pi-mega-compact 0.20.85 → 0.20.86

package/extensions/mega-pipeline/compact.ts

@@ -1,388 +1,15 @@
 /**
- * compact.ts — full compaction pipeline (Trident) + pi no-op prediction.
- *
- * `runCompact` runs the full Trident pipeline (fast-gate aside) and persists a
- * checkpoint. `piCompactWouldNoop` predicts whether pi's `ctx.compact()` would
- * throw a no-op error. Both mutate the shared MegaRuntime (token accounting,
- * ticker, status, events) and are driven by the event + command handlers in
- * mega-events.ts / mega-commands.ts.
- */
-
-import type { ExtensionAPI, ExtensionContext } from "@earendil-works/pi-coding-agent";
-import { sessionEntryToContextMessages } from "@earendil-works/pi-coding-agent";
-import type { AgentMessage } from "@earendil-works/pi-agent-core";
-import { compactSession } from "../../src/engine.js";
-import type { EngineMessage } from "../../src/types.js";
-import { normalizeSessionId } from "../../src/store.js";
-import { repoKey } from "../../src/store/repoKey.js";
-import { estimateBlockTokens } from "../../src/tokens.js";
-import { touchSession, logDaily, incCompactCount, incCacheHitTokens } from "../../src/store/sqlite.js";
-import { consolidateMemories } from "../../src/memory.js";
-import {
-  type MegaRuntime,
-  C,
-  MARKER_TYPE,
-} from "../mega-runtime.js";
-import { resolveRepoRoot, preserveRecentForPressure, type MegaConfig } from "../mega-config.js";
-import { runRaptor } from "../../src/dedup/raptor/index.js";
-import { isRaptorTreeFresh } from "../../src/dedup/raptor/buildHistory.js";
-import { loadDedupConfig } from "../../src/config/dedup.js";
-import { upsertEmbedding as indexUpsertEmbedding } from "../../src/store/vectorIndex.js";
-import { runMemoryReview } from "./memory-review.js";
-import { vectorList } from "../../src/vectorStore.js";
-
-export type RunCompactResult =
-  | { skipped: true }
-  | { skipped: false; result: ReturnType<typeof compactSession>; keepFrom: number; saved: number };
-
-/** Run the full compaction pipeline and persist a checkpoint. Returns the result. */
-export function runCompact(
-  pi: ExtensionAPI,
-  runtime: MegaRuntime,
-  config: MegaConfig,
-  ctx: ExtensionContext,
-  messages: AgentMessage[],
-  opts: { keepFrom?: number; summary?: string; compressionPressure?: number } = {},
-): RunCompactResult {
-  runtime.bindRepo(ctx.cwd);
-  const sid = normalizeSessionId(ctx.sessionManager.getSessionId());
-  runtime.resetRuntime(sid);
-  runtime.rt.sessionId = sid;
-
-  const view = runtime.engineView(messages);
-  // keepFrom deepens with context pressure (Fix E): under high pressure we
-  // compact more of the session, down to the preserveRecentMin floor.
-  const preserve = preserveRecentForPressure(
-    opts.compressionPressure ?? 0,
-    config.preserveRecent,
-    config.preserveRecentMin,
-  );
-  const keepFrom = opts.keepFrom ?? Math.max(0, view.length - preserve);
-  // For very small sessions (fewer messages than preserveRecent), allow
-  // compacting everything except the last message — the user explicitly
-  // requested compaction, so don't refuse it just because the session is short.
-  if (keepFrom <= 0) {
-    if (view.length <= 1) return { skipped: true };
-    // Use the fallback: compact everything except the last message
-    const fallbackKeepFrom = view.length - 1;
-    return doCompact(view, fallbackKeepFrom, opts, sid, config, pi, ctx, runtime);
-  }
-
-  return doCompact(view, keepFrom, opts, sid, config, pi, ctx, runtime);
-}
-
-function doCompact(
-  view: EngineMessage[],
-  keepFrom: number,
-  opts: { keepFrom?: number; summary?: string; compressionPressure?: number },
-  sid: string,
-  config: MegaConfig,
-  pi: ExtensionAPI,
-  ctx: ExtensionContext,
-  runtime: MegaRuntime,
-): RunCompactResult {
-  runtime.pulsing = true; // animate the status line while the (sync) pipeline runs
-  runtime.setEffect?.("pulse", "accent", 1500); // v0.8.3: ambient border pulse during compaction
-  // S21.2: reset the per-compaction memory-op counter so the post-compact
-  // consolidate pass only fires when memory rows actually changed during the
-  // compaction window (turn_end → auto-review may have written some).
