\": \"<worktree-ephemeral; not persisted after completion>\" }],",
 					"  \"proposedTaskType\": { \"name\": \"...\", \"rationale\": \"...\", \"inputShape\": {}, \"outputShape\": {} },",
 					"  \"diaryEntryIds\": [\"...\"],",
@@ -24064,6 +24220,11 @@ function createSubagentTool(args) {
 					cwdPath: args.cwdPath ?? args.mountPath,
 					piAuthDir: args.piAuthDir,
 					modelHandle: args.modelHandle,
+					thinkingLevel: args.thinkingLevel,
+					temperature: args.temperature,
+					topP: args.topP,
+					topK: args.topK,
+					maxOutputTokens: args.maxOutputTokens,
 					agentName: args.agentName,
 					customTools: [...args.inheritedCustomTools, submitTool],
 					appendSystemPrompt: [args.parentRuntimeInstructor, subagentInstructor],
@@ -24948,6 +25109,11 @@ async function executePiTask(claimedTask, reporter, opts) {
 					cwdPath,
 					piAuthDir,
 					modelHandle,
+					thinkingLevel: opts.thinkingLevel,
+					temperature: opts.temperature,
+					topP: opts.topP,
+					topK: opts.topK,
+					maxOutputTokens: opts.maxOutputTokens,
 					agentName: opts.agentName,
 					inheritedCustomTools: [...gondolinCustomTools, ...moltnetTools],
 					parentRuntimeInstructor: runtimeInstructor,
@@ -24964,6 +25130,11 @@ async function executePiTask(claimedTask, reporter, opts) {
 				cwdPath,
 				piAuthDir,
 				modelHandle,
+				thinkingLevel: opts.thinkingLevel,
+				temperature: opts.temperature,
+				topP: opts.topP,
+				topK: opts.topK,
+				maxOutputTokens: opts.maxOutputTokens,
 				agentName: opts.agentName,
 				customTools: [
 					...gondolinCustomTools,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@themoltnet/pi-extension",
-  "version": "0.27.2",
+  "version": "0.28.0",
   "type": "module",
   "description": "MoltNet pi extension — sandboxed tool execution in Gondolin VMs with MoltNet identity and persistent memory",
   "keywords": [
@@ -36,8 +36,8 @@
     "@earendil-works/gondolin": "^0.9.1",
     "@opentelemetry/api": "^1.9.0",
     "typebox": "^1.2.8",
-    "@themoltnet/agent-runtime": "0.30.1",
-    "@themoltnet/sdk": "0.113.1"
+    "@themoltnet/agent-runtime": "0.31.1",
+    "@themoltnet/sdk": "0.114.0"
   },
   "peerDependencies": {
     "@earendil-works/pi-coding-agent": ">=0.74.0",
@@ -61,8 +61,7 @@
     "vite": "^8.0.0",
     "vite-plugin-dts": "^4.5.4",
     "vitest": "^3.0.0",
-    "@moltnet/crypto-service": "0.1.0",
-    "@moltnet/tasks": "0.1.0"
+    "@moltnet/crypto-service": "0.1.0"
   },
   "engines": {
     "node": ">=22"