@phnx-labs/agents-cli 1.22.25 → 1.22.27

package/dist/lib/session/tool-index.js

@@ -1,9 +1,9 @@
 import * as fs from 'fs';
 import { StringDecoder } from 'string_decoder';
-import { getDB } from './db.js';
+import { getDB, maintainSessionSearchIndex } from './db.js';
 import { parseSession } from './parse.js';
 import { TOOL_INDEX_VERSION, TOOL_INDEX_LIMIT_ORDINAL, ToolCallCollector, collectClaudeToolCalls, collectCodexToolCalls, toolCallEvidenceBytes, toolCallsFromEvents, } from './tool-calls.js';
-import { persistToolCalls, purgeToolCalls, toolEvidenceSourcePath, } from './tool-store.js';
+import { canonicalToolLedgerPath, persistToolCalls, purgeToolCalls, toolEvidenceSourcePath, } from './tool-store.js';
 const BACKFILL_MAX_FILES = 25;
 const BACKFILL_MAX_BYTES = 16 * 1024 * 1024;
 const BACKFILL_MAX_IN_MEMORY_SOURCE_BYTES = 16 * 1024 * 1024;
@@ -39,16 +39,71 @@ export function toolSearchRemoteReceiveBudget(envelope) {
     const localBytes = serializedToolSearchEnvelopeBytes(envelope);
     return Math.max(0, TOOL_QUERY_MAX_SERIALIZED_BYTES - TOOL_QUERY_MERGE_OVERHEAD_BYTES - localBytes);
 }
-function needsIndex(db, sessionId, stamp) {
-    const row = db.prepare(`
-    SELECT file_mtime_ms, file_size, extractor_version
+/**
+ * The ledger columns every candidate session is judged on. Deliberately NOT
+ * `parser_state`: this runs once per session in the scan's warm path, and that
+ * column holds a serialized collector snapshot that can reach a megabyte. It is
+ * read separately, only for the sessions that turn out to need indexing.
+ */
+function readToolLedger(db, sessionId) {
+    return db.prepare(`
+    SELECT file_path, file_mtime_ms, file_size, extractor_version, parsed_offset
     FROM tool_scan_ledger WHERE session_id = ?
   `).get(sessionId);
+}
+function readToolParserState(db, sessionId) {
+    const row = db.prepare(`SELECT parser_state FROM tool_scan_ledger WHERE session_id = ?`)
+        .get(sessionId);
+    return row?.parser_state ?? null;
+}
+function needsIndex(row, stamp) {
     return !row
         || row.file_mtime_ms !== stamp.fileMtimeMs
         || row.file_size !== stamp.fileSize
         || row.extractor_version !== TOOL_INDEX_VERSION;
 }
+/**
+ * Where to start reading a session whose transcript changed.
+ *
+ * A live session's transcript is append-only, so re-reading it from byte 0 on
+ * every scan re-parses the entire history to discover the handful of records
+ * that are new — the cost that makes a large session's tool index quadratic in
+ * the number of scans. When the ledger carries a resume point that the current
+ * file still agrees with, the scan reads only the appended bytes and merges the
+ * result (`append`); anything else re-reads the whole file (`replace`).
+ *
+ * Each check below rejects a case where the stored prefix may no longer describe
+ * the file: a harness the streaming parser cannot resume, a different extractor,
+ * no recorded resume point, a source path the ledger row does not describe, a
+ * file that shrank below what was already parsed (a rewrite or truncation, not
+ * an append), or a snapshot that does not read back.
