@iowarp/clio-coder 0.4.2 → 0.4.3

package/src/interactive/worker-stream.ts

@@ -62,7 +62,12 @@ export interface WorkerAttempt {
 export type WorkerResultContract = "pass" | "fail" | "not-reached" | "unmeasured";
 
 /** Result kinds whose parsed envelopes have a dedicated transcript presentation. */
-export type WorkerPresentedResultContract = "debugger-report" | "verifier-report" | "research-report" | "scout-report";
+export type WorkerPresentedResultContract =
+	| "debugger-report"
+	| "verifier-report"
+	| "research-report"
+	| "world-knowledge-report"
+	| "scout-report";
 
 /**
  * Terminal facts a worker block reports. Sealed by the receipt when it can be

package/src/tools/context/index.ts

@@ -181,20 +181,33 @@ function pendingSkillPolicyError(name: string, options: ToolInvokeOptions | unde
 	if (!policy) {
 		return NO_PENDING_SKILL_DENIAL;
 	}
+	const recipeBound = policyIsRecipeBound(policy);
+	// auto-edit / full-auto: the operator already chose to let the model act
+	// without being asked, and a skill only ever narrows the tool surface, so
+	// the model activates an installed skill itself under the same per-run
+	// policy `/skill` produces. A skill that is not installed still fails the
+	// lookup below with the operator-gated marketplace message.
+	if (policy.modelActivation === true && !recipeBound) {
+		if (!policy.loadedSkillNames.has(name)) return null;
+		const window = policy.carriedSurface === true ? "in this session" : "this turn";
+		return `context: skill ${name} already loaded ${window}; continue with the loaded workflow and call ask_user if an interview/choice is needed.`;
+	}
 	const allowed = [...new Set(policy.allowedSkillNames.map((entry) => entry.trim()).filter(Boolean))];
 	if (allowed.length === 0) {
 		return NO_PENDING_SKILL_DENIAL;
 	}
-	const recipeBound = policyIsRecipeBound(policy);
 	if (!allowed.includes(name)) {
-		return recipeBound
-			? `context: this agent run may load only its declared skill(s): ${allowed.join(", ")}.`
-			: `context: this turn has pending skill request(s): ${allowed.join(", ")}. Load only those before doing anything else.`;
+		if (recipeBound) return `context: this agent run may load only its declared skill(s): ${allowed.join(", ")}.`;
+		// A carried surface is a skill the operator activated on an earlier
+		// turn, not a request waiting to be loaded now. Claiming a pending
+		// request here would invite a retry of a load nothing asked for.
+		if (policy.carriedSurface === true) return NO_PENDING_SKILL_DENIAL;
+		return `context: this turn has pending skill request(s): ${allowed.join(", ")}. Load only those before doing anything else.`;
 	}
 	if (policy.loadedSkillNames.has(name)) {
-		return recipeBound
-			? `context: skill ${name} is already loaded in this run; continue with its workflow.`
-			: `context: pending skill ${name} already loaded this turn; continue with the loaded workflow and call ask_user if an interview/choice is needed.`;
+		if (recipeBound) return `context: skill ${name} is already loaded in this run; continue with its workflow.`;
+		const window = policy.carriedSurface === true ? "in this session" : "this turn";
+		return `context: pending skill ${name} already loaded ${window}; continue with the loaded workflow and call ask_user if an interview/choice is needed.`;
 	}
 	return null;
 }
@@ -274,6 +287,7 @@ function renderSkillsList(
 	skills: ReadonlyArray<Skill>,
 	marketplace: ReadonlyArray<MarketplaceSkill>,
 	marketplaceOffered: boolean,
+	modelActivation: boolean,
 ): string {
 	if (skills.length === 0 && marketplace.length === 0) {
 		// A registry that never offers the marketplace (a worker) must not
@@ -308,7 +322,9 @@ function renderSkillsList(
 	// template where they skip conditional prose in the header.
 	lines.push(
 		"",
-		`If one skill above matches the current task, begin your reply with the line \`${SKILL_SUGGESTION_ANCHOR}\` (a comma-separated sequence, in order, when several compose), then continue the task in the same turn without it; only the operator can run it. If none match, do not mention skills.`,
+		modelActivation
+			? `If one skill above matches the current task, load it now with context(scope="skills", name="<name>") and continue in the same turn; at this autonomy level you activate installed skills yourself and do not wait for the operator. Marketplace rows below are not installed and still need the operator. If none match, do not mention skills.`
+			: `If one skill above matches the current task, begin your reply with the line \`${SKILL_SUGGESTION_ANCHOR}\` (a comma-separated sequence, in order, when several compose), then continue the task in the same turn without it; only the operator can run it. If none match, do not mention skills.`,
 	);
 	if (marketplace.length > 0) {
 		// The offer protocol mirrors the marketplace-offer middleware: fixed
@@ -423,7 +439,12 @@ function runSkillsScope(
 		const list = loadSkills({ cwd: cwdFromDeps(deps), ...(deps.getSkillLoaderOptions?.() ?? {}) });
 		const visible = modelVisibleSkills(list.items);
 		const marketplace = marketplaceRowsFor(deps, list.items, options);
-		const rendered = renderSkillsList(visible, marketplace, deps.skillMarketplace !== false);
+		const rendered = renderSkillsList(
+			visible,
+			marketplace,
+			deps.skillMarketplace !== false,
+			options?.pendingSkillPolicy?.modelActivation === true,
+		);
 		// The catalog is bounded but not small: the whole listing must fit the
 		// per-call cap like any observation, and a cut list says so instead of
 		// silently dropping the marketplace tail.

