pi-lens 4.3.0 → 4.4.1

package/dist/clients/lsp/cascade-tier.js

@@ -1,421 +0,0 @@
-/**
- * Tier-aware cascade-lane wait policy (#458, re-scope №2).
- *
- * The cascade/deferred lane (`computeCascadeForFile`'s neighbor-touch fan-out
- * in `clients/dispatch/integration.ts`) actively opens neighbor files against
- * their LSP client and waits up to a per-touch budget (~1000ms cold-snapshot /
- * 2000ms warm) for `textDocument/publishDiagnostics` before deciding the
- * neighbor is clean. For a Tier-3 server — one that is `push-only` AND known
- * to publish NOTHING on a clean→clean transition (see
- * docs/lsp-capability-matrix.md; today that's typescript-language-server, the
- * lone core-set instance) — that wait can never distinguish "clean" from
- * "still analyzing"; it always burns its full budget. Dogfooding measured
- * ~221 such `lsp_diagnostics_timeout` events/day.
- *
- * pi 0.80.6's `agent_settled` quiet window (#483, `clients/quiet-window.ts`)
- * gives the cascade lane a place to resolve that ambiguity OUT of the
- * per-touch budget: fire the touch (didOpen/didChange still happens, so the
- * server starts real work), record it as outstanding, and reconcile against
- * whatever landed in the client's diagnostics cache by the time the agent run
- * goes idle. A touch nothing arrived for by then is recorded `unresolved` —
- * never silently treated as `clean` (the #240 doctrine: a missing answer is
- * not an affirmative answer).
- *
- * This module is deliberately NOT hardcoded to server names for the "should
- * this file skip its in-lane wait" question: it reads the live capability
- * snapshot's `workspaceDiagnosticsSupport.mode` (from
- * `detectWorkspaceDiagnosticsSupport`, cached at `initialize`) and combines it
- * with the `silentOnClean` marker on that server's `DiagnosticStrategy`
- * (`wait-policy/strategies.ts`) — the same per-server behavioral-knowledge table
- * the rest of the LSP layer already uses. A server with no live snapshot yet,
- * or whose mode isn't `push-only`, or that isn't marked `silentOnClean`, is
- * NOT tier-3 — the caller keeps today's full in-lane wait. Fail-safe is
- * always "wait like before".
- *
- * Native TS7 is the cascade-only exception. It is not silent on clean, but
- * its publication does not settle inside the cold-snapshot budget. Cascade
- * classifies it as `collect-later`, sends the same no-wait touch, and
- * reconciles its later per-file push or pull publication. The shared server
- * policy remains `waits`, so main-lane behavior does not change.
- *
- * #524/#529/#541/#558: a server id can now be backed by more than one actual
- * binary — "typescript" is classic typescript-language-server OR TS7's
- * native `tsc --lsp --stdio` (PR #526). PR #526 originally routed the
- * native-ts7 variant through the fail-safe "waits" path because
- * `silentOnClean` had only been measured against the classic server; #541
- * (2026-07-11) briefly lifted that exclusion after a clean-signal probe run
- * appeared to show native-ts7 silent too. A follow-up dual-environment
- * re-measurement (2026-07-12, nightly CI on Linux AND a live local run on
- * Windows dev, same `typescript@7.0.2` both times) found native-ts7 now
- * publishes 2 version-less diagnostic sets on the clean transition
- * (`cleanPubs=2(v:0)`) — it is NOT silent. Classic is unaffected and
- * confirmed still silent (`cleanPubs=0(v:0)`) in the same run. This is
- * therefore an EVIDENCE-BASED revert, not the original unverified caution:
- * native-ts7's clean-signal behavior IS known, and it is "publishes, not
- * silent". The shared classifier still routes a native-ts7 snapshot through
- * `waits`. The cascade-only wrapper routes it through `collect-later` because
- * the measured publication arrives after the in-lane budget. The shared
- * `silentOnClean` flag stays `true` for classic.
- * `scripts/probe-clean-signal.mjs`'s drift check no
- * longer compares native-ts7 rows against the shared marker (it now expects
- * `false` for them explicitly) — see that file's header for the regression
- * watch this sets up for a future TS7 build that becomes silent again.
