@sage-protocol/sage-plugin 0.1.5 → 0.1.6

package/index.js

@@ -5,577 +5,555 @@
 // Now includes RLM feedback appending when steering is detected.
 
 export const SagePlugin = async ({ client, $, directory }) => {
-	const CONFIG = {
-		sageBin: process.env.SAGE_BIN || "sage",
-		suggestLimit: Number.parseInt(process.env.SAGE_SUGGEST_LIMIT || "3", 10),
-		debounceMs: Number.parseInt(
-			process.env.SAGE_SUGGEST_DEBOUNCE_MS || "800",
-			10,
-		),
-		provision: (process.env.SAGE_SUGGEST_PROVISION || "1") === "1",
-		dryRun: (process.env.SAGE_PLUGIN_DRY_RUN || "0") === "1",
-		enableRlmFeedback: (process.env.SAGE_RLM_FEEDBACK || "1") === "1",
-	};
-
-	let promptCaptured = false;
-	let lastInput = "";
-	let lastInjected = "";
-	let timer = null;
-	let runId = 0;
-
-	// Session/model tracking (populated by chat.message hook or session.created event)
-	let currentSessionId = null;
-	let currentModel = null;
-	let assistantParts = []; // accumulate streaming text parts
-
-	// RLM Feedback tracking
-	let lastSuggestion = null;
-	let lastSuggestionTimestamp = null;
-	let lastSuggestionPromptKey = null; // qualified: library/key
-	let lastSuggestionId = null;
-	let lastShownPromptKeys = [];
-	let lastAcceptedFeedbackSent = false;
-	let lastImplicitFeedbackSent = false;
-	const SUGGESTION_CORRELATION_WINDOW_MS = 30000; // 30 second window
-
-	const parsePromptKeyMarkers = (text) => {
-		// Explicit markers only; no fuzzy matching.
-		// Marker format: [[sage:prompt_key=library/key]]
-		const re = /\[\[sage:prompt_key=([^\]]+)\]\]/g;
-		const keys = new Set();
-		for (;;) {
-			const m = re.exec(text);
-			if (!m) break;
-			const key = (m[1] || "").trim();
-			if (key) keys.add(key);
-		}
-		return Array.from(keys);
-	};
-
-	const recordPromptSuggestion = async ({
-		suggestionId,
-		prompt,
-		shownPromptKeys,
-		source,
-		attributesJson,
-	}) => {
-		try {
-			await execSage([
-				"suggest",
-				"prompt",
-				"capture",
-				suggestionId,
-				prompt,
-				"--source",
-				source,
-				"--shown",
-				...shownPromptKeys,
-				...(attributesJson ? ["--attributes-json", attributesJson] : []),
-			]);
-			return true;
-		} catch (e) {
-			await log(
-				"debug",
-				"prompt suggestion capture failed (daemon may be down)",
-				{
-					error: String(e),
-				},
-			);
-			return false;
-		}
-	};
-
-	const recordPromptSuggestionFeedback = async ({ suggestionId, events }) => {
-		try {
-			await execSage([
-				"suggest",
-				"prompt",
-				"feedback",
-				suggestionId,
-				"--events-json",
-				JSON.stringify(events),
-			]);
-			return true;
-		} catch (e) {
-			await log(
-				"debug",
-				"prompt suggestion feedback failed (daemon may be down)",
-				{
-					error: String(e),
-				},
-			);
-			return false;
-		}
-	};
-
-	const log = async (level, message, extra = {}) => {
-		try {
-			if (client?.app?.log) {
-				await client.app.log({
-					service: "sage-plugin",
-					level,
-					message,
-					extra,
-				});
-			} else {
-				console.log(`[sage-plugin:${level}]`, message, extra);
-			}
-		} catch {
-			/* logging should never break the plugin */
-		}
-	};
-
-	const execSage = async (args, env = {}) => {
-		if (CONFIG.dryRun) return "";
-
-		const sageEnv = { ...env, SAGE_SOURCE: "opencode" };
-
-		try {
-			if ($) {
-				// Use OpenCode's $ shell helper for portability
-				const cmd = [CONFIG.sageBin, ...args]
-					.map((a) => `'${a.replace(/'/g, "'\\''")}'`)
-					.join(" ");
-				const result = await $({ env: sageEnv })`${cmd}`;
-				return (result?.stdout ?? result ?? "").toString().trim();
-			}
-			// Fallback to Bun.spawn if $ not available
-			if (typeof Bun !== "undefined") {
-				const proc = Bun.spawn([CONFIG.sageBin, ...args], {
-					env: { ...process.env, ...sageEnv },
-					stdout: "pipe",
-					stderr: "pipe",
-				});
-				const stdout = await new Response(proc.stdout).text();
-				return stdout.trim();
-			}
-			return "";
-		} catch (e) {
-			throw new Error(`sage command failed: ${e.message || e}`);
-		}
-	};
-
-	// Parse suggestion output to extract prompt key
-	const parseSuggestionKey = (suggestionText) => {
-		// Look for patterns like "ultrawork-parallel-orchestration" or similar keys
-		// Format is typically: prompt_name (key: actual-key)
-		const keyMatch = suggestionText.match(/\(key:\s*([^)]+)\)/);
-		if (keyMatch) {
-			return keyMatch[1].trim();
-		}
-
-		// Try to match standalone keys in the text
