@phren/cli/dist/cli/hooks.js

Knowledge layer for AI agents - CLI, MCP server, and data layer

import { moduleEnabled } from "../modules/runtime.js";
// cli-hooks.ts — Thin orchestrator. Delegates to focused modules:
//   shared-retrieval.ts     — shared search, scoring, ranking, snippet selection
//   cli-hooks-citations.ts  — citation parsing and validation
//   cli-hooks-session.ts    — session lifecycle hooks, metrics, background maintenance
//   cli-hooks-output.ts     — hook output formatting
//   cli-hooks-globs.ts      — project glob matching
import { debugLog, sessionMarker, sessionsDir, getPhrenPath, } from "../shared.js";
import { mergeConfig, recordLookupEvents } from "../shared/governance.js";
import { loadIndexForHook, detectProject, } from "../shared/index.js";
import { isProjectHookEnabled } from "../project-config.js";
import { checkConsolidationNeeded, } from "../shared/content.js";
import { buildRobustFtsQuery, extractKeywordEntries, isFeatureEnabled, clampInt, clampFloat, errorMessage, } from "../utils.js";
import { getHooksEnabledPreference } from "../init/init.js";
import { logger } from "../logger.js";
import { isToolHookEnabled, runPrePromptHooks } from "../hooks.js";
import { handleExtractMemories } from "./extract.js";
import { appendAuditLog } from "../shared.js";
import { updateRuntimeHealth } from "../shared/governance.js";
import { getProactivityLevelForTask, getProactivityLevelForFindings } from "../proactivity.js";
import { FINDING_SENSITIVITY_CONFIG } from "./config.js";
import * as fs from "fs";
// ── Re-exports from focused modules ─────────────────────────────────────────
// Citations
export { parseCitations, validateCitation, annotateStale, clearCitationValidCache, } from "./hooks-citations.js";
// Globs
export { getProjectGlobBoost, clearProjectGlobCache, } from "./hooks-globs.js";
// Retrieval
export { detectTaskIntent, filterTaskByPriority, searchDocuments, applyTrustFilter, rankResults, applyRelevanceFloor, promptRarity, DEFAULT_MIN_QUERY_RELEVANCE, selectSnippets, } from "../shared/retrieval.js";
// Output
export { buildHookOutput, } from "./hooks-output.js";
// Session
export { handleHookSessionStart, handleHookStop, handleBackgroundSync, handleHookContext, handleHookTool, trackSessionMetrics, filterConversationInsightsForProactivity, extractToolFindings, filterToolFindingsForProactivity, resolveSubprocessArgs, } from "./hooks-session.js";
// ── Imports for the orchestrator ─────────────────────────────────────────────
import { searchDocumentsAsync, applyTrustFilter, rankResults, applyRelevanceFloor, promptRarity, DEFAULT_MIN_QUERY_RELEVANCE, selectSnippets, detectTaskIntent, wantsTasks } from "../shared/retrieval.js";
import { buildHookOutput } from "./hooks-output.js";
import { getGitContext, trackSessionMetrics, } from "./hooks-session.js";
import { approximateTokens, SNIPPET_OVERHEAD_TOKENS } from "../shared/retrieval.js";
import { resolveRuntimeProfile } from "../runtime-profile.js";
import { handleTaskPromptLifecycle } from "../task/lifecycle.js";
import { bestFindingNodeId } from "../finding-graph-id.js";
import { readKnowsBlock } from "../content/summarize.js";
// Auto-learn from prompts was removed — it learned conversational noise ("bro", "idk", typos)
// as synonyms for high-frequency terms. Manual `phren config synonyms add` still works.
// Future: finding-based co-occurrence mining (phren maintain command, not live hook).