- */
-import { logCascade } from "../cascade-logger.js";
-import { incrementDegradationCount } from "../degradation-ledger.js";
-import { logLatency } from "../latency-logger.js";
-import { normalizeMapKey } from "../path-utils.js";
-import { registerQuietWindowTask } from "../quiet-window.js";
-import { classifyCascadeWaitTier as classifySharedCascadeWaitTier, classifyServerWaitTier, resolvePrimaryServerForWaitPolicy, } from "./wait-policy/classification.js";
-export { classifyServerWaitTier };
-// The observed-latency half of the collect-later policy is shared with CLI
-// runners. Keep this module as the LSP-facing export seam without duplicating
-// the state machine in the cascade classifier.
-export { classifyObservedRunner, observeRunnerLatency, resetObservedRunnerLatency, } from "../dispatch/collect-later-tier.js";
-/**
- * The cascade lane's wait tier for `filePath`. DELEGATES to the shared
- * `wait-policy/classification.ts` rule — this wrapper adds exactly one
- * cascade-only override on top of it (native TS7's push-only snapshot →
- * `collect-later`) and never re-implements the classification itself.
- *
- * #1444 coverage tradeoff: the no-wait touch this tier selects is
- * `clientScope: "primary"`, and the tier itself is decided from the PRIMARY
- * (non-auxiliary) server alone. That tradeoff already existed for
- * `tier3-silent`; the override enlarges the population it applies to — every
- * native-TS7 TypeScript neighbour now takes the no-wait path too, so an
- * auxiliary server configured for those files no longer gets touched in-lane
- * for them (its findings arrive via the next per-edit dispatch, as they
- * already did for classic tier-3 files).
- */
-export function classifyCascadeWaitTier(lspService, filePath, snapshots) {
-    const primary = resolvePrimaryServerForWaitPolicy(filePath, snapshots);
-    if (primary?.serverId === "typescript" &&
-        primary.snapshot?.launchVariant === "native-ts7" &&
-        primary.snapshot.workspaceDiagnosticsSupport?.mode === "push-only") {
-        return "collect-later";
-    }
-    return classifySharedCascadeWaitTier(lspService, filePath, snapshots);
-}
-// --- Kill switch (lazy, memoized — house style per clients/runtime-config.ts /
-// clients/quiet-window.ts's isQuietWindowEnabled) ---
-let _enabledCache;
-/** `PI_LENS_TIER_AWARE_CASCADE=0` disables the whole feature: every cascade
- * touch waits in-lane exactly as it did before #458, no outstanding-touch
- * bookkeeping, no reconcile task registered. */
-export function isTierAwareCascadeEnabled() {
-    if (_enabledCache !== undefined)
-        return _enabledCache;
-    _enabledCache = process.env.PI_LENS_TIER_AWARE_CASCADE !== "0";
-    return _enabledCache;
-}
-/** Test-only: clear the memoized kill-switch read. */
-export function _resetTierAwareCascadeEnabledForTests() {
-    _enabledCache = undefined;
-}
-// Keyed by normalized file path. A later touch for the same file simply
-// replaces the earlier entry (only the most recent touch matters — an
-// older touch's diagnostics, if they ever arrive, are still a strict superset
-// concern the newer touch already re-supersedes via didOpen/didChange).
-//
-// #1899: the registry is drained in full by every reconcile sweep, but nothing
-// bounded it BETWEEN sweeps. The sweep runs on pi's `agent_settled` quiet
-// window, and a session can go a long time without one — dogfood logs show
-// sweeps up to 52 minutes apart. So the two bounds below are the registry's
-// own, independent of when a sweep arrives:
-//
-//   - `MAX_OUTSTANDING_TOUCHES` caps the entry count. Entries are held in
-//     touch order (see `recordOutstandingCascadeTouch`), so eviction takes the
-//     oldest touch at O(1).
-//   - `OUTSTANDING_TOUCH_MAX_AGE_MS` caps how long a touch stays reconcilable.
-//     Past it, the touch is discarded rather than answered: re-injecting a
-//     neighbour error, or clearing a footer entry, from a touch fired that
-//     long ago reports a state the agent has already moved past.
