pi-lens 4.3.0 → 4.4.1

package/dist/clients/mcp/analyze.js

@@ -1,420 +0,0 @@
-/**
- * Host-neutral analysis facade for the MCP path.
- *
- * This is the heart of the "real review loop": it runs the *same* per-edit
- * dispatch pipeline pi-lens runs inside pi (`dispatchLintWithResult`) on a file,
- * and returns a structured, JSON-serializable result — diagnostics plus the
- * latency record for that dispatch, in the same schema pi writes to latency.log.
- *
- * Because the only host coupling is `getFlag` (see host-shim), this runs with no
- * pi process: an MCP server (or a `fresh` worker importing the freshly-built
- * dist) can drive it directly, letting Claude observe a commit's real behavioral
- * + perf impact first-hand rather than inferring it from pasted logs.
- */
-import * as fs from "node:fs";
-import * as path from "node:path";
-import { CacheManager, MCP_TURN_STATE_OWNER_ID } from "../cache-manager.js";
-import { CASCADE_GRAPH_KINDS, dispatchLintWithResult, getLatencyReports, } from "../dispatch/integration.js";
-import { FactStore } from "../dispatch/fact-store.js";
-import { detectFileKind } from "../file-kinds.js";
-import { getDiagnosticTracker } from "../diagnostic-tracker.js";
-import { getLSPService } from "../lsp/index.js";
-import { PathKeyedMap } from "../path-keyed-map.js";
-import { normalizeMapKey } from "../path-utils.js";
-import { loadProjectSnapshot } from "../project-snapshot.js";
-import { buildOrUpdateGraph } from "../review-graph/service.js";
-import { classifyDiagnosticTier, recordDiagnostics } from "../widget-state.js";
-import { deserializeWordIndex, removeWordIndexDocumentAsync, scheduleWordIndexPersist, updateWordIndexDocumentForEdit, WORD_INDEX_MAX_BYTES, } from "../word-index.js";
-import { logWordIndex } from "../word-index-logger.js";
-import { createMcpHost } from "./host-shim.js";
-// #536: module-scoped FactStore for the warm-analyze graph seam, mirroring the
-// per-edit cascade path's own module-level `sessionFacts` singleton
-// (clients/dispatch/integration.ts) — buildOrUpdateGraph's incremental/cached
-// tiers key off a stable FactStore instance across calls, so a fresh FactStore
-// per call would defeat that reuse. Scoped separately from integration.ts's
-// singleton since this file has no dependency on that module's internal state.
-// Subject labels this store's capacity-eviction telemetry distinctly from
-// the other five production FactStore instances (#2243 review round 3, F1).
-const warmGraphFacts = new FactStore("mcp-analyze");
-// #536 rider (issue body: "when #348 phase 2 lands, the word-index per-edit
-// update should ride the SAME seam so both indexes stay warm together"):
-// MCP has no RuntimeCoordinator/`runtime.wordIndex` to hold a live index the
-// way pi's per-edit seam does (clients/dispatch/integration.ts's
-// `updateWordIndexForCascade`, called from clients/runtime-tool-result.ts with
-// `runtime.wordIndex`) — this process-scoped Map is the MCP-side equivalent: a
-// per-cwd live WordIndex, loaded once from the persisted snapshot and mutated
-// in place thereafter, mirroring `runtime.wordIndex`'s lifecycle for a process
-// that has no other place to hold it. `undefined` cached value = "checked,
-// nothing usable" (index missing or pre-phase-2/no-forward-map), distinct from
-// "never checked" (key absent) — avoids re-attempting a snapshot load with no
-// forward index on every single analyze call.