-		const lines = suggestionText.split("\n");
-		for (const line of lines) {
-			// Look for common prompt key patterns
-			const match = line.match(/^\s*[-•*]?\s*([a-z0-9-]+)(?:\s*[-:]\s*|\s*$)/);
-			if (match?.[1]?.includes("-")) {
-				return match[1];
-			}
-		}
-
-		return null;
-	};
-
-	// Append RLM feedback to a prompt
-	const appendRlmFeedback = async (promptKey, feedbackEntry) => {
-		if (!CONFIG.enableRlmFeedback || !promptKey) {
-			return false;
-		}
-
-		try {
-			await log("debug", "appending RLM feedback", {
-				promptKey,
-				feedback: feedbackEntry,
-			});
-
-			const result = await execSage([
-				"suggest",
-				"feedback",
-				promptKey,
-				feedbackEntry,
-				"--source",
-				"opencode-plugin",
-			]);
-
-			if (result) {
-				await log("info", "RLM feedback appended", { promptKey });
-				return true;
-			}
-		} catch (e) {
-			await log("warn", "failed to append RLM feedback", {
-				promptKey,
-				error: String(e),
-			});
-		}
-
-		return false;
-	};
-
-	// Analyze prompt correlation with suggestion
-	const analyzePromptCorrelation = async (userPrompt) => {
-		if (!lastSuggestion || !lastSuggestionTimestamp) {
-			return null;
-		}
-
-		const now = Date.now();
-		const timeDiff = now - lastSuggestionTimestamp;
-
-		// Outside correlation window
-		if (timeDiff > SUGGESTION_CORRELATION_WINDOW_MS) {
-			return null;
-		}
-
-		const suggestionKey = lastSuggestionPromptKey;
-		if (!suggestionKey) {
-			return null;
-		}
-
-		// Check if user prompt matches or differs from suggestion
-		const userPromptLower = userPrompt.toLowerCase().trim();
-		const suggestionLower = lastSuggestion.toLowerCase().trim();
-
-		// Extract keywords from both
-		const userKeywords = userPromptLower.split(/\s+/);
-		const suggestionKeywords = suggestionLower.split(/\s+/);
-
-		// Check for significant overlap
-		const overlap = userKeywords.filter((k) => suggestionKeywords.includes(k));
-		const overlapRatio =
-			overlap.length / Math.max(userKeywords.length, suggestionKeywords.length);
-
-		// Determine correlation type
-		if (overlapRatio > 0.7) {
-			return { type: "accepted", key: suggestionKey, overlap: overlapRatio };
-		}
-		if (overlapRatio > 0.3) {
-			// Steering - user modified the suggestion
-			const addedKeywords = userKeywords.filter(
-				(k) => !suggestionKeywords.includes(k),
-			);
-			const removedKeywords = suggestionKeywords.filter(
-				(k) => !userKeywords.includes(k),
-			);
-
-			return {
-				type: "steered",
-				key: suggestionKey,
-				overlap: overlapRatio,
-				added: addedKeywords,
-				removed: removedKeywords,
-			};
-		}
-		return { type: "rejected", key: suggestionKey, overlap: overlapRatio };
-	};
-
-	const scheduleSuggest = (text) => {
-		lastInput = text;
-		runId += 1;
-		const current = runId;
-
-		if (timer) clearTimeout(timer);
-
-		timer = setTimeout(() => {
-			void (async () => {
-				const prompt = lastInput.trim();
-				if (!prompt) return;
-				if (current !== runId) return;
-				if (prompt === lastInjected) return;
-
-				await log("debug", "running sage suggest", {
-					cwd: directory,
-					prompt_len: prompt.length,
-				});
-
-				try {
-					const args = [
-						"suggest",
-						"skill",
-						prompt,
-						"--format",
-						"json",
-						"--limit",
-						CONFIG.suggestLimit.toString(),
-					];
-					if (CONFIG.provision) args.push("--provision");
-
-					const output = await execSage(args);
-					if (!output) return;
-					if (current !== runId) return;
-
-					let renderedOutput = "";
-					let correlationText = "";
-					let primaryKey = null;
-					let shownKeys = [];
-
-					try {
-						const json = JSON.parse(output);
-						if (
-							json.results &&
-							Array.isArray(json.results) &&
-							json.results.length > 0
-						) {
-							// Extract qualified keys for capture/correlation
-							shownKeys = json.results
-								.map((r) => (r.library ? `${r.library}/${r.key}` : r.key))
-								.filter(Boolean);
-							primaryKey = shownKeys[0] || null;
-
-							// Build correlation text from all results (titles/descriptions/keys)
-							// We exclude full content to keep overlap ratio meaningful
-							correlationText = json.results
-								.map((r) => `${r.name} ${r.description || ""} ${r.key}`)
-								.join(" ");
-
-							// Render output
-							renderedOutput = json.results
-								.map((r) => {
-									const qualifiedKey = r.library
-										? `${r.library}/${r.key}`