-  runtime.memoriesTouchedThisCompaction = 0;
-  const result = compactSession(
-    {
-      sessionId: sid,
-      messages: view,
-      keepFrom,
-      summary: opts.summary,
-      timestamp: Date.now(),
-      onTier: runtime.makeTierCallback(ctx),
-      compressionPressure: opts.compressionPressure,
-    },
-    runtime.store,
-  );
-  runtime.pulsing = false;
-
-  if (result.skipped) return { skipped: true };
-  if (!result.deduped) {
-    runtime.rt.persistedThisSession = true;
-    runtime.rt.lastCheckpointId = result.checkpointId;
-  }
-  runtime.rt.lastCompactedFrom = result.compactedFrom;
-  runtime.rt.lastCompactedTokens = result.tokenEstimate;
-  runtime.rt.dedupAttempts++;
-  // Honest "tokens saved" for this session-instance only:
-  //   new checkpoint      → original − stored
-  //   deduped onto existing → whole original region (nothing new stored)
-  // Resets to 0 on session_start (rt is rebuilt) — so a fresh session shows 0
-  // while the repo's cumulative saved (SQLite meta) keeps the running total.
-  const saved = result.deduped
-    ? result.originalTokenEstimate
-    : Math.max(0, result.originalTokenEstimate - result.tokenEstimate);
-  runtime.rt.tokensSaved += saved;
-  runtime.rt.compactCount += 1;
-  incCompactCount(runtime.currentStateDir);
-  if (result.deduped) { runtime.rt.cacheHitTokens += saved; incCacheHitTokens(saved, runtime.currentStateDir); }
-  runtime.rt.lastCompactAt = Date.now();
-  if (result.deduped) runtime.rt.dedupSkips++;
-  // Grow the rolling "saved" goal so the progress bar always has a fresh
-  // denominator (we don't want it pinned at 100% once we pass an old target).
-  if (runtime.rt.tokensSaved > runtime.savedGoal) runtime.savedGoal = Math.ceil((runtime.rt.tokensSaved * 1.25) / 10_000) * 10_000;
-
-  // Live toolbar activity: what file/region just got compacted or deduped.
-  // Rendered via the rotating ticker line (see snapshot); the ring buffer is
-  // cycled one-per-repaint so the single line scrolls through recent files.
-  const files = result.filesModified ?? [];
-  const fileLabel = files.length
-    ? files.map((f) => f.split("/").pop() ?? f).slice(0, 2).join(", ")
-    : result.regionHash.slice(0, 8);
-  runtime.lastActivityAt = Date.now();
-  // Explain-why line: surfaced while fresh. Pulls the dedup reason (which for
-  // L2 includes the cosine sim) so the user sees WHY a region was kept/dropped.
-  runtime.lastWhy = result.deduped
-    ? `why: deduped@${result.dedupReason ?? "tier"}`
-    : `why: compacted → ${result.checkpointId}`;
-  // Recall/activity ticker: record this event in the ring buffer.
-  const savedK = (saved / 1000).toFixed(1);
-  runtime.pushTicker(
-    result.deduped
-      ? `${C.green}♻${C.reset} deduped ${fileLabel} · ${savedK}k saved`
-      : `${C.cyan}🗜${C.reset} ${result.checkpointId} · +${savedK}k · ${fileLabel}`,
-  );
-  // The per-tier trace has settled into the final outcome — fold it back into
-  // the activity line and stop showing the live trace.
-  runtime.tierTrace = undefined;
-
-  // Record session activity + a daily-log entry in the per-repo SQLite store
-  // (foundation for resume-sessions / daily-log features). Best-effort — never
-  // block a compaction on bookkeeping.
-  try {
-    const root = resolveRepoRoot(ctx.cwd);
-    touchSession(sid, root, runtime.currentStateDir);
-    logDaily(sid, "compact", result.checkpointId, saved, runtime.currentStateDir);
-  } catch {
-    /* non-fatal: stats bookkeeping only */
-  }
-
-  // S21.2: best-effort consolidation of near-duplicate memories for this repo.