+ */
+function planToolScan(db, sessionId, row, sourcePath, stamp, resumable) {
+    const full = { mode: 'replace', startOffset: 0 };
+    if (!resumable || !row)
+        return full;
+    if (row.extractor_version !== TOOL_INDEX_VERSION)
+        return full;
+    if (row.parsed_offset === null)
+        return full;
+    if (row.file_path !== canonicalToolLedgerPath(sourcePath))
+        return full;
+    if (stamp.fileSize < row.file_size || stamp.fileSize < row.parsed_offset)
+        return full;
+    const parserState = readToolParserState(db, sessionId);
+    if (parserState === null)
+        return full;
+    let snapshot;
+    try {
+        snapshot = JSON.parse(parserState);
+    }
+    catch {
+        return full;
+    }
+    if (snapshot?.v !== 1 || !Number.isSafeInteger(snapshot.nextOrdinal))
+        return full;
+    return { mode: 'append', startOffset: row.parsed_offset, snapshot };
+}
 /** Read index completeness from SQLite only; never stat or parse transcripts. */
 export function readToolIndexCoverage(sessions) {
     const db = getDB();
@@ -90,14 +145,23 @@ function backfillLimitCall(session, reason) {
     };
 }
 /** Stream Claude/Codex JSONL without ever retaining an oversized record. */
-async function streamJsonlToolCalls(session) {
-    const collector = new ToolCallCollector();
-    const stream = fs.createReadStream(session.filePath, { highWaterMark: 64 * 1024 });
+async function streamJsonlToolCalls(session, from = { startOffset: 0 }) {
+    const collector = new ToolCallCollector(from.snapshot);
+    const stream = fs.createReadStream(session.filePath, {
+        highWaterMark: 64 * 1024,
+        start: from.startOffset,
+    });
     const decoder = new StringDecoder('utf8');
     let pending = '';
     let pendingBytes = 0;
     let droppingOversizedLine = false;
     let skippedOversizedLine = false;
+    // Byte offset just past the last complete record applied. Only a complete,
+    // newline-terminated record advances it, so resuming here can never re-apply a
+    // record (which would mint a second ordinal for it) nor skip a partial tail.
+    let parsedOffset = from.startOffset;
+    /** Bytes of the record currently being assembled, across chunk boundaries. */
+    let lineBytes = 0;
     const applyLine = (line) => {
         if (!line.trim())
             return;
@@ -119,8 +183,14 @@ async function streamJsonlToolCalls(session) {
             const newline = text.indexOf('\n', start);
             const end = newline >= 0 ? newline : text.length;
             const segment = text.slice(start, end);
+            const segmentBytes = Buffer.byteLength(segment);
+            // Counted outside the drop guard and across chunk boundaries: this is the
+            // record's true size on disk, which is what the resume offset is measured
+            // in. `pendingBytes` cannot stand in for it — that one resets when an
+            // oversized record is dropped, and a record split over two 64 KiB reads
+            // would lose the part carried in from the previous chunk.
+            lineBytes += segmentBytes;
             if (!droppingOversizedLine) {
-                const segmentBytes = Buffer.byteLength(segment);
                 if (pendingBytes + segmentBytes <= BACKFILL_MAX_JSONL_RECORD_BYTES) {
                     pending += segment;
                     pendingBytes += segmentBytes;
@@ -136,6 +206,8 @@ async function streamJsonlToolCalls(session) {
                 break;
             if (!droppingOversizedLine)
                 applyLine(pending);
+            parsedOffset += lineBytes + 1; // + the newline itself
+            lineBytes = 0;
             pending = '';
             pendingBytes = 0;
             droppingOversizedLine = false;
@@ -145,25 +217,48 @@ async function streamJsonlToolCalls(session) {
     for await (const chunk of stream)
         consume(decoder.write(chunk));
     consume(decoder.end());
+    // Snapshot BEFORE the unterminated trailing record, and pair it with an offset
+    // that stops short of that record. The writer may be mid-append, so the record
+    // is indexed now (its evidence is real) but is re-read by the next scan — which
+    // resumes with the same next-ordinal and so re-derives the same ordinals,
+    // making the re-read an idempotent upsert rather than a duplicate.
+    const resume = skippedOversizedLine
+        // A dropped oversized record left the ordinals and the pending map out of
+        // step with the file; nothing here can be resumed from.