-const DEFAULT_WORD_INDEX_IDLE_EVICT_MS = 20 * 60_000;
-const DEFAULT_WORD_INDEX_MAX_WARM_ROOTS = 8;
-const warmWordIndexes = new PathKeyedMap(normalizeMapKey);
-function positiveEnv(name, fallback) {
-    const parsed = Number.parseInt(process.env[name] ?? "", 10);
-    return Number.isSafeInteger(parsed) && parsed > 0 ? parsed : fallback;
-}
-function idleEvictMs() {
-    return positiveEnv("PI_LENS_WORD_INDEX_IDLE_EVICT_MS", DEFAULT_WORD_INDEX_IDLE_EVICT_MS);
-}
-function maxWarmRoots() {
-    return positiveEnv("PI_LENS_WORD_INDEX_MAX_WARM_ROOTS", DEFAULT_WORD_INDEX_MAX_WARM_ROOTS);
-}
-function clearWarmWordIndexTimer(entry) {
-    if (entry.timer)
-        clearTimeout(entry.timer);
-    entry.timer = undefined;
-}
-function evictWarmWordIndex(key, entry, reason) {
-    if (entry.leases > 0 || warmWordIndexes.get(key) !== entry)
-        return false;
-    clearWarmWordIndexTimer(entry);
-    warmWordIndexes.delete(key);
-    logWordIndex({
-        phase: "warm_cache_evicted",
-        cwd: key,
-        trigger: "mcp",
-        reason,
-    });
-    return true;
-}
-function scheduleWarmWordIndexIdleEviction(key, entry) {
-    clearWarmWordIndexTimer(entry);
-    const generation = ++entry.generation;
-    const timer = setTimeout(() => {
-        entry.timer = undefined;
-        if (warmWordIndexes.get(key) !== entry || entry.generation !== generation)
-            return;
-        if (!evictWarmWordIndex(key, entry, "idle"))
-            scheduleWarmWordIndexIdleEviction(key, entry);
-    }, idleEvictMs());
-    timer.unref?.();
-    entry.timer = timer;
-}
-function enforceWarmWordIndexLruCap() {
-    while (warmWordIndexes.size > maxWarmRoots()) {
-        const victim = [...warmWordIndexes.entries()]
-            .filter(([, entry]) => entry.leases === 0)
-            .sort(([, a], [, b]) => a.lastUsedAt - b.lastUsedAt)[0];
-        if (!victim || !evictWarmWordIndex(victim[0], victim[1], "lru"))
-            return;
-    }
-}
-/**
- * Look up (loading from the persisted snapshot on first use per cwd) the warm
- * in-memory word index this analyze facade keeps mutated in place. Exported so
- * `symbolSearch()` (clients/lens-engine.ts) can prefer this live copy over a
- * fresh disk read when one exists for the cwd — otherwise a query immediately
- * following a warm `pilens_analyze` call in the SAME process would read a
- * stale on-disk snapshot until the debounced persist (default 1500ms) flushes.
- */
-export function acquireWarmWordIndex(cwd) {
-    const key = path.resolve(cwd);
-    let entry = warmWordIndexes.get(key);
-    if (!entry) {
-        const snapshot = loadProjectSnapshot(key);
-        const index = deserializeWordIndex(snapshot?.wordIndex) ?? undefined;
-        entry = {
-            index: index?.forward ? index : undefined,
-            timer: undefined,
-            leases: 0,
-            lastUsedAt: Date.now(),
-            generation: 0,
-        };
-        warmWordIndexes.set(key, entry);
-    }
-    entry.leases++;
-    entry.lastUsedAt = Date.now();
-    scheduleWarmWordIndexIdleEviction(key, entry);
-    enforceWarmWordIndexLruCap();
-    let released = false;
-    return {
-        index: entry.index,
-        release() {
-            if (released)
-                return;
-            released = true;
-            entry.leases = Math.max(0, entry.leases - 1);
-            entry.lastUsedAt = Date.now();
-            if (warmWordIndexes.get(key) === entry) {
-                scheduleWarmWordIndexIdleEviction(key, entry);
-                enforceWarmWordIndexLruCap();
-            }
-        },
-    };
-}
-/**
- * Test-only reset — the module-level warm cache otherwise survives across
- * unrelated test cases in the same vitest worker.
- */
-export function _resetWarmWordIndexCacheForTests() {
-    for (const entry of warmWordIndexes.values())
-        clearWarmWordIndexTimer(entry);
-    warmWordIndexes.clear();
-}
-export function _getWarmWordIndexCacheStateForTests() {
-    return {
-        size: warmWordIndexes.size,
-        keys: [...warmWordIndexes.keys()],
-        timers: [...warmWordIndexes.values()].flatMap((entry) => entry.timer ? [entry.timer] : []),
-    };
-}
-// Generous warm-up budgets: a cold language server needs to spawn AND publish
-// diagnostics. The per-edit dispatch runner caps these tightly (spawn budget +
-// 2500ms) for latency; a review tool prioritises completeness, so we pre-warm
-// with room to spare, then the measured dispatch reads the warm cache.
