@phnx-labs/agents-cli 1.22.25 → 1.22.27

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

@@ -21,11 +21,25 @@
  *     the raw file count rather than quietly differing.
  */
 import type { SessionEvent } from './types.js';
+/**
+ * Minimum quiet time after the assistant's last event before a user message
+ * is classified as an **agent silent stall**: the model went quiet on its own
+ * and sat idle until the human resumed. Shorter gaps are normal turn-taking.
+ * 5 minutes is long enough to exclude "user typing the next instruction" and
+ * short enough to catch "went silent mid-task until I said continue."
+ */
+export declare const SILENT_STALL_SECONDS = 300;
 /** Behavioural facets of one session. Serialized as JSON into `session_insights`. */
 export interface InsightFacets {
     toolCounts: Record<string, number>;
     /** Per-model assistant turn counts. `/insights` has no model dimension at all. */
     models: Record<string, number>;
+    /**
+     * Silent stalls attributed to the model that last spoke before the idle gap
+     * (from the nearest preceding `usage`/message model tag). Some models go
+     * quiet more often — this is the laziness split. Key = shortened model id.
+     */
+    silentStallsByModel: Record<string, number>;
     languages: Record<string, number>;
     /** Slash commands the user invoked, by name. */
     slashCommands: Record<string, number>;
@@ -71,6 +85,10 @@ export interface InsightFacets {
     assistantTurns: number;
     toolCount: number;
     errorCount: number;
+    /** Deterministic evidence buckets used by the actions-forward report. */
+    frictionSignals: Record<string, number>;
+    correctionSignals: Record<string, number>;
+    automationSignals: Record<string, number>;
 }
 /**
  * Compute every behavioural facet of one session from its parsed events.
@@ -80,6 +98,11 @@ export interface InsightFacets {
  * can be re-bucketed for a different display timezone without re-parsing.
  */
 export declare function computeInsightFacets(events: SessionEvent[], timezoneOffsetMinutes?: number): InsightFacets;
+/**
+ * Classify a long quiet gap after the assistant's last event as an agent silent stall.
+ * Optionally also marks an explicit resume nudge ("continue", …) after that silence.
+ */
+export declare function classifySilentStall(gapSeconds: number, userText: string, friction: Record<string, number>, corrections: Record<string, number>): void;
 /** Percentile of a numeric sample, nearest-rank. Returns 0 for an empty sample. */
 export declare function percentile(values: number[], p: number): number;
 /** Bucket response gaps for display. Returns every bucket, including empty ones. */
@@ -124,3 +147,17 @@ export declare function topEntries(counts: Record<string, number>, limit: number
     name: string;
     count: number;
 }>;
+export type InsightActionCategory = 'rule' | 'skill' | 'automation' | 'product';
+export interface InsightAction {
+    priority: 'high' | 'medium' | 'low';
+    category: InsightActionCategory;
+    action: string;
+    evidenceCount: number;
+    sampleSessionIds: string[];
+}
+export interface SessionFacetEvidence {
+    id: string;
+    facets: InsightFacets;
+}
+/** Build a stable, evidence-backed action list without exposing transcript text. */
+export declare function buildInsightActions(sessions: SessionFacetEvidence[]): InsightAction[];