async function readStdin() {
    return new Promise((resolve, reject) => {
        const chunks = [];
        process.stdin.on("data", (chunk) => chunks.push(typeof chunk === "string" ? Buffer.from(chunk) : chunk));
        process.stdin.on("end", () => resolve(Buffer.concat(chunks).toString("utf-8")));
        process.stdin.on("error", reject);
    });
}
export function parseHookInput(raw) {
    try {
        const data = JSON.parse(raw);
        const prompt = data.prompt || "";
        if (!prompt.trim())
            return null;
        return { prompt, cwd: data.cwd, sessionId: data.session_id };
    }
    catch (err) {
        debugLog(`parseHookInput: failed to parse hook JSON: ${errorMessage(err)}`);
        return null;
    }
}
// ── handleHookPrompt: orchestrator using extracted stages ────────────────────
export async function handleHookPrompt() {
    const profile = resolveRuntimeProfile(getPhrenPath());
    const stage = { indexMs: 0, searchMs: 0, trustMs: 0, rankMs: 0, selectMs: 0 };
    let raw = "";
    try {
        raw = await readStdin();
    }
    catch (err) {
        if (process.env.PHREN_DEBUG)
            logger.debug("cli-hooks", `hookPrompt stdinRead: ${errorMessage(err)}`);
        process.exit(0);
    }
    const input = parseHookInput(raw);
    if (!input)
        process.exit(0);
    const { prompt, cwd, sessionId } = input;
    if (!getHooksEnabledPreference(getPhrenPath())) {
        appendAuditLog(getPhrenPath(), "hook_prompt", "status=disabled");
        process.exit(0);
    }
    // Check per-tool hook preference (PHREN_HOOK_TOOL is set by session wrappers;
    // Claude hooks always run as "claude" from settings.json)
    const hookTool = process.env.PHREN_HOOK_TOOL || "claude";
    if (!isToolHookEnabled(getPhrenPath(), hookTool)) {
        appendAuditLog(getPhrenPath(), "hook_prompt", `status=tool_disabled tool=${hookTool}`);
        process.exit(0);
    }
    updateRuntimeHealth(getPhrenPath(), { lastPromptAt: new Date().toISOString() });
    const keywordEntries = extractKeywordEntries(prompt);
    const keywords = keywordEntries.join(" ");
    if (!keywords)
        process.exit(0);
    debugLog(`hook-prompt keywords: "${keywords}"`);
    // Session momentum: track topic frequencies within the session
    let hotTopics = [];
    if (sessionId) {
        const topicFile = sessionMarker(getPhrenPath(), `topics-${sessionId}.json`);
        let sessionTopics = {};
        try {
            if (fs.existsSync(topicFile)) {
                sessionTopics = JSON.parse(fs.readFileSync(topicFile, 'utf8'));
            }
        }
        catch { /* ignore parse errors */ }
        for (const kw of keywordEntries) {
            sessionTopics[kw] = (sessionTopics[kw] ?? 0) + 1;
        }
        const topicTmp = `${topicFile}.tmp-${process.pid}`;
        fs.writeFileSync(topicTmp, JSON.stringify(sessionTopics));
        fs.renameSync(topicTmp, topicFile);
        // Find hot topics (3+ mentions this session)
        hotTopics = Object.entries(sessionTopics)
            .filter(([, count]) => count >= 3)
            .map(([topic]) => topic);
    }
    const tIndex0 = Date.now();
    // Never block the prompt on a rebuild: serve the current/stale snapshot and
    // refresh in a detached process (see loadIndexForHook).
    const db = await loadIndexForHook(getPhrenPath(), profile);
    stage.indexMs = Date.now() - tIndex0;
    const intent = detectTaskIntent(prompt);
    const detectedProject = cwd ? detectProject(getPhrenPath(), cwd, profile) : null;
    if (detectedProject)
        debugLog(`Detected project: ${detectedProject}`);
    if (!isProjectHookEnabled(getPhrenPath(), detectedProject, "UserPromptSubmit")) {
        appendAuditLog(getPhrenPath(), "hook_prompt", `status=project_disabled project=${detectedProject}`);
        process.exit(0);
    }
    // Resolve git context AFTER the enabled check — it spawns 3 git processes and is
    // wasted work when the project has hooks disabled.
    const gitCtx = moduleEnabled(getPhrenPath(), "git") ? getGitContext(cwd) : null;
    const resolvedConfig = mergeConfig(getPhrenPath(), detectedProject ?? undefined);
    const safeQuery = buildRobustFtsQuery(keywords, detectedProject, getPhrenPath());
    if (!safeQuery)
        process.exit(0);
    try {
        const tSearch0 = Date.now();
        let rows = await searchDocumentsAsync(db, safeQuery, prompt, keywords, detectedProject, false, getPhrenPath());
        stage.searchMs = Date.now() - tSearch0;
        if (!rows || !rows.length)
            process.exit(0);
        const tTrust0 = Date.now();
        const policy = resolvedConfig.retentionPolicy;
        const memoryTtlDays = Number.parseInt(process.env.PHREN_MEMORY_TTL_DAYS || String(policy.ttlDays), 10);
        const trustResult = applyTrustFilter(rows, Number.isNaN(memoryTtlDays) ? policy.ttlDays : memoryTtlDays, policy.minInjectConfidence, policy.decay, getPhrenPath());
        rows = trustResult.rows;
        // Notes are intentionally available to explicit search, but are personal scratch
        // context and should never be injected into an agent prompt automatically.