-										: r.key;
-									let block = `### ${r.name} (key: ${qualifiedKey})\n`;
-									if (r.library) block += `*Library: ${r.library}*\n`;
-									if (r.description) block += `${r.description}\n`;
-									if (r.content) block += `\n\`\`\`\n${r.content}\n\`\`\`\n`;
-									block += `\n<!-- If you use this suggestion, include marker: [[sage:prompt_key=${qualifiedKey}]] -->\n`;
-									return block;
-								})
-								.join("\n---\n\n");
-						}
-					} catch (e) {
-						// Fallback: If JSON parse fails, assume it might be plain text or broken JSON.
-						// We treat the raw output as the suggestion.
-						renderedOutput = output;
-						primaryKey = parseSuggestionKey(output);
-						correlationText = output;
-					}
-
-					if (!renderedOutput) return;
-
-					const suggestionId =
-						typeof crypto !== "undefined" && crypto.randomUUID
-							? crypto.randomUUID()
-							: `sage-suggest-${Date.now()}-${Math.random().toString(16).slice(2)}`;
-
-					// Store suggestion for correlation tracking
-					lastSuggestion = correlationText;
-					lastSuggestionTimestamp = Date.now();
-					lastSuggestionPromptKey = primaryKey;
-					lastSuggestionId = suggestionId;
-					lastShownPromptKeys = shownKeys;
-					lastAcceptedFeedbackSent = false;
-					lastImplicitFeedbackSent = false;
-
-					// Capture the suggestion to daemon (best-effort)
-					await recordPromptSuggestion({
-						suggestionId,
-						prompt,
-						shownPromptKeys: shownKeys,
-						source: "opencode",
-						attributesJson: JSON.stringify({
-							opencode: {
-								sessionId: currentSessionId,
-								model: currentModel,
-								workspace: directory,
-							},
-						}),
-					});
-
-					await log("debug", "suggestion stored for correlation", {
-						key: lastSuggestionPromptKey,
-						timestamp: lastSuggestionTimestamp,
-					});
-
-					lastInjected = prompt;
-					await client.tui.appendPrompt({
-						body: { text: `\n\n${renderedOutput}\n` },
-					});
-				} catch (e) {
-					await log("warn", "sage suggest failed", { error: String(e) });
-				}
-			})();
-		}, CONFIG.debounceMs);
-	};
-
-	return {
-		// Structured hook: reliable way to capture user prompts with model/session info
-		"chat.message": async (input, output) => {
-			// input: { sessionID, agent, model: {providerID, modelID}, messageID }
-			// output: { message: UserMessage, parts: Part[] }
-			currentSessionId = input?.sessionID ?? currentSessionId;
-			currentModel = input?.model?.modelID ?? currentModel;
-
-			const textParts = (output?.parts ?? []).filter((p) => p.type === "text");
-			const content = textParts.map((p) => p.text ?? "").join("\n");
-			if (!content.trim()) return;
-
-			promptCaptured = true;
-			assistantParts = [];
-
-			// Analyze correlation with previous suggestion
-			const correlation = await analyzePromptCorrelation(content);
-			if (correlation) {
-				await log("debug", "prompt correlation detected", correlation);
-
-				let feedbackEntry = "";
-				const date = new Date().toISOString().split("T")[0];
-
-				switch (correlation.type) {
-					case "accepted":
-						feedbackEntry = `[${date}] Prompt suggestion accepted (overlap: ${(correlation.overlap * 100).toFixed(0)}%)`;
-						break;
-					case "steered": {
-						const added = correlation.added?.slice(0, 3).join(", ") || "none";
-						const removed =
-							correlation.removed?.slice(0, 3).join(", ") || "none";
-						feedbackEntry = `[${date}] User steered from suggestion - Added keywords: "${added}" - Removed: "${removed}"`;
-						break;
-					}
-					case "rejected":
-						feedbackEntry = `[${date}] Prompt suggestion rejected (low overlap: ${(correlation.overlap * 100).toFixed(0)}%)`;
-						break;
-				}
-
-				if (feedbackEntry) {
-					await appendRlmFeedback(correlation.key, feedbackEntry);
-				}
-
-				// Also record prompt-suggestion feedback to daemon (best-effort)
-				if (lastSuggestionId && !lastAcceptedFeedbackSent) {
-					await recordPromptSuggestionFeedback({
-						suggestionId: lastSuggestionId,
-						events: [
-							{
-								kind: correlation.type,
-								prompt_key: correlation.key,
-								confidence: correlation.overlap,
-								features_json: JSON.stringify({ overlap: correlation.overlap }),
-							},
-						],
-					});
-					lastAcceptedFeedbackSent = true;
-				}
-
-				// Keep suggestion state for implicit marker detection on assistant completion.