-// Bounded so a cold analysis can't hang: enough for fast servers (pyright,
-// rust-analyzer, gopls) and a warm typescript-language-server, but NOT enough to
-// fully load a large TS project from cold — that exceeds any per-call budget and
-// is the persistent warm server's job (see the `lsp` honesty signal + Tier 2).
-const WARMUP_CLIENT_WAIT_MS = 10_000;
-const WARMUP_DIAGNOSTICS_WAIT_MS = 6_000;
-/** One diagnostic, flattened to the fields an MCP consumer needs. */
-/**
- * Split the dispatch `warnings` bucket into real warnings vs advisories using
- * the ONE severity-projection policy ({@link classifyDiagnosticTier}, #2414/
- * #2417) so the model-facing counts never present a `hint`/`info`-tier finding
- * as a warning (#2420, remainder of #2414). The dispatch `semantic` axis (which
- * also drives blocking) is intentionally NOT consulted or mutated here: this is
- * a display-count split, not a blocking-decision change. `blockers`
- * (`semantic:"blocking"`) are computed upstream and never enter this bucket, so
- * nothing that blocks can stop blocking through this seam. A non-advisory entry
- * in the bucket (an `error`/`warning`-tier finding routed here as
- * `semantic:"warning"`/`"none"`) stays counted as a warning exactly as before,
- * so a mixed file's real warning count is unchanged.
- */
-export function summarizeWarningTiers(warnings) {
-    let advisories = 0;
-    for (const w of warnings) {
-        if (classifyDiagnosticTier(w) === "advisory")
-            advisories++;
-    }
-    return { warnings: warnings.length - advisories, advisories };
-}
-function toMcpDiagnostic(diagnostic) {
-    return {
-        line: diagnostic.line,
-        column: diagnostic.column,
-        severity: diagnostic.severity,
-        semantic: diagnostic.semantic,
-        tool: diagnostic.tool,
-        rule: diagnostic.rule,
-        code: diagnostic.code,
-        message: diagnostic.message,
-        fixable: diagnostic.fixable,
-        fixSuggestion: diagnostic.fixSuggestion,
-    };
-}
-/**
- * Pre-warm the LSP for a file: spawn the server and wait for it to publish
- * diagnostics, so the subsequent dispatch reads a warm cache instead of a cold
- * (empty) one. Best-effort — failures never block the analysis.
- */
-async function warmLspForFile(absPath, host) {
-    if (host.getFlag("no-lsp"))
-        return;
-    const lspService = getLSPService();
-    if (!lspService.supportsLSP(absPath))
-        return;
-    let content;
-    try {
-        content = fs.readFileSync(absPath, "utf8");
-    }
-    catch {
-        return;
-    }
-    try {
-        await lspService.touchFile(absPath, content, {
-            diagnostics: "document",
-            collectDiagnostics: true,
-            clientScope: "primary",
-            maxClientWaitMs: WARMUP_CLIENT_WAIT_MS,
-            maxDiagnosticsWaitMs: WARMUP_DIAGNOSTICS_WAIT_MS,
-            source: "mcp-warmup",
-        });
-    }
-    catch {
-        // Best-effort warm-up; the dispatch runner still tries on its own.
-    }
-}
-/**
- * Run the dispatch pipeline on `filePath` and return a structured result.
- *
- * Unlike pi's per-edit path this defaults to the *full* analysis (warnings +
- * structural smells, not just blocking errors), pre-warms the LSP so a cold
- * server doesn't under-report, records into the session diagnostic state so the
- * query tools compose, and runs delta-free so a repeated analysis of an
- * unchanged file is a consistent full snapshot rather than "new issues only".
- *
- * The latency report is matched against the dispatches appended *during this
- * call* (we snapshot the report count first), so concurrent callers don't pick
- * up each other's timings.
- */
-export async function analyzeFile(filePath, cwd, options = {}) {
-    const absPath = path.isAbsolute(filePath)
-        ? filePath
-        : path.resolve(cwd, filePath);
-    // no-delta by default → a full snapshot every call (not delta-filtered);
-    // caller flags win over the default.
-    const host = createMcpHost({ "no-delta": true, ...options.flags }, cwd);
-    if (options.warmLsp !== false) {
-        await warmLspForFile(absPath, host);
-    }
-    const reportsBefore = getLatencyReports().length;
-    const start = Date.now();
-    // No telemetryModel/telemetryProvider here (#1448): this MCP facade has no
-    // RuntimeCoordinator to hold a host-reported identity (see the module doc
-    // above), so any worklog entries this dispatch produces get a blank
-    // model/provider — that's correct, not a gap.
