akm-cli 0.9.0-beta.1 → 0.9.0-beta.10

package/dist/indexer/index-writer-lock.js

@@ -0,0 +1,99 @@
+// This Source Code Form is subject to the terms of the Mozilla Public
+// License, v. 2.0. If a copy of the MPL was not distributed with this
+// file, You can obtain one at https://mozilla.org/MPL/2.0/.
+import fs from "node:fs";
+import path from "node:path";
+import { probeLock, releaseLock, releaseLockIfOwned, tryAcquireLockSync } from "../core/file-lock.js";
+import { getDbPath, getIndexWriterLockPath } from "../core/paths.js";
+const INDEX_WRITER_LOCK_STALE_AFTER_MS = 12 * 60 * 60 * 1000;
+const INDEX_WRITER_WAIT_MS = 100;
+const heldLocks = new Map();
+function buildPayload(purpose, pid = process.pid) {
+    return JSON.stringify({
+        pid,
+        purpose,
+        dbPath: getDbPath(),
+        startedAt: new Date().toISOString(),
+    });
+}
+function delay(ms) {
+    return new Promise((resolve) => setTimeout(resolve, ms));
+}
+function throwIfAborted(signal) {
+    if (!signal?.aborted)
+        return;
+    throw signal.reason instanceof Error ? signal.reason : new Error("index writer wait aborted");
+}
+function releaseHeldLock(lockPath) {
+    const held = heldLocks.get(lockPath);
+    if (!held)
+        return;
+    held.depth -= 1;
+    if (held.depth > 0)
+        return;
+    heldLocks.delete(lockPath);
+    process.off("exit", held.exitHandler);
+    releaseLockIfOwned(lockPath, process.pid);
+}
+function retainHeldLock(lockPath) {
+    const existing = heldLocks.get(lockPath);
+    if (existing) {
+        existing.depth += 1;
+        return { lockPath, release: () => releaseHeldLock(lockPath) };
+    }
+    const exitHandler = () => releaseLockIfOwned(lockPath, process.pid);
+    process.on("exit", exitHandler);
+    heldLocks.set(lockPath, { depth: 1, exitHandler });
+    return { lockPath, release: () => releaseHeldLock(lockPath) };
+}
+function detachHeldLock(lockPath) {
+    const held = heldLocks.get(lockPath);
+    if (!held)
+        return;
+    heldLocks.delete(lockPath);
+    process.off("exit", held.exitHandler);
+}
+export async function acquireIndexWriterLease(options) {
+    const mode = options.mode ?? "wait";
+    const lockPath = getIndexWriterLockPath();
+    fs.mkdirSync(path.dirname(lockPath), { recursive: true });
+    if (heldLocks.has(lockPath)) {
+        return retainHeldLock(lockPath);
+    }
+    while (true) {
+        throwIfAborted(options.signal);
+        if (tryAcquireLockSync(lockPath, buildPayload(options.purpose))) {
+            return retainHeldLock(lockPath);
+        }
+        const probe = probeLock(lockPath, { staleAfterMs: INDEX_WRITER_LOCK_STALE_AFTER_MS });
+        if (probe.state === "held" && probe.holderPid === process.pid) {
+            return retainHeldLock(lockPath);
+        }
+        if (probe.state === "stale") {
+            releaseLock(lockPath);
+            continue;
+        }
+        if (mode === "try")
+            return undefined;
+        await delay(INDEX_WRITER_WAIT_MS);
+    }
+}
+export async function withIndexWriterLease(options, run) {
+    const lease = await acquireIndexWriterLease(options);
+    if (!lease) {
+        throw new Error(`index writer lease unavailable for ${options.purpose}`);
+    }
+    try {
+        return await run();
+    }
+    finally {
+        lease.release();
+    }
+}
+export function handoffIndexWriterLeaseToPid(lease, pid, purpose) {
+    fs.writeFileSync(lease.lockPath, buildPayload(purpose, pid), "utf8");
+    detachHeldLock(lease.lockPath);
+}
+export function probeIndexWriterLease() {
+    return probeLock(getIndexWriterLockPath(), { staleAfterMs: INDEX_WRITER_LOCK_STALE_AFTER_MS });
+}

package/dist/indexer/indexer.js

@@ -12,6 +12,7 @@ import { resolveIndexPassLLM } from "../llm/index-passes.js";
 import { takeWorkflowDocument } from "../workflows/runtime/document-cache.js";
 import { clearStaleCacheEntries, closeDatabase, deleteEntriesByDir, deleteEntriesByIds, deleteEntriesByStashDir, deleteIndexDirStatesByStashDir, getAllEntriesForEmbedding, getEmbeddableEntryCount, getEmbeddingCount, getEntriesByDir, getEntryCount, getIndexDirState, getMeta, isVecAvailable, openDatabase, openExistingDatabase, rebuildFts, relinkUsageEvents, setMeta, upsertEmbedding, upsertEntry, upsertIndexDirState, upsertUtilityScore, upsertWorkflowDocument, warnIfVecMissing, } from "./db/db.js";
 import { deleteStoredGraph } from "./db/graph-db.js";
+import { withIndexWriterLease } from "./index-writer-lock.js";