-  // Runs after the per-repo stats touch so `consolidateMemories` can use the
-  // same stateDir. Non-fatal — a failed consolidate never blocks a compaction.
-  // Only runs when new memory ops landed in this pass (otherwise the prior
-  // compaction's consolidate already had its shot — re-running would just
-  // touch every row again with no merges).
-  if (!result.deduped && runtime.memoriesTouchedThisCompaction > 0) {
-    try {
-      const root = resolveRepoRoot(ctx.cwd);
-      void consolidateMemories(runtime.currentStateDir, root).then(
-        (n) => {
-          if (n > 0) runtime.pushTicker(`${C.green}∫${C.reset} consolidated ${n} memory dup${n === 1 ? "" : "s"}`);
-        },
-        () => {
-          /* swallow: consolidate failures must never surface to the user */
-        },
-      );
-    } catch {
-      /* non-fatal */
-    }
-  }
-
-  // S24 review-on-compact: when pressure is high, the just-compacted region is
-  // exactly the context worth remembering, so review it immediately rather than
-  // waiting for the next turn-cadence tick. Uses the shared runMemoryReview
-  // helper (fire-and-forget; doCompact is sync). Best-effort + non-fatal. Only
-  // fires above the `high` band so low-pressure compactions don't pay the cost.
-  if (!result.deduped && config.memoryAutoReview && runtime.pressureBand !== "low" && runtime.pressureBand !== "medium") {
-    void runMemoryReview(runtime, view, "pressure");
-  }
-
-  // Sentinel marker: a non-LLM bookkeeping entry so subsequent triggers can
-  // skip re-vectorizing an already-compacted region (zero token cost).
-  // v0.8.6: gate on !result.deduped so the marker ONLY lands when a genuinely
-  // new checkpoint was created. Without this, every dedup re-fire appended a
-  // fresh sentinel to the real transcript, bloating it and perturbing the
-  // provider KV-cache prefix (the alternating cache-miss regression). Matches
-  // the RAPTOR + vector-index blocks above, which are already !deduped-gated.
-  if (!result.deduped) {
-    pi.appendEntry(MARKER_TYPE, {
-      checkpointId: result.checkpointId,
-      regionHash: result.regionHash,
-      tokenEstimate: result.tokenEstimate,
-      deduped: result.deduped,
-    });
-  }
-
-  // Fix D: refresh the RAPTOR tree for this session so live recall (search) can
-  // serve high-level summaries. Best-effort + non-fatal: never block compaction.
-  // Budget-guarded (RAPTOR_BUDGET_MS) so it can't hang a large session.
-  if (config.raptorEnabled && !result.deduped) {
-    try {
-      const dd = loadDedupConfig();
-      const all = vectorList(runtime.store, sid);
-      const leaves = all.map((cp) => ({
-        id: cp.checkpointId,
-        messages: [],
-        sourceText: cp.normalizedText ?? cp.summary ?? cp.regionHash,
-        embedding: cp.embedding,
-      }));
-      if (leaves.length >= 2) {
-        // S42D: skip the rebuild when the last build is fresh (within
-        // RAPTOR_FRESHNESS_HOURS) and the checkpoint count hasn't drifted by
-        // more than 20%. avoids re-clustering on every compaction when the
-        // tree is still representative. 0 disables (always rebuild).
-        if (
-          dd.RAPTOR_FRESHNESS_HOURS > 0 &&
-          isRaptorTreeFresh(sid, runtime.currentStateDir, dd.RAPTOR_FRESHNESS_HOURS, all.length)
-        ) {
-          runtime.logger?.info("raptor_skip_fresh", { sessionId: sid });
-        } else {
-          // S25: stamp the tree with the newest checkpoint epoch so the
-          // freshness guard in raptorSearchHits can reject stale trees after a
-          // later compaction adds newer checkpoints.