-//
-// Both constants are sized from the 2026-08-09..20 dogfood window (cascade.log,
-// 67 sweeps / 756 entries): the largest single sweep held 53 entries, and a
-// 15-minute age bound retains 72 of the 80 outcomes that ever resolved (90%)
-// while discarding the tail that resolved after ~26-52 minutes.
-const MAX_OUTSTANDING_TOUCHES = 256;
-const OUTSTANDING_TOUCH_MAX_AGE_MS = 15 * 60_000;
-const _outstandingTouches = new Map();
-/** Entries dropped by the age bound since the last sweep reported. */
-let _expiredSinceLastSweep = 0;
-/** Entries dropped by the size cap since the last sweep reported. */
-let _evictedSinceLastSweep = 0;
-/**
- * Drop every entry older than `OUTSTANDING_TOUCH_MAX_AGE_MS`. The map is held
- * in touch order, so the expired entries are a prefix and the walk stops at
- * the first live one.
- */
-function pruneExpiredOutstandingTouches(now) {
-    for (const [key, touch] of _outstandingTouches) {
-        if (now - touch.touchedAt <= OUTSTANDING_TOUCH_MAX_AGE_MS)
-            break;
-        _outstandingTouches.delete(key);
-        _expiredSinceLastSweep++;
-    }
-}
-/**
- * Record a Tier-3 cascade touch that skipped its in-lane wait. Called right
- * after the (still-performed) didOpen/didChange notify, before returning
- * without waiting. `touchedAt` must be sampled BEFORE the notify (see the
- * field doc) so the reconcile comparison can never misread a publish that
- * raced the record as pre-touch.
- *
- * #1899: also enforces the registry's two bounds. The `delete` before the
- * `set` is load-bearing — `Map.set` on an existing key keeps the key's
- * ORIGINAL insertion position, so without it a re-touched file would keep a
- * stale position and both the age prefix walk and the oldest-first eviction
- * would pick the wrong entry.
- */
-export function recordOutstandingCascadeTouch(entry) {
-    const key = normalizeMapKey(entry.filePath);
-    _outstandingTouches.delete(key);
-    // `Date.now()`, not `entry.touchedAt`: the caller samples `touchedAt` before
-    // its notify, so it trails real time, and the age bound is a statement about
-    // the clock rather than about the newest touch's own stamp.
-    pruneExpiredOutstandingTouches(Date.now());
-    _outstandingTouches.set(key, entry);
-    while (_outstandingTouches.size > MAX_OUTSTANDING_TOUCHES) {
-        const oldest = _outstandingTouches.keys().next();
-        if (oldest.done)
-            break;
-        _outstandingTouches.delete(oldest.value);
-        _evictedSinceLastSweep++;
-        // Bounded by construction: one ledger tally per session, not per
-        // eviction event. The subject is the cap itself rather than the evicted
-        // path — the discriminating fact is "this session overflowed the tier-3
-        // backlog", and per-path subjects would unbound the ledger group.
-        incrementDegradationCount({
-            kind: "cascade-tier3-backlog-evicted",
-            subject: `cap=${MAX_OUTSTANDING_TOUCHES}`,
-            reason: "tier-3 outstanding-touch registry hit its cap before a quiet-window reconcile drained it; the oldest touch was dropped unanswered",
-        });
-    }
-}
-/**
- * #1910: session-boundary reset for the outstanding-touch registry and its
- * sweep-scoped `_expiredSinceLastSweep`/`_evictedSinceLastSweep` counters.
- * Both predate #1899 and were never wired into `handleSessionStart` — a
- * session replacement inherited the prior session's outstanding touches, and
- * a stray eviction/expiry landing between a sweep and the boundary attributed
- * its count to the NEXT session's first reconcile gauge. Wired primary-only,
- * same as every other entry in that reset block: a concurrently-live
- * secondary never reaches `handleSessionStart` at all (the #473 guard in
- * index.ts), so it can never race this reset against the still-live
- * primary's outstanding touches.
- */
-export function resetCascadeTierSessionState() {
-    _outstandingTouches.clear();
-    _expiredSinceLastSweep = 0;
-    _evictedSinceLastSweep = 0;
-}
-/** Test-only: clear the outstanding-touch registry between test cases. */
-export function _resetOutstandingCascadeTouchesForTests() {
-    resetCascadeTierSessionState();
-}
-/** Test-only: peek at the registry without mutating it. */
-export function _getOutstandingCascadeTouchesForTests() {
-    return [..._outstandingTouches.values()];
-}
-/**
- * Test-only: peek at the sweep-scoped expired/evicted counters without
- * mutating them. Exists so a probe of `resetCascadeTierSessionState` can
- * assert on THESE counters directly, rather than only on the registry map —
- * a future counter added here and left out of the reset would otherwise
- * stay invisible to any probe that only checks the map went empty.