-			}
-
-			try {
-				await execSage(["capture", "hook", "prompt"], {
-					SAGE_SESSION_ID: currentSessionId ?? "",
-					SAGE_MODEL: currentModel ?? "",
-					SAGE_WORKSPACE: directory ?? "",
-				});
-			} catch (e) {
-				await log("warn", "capture prompt failed", { error: String(e) });
-				promptCaptured = false;
-			}
-		},
-
-		event: async ({ event }) => {
-			const { type: eventType, properties } = event;
-
-			switch (eventType) {
-				case "message.part.updated": {
-					// OpenCode schema: { part: { id, sessionID, messageID, type, text }, delta? }
-					const part = properties?.part;
-					if (part?.type === "text" && promptCaptured) {
-						// Accumulate assistant text parts during streaming
-						assistantParts.push(part.text ?? "");
-					}
-					break;
-				}
-
-				case "message.updated": {
-					// OpenCode schema: { info: { id, sessionID, role, modelID, providerID, cost, tokens: {input, output, reasoning, cache} } }
-					const info = properties?.info;
-					if (info?.role === "assistant" && promptCaptured) {
-						const responseText = assistantParts.join("");
-						if (responseText.trim()) {
-							// If assistant explicitly marks one suggested prompt key as used, record implicitly_helpful.
-							if (
-								lastSuggestionId &&
-								lastSuggestionTimestamp &&
-								!lastImplicitFeedbackSent &&
-								Date.now() - lastSuggestionTimestamp <=
-									SUGGESTION_CORRELATION_WINDOW_MS
-							) {
-								const marked = parsePromptKeyMarkers(responseText);
-								const allowed = new Set(lastShownPromptKeys || []);
-								const matched = marked.filter((k) => allowed.has(k));
-								if (matched.length === 1) {
-									await recordPromptSuggestionFeedback({
-										suggestionId: lastSuggestionId,
-										events: [
-											{
-												kind: "implicitly_helpful",
-												prompt_key: matched[0],
-												confidence: 1.0,
-												features_json: JSON.stringify({ marker: true }),
-											},
-										],
-									});
-									lastImplicitFeedbackSent = true;
-								}
-							}
-
-							try {
-								await execSage(["capture", "hook", "response"], {
-									SAGE_SESSION_ID: info.sessionID ?? currentSessionId ?? "",
-									SAGE_MODEL: info.modelID ?? currentModel ?? "",
-									TOKENS_INPUT: String(info.tokens?.input ?? ""),
-									TOKENS_OUTPUT: String(info.tokens?.output ?? ""),
-								});
-							} catch (e) {
-								await log("warn", "capture response failed", {
-									error: String(e),
-								});
-							}
-						}
-						promptCaptured = false;
-						assistantParts = [];
-
-						// Clear suggestion tracking once we've had a full assistant completion after it.