 import { applyCuratedFrontmatter, applyWikiFrontmatter, generateMetadataFlat, isEnrichmentComplete, isWorkflowSkipWarning, loadStashFile, shouldIndexStashFile, } from "./passes/metadata.js";
 import { buildSearchText } from "./search/search-fields.js";
 import { classifySemanticFailure, clearSemanticStatus, deriveSemanticProviderFingerprint, writeSemanticStatus, } from "./search/semantic-status.js";
@@ -225,119 +226,121 @@ function runCleanPass(db, dryRun) {
 }
 // ── Indexer ──────────────────────────────────────────────────────────────────
 export async function akmIndex(options) {
-    const stashDir = options?.stashDir || resolveStashDir();
-    const onProgress = options?.onProgress ?? (() => { });
-    const signal = options?.signal;
-    const reEnrich = options?.reEnrich === true;
-    const full = options?.full === true;
-    const clean = options?.clean === true;
-    const dryRun = options?.dryRun === true;
-    // Load config and resolve all stash sources
-    const { loadConfig } = await import("../core/config/config.js");
-    const config = loadConfig();
-    // One-time, read-only guard: warn if the writable stash still holds an
-    // un-migrated `vaults/` directory. In 0.9.0 the indexer skips `vaults/`
-    // entirely, so an unmigrated vault's `.env` data would silently never be
-    // indexed. Non-destructive — only stats, never reads/writes/deletes.
-    const { warnOnUnmigratedVaults } = await import("./usage/unmigrated-vaults-guard.js");
-    warnOnUnmigratedVaults(stashDir);
-    // Ensure git stash caches are extracted before resolving stash dirs,
-    // so their content directories exist on disk for the walker to discover.
-    const { ensureSourceCaches, resolveSourceEntries } = await import("./search/search-source.js");
-    await ensureSourceCaches(config, { force: full });
-    const allSourceEntries = resolveSourceEntries(stashDir, config);
-    const allSourceDirs = allSourceEntries.map((s) => s.path);
-    const t0 = Date.now();
-    // Open database — pass embedding dimension from config if available
-    const dbPath = getDbPath();
-    const embeddingDim = config.embedding?.dimension;
-    const db = openDatabase(dbPath, embeddingDim ? { embeddingDim } : undefined);
-    try {
-        // Determine incremental vs full mode
-        const prevStashDir = getMeta(db, "stashDir");
-        const prevBuiltAt = getMeta(db, "builtAt");
-        const isIncremental = !full && prevStashDir === stashDir && !!prevBuiltAt;
-        const builtAtMs = isIncremental && prevBuiltAt ? new Date(prevBuiltAt).getTime() : 0;
-        // Assemble the run context
-        const ctx = {
-            db,
-            config,
-            sources: allSourceEntries,
-            sourceDirs: allSourceDirs,
-            full,
-            reEnrich,
-            stashDir,
-            onProgress,
-            signal,
-            timing: {
-                t0,
-                tWalkStart: t0,
-                tWalkEnd: t0,
-                tLlmEnd: t0,
-                tFtsEnd: t0,
-                tEmbedEnd: t0,
-            },
-            isIncremental,
-            builtAtMs,
-            hadRemovedSources: false,
-            scannedDirs: 0,
-            skippedDirs: 0,
-            generatedCount: 0,
-            walkWarnings: [],
-            dirsNeedingLlm: [],
-            embeddingResult: null,
-            graphExtractionResult: null,
-        };
-        onProgress({
-            phase: "summary",
-            message: buildIndexSummaryMessage({
+    return withIndexWriterLease({ purpose: "akm-index", signal: options?.signal }, async () => {
+        const stashDir = options?.stashDir || resolveStashDir();
+        const onProgress = options?.onProgress ?? (() => { });
+        const signal = options?.signal;
+        const reEnrich = options?.reEnrich === true;
+        const full = options?.full === true;
+        const clean = options?.clean === true;
+        const dryRun = options?.dryRun === true;
+        // Load config and resolve all stash sources
+        const { loadConfig } = await import("../core/config/config.js");
+        const config = loadConfig();
+        // One-time, read-only guard: warn if the writable stash still holds an
+        // un-migrated `vaults/` directory. In 0.9.0 the indexer skips `vaults/`
+        // entirely, so an unmigrated vault's `.env` data would silently never be
+        // indexed. Non-destructive — only stats, never reads/writes/deletes.