-    const result = await dispatchLintWithResult(absPath, cwd, host, undefined, undefined, {
-        blockingOnly: options.blockingOnly ?? false,
-    });
-    const durationMs = Date.now() - start;
-    if (options.record !== false) {
-        // Mirror pipeline.ts's recording so pilens_diagnostics (mode=all) and
-        // pilens_health see what this analysis found. #3160 F1: `absPath` here
-        // is derived raw from the agent-supplied `file` arg (path.resolve above)
-        // — unlike pipeline.ts, which always records ctx.filePath, already
-        // canonicalized to on-disk casing by createDispatchContext (#2016/
-        // #3098). Normalize before recording so a mis-cased analyze call lands
-        // under the SAME widget-state key a canonical per-edit write already
-        // uses, instead of an orphaned key no reader (raw or normalized) can
-        // reach.
-        recordDiagnostics(normalizeMapKey(absPath), result.diagnostics);
-        if (result.diagnostics.length > 0) {
-            getDiagnosticTracker().trackShown(result.diagnostics);
-        }
-    }
-    if (options.registerTurnState) {
-        // Full-file range, importsChanged=true (conservative → dep/knip re-check
-        // broadly). Clear any stale pi session ID because the MCP Stop route owns
-        // this workspace-scoped worklist independently. Best-effort.
-        try {
-            const lineCount = fs.readFileSync(absPath, "utf8").split("\n").length;
-            new CacheManager().addModifiedRange(absPath, { start: 1, end: lineCount }, true, cwd, options.ownerId ?? MCP_TURN_STATE_OWNER_ID, "mcp");
-        }
-        catch {
-            // unreadable — skip turn-state registration
-        }
-    }
-    if (options.updateGraph && !result.hasBlockers) {
-        // #536: maintain the review graph on a successful warm analysis — the same
-        // call pi's per-edit cascade path makes (computeCascadeForFile), gated the
-        // same way (CASCADE_GRAPH_KINDS: only languages the graph actually models,
-        // and skipped when the file has blockers, matching the cascade path's own
-        // "primary_has_blockers" skip). buildOrUpdateGraph owns its own debounced
-        // persist + seq machinery internally — this is the ONLY call needed; no
-        // separate persist/flush step. Best-effort: a graph build failure must
-        // never fail the analysis itself (the diagnostics are already computed).
-        const fileKind = detectFileKind(absPath);
-        if (fileKind && CASCADE_GRAPH_KINDS.has(fileKind)) {
-            try {
-                await buildOrUpdateGraph(cwd, [absPath], warmGraphFacts);
-            }
-            catch {
-                // Best-effort — the graph update is additive; a failure here must not
-                // surface as an analyze failure.
-            }
-        }
-        // #536 rider: ride the SAME seam for the word index (#348 phase 2's
-        // per-edit primitive), mirroring pi's `updateWordIndexForCascade`
-        // (clients/dispatch/integration.ts) rule-for-rule rather than reusing it
-        // directly — that function is module-private and reads its file-content
-        // argument from the pipeline's already-read buffer, whereas this seam
-        // reads the file itself (no pipeline hook here). Same rules: a cached
-        // index with no `forward` map (or none loaded) ⇒ no-op (no incremental
-        // primitive available — the eventual full rebuild covers it); an
-        // oversized file is REMOVED, never partially indexed; the update is
-        // cooperative and serialized through the index's own operation queue
-        // (#2067 AC4), so this seam no longer holds the event loop for a whole
-        // replacement either; a successful update schedules the SAME debounced persist
-        // (`scheduleWordIndexPersist`, `PI_LENS_WORD_INDEX_PERSIST_DEBOUNCE_MS`)
-        // pi's path uses — no second persist mechanism.
-        //
-        // Key shape: `absPath` (`path.resolve` of the tool-input path). Since #1025
-        // the word index's path maps fold every key through `wordIndexKey`
-        // INTERNALLY (PathKeyedMap), so this edit-form key and the build path's
-        // walk-derived key converge on the same entry regardless of casing/
-        // separator — the old duplicate-orphan hazard is gone at the map layer.