        rows = rows.filter((row) => row.type !== "notes");
        // Tasks are a backlog, not memory: only when the prompt is about the work.
        if (!wantsTasks(prompt, intent))
            rows = rows.filter((row) => row.type !== "task");
        stage.trustMs = Date.now() - tTrust0;
        if (!rows.length)
            process.exit(0);
        const findingsProactivity = resolvedConfig.proactivity.findings
            ?? resolvedConfig.proactivity.base
            ?? getProactivityLevelForFindings(getPhrenPath());
        if (moduleEnabled(getPhrenPath(), "git") && isFeatureEnabled("PHREN_FEATURE_AUTO_EXTRACT", true) && findingsProactivity !== "low" && sessionId && detectedProject && cwd) {
            const marker = sessionMarker(getPhrenPath(), `extracted-${sessionId}-${detectedProject}`);
            if (!fs.existsSync(marker)) {
                try {
                    await handleExtractMemories(detectedProject, cwd, true, sessionId, "hook");
                    fs.writeFileSync(marker, "");
                }
                catch (err) {
                    debugLog(`auto-extract failed for ${detectedProject}: ${errorMessage(err)}`);
                }
            }
        }
        const tRank0 = Date.now();
        rows = rankResults(rows, intent, gitCtx, detectedProject, getPhrenPath(), db, cwd, keywords);
        stage.rankMs = Date.now() - tRank0;
        if (!rows.length)
            process.exit(0);
        // Relevance floor: inject signal or nothing. Ranking scores on priors too,
        // so a doc can rank well with no real tie to this prompt — drop those rather
        // than pad to a quota with noise (env PHREN_MIN_QUERY_RELEVANCE=0 disables).
        const relevanceFloor = clampFloat(process.env.PHREN_MIN_QUERY_RELEVANCE, DEFAULT_MIN_QUERY_RELEVANCE, 0, 1);
        const preFloorCount = rows.length;
        debugLog(`relevance-floor candidates: ${rows.slice(0, 12).map((row) => `${row.project}/${row.filename}`).join(", ")}`);
        rows = applyRelevanceFloor(rows, keywords, gitCtx, detectedProject, relevanceFloor, promptRarity(db, keywords));
        if (rows.length !== preFloorCount) {
            debugLog(`relevance-floor: ${preFloorCount} -> ${rows.length} rows (floor=${relevanceFloor})`);
        }
        if (!rows.length)
            process.exit(0);
        let safeTokenBudget = clampInt(process.env.PHREN_CONTEXT_TOKEN_BUDGET, 550, 180, 10000);
        const safeLineBudget = clampInt(process.env.PHREN_CONTEXT_SNIPPET_LINES, 6, 2, 100);
        const safeCharBudget = clampInt(process.env.PHREN_CONTEXT_SNIPPET_CHARS, 520, 120, 10000);
        // Session momentum: boost token budget for hot topics
        if (hotTopics.length > 0) {
            safeTokenBudget = Math.min(Math.floor(safeTokenBudget * 1.3), parseInt(process.env.PHREN_MAX_INJECT_TOKENS ?? '2000', 10));
        }
        const tSelect0 = Date.now();
        let { selected, usedTokens } = selectSnippets(rows, keywords, safeTokenBudget, safeLineBudget, safeCharBudget);
        stage.selectMs = Date.now() - tSelect0;
        if (!selected.length)
            process.exit(0);
        // Once per session per project: the "What phren knows" paragraph from
        // summary.md, so the agent gets the shape of a project's archive before
        // individual bullets. Written by `phren maintain summarize`; absent until then.