-						if (
-							lastSuggestionTimestamp &&
-							Date.now() - lastSuggestionTimestamp >
-								SUGGESTION_CORRELATION_WINDOW_MS
-						) {
-							lastSuggestion = null;
-							lastSuggestionTimestamp = null;
-							lastSuggestionPromptKey = null;
-							lastSuggestionId = null;
-							lastShownPromptKeys = [];
-							lastAcceptedFeedbackSent = false;
-							lastImplicitFeedbackSent = false;
-						}
-					}
-					break;
-				}
-
-				case "session.created": {
-					// OpenCode schema: { info: { id, parentID, directory, title, ... } }
-					const info = properties?.info;
-					currentSessionId = info?.id ?? null;
-					promptCaptured = false;
-					assistantParts = [];
-					await log("info", "session created", {
-						sessionId: currentSessionId ?? "unknown",
-						isSubagent: info?.parentID != null,
-						cwd: directory,
-					});
-					break;
-				}
-
-				case "tui.prompt.append": {
-					const text = properties?.text ?? "";
-					if (text.trim()) {
-						scheduleSuggest(text);
-					}
-					break;
-				}
-			}
-		},
-	};
+  const CONFIG = {
+    sageBin: process.env.SAGE_BIN || "sage",
+    suggestLimit: Number.parseInt(process.env.SAGE_SUGGEST_LIMIT || "3", 10),
+    debounceMs: Number.parseInt(process.env.SAGE_SUGGEST_DEBOUNCE_MS || "800", 10),
+    provision: (process.env.SAGE_SUGGEST_PROVISION || "1") === "1",
+    dryRun: (process.env.SAGE_PLUGIN_DRY_RUN || "0") === "1",
+    enableRlmFeedback: (process.env.SAGE_RLM_FEEDBACK || "1") === "1",
+  };
+
+  let promptCaptured = false;
+  let lastInput = "";
+  let lastInjected = "";
+  let timer = null;
+  let runId = 0;
+
+  // Session/model tracking (populated by chat.message hook or session.created event)
+  let currentSessionId = null;
+  let currentModel = null;
+  let assistantParts = []; // accumulate streaming text parts
+
+  // RLM Feedback tracking
+  let lastSuggestion = null;
+  let lastSuggestionTimestamp = null;
+  let lastSuggestionPromptKey = null; // qualified: library/key
+  let lastSuggestionId = null;
+  let lastShownPromptKeys = [];
+  let lastAcceptedFeedbackSent = false;
+  let lastImplicitFeedbackSent = false;
+  const SUGGESTION_CORRELATION_WINDOW_MS = 30000; // 30 second window
+
+  const parsePromptKeyMarkers = (text) => {
+    // Explicit markers only; no fuzzy matching.
+    // Marker format: [[sage:prompt_key=library/key]]
+    const re = /\[\[sage:prompt_key=([^\]]+)\]\]/g;
+    const keys = new Set();
+    for (;;) {
+      const m = re.exec(text);
+      if (!m) break;
+      const key = (m[1] || "").trim();
+      if (key) keys.add(key);
+    }
+    return Array.from(keys);
+  };
+
+  const recordPromptSuggestion = async ({
+    suggestionId,
+    prompt,
+    shownPromptKeys,
+    source,
+    attributesJson,
+  }) => {
+    try {
+      await execSage([
+        "suggest",
+        "prompt",
+        "capture",
+        suggestionId,
+        prompt,
+        "--source",
+        source,
+        "--shown",
+        ...shownPromptKeys,
+        ...(attributesJson ? ["--attributes-json", attributesJson] : []),
+      ]);
+      return true;
+    } catch (e) {
+      await log("debug", "prompt suggestion capture failed (daemon may be down)", {
+        error: String(e),
+      });
+      return false;
+    }
+  };
+
+  const recordPromptSuggestionFeedback = async ({ suggestionId, events }) => {
+    try {
+      await execSage([
+        "suggest",
+        "prompt",
+        "feedback",
+        suggestionId,
+        "--events-json",
+        JSON.stringify(events),
+      ]);
+      return true;
+    } catch (e) {
+      await log("debug", "prompt suggestion feedback failed (daemon may be down)", {
+        error: String(e),
+      });
+      return false;
+    }
+  };
+
+  const log = async (level, message, extra = {}) => {
+    try {
+      if (client?.app?.log) {
+        await client.app.log({
+          service: "sage-plugin",
+          level,
+          message,
+          extra,
+        });
+      } else {
+        console.log(`[sage-plugin:${level}]`, message, extra);
+      }
+    } catch {
+      /* logging should never break the plugin */
+    }
+  };
+
+  const execSage = async (args, env = {}) => {
+    if (CONFIG.dryRun) return "";
+
+    const sageEnv = { ...env, SAGE_SOURCE: "opencode" };
+
+    try {
+      if ($) {
+        // Use OpenCode's $ shell helper for portability
+        const cmd = [CONFIG.sageBin, ...args].map((a) => `'${a.replace(/'/g, "'\\''")}'`).join(" ");
+        const result = await $({ env: sageEnv })`${cmd}`;
+        return (result?.stdout ?? result ?? "").toString().trim();
+      }
+      // Fallback to Bun.spawn if $ not available
+      if (typeof Bun !== "undefined") {
+        const proc = Bun.spawn([CONFIG.sageBin, ...args], {
+          env: { ...process.env, ...sageEnv },
+          stdout: "pipe",
+          stderr: "pipe",
+        });
+        const stdout = await new Response(proc.stdout).text();
+        return stdout.trim();
+      }
+      return "";
+    } catch (e) {
+      throw new Error(`sage command failed: ${e.message || e}`);
+    }
+  };
+
+  // Parse suggestion output to extract prompt key
+  const parseSuggestionKey = (suggestionText) => {
+    // Look for patterns like "ultrawork-parallel-orchestration" or similar keys
+    // Format is typically: prompt_name (key: actual-key)
+    const keyMatch = suggestionText.match(/\(key:\s*([^)]+)\)/);
+    if (keyMatch) {
+      return keyMatch[1].trim();
+    }
+
+    // Try to match standalone keys in the text