+        ? null
+        : { parserState: JSON.stringify(collector.snapshot()), parsedOffset };
     if (!droppingOversizedLine && pending.length > 0)
         applyLine(pending);
     const calls = collector.drainChanged();
     if (skippedOversizedLine && !calls.some((call) => call.ordinal === TOOL_INDEX_LIMIT_ORDINAL)) {
         calls.push(backfillLimitCall(session, 'At least one JSONL record exceeded the 1 MiB tool-backfill parser limit.'));
     }
-    return calls;
+    return { calls, resume };
+}
+/** True for the harnesses whose transcript the streaming parser can resume. */
+function isResumableToolSource(agent) {
+    return agent === 'claude' || agent === 'codex';
 }
-async function toolCallsForBackfill(session, sourceBytes) {
-    if (session.agent === 'claude' || session.agent === 'codex') {
+async function toolCallsForBackfill(session, sourceBytes, from = { startOffset: 0 }) {
+    if (isResumableToolSource(session.agent)) {
         if (sourceBytes > BACKFILL_MAX_STREAM_SOURCE_BYTES) {
-            return [backfillLimitCall(session, 'Transcript exceeds the 64 MiB safe streaming tool-backfill limit.')];
+            return {
+                calls: [backfillLimitCall(session, 'Transcript exceeds the 64 MiB safe streaming tool-backfill limit.')],
+                resume: null,
+            };
         }
-        return streamJsonlToolCalls(session);
+        return streamJsonlToolCalls(session, from);
     }
     if (sourceBytes > BACKFILL_MAX_IN_MEMORY_SOURCE_BYTES) {
-        return [backfillLimitCall(session, 'Transcript exceeds the 16 MiB safe in-memory tool-backfill parser limit.')];
+        return {
+            calls: [backfillLimitCall(session, 'Transcript exceeds the 16 MiB safe in-memory tool-backfill parser limit.')],
+            resume: null,
+        };
     }
-    return toolCallsFromEvents(parseSession(session.filePath, session.agent));
+    // Every other harness is parsed whole into memory by parseSession, which
+    // exposes no byte offset to resume from — so these stay full replaces and
+    // record no resume point, rather than storing one this path cannot honour.
+    return { calls: toolCallsFromEvents(parseSession(session.filePath, session.agent)), resume: null };
 }
 /**
  * Fill one bounded chunk of the independent tool index. A warm call performs
@@ -179,6 +274,7 @@ export async function ensureToolIndex(sessions, limits = {}) {
         if (!session.filePath)
             continue;
         const sourcePath = toolEvidenceSourcePath(session.filePath, session.agent);
+        const ledger = readToolLedger(db, session.id);
         const mustStatSource = limits.verifySourceStamps || sourcePath !== session.filePath;
         const indexed = !mustStatSource
             ? db.prepare(`
@@ -199,8 +295,15 @@ export async function ensureToolIndex(sessions, limits = {}) {
                 continue;
             }
         }
-        if (needsIndex(db, session.id, stamp))
-            pending.push({ session, stamp });
+        if (!needsIndex(ledger, stamp))
+            continue;
+        const plan = planToolScan(db, session.id, ledger, sourcePath, stamp, isResumableToolSource(session.agent));
+        pending.push({
+            session,
+            stamp,
+            plan,
+            readBytes: Math.max(0, stamp.fileSize - plan.startOffset),
+        });
     }
     let indexedFiles = 0;
     let indexedCalls = 0;
@@ -212,13 +315,16 @@ export async function ensureToolIndex(sessions, limits = {}) {
         // The byte budget is a batch boundary, not a correctness boundary. Admit
         // one oversized transcript by itself so it can never wedge the ledger or
         // silently disappear from results; the next invocation resumes afterward.
-        if (attemptedFiles > 0 && consumedBytes + item.stamp.fileSize > maxBytes)
+        // Budgeted on the bytes this scan reads, not the file's size: a resumed
+        // session costs only its appended tail, so a batch can cover far more
+        // growing sessions than it could when every one was re-read whole.
+        if (attemptedFiles > 0 && consumedBytes + item.readBytes > maxBytes)
             break;
         attemptedFiles++;
-        consumedBytes += item.stamp.fileSize;
+        consumedBytes += item.readBytes;
         try {
-            const calls = await toolCallsForBackfill(item.session, item.stamp.fileSize);
-            persistToolCalls(db, item.session, calls, item.stamp);
+            const { calls, resume } = await toolCallsForBackfill(item.session, item.stamp.fileSize, item.plan);
+            persistToolCalls(db, item.session, calls, item.stamp, { mode: item.plan.mode, resume });
             indexedFiles++;
             indexedCalls += calls.length;
         }
@@ -226,6 +332,11 @@ export async function ensureToolIndex(sessions, limits = {}) {
             skippedFiles++;
         }
     }
+    // The scan just wrote a batch of FTS segments; pay a bounded slice of the
+    // merge they need so the index converges here instead of degrading until
+    // someone runs `agents sessions optimize` by hand (RUSH-2208).