        if (detectedProject && sessionId) {
            const knowsMarker = sessionMarker(getPhrenPath(), `knows-${sessionId}-${detectedProject}`);
            if (!fs.existsSync(knowsMarker)) {
                const knows = readKnowsBlock(getPhrenPath(), detectedProject);
                if (knows) {
                    const text = knows.text.length > 520 ? `${knows.text.slice(0, 519)}…` : knows.text;
                    selected = [{ doc: { project: detectedProject, filename: "summary.md", type: "summary", content: knows.text, path: knows.path }, snippet: text, key: `knows:${detectedProject}` }, ...selected.filter((s) => s.key !== `knows:${detectedProject}`)];
                    usedTokens += approximateTokens(text) + SNIPPET_OVERHEAD_TOKENS;
                    try {
                        fs.writeFileSync(knowsMarker, "");
                    }
                    catch (err) {
                        debugLog(`knows marker: ${errorMessage(err)}`);
                    }
                }
            }
        }
        // Injection budget: cap total injected tokens across all content
        const maxInjectTokens = clampInt(process.env.PHREN_MAX_INJECT_TOKENS, 2000, 200, 20000);
        let budgetSelected = selected;
        let budgetUsedTokens = usedTokens;
        if (budgetUsedTokens > maxInjectTokens) {
            const priorityOrder = (s) => {
                if (s.doc.type === "findings")
                    return 0;
                if (s.doc.type === "canonical")
                    return 1;
                if (s.doc.type === "summary" || s.doc.type === "claude")
                    return 2;
                if (s.doc.type === "reference")
                    return 4;
                return 3;
            };
            const sorted = [...budgetSelected].sort((a, b) => priorityOrder(a) - priorityOrder(b));
            const kept = [];
            let runningTokens = 36;
            for (const s of sorted) {
                const est = approximateTokens(s.snippet) + SNIPPET_OVERHEAD_TOKENS;
                if (runningTokens + est <= maxInjectTokens) {
                    kept.push(s);
                    runningTokens += est;
                }
            }
            budgetSelected = kept;
            budgetUsedTokens = runningTokens;
            debugLog(`injection-budget: trimmed ${selected.length} -> ${kept.length} snippets to fit ${maxInjectTokens} token budget`);
        }
        // Live activity: what a prompt pulled in is a recall as much as a search
        // hit is, and it is by far the most common kind. Recording it here is what
        // lets the terminal graph's watch mode, and anything else tailing the log,
        // light up while an agent works. Best-effort: logging must never break
        // the hook.
        try {
            const at = new Date().toISOString();
            // The prompt, not the expanded keyword string: the feed shows this to a
            // person, and "retry retry backoff backoff jitter" is not what they asked.
            const asked = prompt.replace(/\s+/g, " ").trim().slice(0, 160);
            recordLookupEvents(getPhrenPath(), budgetSelected.map((s) => {
                const nodeId = s.doc.type === "findings" ? bestFindingNodeId(s.doc.project, s.doc.content, keywords) : null;
                return {
                    at,
                    query: asked,
                    project: s.doc.project,
                    filename: s.doc.filename,
                    type: s.doc.type,
                    path: s.doc.path,
                    snippet: s.snippet,
                    source: "hook",
                    ...(nodeId ? { nodeId } : {}),
                    ...(sessionId ? { session: sessionId } : {}),
                };
            }));
        }
        catch (err) {
            debugLog(`hook-prompt lookup-events: ${errorMessage(err)}`);
        }
        const parts = buildHookOutput(budgetSelected, budgetUsedTokens, intent, gitCtx, detectedProject, stage, safeTokenBudget, getPhrenPath(), sessionId, keywords);
        const taskLevel = resolvedConfig.proactivity.tasks
            ?? resolvedConfig.proactivity.base
            ?? getProactivityLevelForTask(getPhrenPath());
        const taskLifecycle = moduleEnabled(getPhrenPath(), "tasks") ? handleTaskPromptLifecycle({
            phrenPath: getPhrenPath(),
            prompt,
            project: detectedProject,
            sessionId,
            intent,
            taskLevel,
        }) : { noticeLines: [] };
        if (taskLifecycle.noticeLines.length > 0) {
            parts.push("");
            parts.push(...taskLifecycle.noticeLines);
        }
        // Inject finding sensitivity agent instruction
        try {
            const sensitivity = resolvedConfig.findingSensitivity ?? "balanced";
            const sensitivityConfig = FINDING_SENSITIVITY_CONFIG[sensitivity];
            if (sensitivityConfig) {
                parts.push("");
                parts.push(`[phren finding-sensitivity=${sensitivity}] ${sensitivityConfig.agentInstruction}`);
            }
        }
        catch {
            // ignore — non-fatal
        }
        // Add budget info to trace
        if (parts.length > 0) {
            const traceIdx = parts.findIndex(p => p.includes("trace:"));
            if (traceIdx !== -1) {
                parts[traceIdx] = parts[traceIdx].replace(/tokens/, `budget=${budgetUsedTokens}/${maxInjectTokens};tokens`);
            }
        }
        if (sessionId) {
            trackSessionMetrics(getPhrenPath(), sessionId, budgetSelected);
        }
        // Reads stay side-effect free: trust filter output informs ranking/snippets now,
        // while queue/audit mutation is deferred to explicit governance maintenance.