+        const { warnOnUnmigratedVaults } = await import("./usage/unmigrated-vaults-guard.js");
+        warnOnUnmigratedVaults(stashDir);
+        // Ensure git stash caches are extracted before resolving stash dirs,
+        // so their content directories exist on disk for the walker to discover.
+        const { ensureSourceCaches, resolveSourceEntries } = await import("./search/search-source.js");
+        await ensureSourceCaches(config, { force: full });
+        const allSourceEntries = resolveSourceEntries(stashDir, config);
+        const allSourceDirs = allSourceEntries.map((s) => s.path);
+        const t0 = Date.now();
+        // Open database — pass embedding dimension from config if available
+        const dbPath = getDbPath();
+        const embeddingDim = config.embedding?.dimension;
+        const db = openDatabase(dbPath, embeddingDim ? { embeddingDim } : undefined);
+        try {
+            // Determine incremental vs full mode
+            const prevStashDir = getMeta(db, "stashDir");
+            const prevBuiltAt = getMeta(db, "builtAt");
+            const isIncremental = !full && prevStashDir === stashDir && !!prevBuiltAt;
+            const builtAtMs = isIncremental && prevBuiltAt ? new Date(prevBuiltAt).getTime() : 0;
+            // Assemble the run context
+            const ctx = {
+                db,
+                config,
+                sources: allSourceEntries,
+                sourceDirs: allSourceDirs,
+                full,
+                reEnrich,
+                stashDir,
+                onProgress,
+                signal,
+                timing: {
+                    t0,
+                    tWalkStart: t0,
+                    tWalkEnd: t0,
+                    tLlmEnd: t0,
+                    tFtsEnd: t0,
+                    tEmbedEnd: t0,
+                },
+                isIncremental,
+                builtAtMs,
+                hadRemovedSources: false,
+                scannedDirs: 0,
+                skippedDirs: 0,
+                generatedCount: 0,
+                walkWarnings: [],
+                dirsNeedingLlm: [],
+                embeddingResult: null,
+                graphExtractionResult: null,
+            };
+            onProgress({
+                phase: "summary",
+                message: buildIndexSummaryMessage({
+                    mode: isIncremental ? "incremental" : "full",
+                    sourcesCount: allSourceDirs.length,
+                    semanticSearchMode: config.semanticSearchMode,
+                    embeddingProvider: getEmbeddingProvider(config.embedding),
+                    llmEnabled: !!resolveIndexPassLLM("enrichment", config),
+                    vecAvailable: isVecAvailable(db),
+                }),
+            });
+            // ── Phase sequence ───────────────────────────────────────────────────────
+            await runSourceCachePhase(ctx);
+            await runWalkPhase(ctx);
+            await runEmbeddingPhase(ctx);
+            await runFinalizePhase(ctx);
+            // ────────────────────────────────────────────────────────────────────────
+            const { _verification: verification, _totalEntries: totalEntries } = ctx;
+            const { timing } = ctx;
+            // ── Clean pass ───────────────────────────────────────────────────────────
+            // After the normal index completes, remove entries whose source files no
+            // longer exist on disk. Remote entries (empty file_path) are skipped.