+    const lines = suggestionText.split("\n");
+    for (const line of lines) {
+      // Look for common prompt key patterns
+      const match = line.match(/^\s*[-•*]?\s*([a-z0-9-]+)(?:\s*[-:]\s*|\s*$)/);
+      if (match?.[1]?.includes("-")) {
+        return match[1];
+      }
+    }
+
+    return null;
+  };
+
+  // Append RLM feedback to a prompt
+  const appendRlmFeedback = async (promptKey, feedbackEntry) => {
+    if (!CONFIG.enableRlmFeedback || !promptKey) {
+      return false;
+    }
+
+    try {
+      await log("debug", "appending RLM feedback", {
+        promptKey,
+        feedback: feedbackEntry,
+      });
+
+      const result = await execSage([
+        "suggest",
+        "feedback",
+        promptKey,
+        feedbackEntry,
+        "--source",
+        "opencode-plugin",
+      ]);
+
+      if (result) {
+        await log("info", "RLM feedback appended", { promptKey });
+        return true;
+      }
+    } catch (e) {
+      await log("warn", "failed to append RLM feedback", {
+        promptKey,
+        error: String(e),
+      });
+    }
+
+    return false;
+  };
+
+  // Analyze prompt correlation with suggestion
+  const analyzePromptCorrelation = async (userPrompt) => {
+    if (!lastSuggestion || !lastSuggestionTimestamp) {
+      return null;
+    }
+
+    const now = Date.now();
+    const timeDiff = now - lastSuggestionTimestamp;
+
+    // Outside correlation window
+    if (timeDiff > SUGGESTION_CORRELATION_WINDOW_MS) {
+      return null;
+    }
+
+    const suggestionKey = lastSuggestionPromptKey;
+    if (!suggestionKey) {
+      return null;
+    }
+
+    // Check if user prompt matches or differs from suggestion
+    const userPromptLower = userPrompt.toLowerCase().trim();
+    const suggestionLower = lastSuggestion.toLowerCase().trim();
+
+    // Extract keywords from both
+    const userKeywords = userPromptLower.split(/\s+/);
+    const suggestionKeywords = suggestionLower.split(/\s+/);
+
+    // Check for significant overlap
+    const overlap = userKeywords.filter((k) => suggestionKeywords.includes(k));
+    const overlapRatio = overlap.length / Math.max(userKeywords.length, suggestionKeywords.length);
+
+    // Determine correlation type
+    if (overlapRatio > 0.7) {
+      return { type: "accepted", key: suggestionKey, overlap: overlapRatio };
+    }
+    if (overlapRatio > 0.3) {
+      // Steering - user modified the suggestion
+      const addedKeywords = userKeywords.filter((k) => !suggestionKeywords.includes(k));
+      const removedKeywords = suggestionKeywords.filter((k) => !userKeywords.includes(k));
+
+      return {
+        type: "steered",
+        key: suggestionKey,
+        overlap: overlapRatio,
+        added: addedKeywords,
+        removed: removedKeywords,
+      };
+    }
+    return { type: "rejected", key: suggestionKey, overlap: overlapRatio };
+  };
+
+  const scheduleSuggest = (text) => {
+    lastInput = text;
+    runId += 1;
+    const current = runId;
+
+    if (timer) clearTimeout(timer);
+
+    timer = setTimeout(() => {
+      void (async () => {
+        const prompt = lastInput.trim();
+        if (!prompt) return;
+        if (current !== runId) return;
+        if (prompt === lastInjected) return;
+
+        await log("debug", "running sage suggest", {
+          cwd: directory,
+          prompt_len: prompt.length,
+        });
+
+        try {
+          const args = [
+            "suggest",
+            "skill",
+            prompt,
+            "--format",
+            "json",
+            "--limit",
+            CONFIG.suggestLimit.toString(),
+          ];
+          if (CONFIG.provision) args.push("--provision");
+
+          const output = await execSage(args);
+          if (!output) return;
+          if (current !== runId) return;
+
+          let renderedOutput = "";
+          let correlationText = "";
+          let primaryKey = null;
+          let shownKeys = [];
+
+          try {
+            const json = JSON.parse(output);
+            if (json.results && Array.isArray(json.results) && json.results.length > 0) {
+              // Extract qualified keys for capture/correlation
+              shownKeys = json.results
+                .map((r) => (r.library ? `${r.library}/${r.key}` : r.key))
+                .filter(Boolean);
+              primaryKey = shownKeys[0] || null;
+
+              // Build correlation text from all results (titles/descriptions/keys)
+              // We exclude full content to keep overlap ratio meaningful
+              correlationText = json.results
+                .map((r) => `${r.name} ${r.description || ""} ${r.key}`)
+                .join(" ");
+
+              // Render output
+              renderedOutput = json.results
+                .map((r) => {
+                  const qualifiedKey = r.library ? `${r.library}/${r.key}` : r.key;
+                  let block = `### ${r.name} (key: ${qualifiedKey})\n`;
+                  if (r.library) block += `*Library: ${r.library}*\n`;
+                  if (r.description) block += `${r.description}\n`;
+                  if (r.content) block += `\n\`\`\`\n${r.content}\n\`\`\`\n`;
+                  block += `\n<!-- If you use this suggestion, include marker: [[sage:prompt_key=${qualifiedKey}]] -->\n`;
+                  return block;
+                })
+                .join("\n---\n\n");
+            }
+          } catch (e) {
+            // Fallback: If JSON parse fails, assume it might be plain text or broken JSON.