-          const builtAt = all.length > 0 ? Math.max(...all.map((c) => c.timestamp)) : Date.now();
-          runRaptor(
-            leaves,
-            {
-              stateDir: runtime.currentStateDir,
-              sessionId: sid,
-              budgetMs: dd.RAPTOR_BUDGET_MS,
-              clustersPerLevel: dd.RAPTOR_CLUSTERS_PER_LEVEL,
-              consistencyThreshold: dd.RAPTOR_CONSISTENCY,
-              logger: runtime.logger,
-              builtAt: Number.isFinite(builtAt) ? builtAt : Date.now(),
-            },
-          );
-        }
-      }
-    } catch {
-      /* non-fatal: tree refresh never blocks a compaction */
-    }
-  }
-
-  // Slice 2: best-effort mirror of the new checkpoint into the async global
-  // PGlite/HNSW vector index. Fires once per compaction (not per-add), so the
-  // shared global dir is never hammered by concurrent test workers.
-  // Non-fatal: a WASM init failure degrades to the sync scan silently.
-  if (!result.deduped) {
-    try {
-      const all = vectorList(runtime.store, sid);
-      const latest = all.find((cp) => cp.checkpointId === result.checkpointId);
-      if (latest?.embedding) {
-        void indexUpsertEmbedding(
-          repoKey(runtime.currentStateDir),
-          sid,
-          latest.checkpointId,
-          latest.embedding,
-        ).catch(() => {
-          /* non-fatal: index refresh never blocks a compaction */
-        });
-      }
-    } catch {
-      /* non-fatal: index refresh never blocks a compaction */
-    }
-  }
-
-  runtime.setStatus(
-    ctx,
-    runtime.rt.persistedThisSession
-      ? `mega-compact: ${result.checkpointId} · ${saved} tok saved`
-      : `mega-compact: ready`,
-  );
-  runtime.logger.info("compact", {
-    sessionId: sid,
-    checkpointId: result.checkpointId ?? "(deduped)",
-    deduped: result.deduped,
-    tokenEstimate: saved,
-    compactedFrom: result.compactedFrom,
-  });
-  runtime.dashboard.event("compact", {
-    sessionId: sid,
-    checkpointId: result.checkpointId ?? "(deduped)",
-    deduped: result.deduped,
-    tokenEstimate: saved,
-    compactedFrom: result.compactedFrom,
-  });
-  runtime.snapshot(ctx);
-  return { skipped: false, result, keepFrom, saved };
-}
-
-/**
- * Predict whether pi's `ctx.compact()` would throw a no-op error — "Already
- * compacted" or "Nothing to compact (session too small)" — so the auto-trigger
- * can SKIP the call instead of surfacing a hard, user-facing error.
- *
- * Why we can't intercept or suppress it: pi's public `compact()` computes
- * `prepareCompaction()` and throws *before* it emits `session_before_compact`,
- * so our handler there never runs on the no-op path. And `ctx.compact()`'s
- * `onError` callback fires only AFTER pi has already emitted a `compaction_end`
- * event carrying the error message (which the interactive UI renders) — so
- * `onError` cannot mute it either. The only robust fix is to not call
- * `ctx.compact()` when pi would no-op. (pi's own `_runAutoCompaction` path is
- * silent on this same condition; the public path we're forced through is the
- * one that throws.)
- *
- * Skipping is correct, not a compromise: by the time this runs, `runCompact()`
- * has already persisted the recall checkpoint (Path A). The durable on-disk
- * trim is only useful when pi can actually summarize a region; a transcript
- * under pi's `keepRecentTokens` budget is small enough that reloading it on
- * resume isn't a token-growth problem, so the durable trim is unnecessary
- * there anyway.
- *
- * Mirrors pi's `prepareCompaction()` return-undefined conditions (compaction.js):
- *  (1) last entry is a compaction      → "Already compacted"
- *  (2) <2 cut-point messages since the last compaction → nothing to summarize
- *      (a cut point = any non-toolResult message — user/assistant/bash/custom/
- *       branchSummary/compactionSummary — matching pi's isCutPointMessage)
- *  (3) transcript tokens since the last compaction < keepRecentTokens → pi
- *      keeps everything → nothing to summarize
- * `keepRecentTokens` isn't readable from the extension API, so (3) uses the pi
- * default (20000) as a conservative floor; raise it via
- * `MEGACOMPACT_DURABLE_TRIM_FLOOR` if you raise pi's `compact.keepRecentTokens`.
- *
- * Best-effort: on any read error returns true (skip) — skipping a durable trim
- * is always safe; calling `ctx.compact()` on a no-op throws to the user.
+ * compact.ts — delegate-shell for the compaction pipeline (3WF-2 split).