        if (trustResult.queueItems.length > 0 || trustResult.auditEntries.length > 0) {
            debugLog(`hook-prompt deferred trust governance items=${trustResult.queueItems.length} audit=${trustResult.auditEntries.length}`);
        }
        const noticeFile = sessionId ? sessionMarker(getPhrenPath(), `noticed-${sessionId}`) : null;
        const alreadyNoticed = noticeFile ? fs.existsSync(noticeFile) : false;
        if (!alreadyNoticed) {
            // Defer stale session marker cleanup to after output — it's I/O-heavy and not needed for results
            setImmediate(() => {
                try {
                    const cutoff = Date.now() - 86400000;
                    const sessDir = sessionsDir(getPhrenPath());
                    if (fs.existsSync(sessDir)) {
                        for (const f of fs.readdirSync(sessDir)) {
                            if (!f.startsWith("noticed-") && !f.startsWith("extracted-"))
                                continue;
                            const fp = `${sessDir}/${f}`;
                            try {
                                if (fs.statSync(fp).mtimeMs < cutoff)
                                    fs.unlinkSync(fp);
                            }
                            catch { /* ignore per-file errors */ }
                        }
                    }
                    // Also clean stale markers from the phren root
                    for (const f of fs.readdirSync(getPhrenPath())) {
                        if (!f.startsWith(".noticed-") && !f.startsWith(".extracted-"))
                            continue;
                        const fp = `${getPhrenPath()}/${f}`;
                        try {
                            fs.unlinkSync(fp);
                        }
                        catch { /* ignore */ }
                    }
                }
                catch (err) {
                    debugLog(`stale notice cleanup failed: ${errorMessage(err)}`);
                }
            });
            const needed = checkConsolidationNeeded(getPhrenPath(), profile);
            if (needed.length > 0) {
                const notices = needed.map((n) => {
                    const since = n.lastConsolidated ? ` since ${n.lastConsolidated}` : "";
                    return `  ${n.project}: ${n.entriesSince} new findings${since}`;
                });
                parts.push(`\u25c8 phren \u00b7 consolidation ready`);
                parts.push(`<phren-notice>`);
                parts.push(`Findings ready for consolidation:`);
                parts.push(notices.join("\n"));
                parts.push(`Run phren-consolidate when ready.`);
                parts.push(`</phren-notice>`);
            }
            if (noticeFile) {
                try {
                    fs.writeFileSync(noticeFile, "");
                }
                catch (err) {
                    if (process.env.PHREN_DEBUG)
                        logger.debug("cli-hooks", `hookPrompt noticeFileWrite: ${errorMessage(err)}`);
                }
            }
        }
        const totalMs = stage.indexMs + stage.searchMs + stage.trustMs + stage.rankMs + stage.selectMs;
        const slowThreshold = Number.parseInt(process.env.PHREN_SLOW_FS_WARN_MS || "3000", 10) || 3000;
        if (totalMs > slowThreshold) {
            debugLog(`slow-fs: hook-prompt took ${totalMs}ms (index=${stage.indexMs} search=${stage.searchMs} trust=${stage.trustMs} rank=${stage.rankMs} select=${stage.selectMs})`);
            process.stderr.write(`phren: hook-prompt took ${totalMs}ms, check if ~/.phren is on a slow or network filesystem\n`);
        }
        // Run pre-prompt custom hooks and prepend their output
        const prePromptOutput = runPrePromptHooks(getPhrenPath(), raw);
        if (prePromptOutput)
            parts.unshift(prePromptOutput);
        console.log(parts.join("\n"));
    }
    catch (err) {
        const msg = errorMessage(err);
        process.stdout.write(`\n<phren-error>phren hook failed: ${msg}. Check ~/.phren/.runtime/debug.log for details.</phren-error>\n`);
        debugLog(`hook-prompt error: ${msg}`);
        process.exit(0);
    }
}