- */
-export function _getCascadeTierSweepCountersForTests() {
-    return { expired: _expiredSinceLastSweep, evicted: _evictedSinceLastSweep };
-}
-/**
- * Reconcile every outstanding Tier-3 touch against the LSP client's current
- * diagnostics cache. For each:
- *   - If the client holds a PER-FILE diagnostics entry for the touched file
- *     whose publish timestamp (`getAllDiagnostics()`'s `ts` — the max of the
- *     push/pull timestamps for that file, client.ts) is newer than the
- *     touch's pre-notify `touchedAt`, something published for THAT FILE since
- *     the touch — record `resolved-found` (diagnostics present) or
- *     `resolved-clean` (empty, but PROVEN empty by an actual publish for that
- *     file after the touch). A client-WIDE signal is deliberately not used:
- *     it advances on any file's publish, so it could falsely "prove" a silent
- *     neighbor clean when a sibling neighbor published (#240).
- *   - If nothing published for the file by settle time, record `unresolved` —
- *     per the #240 doctrine this is NEVER treated as clean.
- *
- * Client lookup is WARM-ONLY (`getWarmClientForFile`): the quiet window must
- * never resurrect an idle-reaped server (a full tsserver spawn + cold index)
- * just to write a log line. A warm-miss ⇒ `unresolved`.
- *
- * Always drains the whole registry (each entry is independently resolved;
- * one entry's client lookup failing doesn't block the rest) and never
- * throws — callers (the quiet-window task) must be fail-safe.
- */
-export async function reconcileOutstandingCascadeTouches(lspService) {
-    const outcomes = [];
-    // #1899: the age bound applies at sweep time too, not only at record time —
-    // a registry that received no further touches would otherwise hand the sweep
-    // entries of unbounded age.
-    pruneExpiredOutstandingTouches(Date.now());
-    const entries = [..._outstandingTouches.entries()];
-    _outstandingTouches.clear();
-    for (const [key, touch] of entries) {
-        const ageMs = Date.now() - touch.touchedAt;
-        try {
-            const spawned = await lspService.getWarmClientForFile(touch.filePath);
-            if (!spawned || spawned.client.serverId !== touch.serverId) {
-                outcomes.push({
-                    filePath: touch.filePath,
-                    serverId: touch.serverId,
-                    outcome: "unresolved",
-                    unresolvedReason: spawned ? "server-mismatch" : "warm-miss",
-                    ageMs,
-                });
-                continue;
-            }
-            const entry = spawned.client
-                .getAllDiagnostics()
-                .get(normalizeMapKey(touch.filePath));
-            if (!entry || entry.ts <= touch.touchedAt) {
-                // No per-file publish since the touch (or ever) — a missing answer
-                // is not a clean answer.
-                outcomes.push({
-                    filePath: touch.filePath,
-                    serverId: touch.serverId,
-                    outcome: "unresolved",
-                    unresolvedReason: entry ? "no-publish-since-touch" : "no-publish",
-                    ageMs,
-                });
-                continue;
-            }
-            const found = entry.diags.length > 0;
-            outcomes.push({
-                filePath: touch.filePath,
-                serverId: touch.serverId,
-                outcome: found ? "resolved-found" : "resolved-clean",
-                ageMs,
-                diagnosticCount: entry.diags.length,
-                publishedAt: entry.ts,
-                // #1023: carry the diagnostics so the task can re-surface them.
-                ...(found && { diagnostics: entry.diags }),
-            });
-        }
-        catch (err) {
-            outcomes.push({
-                filePath: touch.filePath,
-                serverId: touch.serverId,
-                outcome: "unresolved",
-                unresolvedReason: "error",
-                ageMs,
-            });
-            logLatency({
-                type: "phase",
-                phase: "cascade_tier3_reconcile_error",
-                filePath: key,
-                durationMs: 0,
-                metadata: { error: String(err) },
-            });
-        }
-    }
-    return outcomes;
-}
-let _reconcileTaskRegistered = false;
-/**
- * Register the Tier-3 reconcile task with the quiet-window scheduler
- * (`clients/quiet-window.ts`). Idempotent — safe to call more than once
- * (e.g. multiple extension activations in tests).