+ *
+ * This file is now a thin barrel: the real bodies live in `./compact/` so the
+ * module stays well under the extensions/ 400-line soft cap. The public API is
+ * UNCHANGED — every existing importer (via `../mega-pipeline.js`, which does
+ * `export * from "./mega-pipeline/compact.js"`) keeps working with ZERO changes
+ * at their call sites. Exports preserved exactly:
+ *   - `RunCompactResult` (type)
+ *   - `runCompact`
+ *   - `piCompactWouldNoop`
  */
-export function piCompactWouldNoop(ctx: ExtensionContext): boolean {
-  try {
-    const branch = ctx.sessionManager.getBranch();
-    if (branch.length === 0) return true;
-    // (1) already compacted — pi throws "Already compacted"
-    if (branch[branch.length - 1].type === "compaction") return true;
-    // boundaryStart = index just after the most recent compaction entry (or 0)
-    let boundaryStart = 0;
-    for (let i = branch.length - 1; i >= 0; i--) {
-      if (branch[i].type === "compaction") { boundaryStart = i + 1; break; }
-    }
-    let cutPoints = 0;
-    let tokens = 0;
-    for (let i = boundaryStart; i < branch.length; i++) {
-      const e = branch[i];
-      if (e.type === "compaction") continue;
-      let isCut = false;
-      for (const m of sessionEntryToContextMessages(e)) {
-        // pi's isCutPointMessage: every role except toolResult
-        if ((m as { role?: string }).role !== "toolResult") isCut = true;
-        const c = (m as { content?: unknown }).content;
-        const text =
-          typeof c === "string" ? c
-          : Array.isArray(c)
-            ? (c as { text?: string }[]).map((b) => b?.text ?? "").join(" ")
-            : "";
-        if (text) tokens += estimateBlockTokens(text);
-      }
-      if (isCut) cutPoints++;
-    }
-    // (2) need >=2 cut points so the kept cut isn't the first message
-    if (cutPoints < 2) return true;
-    // (3) transcript under pi's keepRecentTokens budget → pi keeps everything
-    if (tokens < durableTrimFloorTokens()) return true;
-    return false;
-  } catch {
-    return true; // safe: skip the durable trim rather than risk a user-facing throw
-  }
-}
 
-/** pi's default keepRecentTokens (compaction settings). Override with
- *  MEGACOMPACT_DURABLE_TRIM_FLOOR if you raise pi's compact.keepRecentTokens. */
-function durableTrimFloorTokens(): number {
-  const raw = process.env.MEGACOMPACT_DURABLE_TRIM_FLOOR;
-  if (raw !== undefined && Number.isFinite(Number(raw))) return Number(raw);
-  return 20_000;
-}
+export { runCompact, type RunCompactResult } from "./compact/run.js";
+export { piCompactWouldNoop } from "./compact/noop.js";

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-mega-compact",
-  "version": "0.20.85",
+  "version": "0.20.86",
   "description": "Layered, local, vector-backed context compressor for pi — supersede/collapse/cluster compaction with deduped inline recall.",
   "type": "module",
   "license": "BSD-3-Clause",

package/src/failback/compact.ts

@@ -0,0 +1,122 @@
+/**
+ * src/failback/compact.ts — 3WF-2 candidate-veto + vote module (pure, advisory).
+ *
+ * The production bug this fixes: compaction "succeeded" (a checkpoint was
+ * persisted, `saved` grew) while the LIVE WINDOW (`currentTokens`) never
+ * shrank — because `saved` is a cumulative SQLite total, not the working-set
+ * delta. This module builds competing summary candidates (extractive vs
+ * cluster/raptor) and VOTES which one, if any, is worth replacing the
+ * supersede-only result. It is purely advisory/observational: it never mutates
+ * a checkpoint, never overwrites `result.summary`, and returning `null` means
+ * "keep the supersede-only result" — the caller must NOT substitute a summary.
+ *
+ * Pure: no store mutation, no I/O, no network, no console.*. pi-agnostic
+ * (imports only from src/). Designed for the 3WF umbrella flag gate at the
+ * extension layer (see extensions/mega-pipeline/compact/vote.ts).