+            let cleanResult;
+            if (clean) {
+                cleanResult = runCleanPass(db, dryRun);
+            }
+            // ────────────────────────────────────────────────────────────────────────
+            return {
+                stashDir,
+                totalEntries,
+                generatedMetadata: ctx.generatedCount,
+                indexPath: dbPath,
                 mode: isIncremental ? "incremental" : "full",
-                sourcesCount: allSourceDirs.length,
-                semanticSearchMode: config.semanticSearchMode,
-                embeddingProvider: getEmbeddingProvider(config.embedding),
-                llmEnabled: !!resolveIndexPassLLM("enrichment", config),
-                vecAvailable: isVecAvailable(db),
-            }),
-        });
-        // ── Phase sequence ───────────────────────────────────────────────────────
-        await runSourceCachePhase(ctx);
-        await runWalkPhase(ctx);
-        await runEmbeddingPhase(ctx);
-        await runFinalizePhase(ctx);
-        // ────────────────────────────────────────────────────────────────────────
-        const { _verification: verification, _totalEntries: totalEntries } = ctx;
-        const { timing } = ctx;
-        // ── Clean pass ───────────────────────────────────────────────────────────
-        // After the normal index completes, remove entries whose source files no
-        // longer exist on disk. Remote entries (empty file_path) are skipped.
-        let cleanResult;
-        if (clean) {
-            cleanResult = runCleanPass(db, dryRun);
+                directoriesScanned: ctx.scannedDirs,
+                directoriesSkipped: ctx.skippedDirs,
+                ...(ctx.walkWarnings.length > 0 ? { warnings: ctx.walkWarnings } : {}),
+                verification,
+                timing: {
+                    totalMs: Date.now() - timing.t0,
+                    walkMs: timing.tWalkEnd - timing.tWalkStart,
+                    llmMs: timing.tLlmEnd - timing.tWalkEnd,
+                    embedMs: timing.tEmbedEnd - timing.tLlmEnd,
+                    ftsMs: timing.tFtsEnd - timing.tEmbedEnd,
+                },
+                ...(cleanResult !== undefined ? { clean: cleanResult } : {}),
+            };
         }
-        // ────────────────────────────────────────────────────────────────────────
-        return {
-            stashDir,
-            totalEntries,
-            generatedMetadata: ctx.generatedCount,
-            indexPath: dbPath,
-            mode: isIncremental ? "incremental" : "full",
-            directoriesScanned: ctx.scannedDirs,
-            directoriesSkipped: ctx.skippedDirs,
-            ...(ctx.walkWarnings.length > 0 ? { warnings: ctx.walkWarnings } : {}),
-            verification,
-            timing: {
-                totalMs: Date.now() - timing.t0,
-                walkMs: timing.tWalkEnd - timing.tWalkStart,
-                llmMs: timing.tLlmEnd - timing.tWalkEnd,
-                embedMs: timing.tEmbedEnd - timing.tLlmEnd,
-                ftsMs: timing.tFtsEnd - timing.tEmbedEnd,
-            },
-            ...(cleanResult !== undefined ? { clean: cleanResult } : {}),
-        };
-    }
-    finally {
-        closeDatabase(db);
-    }
+        finally {
+            closeDatabase(db);
+        }
+    });
 }
 // ── Extracted helpers for indexing ────────────────────────────────────────────
 async function indexEntries(db, allSourceEntries, isIncremental, builtAtMs, hadRemovedSources, doFullDelete = false, onProgress) {

package/dist/indexer/passes/memory-inference.js

@@ -119,6 +119,26 @@ export async function runMemoryInferencePass(ctx) {
         // 2026-05-26).
         if (signal?.aborted)
             return { aborted: true };
+        // Pre-check (#588): when `<parent>.derived.md` is already on disk the
+        // inference is by definition complete — the parent only looks pending
+        // because `markParentProcessed` never ran (process killed between the
+        // child write and the mark) or the child was created externally (e.g.
+        // consolidation). Skip the LLM/cache call entirely and mark the parent
+        // so it never re-pends. Before this check, production measurements
+        // showed ~55% of the pass's LLM budget re-deriving such parents only to
+        // discover the existing child after the fact.