- */
-export function registerCascadeTierReconcileTask(getLspService, options = {}) {
-    if (_reconcileTaskRegistered)
-        return;
-    _reconcileTaskRegistered = true;
-    registerQuietWindowTask("cascade_tier3_reconcile", async () => {
-        if (!isTierAwareCascadeEnabled())
-            return;
-        const outcomes = await reconcileOutstandingCascadeTouches(getLspService());
-        // #1023: re-inject each resolved-found neighbor error so it reaches the
-        // agent (previously logs-only). #1444: hand each resolved-CLEAN outcome to
-        // the footer reconcile for the mirror-image case. Isolated per-outcome — a
-        // throwing callback must not drop the log line or the sibling deliveries.
-        for (const o of outcomes) {
-            try {
-                if (o.outcome === "resolved-found" && o.diagnostics?.length) {
-                    options.onResolvedFound?.({
-                        filePath: o.filePath,
-                        serverId: o.serverId,
-                        diagnostics: o.diagnostics,
-                        ...(o.publishedAt !== undefined
-                            ? { publishedAt: o.publishedAt }
-                            : {}),
-                    });
-                }
-                else if (o.outcome === "resolved-clean" && o.publishedAt != null) {
-                    // #1444: the stale-footer half of the same honesty problem — the
-                    // neighbour proved clean, but only after the in-lane wait was
-                    // skipped, so nothing has cleared its earlier error entries.
-                    options.onResolvedClean?.({
-                        filePath: o.filePath,
-                        serverId: o.serverId,
-                        publishedAt: o.publishedAt,
-                    });
-                }
-            }
-            catch {
-                // best-effort surfacing; the log below is the durable record.
-            }
-        }
-        let resolvedFound = 0;
-        let resolvedClean = 0;
-        let unresolved = 0;
-        let ageSumMs = 0;
-        let maxAgeMs = 0;
-        const unresolvedByReason = {};
-        for (const o of outcomes) {
-            if (o.outcome === "resolved-found")
-                resolvedFound++;
-            else if (o.outcome === "resolved-clean")
-                resolvedClean++;
-            else {
-                unresolved++;
-                const reason = o.unresolvedReason ?? "no-publish";
-                unresolvedByReason[reason] = (unresolvedByReason[reason] ?? 0) + 1;
-            }
-            ageSumMs += o.ageMs;
-            if (o.ageMs > maxAgeMs)
-                maxAgeMs = o.ageMs;
-        }
-        const avgAgeMs = outcomes.length
-            ? Math.round(ageSumMs / outcomes.length)
-            : 0;
-        // #1899: emitted on EVERY sweep, including the empty one. The backlog's
-        // size was previously observable only when it was non-empty, so "the
-        // queue is drained" and "the sweep never ran" read identically in
-        // cascade.log. This record is the backlog gauge: what the sweep drained,
-        // what the two bounds dropped since the last sweep, and what the bounds
-        // are, so a reader never has to know the constants to judge the numbers.
-        logCascade({
-            phase: "cascade_tier3_reconcile",
-            filePath: "<quiet-window>",
-            metadata: {
-                count: outcomes.length,
-                resolvedFound,
-                resolvedClean,
-                unresolved,
-                unresolvedByReason,
-                avgAgeMs,
-                maxAgeMs,
-                expired: _expiredSinceLastSweep,
-                evicted: _evictedSinceLastSweep,
-                capacity: MAX_OUTSTANDING_TOUCHES,
-                maxAgeCapMs: OUTSTANDING_TOUCH_MAX_AGE_MS,
-                outcomes,
-            },
-        });
-        _expiredSinceLastSweep = 0;
-        _evictedSinceLastSweep = 0;
-    });
-}
-/** Test-only: undo registerCascadeTierReconcileTask's idempotency guard. */
-export function _resetCascadeTierReconcileRegistrationForTests() {
-    _reconcileTaskRegistered = false;
-}