+ */
+
+import type { EngineMessage } from "../types.js";
+import { collectRecentUserRequests, summarizeMessages } from "../compact.js";
+import { summarizeCluster } from "../dedup/raptor/summarizer.js";
+import { estimateBlockTokens } from "../tokens.js";
+import type { CompactCandidate } from "./types.js";
+
+/**
+ * Default floor (tokens of net reduction) below which a candidate vote is
+ * REJECTED, returning `null` (keep the supersede-only result).
+ *
+ * Rationale: a candidate that reduces the region by fewer than 1 token is not
+ * meaningfully smaller than the compacted region it would replace — swapping
+ * the supersede-only result for it buys nothing and only adds a (possibly
+ * less faithful) summary. The floor therefore requires the voted summary to be
+ * STRICTLY smaller than the compacted region. Set to 1 (minimally defensible:
+ * the summary must actually be smaller). Overridable via `opts.floor`.
+ */
+export const DEFAULT_VOTE_FLOOR_TOKENS = 1;
+
+/** Strip a trailing ellipsis/truncation marker from a needle before containment. */
+function stripEllipsis(s: string): string {
+	// collectRecentUserRequests truncates to 160 chars via compact.ts's truncate,
+	// which appends the U+2026 ellipsis when it cuts. Drop it for a fair test.
+	return s.replace(/…\s*$/u, "").trim();
+}
+
+/** Normalize for containment: collapse whitespace, lowercase. */
+function normalize(s: string): string {
+	return s.replace(/\s+/g, " ").trim().toLowerCase();
+}
+
+/** True when `summary` contains the content of EVERY recent user request. */
+export function signalPreserved(summary: string, messages: EngineMessage[]): boolean {
+	const requests = collectRecentUserRequests(messages, 3);
+	if (requests.length === 0) return true; // nothing to preserve
+	const haystack = normalize(summary);
+	return requests.every((r) => haystack.includes(normalize(stripEllipsis(r))));
+}
+
+/**
+ * Build the two competing candidates (extractive + cluster/raptor) for a
+ * compacted message region. Degenerate (empty/whitespace-only) summaries are
+ * VETOED — never returned. Both candidates use estimateBlockTokens(summary) for
+ * a single consistent token basis so the vote compares like with like (the
+ * cluster path's own tokenEstimate is intentionally ignored for fairness).
+ */
+export function buildCandidates(messages: EngineMessage[]): CompactCandidate[] {
+	const out: CompactCandidate[] = [];
+
+	const extractive = summarizeMessages(messages);
+	if (extractive.trim().length > 0) {
+		out.push({
+			source: "extractive",
+			summary: extractive,
+			tokenEstimate: estimateBlockTokens(extractive),
+			signalPreserved: signalPreserved(extractive, messages),
+		});
+	}
+
+	// summarizeCluster returns deterministic extractive when MEGACOMPACT_RAPTOR_MODEL
+	// is unset and the local-only Ollama variant when set — so using it makes the
+	// Ollama path an insertion that adds NO new LLM call site for the on-by-default
+	// extraction. No behavior change for the default config.
+	const cluster = summarizeCluster(messages).summary;
+	if (cluster.trim().length > 0) {
+		out.push({
+			source: "cluster",
+			summary: cluster,
+			tokenEstimate: estimateBlockTokens(cluster),
+			signalPreserved: signalPreserved(cluster, messages),
+		});
+	}
+
+	return out;
+}
+
+/**
+ * Vote the best candidate. `score = reduction * (signalPreserved ? 1 : 0.5)`
+ * where `reduction = tokensBefore - candidate.tokenEstimate`. Ties resolve to
+ * the EARLIER (extractive) candidate for determinism. Returns `null` when the
+ * winner's score is below `opts.floor` (default DEFAULT_VOTE_FLOOR_TOKENS) —
+ * caller MUST keep the supersede-only result and must NOT substitute a summary.
+ */
+export function voteCandidate(
+	messages: EngineMessage[],
+	tokensBefore: number,
+	opts: { floor?: number } = {},
+): CompactCandidate | null {
+	const floor = opts.floor ?? DEFAULT_VOTE_FLOOR_TOKENS;
+	const candidates = buildCandidates(messages);
+	let best: CompactCandidate | null = null;
+	let bestScore = -Infinity;
+	for (const c of candidates) {
+		const reduction = tokensBefore - c.tokenEstimate;
+		const score = reduction * (c.signalPreserved ? 1 : 0.5);
+		// Earlier candidate wins ties (strict > keeps insertion order = extractive first).