+    if (indexedFiles > 0)
+        maintainSessionSearchIndex(db);
     const remainingFiles = Math.max(0, pending.length - attemptedFiles);
     const limitedSessionIds = new Set();
     const sessionIds = sessions.map((session) => session.id);

package/dist/lib/session/tool-store.d.ts

@@ -8,8 +8,32 @@ export declare function toolEvidenceSourcePath(filePath: string, agent: string):
 export declare function purgeToolCalls(db: Database.Database, sessionId: string): void;
 /** Purge deleted direct children when a transcript directory's stamp changes. */
 export declare function purgeMissingToolCallsInDirectory(db: Database.Database, dirPath: string, currentFilePaths: string[]): number;
-/** Persist one parser batch and its file stamp atomically. */
+/** The resume point a later incremental scan starts from. */
+export interface ToolScanResumePoint {
+    /** Serialized ToolCallCollector snapshot at `parsedOffset`. */
+    parserState: string;
+    /** Byte offset just past the last complete record consumed. */
+    parsedOffset: number;
+}
+export interface PersistToolCallsOptions {
+    /**
+     * `replace` drops the session's stored evidence first — correct for a parse
+     * that started at byte 0. `append` merges the batch into what is already
+     * stored and requires an existing ledger row; use it only for a parse that
+     * resumed from that row's `parsedOffset`.
+     */
+    mode?: 'replace' | 'append';
+    /**
+     * Where a later scan may resume. Omitted (or null) clears any stored resume
+     * point, which forces the next scan of this session to re-read from byte 0 —
+     * the correct outcome whenever the parse could not cover the whole prefix
+     * (an oversized record, a size-capped transcript, a non-streaming harness).
+     */
+    resume?: ToolScanResumePoint | null;
+    maxSessionBytes?: number;
+}
+/** Persist one parser batch, its file stamp, and its resume point atomically. */
 export declare function persistToolCalls(db: Database.Database, session: SessionMeta, calls: IndexedToolCall[], sourceStamp: {
     fileMtimeMs: number;
     fileSize: number;
-}, mode?: 'replace' | 'append', maxSessionBytes?: number): void;
+}, options?: PersistToolCallsOptions): void;

package/dist/lib/session/tool-store.js

@@ -28,11 +28,17 @@ export function toolEvidenceSourcePath(filePath, agent) {
     return filePath;
 }
 function deleteSessionCalls(db, sessionId) {
-    const keys = db.prepare(`SELECT call_key FROM tool_calls WHERE session_id = ?`).all(sessionId);
-    for (const { call_key } of keys) {
-        db.prepare(`DELETE FROM tool_call_programs WHERE call_key = ?`).run(call_key);
-        db.prepare(`DELETE FROM tool_program_occurrences WHERE call_key = ?`).run(call_key);
-        db.prepare(`DELETE FROM tool_call_text WHERE call_key = ?`).run(call_key);
+    const rows = db.prepare(`SELECT rowid, call_key FROM tool_calls WHERE session_id = ?`)
+        .all(sessionId);
+    const deletePrograms = db.prepare(`DELETE FROM tool_call_programs WHERE call_key = ?`);
+    const deleteOccurrences = db.prepare(`DELETE FROM tool_program_occurrences WHERE call_key = ?`);
+    // Addressed by rowid, never by the UNINDEXED call_key — see the tool_call_text
+    // schema comment in db.ts. A call_key predicate scans the whole FTS index.
+    const deleteText = db.prepare(`DELETE FROM tool_call_text WHERE rowid = ?`);
+    for (const { rowid, call_key } of rows) {
+        deletePrograms.run(call_key);
+        deleteOccurrences.run(call_key);
+        deleteText.run(rowid);
     }
     db.prepare(`DELETE FROM tool_calls WHERE session_id = ?`).run(sessionId);
 }
@@ -61,8 +67,11 @@ export function purgeMissingToolCallsInDirectory(db, dirPath, currentFilePaths)
     }
     return purged;
 }
-/** Persist one parser batch and its file stamp atomically. */
-export function persistToolCalls(db, session, calls, sourceStamp, mode = 'replace', maxSessionBytes = TOOL_SESSION_EVIDENCE_MAX_BYTES) {
+/** Persist one parser batch, its file stamp, and its resume point atomically. */
+export function persistToolCalls(db, session, calls, sourceStamp, options = {}) {
+    const mode = options.mode ?? 'replace';
+    const maxSessionBytes = options.maxSessionBytes ?? TOOL_SESSION_EVIDENCE_MAX_BYTES;
+    const resume = options.resume ?? null;
     const sourcePath = toolEvidenceSourcePath(session.filePath, session.agent);
     const insertCall = db.prepare(`
     INSERT INTO tool_calls (
@@ -89,16 +98,19 @@ export function persistToolCalls(db, session, calls, sourceStamp, mode = 'replac
     INSERT INTO tool_program_occurrences (call_key, occurrence_ordinal, program, role)
     VALUES (?, ?, ?, ?)