package/src/tools/dispatch-arguments.ts

@@ -50,6 +50,7 @@ const AUTO_BASELINE_BY_TASK_TYPE: Readonly<Record<AgentTaskType, string>> = {
 	// Read-only shapes.
 	code_review: "verifier",
 	research: "researcher",
+	world_knowledge: "world-knowledge",
 	code_read: "scout",
 	unknown: "scout",
 };

package/src/tools/dispatch-event-text.ts

@@ -1,3 +1,4 @@
+import { gatewayRoutingObservationFromRecord, gatewayRoutingObservationLabel } from "../core/gateway-routing.js";
 import { responseModelIdObservationFromRecord } from "../core/response-model-id.js";
 import { durableAssistantTextFromEvent } from "../domains/dispatch/event-pump.js";
 import type { RunReceipt } from "../domains/dispatch/types.js";
@@ -23,3 +24,12 @@ export function receiptResponseModelIdObservationLabel(receipt: Pick<RunReceipt,
 	});
 	return `response_model_id_observation=${labels.join(",")}`;
 }
+
+/** Human projection of gateway routes carried by a receipt. */
+export function receiptGatewayRoutingLabel(receipt: Pick<RunReceipt, "upstreamResponses">): string | null {
+	const observations = (receipt.upstreamResponses ?? [])
+		.map((response) => gatewayRoutingObservationFromRecord(response as unknown as Record<string, unknown>))
+		.filter((observation) => observation !== null);
+	if (observations.length === 0) return null;
+	return observations.map(gatewayRoutingObservationLabel).join(" | ");
+}

package/src/tools/dispatch-plan.ts

@@ -637,6 +637,7 @@ const RESULT_CONTRACT_KINDS = new Set<ResultContract["kind"]>([
 	"council-report",
 	"debugger-report",
 	"research-report",
+	"world-knowledge-report",
 	"mutation-report",
 	"provenance-report",
 	"oracle-report",

package/src/tools/dispatch-runner.ts

@@ -2153,6 +2153,18 @@ export async function runDispatchTool(
 	if (options?.signal?.aborted) return { kind: "error", message: "dispatch: aborted" };
 	const maxOutputBytes = snapshot.maxOutputBytes;
 	const timeoutMs = snapshot.timeoutMs;
+	// All modes, including derived Scout/review runs, share the invocation's
+	// admission bounds. Keep these off the serializable approved request.
+	const dispatch = deps.dispatch;
+	const preparation = {
+		...(options?.signal === undefined ? {} : { signal: options.signal }),
+		...(timeoutMs === undefined ? {} : { deadlineAt: Date.now() + timeoutMs }),
+	};
+	deps.dispatch = {
+		...dispatch,
+		dispatch: (request, observer) => dispatch.dispatch(request, observer, preparation),
+		dispatchBatch: (requests) => dispatch.dispatchBatch(requests, preparation),
+	};
 	let review = snapshot.review === undefined ? undefined : structuredClone(snapshot.review);
 	let council = snapshot.council === undefined ? undefined : structuredClone(snapshot.council);
 