+        if (fs.existsSync(derivedChildPath(record))) {
+            markParentProcessed(record);
+            return {
+                skipped: false,
+                splitParent: false,
+                written: 0,
+                fromCache: false,
+                retryAttempts: 0,
+                childExists: true,
+                precheck: true,
+            };
+        }
         // Incremental cache: skip LLM call when body hash is unchanged and
         // --re-enrich was not requested. The cache ref is the absolute file path.
         const validate = (raw) => {
@@ -171,23 +191,30 @@ export async function runMemoryInferencePass(ctx) {
             return { skipped: false, splitParent: true, written: writeOutcome.written, fromCache, retryAttempts };
         }
         // LLM produced a valid derived draft but no file was written — either
-        // because `<parent>.derived.md` already exists on disk or
-        // `writeAssetToSource` threw. Categorise as `childExists` so the
-        // attempt is accounted for in health metrics rather than vanishing
-        // into the freshAttempts denominator.
+        // because `<parent>.derived.md` appeared on disk after the pre-check
+        // above (a rare mid-flight race) or `writeAssetToSource` threw.
+        // Categorise as `childExists` so the consumed attempt is accounted for
+        // in health metrics rather than vanishing into the freshAttempts
+        // denominator.
         //
-        // When the child already exists on disk the inference is, by definition,
-        // already complete — so mark the parent processed here too (#550).
-        // Without this, `isPendingMemory()` re-queues the same parent every run
-        // (the `written > 0` path was previously the only site that marks it),
-        // causing permanent re-queueing and wasted LLM calls. A genuine write
-        // *failure* (`writeAssetToSource` threw) must NOT mark the parent — it
-        // should be retried next run — so we key off the explicit `childExists`
-        // outcome rather than the conflated `written === 0`.
+        // When the child exists the inference is, by definition, complete — so
+        // mark the parent processed here too (#550), otherwise
+        // `isPendingMemory()` re-queues the same parent every run. A genuine
+        // write *failure* (`writeAssetToSource` threw) must NOT mark the parent
+        // — it should be retried next run — so we key off the explicit
+        // `childExists` outcome rather than the conflated `written === 0`.
         if (writeOutcome.childExists) {
             markParentProcessed(record);
         }
-        return { skipped: false, splitParent: false, written: 0, fromCache, retryAttempts, childExists: true };
+        return {
+            skipped: false,
+            splitParent: false,
+            written: 0,
+            fromCache,
+            retryAttempts,
+            childExists: true,
+            precheck: false,
+        };
     }, 
     // Default concurrency of 4 for cloud APIs. Set `llm.concurrency: 1`
     // in config.json for local model servers (LM Studio, Ollama).
@@ -224,11 +251,16 @@ export async function runMemoryInferencePass(ctx) {
             result.writtenFacts += res.written;
         }
         else if ("childExists" in res && res.childExists) {
-            // LLM call was consumed but the derived file already existed (or the
-            // write threw). Track separately so this category is observable in
-            // health output and stops bleeding into the freshAttempts denominator.
+            // Derived child already on disk. Track separately so this category is
+            // observable in health output and stops bleeding into the
+            // freshAttempts denominator. Pre-check skips (#588) are the routine
+            // self-healing path — no LLM attempt was consumed and the parent has
+            // been marked processed — so only the rare post-LLM case (mid-flight
+            // race or write failure) warrants a per-ref warning.
             result.skippedChildExists += 1;
-            warn(`memory inference: derived child for ${pending[i]?.ref ?? "<unknown>"} already existed or write failed; counted as skippedChildExists`);
+            if (!res.precheck) {
+                warn(`memory inference: derived child for ${pending[i]?.ref ?? "<unknown>"} already existed or write failed; counted as skippedChildExists`);
+            }
         }
         else {
             // The per-record state machine should cover every outcome. A hit here
@@ -324,6 +356,14 @@ function toMemoryName(memoriesDir, filePath) {
     // user has organised under memories/.
     return rel.replace(/\\/g, "/").replace(/\.md$/i, "");
 }
+/**
+ * Absolute path of the derived child for a parent memory. Single source of
+ * truth for the `<parent>.derived.md` naming convention — used both by the
+ * pre-LLM existence check (#588) and the write path.