+            // We treat the raw output as the suggestion.
+            renderedOutput = output;
+            primaryKey = parseSuggestionKey(output);
+            correlationText = output;
+          }
+
+          if (!renderedOutput) return;
+
+          const suggestionId =
+            typeof crypto !== "undefined" && crypto.randomUUID
+              ? crypto.randomUUID()
+              : `sage-suggest-${Date.now()}-${Math.random().toString(16).slice(2)}`;
+
+          // Store suggestion for correlation tracking
+          lastSuggestion = correlationText;
+          lastSuggestionTimestamp = Date.now();
+          lastSuggestionPromptKey = primaryKey;
+          lastSuggestionId = suggestionId;
+          lastShownPromptKeys = shownKeys;
+          lastAcceptedFeedbackSent = false;
+          lastImplicitFeedbackSent = false;
+
+          // Capture the suggestion to daemon (best-effort)
+          await recordPromptSuggestion({
+            suggestionId,
+            prompt,
+            shownPromptKeys: shownKeys,
+            source: "opencode",
+            attributesJson: JSON.stringify({
+              opencode: {
+                sessionId: currentSessionId,
+                model: currentModel,
+                workspace: directory,
+              },
+            }),
+          });
+
+          await log("debug", "suggestion stored for correlation", {
+            key: lastSuggestionPromptKey,
+            timestamp: lastSuggestionTimestamp,
+          });
+
+          lastInjected = prompt;
+          await client.tui.appendPrompt({
+            body: { text: `\n\n${renderedOutput}\n` },
+          });
+        } catch (e) {
+          await log("warn", "sage suggest failed", { error: String(e) });
+        }
+      })();
+    }, CONFIG.debounceMs);
+  };
+
+  return {
+    // Structured hook: reliable way to capture user prompts with model/session info
+    "chat.message": async (input, output) => {
+      // input: { sessionID, agent, model: {providerID, modelID}, messageID }
+      // output: { message: UserMessage, parts: Part[] }
+      currentSessionId = input?.sessionID ?? currentSessionId;
+      currentModel = input?.model?.modelID ?? currentModel;
+
+      const textParts = (output?.parts ?? []).filter((p) => p.type === "text");
+      const content = textParts.map((p) => p.text ?? "").join("\n");
+      if (!content.trim()) return;
+
+      promptCaptured = true;
+      assistantParts = [];
+
+      // Analyze correlation with previous suggestion
+      const correlation = await analyzePromptCorrelation(content);
+      if (correlation) {
+        await log("debug", "prompt correlation detected", correlation);
+
+        let feedbackEntry = "";
+        const date = new Date().toISOString().split("T")[0];
+
+        switch (correlation.type) {
+          case "accepted":
+            feedbackEntry = `[${date}] Prompt suggestion accepted (overlap: ${(correlation.overlap * 100).toFixed(0)}%)`;
+            break;
+          case "steered": {
+            const added = correlation.added?.slice(0, 3).join(", ") || "none";
+            const removed = correlation.removed?.slice(0, 3).join(", ") || "none";
+            feedbackEntry = `[${date}] User steered from suggestion - Added keywords: "${added}" - Removed: "${removed}"`;
+            break;
+          }
+          case "rejected":
+            feedbackEntry = `[${date}] Prompt suggestion rejected (low overlap: ${(correlation.overlap * 100).toFixed(0)}%)`;
+            break;
+        }
+
+        if (feedbackEntry) {
+          await appendRlmFeedback(correlation.key, feedbackEntry);
+        }
+
+        // Also record prompt-suggestion feedback to daemon (best-effort)
+        if (lastSuggestionId && !lastAcceptedFeedbackSent) {
+          await recordPromptSuggestionFeedback({
+            suggestionId: lastSuggestionId,
+            events: [
+              {
+                kind: correlation.type,
+                prompt_key: correlation.key,
+                confidence: correlation.overlap,
+                features_json: JSON.stringify({ overlap: correlation.overlap }),
+              },
+            ],
+          });
+          lastAcceptedFeedbackSent = true;
+        }
+
+        // Keep suggestion state for implicit marker detection on assistant completion.