package/src/tools/registry.ts

@@ -527,9 +527,12 @@ export function createRegistry(deps: RegistryDeps): ToolRegistry {
 		}
 		// Stage 1.5, the skill tool surface: a loaded SKILL.md that declares
 		// allowed-tools or disallowed-tools narrows the surface until the
-		// policy's lifetime ends (turn end for the main agent, run end for
-		// workers). Narrowing only blocks; it never grants, and an out-of-
-		// surface call blocks terminally instead of parking for confirmation.
+		// policy's lifetime ends. Interactively that is the session: the
+		// narrowing stays armed across the operator's later turns until a
+		// different skill replaces it, `/skill off` clears it, or the session
+		// ends. A worker keeps the run-scoped lifetime. Narrowing only blocks;
+		// it never grants, and an out-of-surface call blocks terminally
+		// instead of parking for confirmation.
 		const surfaceViolation = evaluateSkillToolSurface(options?.pendingSkillPolicy, call.tool);
 		if (surfaceViolation) {
 			const verdict = skillSurfaceBlockedVerdict(decision, call.tool, surfaceViolation);
@@ -1090,10 +1093,13 @@ function skillSurfaceBlockedVerdict(
 	tool: string,
 	violation: SkillToolSurfaceViolation,
 ): Extract<RegistryVerdict, { kind: "blocked" }> {
+	const lifetime = violation.carriedSurface
+		? "The narrowing stays active for the rest of the session, across the operator's later turns, until a different skill replaces it or the operator clears it with /skill off."
+		: "The narrowing ends when the skill policy's turn or worker run ends.";
 	const reason =
 		violation.disallowedBy.length > 0
-			? `${tool} is disallowed by the active skill(s) ${violation.disallowedBy.join(", ")} (disallowed-tools). The narrowing ends when the skill policy's turn or worker run ends. Work within the skill workflow; if it genuinely needs this step, use ask_user when available or state the blocker in your reply.`
-			: `${tool} is outside the tool surface declared by the active skill(s) ${violation.skills.join(", ")}. Tools are narrowed to: ${(violation.mergedAllowedTools ?? []).join(", ")} (plus context and ask_user). The narrowing ends when the skill policy's turn or worker run ends. Work within the skill workflow; if it genuinely needs this step, use ask_user when available or state the blocker in your reply.`;
+			? `${tool} is disallowed by the active skill(s) ${violation.disallowedBy.join(", ")} (disallowed-tools). ${lifetime} Work within the skill workflow; if it genuinely needs this step, use ask_user when available or state the blocker in your reply.`
+			: `${tool} is outside the tool surface declared by the active skill(s) ${violation.skills.join(", ")}. Tools are narrowed to: ${(violation.mergedAllowedTools ?? []).join(", ")} (plus context and ask_user). ${lifetime} Work within the skill workflow; if it genuinely needs this step, use ask_user when available or state the blocker in your reply.`;
 	const blocked: SafetyDecision = {
 		kind: "block",
 		classification: decision.classification,

package/src/tools/worker-evidence.ts

@@ -9,7 +9,7 @@ import type { ReceiptIntegrityResult } from "../domains/dispatch/receipt-integri
 import type { RunReceipt, RunReceiptVerification } from "../domains/dispatch/types.js";
 import { formatTrustAxes, formatTrustSummary } from "../domains/evidence/trust-projection.js";
 import { adaptRunReceiptTrustStatus, type CanonicalTrustStatus } from "../domains/evidence/trust-status.js";
-import { receiptResponseModelIdObservationLabel } from "./dispatch-event-text.js";
+import { receiptGatewayRoutingLabel, receiptResponseModelIdObservationLabel } from "./dispatch-event-text.js";
 
 /** Matches a source citation the parent can independently spot-check. */
 const SOURCE_CITATION_PATTERN = /([\w./~-]+):(\d+)/g;
@@ -145,6 +145,7 @@ export function receiptEvidenceLabels(
 		}${receipt.projectContext.contentHash !== undefined ? ` sha256:${receipt.projectContext.contentHash}` : ""}`;
 	}
 	const responseModelIdObservation = receiptResponseModelIdObservationLabel(receipt);
+	const gatewayRouting = receiptGatewayRoutingLabel(receipt);
 	const budget =
 		receipt.budget === undefined
 			? []
@@ -160,6 +161,7 @@ export function receiptEvidenceLabels(
 		`evidence_verification=${verification.state}/${verification.basis}`,
 		`host_verification=${receipt.hostVerification?.status ?? "not_requested"}`,
 		...(responseModelIdObservation ? [responseModelIdObservation] : []),
+		...(gatewayRouting ? [gatewayRouting] : []),
 		...budget,
 		...receiptAdmissionLabels(receipt),
 		...receiptActivityLabels(receipt, status),

package/src/worker/spec-contract.ts

@@ -213,6 +213,7 @@ const RUNTIME_API_FAMILIES = [
 	"embeddings-http",
 	"claude-agent-sdk",
 	"claude-code-subprocess",
+	"external-agent-subprocess",
 ] as const satisfies ReadonlyArray<RuntimeApiFamily>;
 const RUNTIME_AUTHS = [
 	"api-key",
@@ -412,6 +413,9 @@ function validateCapabilityPatch(value: unknown, source: string): void {
 	for (const key of ["contextWindow", "maxTokens"] as const) {
 		readOptionalNumber(caps, key, source);
 	}
+	if (caps.thinkingControlRuntime !== undefined) {
+		readEnum(caps.thinkingControlRuntime, `${source}.thinkingControlRuntime`, ["lmstudio", "llamacpp"]);
+	}
 	for (const key of ["toolCallFormat", "thinkingFormat", "structuredOutputs"] as const) {
 		readOptionalString(caps, key, source);
 	}