+ */
+function derivedChildPath(parent) {
+    return path.join(parent.stashRoot, "memories", `${parent.name}.derived.md`);
+}
 async function writeDerivedMemory(parent, derived) {
     const writeTarget = {
         kind: "filesystem",
@@ -338,11 +378,10 @@ async function writeDerivedMemory(parent, derived) {
     };
     const childName = `${parent.name}.derived`;
     const childRefStr = `memory:${childName}`;
-    const childPath = path.join(parent.stashRoot, "memories", `${childName}.md`);
-    if (fs.existsSync(childPath)) {
-        // The derived child is already on disk — inference for this parent is
-        // complete. Report `childExists` so the caller marks the parent processed
-        // (#550) instead of re-queueing it forever.
+    if (fs.existsSync(derivedChildPath(parent))) {
+        // The derived child appeared on disk after the caller's pre-check (#588)
+        // — a rare mid-flight race. Report `childExists` so the caller marks the
+        // parent processed (#550) instead of re-queueing it forever.
         return { written: 0, childExists: true };
     }
     try {

package/dist/integrations/harnesses/claude/session-log.js

@@ -5,7 +5,19 @@ import fs from "node:fs";
 import os from "node:os";
 import path from "node:path";
 import { extractInlineRefMentions } from "../../session-logs/inline-refs.js";
-const CLAUDE_PROJECTS_DIR = path.join(os.homedir(), ".claude", "projects");
+/**
+ * Root directory holding Claude Code's per-project JSONL session logs.
+ *
+ * Resolved per call (not memoized at module load) so the `AKM_CLAUDE_PROJECTS_DIR`
+ * override can be set after import. The override exists so tests — and the
+ * isolated-storage sandbox — can point the scan at an empty fixture directory
+ * instead of the real `~/.claude/projects`, which on an actively-used machine
+ * holds many large session files and would make `akm health` (which scans it
+ * synchronously) slow and non-hermetic.
+ */
+function claudeProjectsDir() {
+    return process.env.AKM_CLAUDE_PROJECTS_DIR ?? path.join(os.homedir(), ".claude", "projects");
+}
 /**
  * Parse a single Claude Code JSONL event into a normalized {@link SessionEvent}.
  * Returns `undefined` for events that don't carry textual content (file
@@ -93,11 +105,11 @@ export class ClaudeCodeProvider {
     // HARNESS_BY_ID.get("claude").runtimeId.
     name = "claude-code";
     isAvailable() {
-        return fs.existsSync(CLAUDE_PROJECTS_DIR);
+        return fs.existsSync(claudeProjectsDir());
     }
     *readEvents(input) {
         try {
-            for (const jsonlPath of this.#walkJsonl(CLAUDE_PROJECTS_DIR)) {
+            for (const jsonlPath of this.#walkJsonl(claudeProjectsDir())) {
                 const stat = fs.statSync(jsonlPath);
                 if (stat.mtimeMs < input.sinceMs)
                     continue;
@@ -128,7 +140,7 @@ export class ClaudeCodeProvider {
         }
     }
     listSessions(input = {}) {
-        const root = input.location ?? CLAUDE_PROJECTS_DIR;
+        const root = input.location ?? claudeProjectsDir();
         const sinceMs = input.sinceMs ?? 0;
         const summaries = [];
         try {
@@ -286,7 +298,7 @@ export class ClaudeCodeProvider {
                 const full = path.join(dir, entry.name);
                 if (entry.isDirectory())
                     yield* this.#walkJsonl(full);
-                else if (entry.name.endsWith(".jsonl"))
+                else if (entry.name.endsWith(".jsonl") && entry.name !== "journal.jsonl")
                     yield full;
             }
         }

package/dist/llm/client.js

@@ -14,6 +14,7 @@ import { fetchWithTimeout } from "../core/common.js";
 import { resolveSecret } from "../core/config/config.js";
 import { escapeJsonStringControls, parseJsonResponse, stripCodeFences, stripThinkBlocks } from "../core/parse.js";
 import { warnVerbose } from "../core/warn.js";
+import { emitLlmUsage, extractUsageTokens } from "./usage-telemetry.js";
 // Re-export shared parse utilities so existing importers of `client.ts` continue
 // to resolve `parseJsonResponse` and `parseEmbeddedJsonResponse` from this module.