+      }
+
+      try {
+        await execSage(["capture", "hook", "prompt"], {
+          // Capture hook expects the prompt via stdin JSON (Claude Code) or env vars.
+          // OpenCode plugin uses env vars.
+          PROMPT: content,
+          SAGE_SESSION_ID: currentSessionId ?? "",
+          SAGE_MODEL: currentModel ?? "",
+          SAGE_WORKSPACE: directory ?? "",
+        });
+      } catch (e) {
+        await log("warn", "capture prompt failed", { error: String(e) });
+        promptCaptured = false;
+      }
+    },
+
+    event: async ({ event }) => {
+      const { type: eventType, properties } = event;
+
+      switch (eventType) {
+        case "message.part.updated": {
+          // OpenCode schema: { part: { id, sessionID, messageID, type, text }, delta? }
+          const part = properties?.part;
+          if (part?.type === "text" && promptCaptured) {
+            // Accumulate assistant text parts during streaming
+            assistantParts.push(part.text ?? "");
+          }
+          break;
+        }
+
+        case "message.updated": {
+          // OpenCode schema: { info: { id, sessionID, role, modelID, providerID, cost, tokens: {input, output, reasoning, cache} } }
+          const info = properties?.info;
+          if (info?.role === "assistant" && promptCaptured) {
+            const responseText = assistantParts.join("");
+            if (responseText.trim()) {
+              // If assistant explicitly marks one suggested prompt key as used, record implicitly_helpful.
+              if (
+                lastSuggestionId &&
+                lastSuggestionTimestamp &&
+                !lastImplicitFeedbackSent &&
+                Date.now() - lastSuggestionTimestamp <= SUGGESTION_CORRELATION_WINDOW_MS
+              ) {
+                const marked = parsePromptKeyMarkers(responseText);
+                const allowed = new Set(lastShownPromptKeys || []);
+                const matched = marked.filter((k) => allowed.has(k));
+                if (matched.length === 1) {
+                  await recordPromptSuggestionFeedback({
+                    suggestionId: lastSuggestionId,
+                    events: [
+                      {
+                        kind: "implicitly_helpful",
+                        prompt_key: matched[0],
+                        confidence: 1.0,
+                        features_json: JSON.stringify({ marker: true }),
+                      },
+                    ],
+                  });
+                  lastImplicitFeedbackSent = true;
+                }
+              }
+
+              try {
+                await execSage(["capture", "hook", "response"], {
+                  SAGE_SESSION_ID: info.sessionID ?? currentSessionId ?? "",
+                  SAGE_MODEL: info.modelID ?? currentModel ?? "",
+                  TOKENS_INPUT: String(info.tokens?.input ?? ""),
+                  TOKENS_OUTPUT: String(info.tokens?.output ?? ""),
+                  // Pass the actual response content for capture completion
+                  SAGE_RESPONSE: responseText,
+                });
+              } catch (e) {
+                await log("warn", "capture response failed", {
+                  error: String(e),
+                });
+              }
+            }
+            promptCaptured = false;
+            assistantParts = [];
+
+            // Clear suggestion tracking once we've had a full assistant completion after it.
+            if (
+              lastSuggestionTimestamp &&
+              Date.now() - lastSuggestionTimestamp > SUGGESTION_CORRELATION_WINDOW_MS
+            ) {
+              lastSuggestion = null;
+              lastSuggestionTimestamp = null;
+              lastSuggestionPromptKey = null;
+              lastSuggestionId = null;
+              lastShownPromptKeys = [];
+              lastAcceptedFeedbackSent = false;
+              lastImplicitFeedbackSent = false;
+            }
+          }
+          break;
+        }
+
+        case "session.created": {
+          // OpenCode schema: { info: { id, parentID, directory, title, ... } }
+          const info = properties?.info;
+          currentSessionId = info?.id ?? null;
+          promptCaptured = false;
+          assistantParts = [];
+          await log("info", "session created", {
+            sessionId: currentSessionId ?? "unknown",
+            isSubagent: info?.parentID != null,
+            cwd: directory,
+          });
+          break;
+        }
+
+        case "tui.prompt.append": {
+          const text = properties?.text ?? "";
+          if (text.trim()) {
+            scheduleSuggest(text);
+          }
+          break;
+        }
+      }
+    },
+  };
 };
 
 export default SagePlugin;