   `);
-    const insertText = db.prepare(`INSERT INTO tool_call_text (call_key, tool, input, output, error) VALUES (?, ?, ?, ?, ?)`);
+    // tool_call_text rows are addressed by the rowid of the tool_calls row they
+    // describe (db.ts schema comment): its UNINDEXED call_key cannot be seeked.
+    const insertText = db.prepare(`INSERT INTO tool_call_text (rowid, call_key, tool, input, output, error) VALUES (?, ?, ?, ?, ?, ?)`);
+    const callRowid = db.prepare(`SELECT rowid FROM tool_calls WHERE call_key = ?`);
     const deletePrograms = db.prepare(`DELETE FROM tool_call_programs WHERE call_key = ?`);
     const deleteOccurrences = db.prepare(`DELETE FROM tool_program_occurrences WHERE call_key = ?`);
-    const deleteText = db.prepare(`DELETE FROM tool_call_text WHERE call_key = ?`);
+    const deleteText = db.prepare(`DELETE FROM tool_call_text WHERE rowid = ?`);
     const deleteCall = db.prepare(`DELETE FROM tool_calls WHERE call_key = ?`);
     const writeLedger = db.prepare(`
     INSERT INTO tool_scan_ledger (
       session_id, file_path, file_mtime_ms, file_size, extractor_version, indexed_at, call_count,
-      evidence_bytes
-    ) VALUES (?, ?, ?, ?, ?, ?, ?, ?)
+      evidence_bytes, parser_state, parsed_offset
+    ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
     ON CONFLICT(session_id) DO UPDATE SET
       file_path = excluded.file_path,
       file_mtime_ms = excluded.file_mtime_ms,
@@ -106,7 +118,9 @@ export function persistToolCalls(db, session, calls, sourceStamp, mode = 'replac
       extractor_version = excluded.extractor_version,
       indexed_at = excluded.indexed_at,
       call_count = excluded.call_count,
-      evidence_bytes = excluded.evidence_bytes
+      evidence_bytes = excluded.evidence_bytes,
+      parser_state = excluded.parser_state,
+      parsed_offset = excluded.parsed_offset
   `);
     const txn = db.transaction(() => {
         if (mode === 'replace')
@@ -119,7 +133,7 @@ export function persistToolCalls(db, session, calls, sourceStamp, mode = 'replac
             throw new Error('Append tool evidence requires an existing tool ledger row.');
         }
         if (mode === 'append' && calls.length === 0) {
-            writeLedger.run(session.id, ledgerPath, sourceStamp.fileMtimeMs, sourceStamp.fileSize, TOOL_INDEX_VERSION, Date.now(), priorLedger.call_count, priorLedger.evidence_bytes);
+            writeLedger.run(session.id, ledgerPath, sourceStamp.fileMtimeMs, sourceStamp.fileSize, TOOL_INDEX_VERSION, Date.now(), priorLedger.call_count, priorLedger.evidence_bytes, resume?.parserState ?? null, resume?.parsedOffset ?? null);
             return;
         }
         const limitKey = toolCallKey(session.id, TOOL_INDEX_LIMIT_ORDINAL);
@@ -129,9 +143,11 @@ export function persistToolCalls(db, session, calls, sourceStamp, mode = 'replac
         const priorLimit = mode === 'append'
             ? existingSize.get(session.id, TOOL_INDEX_LIMIT_ORDINAL)
             : undefined;
+        const limitRow = callRowid.get(limitKey);
         deletePrograms.run(limitKey);
         deleteOccurrences.run(limitKey);
-        deleteText.run(limitKey);
+        if (limitRow)
+            deleteText.run(limitRow.rowid);
         deleteCall.run(limitKey);
         const existingRows = mode === 'append'
             ? calls.map((call) => ({
@@ -182,17 +198,20 @@ export function persistToolCalls(db, session, calls, sourceStamp, mode = 'replac
         for (const call of accepted) {
             const key = toolCallKey(session.id, call.ordinal);
             insertCall.run(key, session.id, call.ordinal, call.sourceCallId ?? null, call.timestamp, call.tool, call.input, call.outcome, call.exitCode ?? null, call.statusCode ?? null, call.errorCode ?? null, call.output ?? null, call.error ?? null, call.parseError ?? null, toolCallEvidenceBytes(call));
+            // The upsert above preserves the rowid of a call it updated, so this is
+            // the same rowid the existing text row (if any) was written under.
