chore(sf): auto-prompts residual sweep

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

src/resources/extensions/sf/auto-prompts.ts

@@ -1776,6 +1776,9 @@ export async function buildPlanMilestonePrompt(
 		relMilestoneFile(base, mid, "RESEARCH"),
 	);
 	const secretsOutputPath = join(base, relMilestoneFile(base, mid, "SECRETS"));
+	const inlinedContextWithFraming = framingBlock
+		? `${framingBlock}\n\n${inlinedContext}`
+		: inlinedContext;
 	return loadPrompt("plan-milestone", {
 		workingDirectory: base,
 		milestoneId: mid,
@@ -1786,7 +1789,7 @@ export async function buildPlanMilestonePrompt(
 		researchOutputPath,
 		outputPath: join(base, outputRelPath),
 		secretsOutputPath,
-		inlinedContext,
+		inlinedContext: inlinedContextWithFraming,
 		sourceFilePaths: buildSourceFilePaths(base, mid),
 		skillActivation: buildSkillActivationBlock({
 			base,

src/resources/extensions/sf/bootstrap/judgment-tools.ts

@@ -0,0 +1,70 @@
+/**
+ * Judgment tools — expose sf_log_judgment to the agent in autonomous mode.
+ *
+ * The agent is instructed (via the system prompt) to call this tool when
+ * making non-trivial calls so the user can review reasoning at milestone close.
+ */
+
+import { Type } from "@sinclair/typebox";
+import type { ExtensionAPI } from "@singularity-forge/pi-coding-agent";
+
+import { appendJudgment } from "../judgment-log.js";
+
+export function registerJudgmentTools(pi: ExtensionAPI): void {
+	pi.registerTool({
+		name: "sf_log_judgment",
+		label: "Log Judgment",
+		description:
+			"Record an agent judgment call for user review at milestone close. " +
+			"Call this when choosing between alternatives at an ambiguous decision point. " +
+			"Does NOT delay or block work — pure append-only side-effect.",
+		promptSnippet:
+			"Log a judgment call: decision taken, alternatives considered, reasoning, confidence",
+		promptGuidelines: [
+			"Call whenever you choose one approach over plausible alternatives.",
+			"Set confidence=low when the decision is speculative or you lacked context.",
+			"Set confidence=high when you had strong evidence for the choice.",
+			"This call is fire-and-forget — never wait for it or re-read the log mid-task.",
+		],
+		parameters: Type.Object({
+			unitId: Type.String({
+				description:
+					"Current unit ID (e.g. M001/S01/T01). Use the active task ID.",
+			}),
+			decision: Type.String({
+				description: "Short description of the decision taken (1-2 sentences).",
+			}),
+			alternatives: Type.Array(Type.String(), {
+				description: "Alternatives that were considered but not chosen.",
+			}),
+			reasoning: Type.String({
+				description:
+					"Why this decision was made over the alternatives (1-3 sentences).",
+			}),
+			confidence: Type.Union(
+				[Type.Literal("low"), Type.Literal("medium"), Type.Literal("high")],
+				{
+					description:
+						"Agent confidence in the decision: low = speculative, medium = reasonable, high = well-evidenced.",
+				},
+			),
+		}),
+		execute: async (_toolCallId, params, _signal, _onUpdate, _ctx) => {
+			appendJudgment(process.cwd(), {
+				unitId: params.unitId,
+				decision: params.decision,
+				alternatives: params.alternatives,
+				reasoning: params.reasoning,
+				confidence: params.confidence,
+			});
+			return {
+				content: [
+					{
+						type: "text" as const,
+						text: `Judgment logged for unit ${params.unitId}: "${params.decision}" (confidence: ${params.confidence})`,
+					},
+				],
+			};
+		},
+	});
+}