-        const warmLease = acquireWarmWordIndex(cwd);
-        const warmIndex = warmLease.index;
-        try {
-            if (warmIndex) {
-                try {
-                    const content = fs.readFileSync(absPath, "utf8");
-                    const byteLength = Buffer.byteLength(content, "utf-8");
-                    if (byteLength > WORD_INDEX_MAX_BYTES) {
-                        await removeWordIndexDocumentAsync(warmIndex, absPath);
-                    }
-                    else {
-                        await updateWordIndexDocumentForEdit(warmIndex, {
-                            path: absPath,
-                            content,
-                        });
-                    }
-                    scheduleWordIndexPersist(cwd, warmIndex);
-                }
-                catch {
-                    // unreadable/deleted, or an update failure — best-effort, same as the
-                    // graph update above.
-                }
-            }
-        }
-        finally {
-            warmLease.release();
-        }
-    }
-    // dispatchForFile appended a latency report during the call above. Match the
-    // newly-added report for this exact path; fall back to the most recent new
-    // report if the path normalization differs.
-    const newReports = getLatencyReports().slice(reportsBefore);
-    const latencyReport = newReports.find((report) => path.resolve(report.filePath) === absPath) ??
-        newReports[newReports.length - 1];
-    const lspRunner = latencyReport?.runners.find((runner) => runner.runnerId === "lsp");
-    const lsp = lspRunner
-        ? {
-            // `deferred` means the runner entered the LSP path but its auxiliary
-            // coverage is delivered later, so it still counts as ran here. The
-            // status remains available to surfaces that need the coverage detail.
-            ran: lspRunner.status !== "skipped" &&
-                lspRunner.status !== "when_skipped" &&
-                lspRunner.status !== "test_file_skipped",
-            status: lspRunner.status,
-            diagnosticCount: lspRunner.diagnosticCount,
-            durationMs: lspRunner.durationMs,
-        }
-        : undefined;
-    return {
-        filePath: absPath,
-        cwd,
-        fileKind: latencyReport?.fileKind,
-        durationMs,
-        hasBlockers: result.hasBlockers,
-        counts: {
-            diagnostics: result.diagnostics.length,
-            blockers: result.blockers.length,
-            // #2420: `result.warnings` is the dispatch warnings bucket
-            // (`semantic:"warning"|"none"`), which still carries `hint`/`info`-tier
-            // findings folded in by the runner's severity→semantic map. Project
-            // them out through the shared tier policy so the model sees real
-            // warnings and advisories separately.
-            ...summarizeWarningTiers(result.warnings),
-            fixed: result.fixed.length,
-        },
-        lsp,
-        diagnostics: result.diagnostics.map(toMcpDiagnostic),
-        latency: latencyReport
-            ? {
-                totalDurationMs: latencyReport.totalDurationMs,
-                stoppedEarly: latencyReport.stoppedEarly,
-                runners: latencyReport.runners.map((runner) => ({
-                    runnerId: runner.runnerId,
-                    durationMs: runner.durationMs,
-                    status: runner.status,
-                    diagnosticCount: runner.diagnosticCount,
-                })),
-            }
-            : undefined,
-    };
-}

package/dist/clients/mcp/host-shim.js

@@ -1,31 +0,0 @@
-/**
- * Host shim for the MCP path.
- *
- * pi-lens's dispatch core only couples to the host through a single-method
- * interface — `PiAgentAPI.getFlag`. When pi-lens runs *inside pi*, that flag
- * resolver is backed by pi's CLI flags + project/global config. When it runs as
- * an MCP server (no pi process, no CLI flags), we back it by project/global
- * config + env + optional per-call overrides instead. This is the *entire* host
- * coupling the MCP path has to satisfy — everything else under `clients/` is
- * host-neutral.
- */
-import { loadPiLensGlobalConfig, resolvePiLensFlag } from "../lens-config.js";
-import { loadPiLensProjectConfig } from "../project-lens-config.js";
-/**
- * Build a `PiAgentAPI` for the MCP path. `overrides` lets a single MCP tool call
- * pin flags for that analysis (e.g. `no-lsp: true` to bench the non-LSP path)
- * without mutating config. Precedence: override → project config → global
- * config default.
- */
-export function createMcpHost(overrides, projectRoot = process.cwd()) {
-    const config = loadPiLensGlobalConfig();
-    const projectConfig = loadPiLensProjectConfig(projectRoot);
-    return {
-        getFlag(name, filePath) {
-            if (overrides && Object.hasOwn(overrides, name)) {
-                return overrides[name];
-            }
-            return resolvePiLensFlag(name, undefined, config, projectConfig, filePath, projectRoot);
-        },
-    };
-}