+            const { rowid } = callRowid.get(key);
             deletePrograms.run(key);
             deleteOccurrences.run(key);
-            deleteText.run(key);
+            deleteText.run(rowid);
             for (const program of call.programs)
                 insertProgram.run(key, program);
             call.programOccurrences.forEach((occurrence, occurrenceOrdinal) => {
                 insertOccurrence.run(key, occurrenceOrdinal, occurrence.program, occurrence.role);
             });
-            insertText.run(key, call.tool, call.input, call.output ?? '', call.error ?? '');
+            insertText.run(rowid, key, call.tool, call.input, call.output ?? '', call.error ?? '');
         }
-        writeLedger.run(session.id, ledgerPath, sourceStamp.fileMtimeMs, sourceStamp.fileSize, TOOL_INDEX_VERSION, Date.now(), count, storedBytes);
+        writeLedger.run(session.id, ledgerPath, sourceStamp.fileMtimeMs, sourceStamp.fileSize, TOOL_INDEX_VERSION, Date.now(), count, storedBytes, resume?.parserState ?? null, resume?.parsedOffset ?? null);
     });
     txn();
 }

package/dist/lib/session/types.d.ts

@@ -117,6 +117,24 @@ export interface TeamOrigin {
      * meta dir has aged past the teams cleanup window.
      */
     parentSessionId?: string;
+    /**
+     * Spawn time (`started_at` in the teammate meta.json) — when `agents teams`
+     * launched this teammate. Distinct from the session's own `timestamp` (first
+     * transcript line), and present even before the harness writes a transcript.
+     * Absent for the entrypoint-only fallback (no meta record).
+     */
+    startedAt?: string;
+    /**
+     * How this origin was established, which is what separates a real
+     * `agents teams` teammate from a plain SDK sub-agent (a `Task` / `Agent()`
+     * spawn). Both carry the `sdk-cli` entrypoint that sets `isTeamOrigin`, so the
+     * entrypoint flag alone cannot tell them apart — only a teammate has a
+     * `meta.json` under the teams agents dir.
+     *   - `'meta'`     — a teammate: read from its `meta.json` record.
+     *   - `'entrypoint'` — an SDK spawn with no team record (a sub-agent, or a
+     *                      teammate whose meta dir aged past the cleanup window).
+     */
+    source?: 'meta' | 'entrypoint';
 }
 /** Lightweight metadata for a discovered session, used in listings and pickers. */
 export interface SessionMeta {
@@ -144,8 +162,23 @@ export interface SessionMeta {
     tokenCount?: number;
     /** Real generated (output) tokens — excludes cache-read/-write context (issue: `agents output`). */
     outputTokens?: number;
+    /**
+     * Uncached input tokens, cache-read tokens, and cache-write (cache-creation)
+     * tokens — the burn split `agents output` reports, kept only for harnesses that
+     * record a per-message cache split (Claude/Codex/Gemini/Droid). Undefined for
+     * harnesses that expose no split (RUSH-2287).
+     */
+    inputTokens?: number;
+    cacheReadTokens?: number;
+    cacheWriteTokens?: number;
     /** Total USD cost, computed at scan time from per-model token usage (issue #323). */
     costUsd?: number;
+    /**
+     * USD cost priced as if caching were off — cache read/write billed at the full
+     * input rate. Backs `agents output --pricing no-cache` (RUSH-2287). Undefined
+     * when the harness records no cache split (then no-cache == actual by definition).
+     */
+    costUsdNoCache?: number;
     /** Wall-clock duration in ms (lastTs − firstTs), persisted at scan time. */
     durationMs?: number;
     /** Underlying LLM model observed in the transcript, when the agent records one. */

package/dist/lib/smart-launch.d.ts

@@ -1,11 +1,13 @@
 /**
- * Device affinity for `--device auto` / `--host auto`.
+ * Device placement for `--device auto` / `--host auto`.