+		if (score > bestScore) {
+			bestScore = score;
+			best = c;
+		}
+	}
+	if (best === null) return null;
+	if (bestScore < floor) return null;
+	return best;
+}

package/src/failback/types.ts

@@ -44,3 +44,47 @@ export interface GuardOpts {
 	/** Max hits to recall (mirrors autoInlineK). */
 	limit: number;
 }
+
+/** A competing compaction summary candidate produced by the 3-source vote. */
+export interface CompactCandidate {
+	/** Which generator produced this candidate (structural telemetry label —
+	 *  never infer the source by sniffing the summary text). */
+	source: "extractive" | "cluster";
+	/** The candidate summary text (extractive or cluster/raptor variant). */
+	summary: string;
+	/** Estimated token cost of the candidate summary (estimateBlockTokens basis). */
+	tokenEstimate: number;
+	/** True when the summary preserves every recent user request signal. */
+	signalPreserved: boolean;
+}
+
+/**
+ * The measured reduction verdict across consecutive `context` events in the
+ * LIVE WINDOW (the model's current working tokens), NOT the stored-checkpoint
+ * `saved` metric. The live-window `currentTokens` delta is the real signal of
+ * whether compaction actually freed working context; the stored `saved` field
+ * is a cumulative SQLite total that can look healthy while the live window is
+ * unchanged — the false metric behind the production thrash bug this sprint
+ * fixes. `liveBefore`/`liveAfter` are the live-window token counts bracketing
+ * the compaction.
+ */
+export interface ReductionVerdict {
+	/** True when the live window measurably shrank after compaction. */
+	effective: boolean;
+	/** Live-window token count before the compaction event. */
+	liveBefore: number;
+	/** Live-window token count after the compaction event. */
+	liveAfter: number;
+}
+
+/**
+ * State of the compaction thrash guard. Arms after a compaction that produced
+ * no live-window reduction, so we do not re-fire into a window that cannot
+ * shrink. Re-arms or clears as the live window grows again.
+ */
+export interface ThrashGuardState {
+	/** Live-window token count below which re-firing is refused (guard active). */
+	blockedUntilTokens: number;
+	/** ms epoch at which the guard was armed. */
+	armedAt: number;
+}

package/src/store/sqlite/meta.ts

@@ -55,6 +55,42 @@ export function getMetaNumber(key: string, stateDir: string = getStateDir()): nu
   return Number.isFinite(n) ? n : 0;
 }
 
+/**
+ * Upsert a single numeric meta key to an absolute value (NOT a cumulative
+ * counter). Follows the `addTokensSaved` INSERT-ON-CONFLICT pattern with a
+ * fully parameterized query (PREVENT-002: no SQL string concat — `key` and
+ * `value` are both bound, never interpolated). The only write is the standard
+ * ON CONFLICT upsert of THIS key's own value; no other key is touched, no
+ * DELETE is issued.
+ *
+ * Used by the 3WF-2 ThrashGuard to persist exactly two keys:
+ *  - `thrasguard.baseline_tokens` — the live-window token count at the moment
+ *    an ineffective compaction was observed (the baseline the guard re-arms from).
+ *  - `thrasguard.blocked_until` — the live-window token count below which
+ *    re-firing is refused (guard active).
+ *
+ * Non-finite input (NaN / ±Infinity) is rejected: the extension must never
+ * persist a non-number into the meta table (getMetaNumber would read it back as
+ * 0), so we return early, non-fatal. Best-effort: any store failure is swallowed.
+ */
+export function setMetaNumber(
+  key: string,
+  value: number,
+  stateDir: string = getStateDir(),
+): void {
+  if (key.length === 0) return;
+  if (!Number.isFinite(value)) return;
+  try {
+    const db = openStore(stateDir);
+    db.prepare(
+      `INSERT INTO meta(key, value) VALUES(?, ?)
+       ON CONFLICT(key) DO UPDATE SET value = ?`,
+    ).run(key, String(value), String(value));
+  } catch {
+    /* non-fatal: meta writes never break the agent loop */
+  }
+}
+
 /** Atomically add `delta` to an integer meta counter. */
 function incMeta(key: string, delta: number, stateDir: string = getStateDir()): void {
   if (!(delta > 0)) return;