 export { escapeJsonStringControls, parseEmbeddedJsonResponse, parseJsonResponse, stripCodeFences, stripThinkBlocks, } from "../core/parse.js";
@@ -118,9 +119,23 @@ function looksLikeContextOverflow(message) {
 /**
  * Decide whether a first-attempt {@link LlmCallError} is eligible for a single
  * retry. Retryable: HTTP 5xx (`provider_error` with statusCode >= 500) and
- * `network_error` whose message looks like a transient connection reset
- * (ECONNRESET / EPIPE / "fetch failed"). NOT retryable: 4xx, `rate_limited`
- * (429), `timeout`, `parse_error`, and context-overflow-classified errors.
+ * `network_error` whose message looks like a transient connection drop.
+ * NOT retryable: 4xx, `rate_limited` (429), `timeout`, `parse_error`, and
+ * context-overflow-classified errors.
+ *
+ * The connection-drop heuristic covers the substrings emitted across runtimes
+ * for a mid-flight socket close:
+ *  - `ECONNRESET` / `EPIPE` — Node/libuv socket reset codes
+ *  - `fetch failed` — undici's generic wrapper message
+ *  - `socket connection was closed` — Bun's message for a dropped connection
+ *    (e.g. "The socket connection was closed unexpectedly.")
+ *  - `terminated` / `other side closed` — undici's phrasings for the same
+ *
+ * These all describe a transient transport failure where a second attempt can
+ * legitimately succeed, which is exactly the case a single bounded retry is
+ * meant to absorb. Before this list was widened, Bun's "socket connection was
+ * closed unexpectedly" fell through unretried and surfaced as a recurring
+ * failure in the improve/reflect and capability-probe flows.
  */
 function isRetryable(err) {
     if (looksLikeContextOverflow(err.message))
@@ -130,7 +145,12 @@ function isRetryable(err) {
     }
     if (err.code === "network_error") {
         const lower = err.message.toLowerCase();
-        return lower.includes("econnreset") || lower.includes("epipe") || lower.includes("fetch failed");
+        return (lower.includes("econnreset") ||
+            lower.includes("epipe") ||
+            lower.includes("fetch failed") ||
+            lower.includes("socket connection was closed") ||
+            lower.includes("terminated") ||
+            lower.includes("other side closed"));
     }
     return false;
 }
@@ -179,6 +199,10 @@ async function chatCompletionAttempt(config, messages, options, timeoutMs) {
     const responseFormat = options?.responseSchema && config.supportsJsonSchema
         ? { response_format: { type: "json_schema", json_schema: { schema: options.responseSchema, strict: true } } }
         : {};
+    // Wall-clock start for per-call usage telemetry (#576). Captured here so the
+    // emitted duration covers the full request/response/parse cycle of a single
+    // attempt, not the retry-wrapping `chatCompletion`.
+    const requestStartedAt = Date.now();
     let response;
     try {
         response = await fetchWithTimeout(config.endpoint, {
@@ -241,6 +265,16 @@ async function chatCompletionAttempt(config, messages, options, timeoutMs) {
     catch {
         throw new LlmCallError(`LLM response was not valid JSON ${config.endpoint}: ${redactErrorBody(rawOkBody)}`, "parse_error", response.status);
     }
+    // Per-call usage telemetry (#576). Best-effort and fully isolated: a missing
+    // or garbled usage block still records duration + model, and a throwing sink
+    // can never fail the call (emitLlmUsage swallows its own errors). The stage
+    // is supplied ambiently by emitLlmUsage; no `stage` param is threaded here.
+    emitLlmUsage({
+        model: typeof json.model === "string" && json.model ? json.model : config.model,
+        durationMs: Date.now() - requestStartedAt,
+        finishReason: typeof json.choices?.[0]?.finish_reason === "string" ? json.choices[0].finish_reason : undefined,
+        ...extractUsageTokens(json.usage),
+    });
     const content = (json.choices?.[0]?.message?.content ?? "").trim();
     const reasoning = (json.choices?.[0]?.message?.reasoning_content ?? "").trim();
     return content || reasoning;