  *
- * Picks a host from 14d session usage (launches by machine), weighted sample
- * among online devices. Most-used has highest probability — not a hard lock.
+ * Explicit auto placement probes live fleet health and harness readiness.
+ * Historical affinity remains for generic host resolution callers.
  * Account selection is separate (`--strategy balanced` / rotate.ts).
  */
 import { type AffinityRow } from './session/db.js';
+import { type DevicePlacementSignal } from './teams/scheduler.js';
+import type { AgentType } from './teams/agents.js';
 /** Peakiness of usage weights; >1 amplifies the most-used option. */
 export declare const DEFAULT_AFFINITY_ALPHA = 1.3;
 export interface WeightedCandidate {
@@ -45,6 +47,29 @@ export interface DeviceAffinityPlan {
     deviceCandidates: WeightedCandidate[];
     pickedDeviceKey: string | null;
 }
+/** Live placement plan used by the explicit `--device auto` surface. */
+export interface DeviceAutoPlan {
+    /** null means the local machine won the comparison. */
+    host: string | null;
+    candidates: Array<{
+        key: string;
+        loadPercent?: number;
+        installed?: boolean;
+        signedIn?: boolean;
+    }>;
+    pickedDeviceKey: string;
+}
+/**
+ * Pick the least-loaded healthy device that can run `agent` when the harness is
+ * known. `run auto` omits the agent and ranks the same live health/load signals
+ * without an installed-harness filter. The local machine participates in the
+ * same probe; a fully unusable fleet degrades to local instead of stranding a run.
+ */
+export declare function resolveDeviceAuto(agent?: string, opts?: {
+    eligibleHosts?: string[];
+    localMachine?: string;
+    probe?: (pool: string[], agent?: AgentType) => Promise<Map<string, DevicePlacementSignal>>;
+}): Promise<DeviceAutoPlan>;
 /**
  * Resolve host for `--device auto`. Does NOT pick harness or accounts.
  */
@@ -81,6 +106,7 @@ export type DeviceAutoApplyResult = {
  * returns `skipped` — never throws; callers must not hard-exit.
  */
 export declare function applyDeviceAutoToOptions(options: DeviceAutoHostOptions, deps?: {
-    resolve?: () => DeviceAffinityPlan;
+    resolve?: () => DeviceAutoPlan | Promise<DeviceAutoPlan>;
+    agent?: string;
     accountPickerRequested?: boolean;
-}): DeviceAutoApplyResult;
+}): Promise<DeviceAutoApplyResult>;

package/dist/lib/smart-launch.js

@@ -1,14 +1,16 @@
 /**
- * Device affinity for `--device auto` / `--host auto`.
+ * Device placement for `--device auto` / `--host auto`.
  *
- * Picks a host from 14d session usage (launches by machine), weighted sample
- * among online devices. Most-used has highest probability — not a hard lock.
+ * Explicit auto placement probes live fleet health and harness readiness.
+ * Historical affinity remains for generic host resolution callers.
  * Account selection is separate (`--strategy balanced` / rotate.ts).
  */
 import { queryAffinityRollup } from './session/db.js';
 import { localMachineId } from './session/origin-machine.js';
 import { loadDevicesSync } from './devices/registry.js';
 import { normalizeHost } from './machine-id.js';
+import { probePoolSignals } from './teams/placement-probe.js';
+import { pickBestDevice } from './teams/scheduler.js';
 /** Peakiness of usage weights; >1 amplifies the most-used option. */
 export const DEFAULT_AFFINITY_ALPHA = 1.3;
 /**
@@ -62,6 +64,39 @@ export function listOnlineDeviceNames(localName = localMachineId()) {
     }
     return [...names];
 }
+/**
+ * Pick the least-loaded healthy device that can run `agent` when the harness is
+ * known. `run auto` omits the agent and ranks the same live health/load signals
+ * without an installed-harness filter. The local machine participates in the
+ * same probe; a fully unusable fleet degrades to local instead of stranding a run.
+ */
+export async function resolveDeviceAuto(agent, opts = {}) {
+    const local = normalizeHost(opts.localMachine ?? localMachineId());
+    const pool = [...new Set((opts.eligibleHosts ?? listOnlineDeviceNames(local)).map(normalizeHost))];
+    if (!pool.includes(local))
+        pool.push(local);
+    try {
+        const signals = await (opts.probe ?? probePoolSignals)(pool, agent);
+        const picked = pickBestDevice(pool, [], { signals, agentLabel: agent });
+        return {
+            host: picked === local ? null : picked,
+            pickedDeviceKey: picked,
+            candidates: pool.map((key) => ({
+                key,
+                loadPercent: signals.get(key)?.loadPercent,
+                installed: signals.get(key)?.installed,
+                signedIn: signals.get(key)?.signedIn,
+            })),
+        };
+    }
+    catch {
+        return {
+            host: null,
+            pickedDeviceKey: local,
+            candidates: [{ key: local }],
+        };
+    }
+}
 /**
  * Resolve host for `--device auto`. Does NOT pick harness or accounts.
  */
@@ -120,7 +155,7 @@ const HOST_SLOTS = ['host', 'device', 'on', 'computer'];
  * Mutates `options` in place. On affinity failure, clears auto slots (local) and
  * returns `skipped` — never throws; callers must not hard-exit.
  */
-export function applyDeviceAutoToOptions(options, deps = {}) {
+export async function applyDeviceAutoToOptions(options, deps = {}) {
     let deprecationSmart = false;
     if (options.smart) {
         const anyHost = HOST_SLOTS.some((k) => typeof options[k] === 'string' && options[k].trim() !== '');
@@ -133,7 +168,8 @@ export function applyDeviceAutoToOptions(options, deps = {}) {
         return { attempted: false, deprecationSmart };
     }
     try {
-        const plan = (deps.resolve ?? (() => resolveDeviceAffinity({})))();
+        const resolve = deps.resolve ?? (() => resolveDeviceAuto(deps.agent));
+        const plan = await resolve();
         const concrete = plan.host; // null = local
         for (const k of HOST_SLOTS) {
             if (isDeviceAuto(options[k])) {
@@ -145,9 +181,9 @@ export function applyDeviceAutoToOptions(options, deps = {}) {
             options.balanced = true;
         }
         const hostLabel = concrete ?? 'local';
-        const deviceHint = plan.deviceCandidates
+        const deviceHint = plan.candidates
             .slice(0, 4)
-            .map((c) => `${c.key}:${c.launches}`)
+            .map((c) => `${c.key}:${c.loadPercent === undefined ? '?' : `${Math.round(c.loadPercent)}%`}`)
             .join(', ');
         const acctNote = accountPickerRequested ? 'accounts=picker' : 'accounts=balanced';
         return {

package/dist/lib/ssh-exec.d.ts

@@ -28,6 +28,31 @@ export declare function shellQuote(s: string): string;
  * long-lived `-N` tunnel can't leave a zombie ssh + socket pinned on the laptop.
  */
 export declare const SSH_OPTS: readonly string[];
+/**
+ * Hard ceiling on the stdout one peer may return before its capture is aborted.
+ * A cross-machine fan-out streams every peer's `agents … --json` into memory in
+ * parallel, so an unbounded buffer means one runaway peer (a corrupt or
+ * pathologically large payload) can exhaust the caller's heap and take the whole
+ * sweep down with it — RUSH-2065 observed ~170MB retained across a single
+ * `--active` gather. 16 MiB is far above any legitimate JSON listing yet small
+ * enough that N peers in flight stay bounded. A peer that overflows is treated as
+ * unreachable rather than trusted with partial output.
+ */
+export declare const REMOTE_STDOUT_MAX_BYTES: number;
+/**
+ * Accumulate SSH stdout as UTF-8 without corrupting a multi-byte code point that
+ * a chunk boundary splits — the streaming `on('data')` callbacks hand us raw
+ * `Buffer`s, and a naive `buf.toString()` per chunk mangles any character the
+ * kernel cut in half. `StringDecoder` holds the trailing partial bytes until the
+ * next chunk completes them.
+ */
+export declare class RemoteUtf8Accumulator {
+    private readonly decoder;
+    private value;
+    write(chunk: Buffer): void;
+    end(): string;
+    current(): string;
+}
 export declare function controlOpts(): string[];
 /**
  * Compose an ssh connection-option prefix.
@@ -128,6 +153,8 @@ export interface SshStreamOptions {
      * interactive credential-copy path (RUSH-1767).
      */
     hostKeyOpts?: string[];
+    /** Additional OpenSSH argv placed before the target (for example `-i <path>`). */
+    extraSshArgs?: string[];
 }
 /**
  * Foreground counterpart to `sshExec`: run `remoteCmd` on `target` with the