package/dist/lib/session/insights.js

@@ -47,24 +47,42 @@ const ERROR_CATEGORIES = [
     [['exit code', 'command failed', 'error:'], 'Command Failed'],
 ];
 /**
- * Gaps longer than this are someone leaving and coming back, not a reply latency.
- * Counted separately rather than silently dropped.
+ * Gaps longer than this leave the reply-latency percentiles (someone left for
+ * lunch / overnight). They still count as silent stalls when the assistant
+ * was the last speaker — the agent had already stopped before the user left.
  */
 const GAP_CEILING_SECONDS = 3600;
+/**
+ * Minimum quiet time after the assistant's last event before a user message
+ * is classified as an **agent silent stall**: the model went quiet on its own
+ * and sat idle until the human resumed. Shorter gaps are normal turn-taking.
+ * 5 minutes is long enough to exclude "user typing the next instruction" and
+ * short enough to catch "went silent mid-task until I said continue."
+ */
+export const SILENT_STALL_SECONDS = 300;
 /** Response-gap buckets, in ascending order. Upper bound is exclusive. */
 const GAP_BUCKETS = [
     ['<10s', 10], ['10-30s', 30], ['30s-1m', 60], ['1-2m', 120],
     ['2-5m', 300], ['5-15m', 900], ['15-60m', Infinity],
 ];
+/** Silent-stall duration buckets (agent idle after its last event). */
+const SILENT_STALL_BUCKETS = [
+    // label, min inclusive, max exclusive (Infinity = open)
+    ['silent stall: 5-15m', 300, 900],
+    ['silent stall: 15-60m', 900, 3600],
+    ['silent stall: 1h+', 3600, Infinity],
+];
 function emptyFacets() {
     return {
-        toolCounts: {}, models: {}, languages: {}, slashCommands: {}, errorCategories: {},
+        toolCounts: {}, models: {}, silentStallsByModel: {}, languages: {},
+        slashCommands: {}, errorCategories: {},
         interruptions: 0, responseGaps: [], gapsOverCeiling: 0,
         linesTouchedBefore: 0, linesTouchedAfter: 0, editingToolCalls: 0,
         filesCreated: 0, filesModified: 0, filesDeleted: 0, gitCommits: 0, gitPushes: 0,
         shellCommandsSeen: 0,
         messageHours: new Array(24).fill(0), userTurns: 0, assistantTurns: 0,
         toolCount: 0, errorCount: 0,
+        frictionSignals: {}, correctionSignals: {}, automationSignals: {},
     };
 }
 function bump(map, key, by = 1) {
@@ -141,10 +159,19 @@ export function computeInsightFacets(events, timezoneOffsetMinutes = new Date().
     // reported p50 by 63% (143s against a true 88s), because fast replies are common and
     // dropping them all shifts the median right. A 0-second reply is a real reply.
     //
-    // The upper bound stays: past an hour the user went away and came back, which is not
-    // a reply latency. It censors 5.5% of gaps, and `gapsOverCeiling` reports how many so
-    // the number is never quietly truncated.
+    // The upper bound stays for *percentiles only*: past an hour the user often left the
+    // desk, which is not "reply latency." Those gaps still feed silent-stall classification
+    // (the agent had already stopped). `gapsOverCeiling` reports how many so percentiles
+    // are never quietly truncated.
+    //
+    // Silent stall (agent idle): when the gap after the assistant's last event is long
+    // enough that the model clearly stopped mid-session and waited for a human nudge
+    // ("continue", "keep going", or any later message after minutes of silence). This is
+    // the inverse framing of reply latency: same timestamps, attributed to the agent.
     let lastAssistantTs = null;
+    /** Shortened model id of the last assistant activity (for stall attribution). */
+    let lastAssistantModel = null;
+    let lastFailedTool = null;
     for (const e of events) {
         const ts = new Date(e.timestamp).getTime();
         const hasTs = !Number.isNaN(ts);
@@ -153,22 +180,44 @@ export function computeInsightFacets(events, timezoneOffsetMinutes = new Date().
                 f.interruptions++;
                 break;
             case 'usage':
-                // shortenModel so the label matches `agents sessions <id>` and `trends`
+                // shortenModel so the label matches `agents sessions <id>` and `insights mix`
                 // rather than printing the raw id beside their shortened one.
-                if (e.model)
-                    bump(f.models, shortenModel(e.model));
+                if (e.model) {
+                    const m = shortenModel(e.model);
+                    bump(f.models, m);
+                    // Usage rows are the reliable model tag for Claude turns; keep as
+                    // "last model" even when the timestamp is missing so a following
+                    // tool_use still attributes a stall correctly.
+                    lastAssistantModel = m;
+                }
                 break;
             case 'error':
                 bump(f.errorCategories, categorizeError(e.content ?? e.output ?? ''));
+                if (e.tool && e.tool === lastFailedTool)
+                    bump(f.frictionSignals, `failed tool loop: ${e.tool}`);
+                lastFailedTool = e.tool ?? null;
+                classifyFriction(e.content ?? e.output ?? '', f.frictionSignals);
+                break;
+            case 'tool_result':
+                if (e.success !== false && e.outcome !== 'error')
+                    lastFailedTool = null;
                 break;
             case 'message':
                 if (e.role === 'assistant') {
                     if (hasTs)
                         lastAssistantTs = ts;
+                    if (e.model)
+                        lastAssistantModel = shortenModel(e.model);
                     break;
                 }
                 if (e.role !== 'user')
                     break;
+                // Synthetic user rows (stop-hook feedback, injected meta) are not a human
+                // resume — skip correction/stall classification and leave lastAssistantTs so
+                // the next real user message still measures the full idle gap.
+                if (e._synthetic)
+                    break;
+                classifyCorrection(e.content ?? '', f.correctionSignals);
                 if (hasTs) {
                     // Local-time hour. parse.ts falls back to `new Date()` for a record with no
                     // timestamp; those are indistinguishable here, but they are rare and would
@@ -184,6 +233,11 @@ export function computeInsightFacets(events, timezoneOffsetMinutes = new Date().
                             f.responseGaps.push(gap);
                         else if (gap >= GAP_CEILING_SECONDS)
                             f.gapsOverCeiling++;
+                        // Agent silent stall: model stopped; session sat idle until this message.
+                        if (gap >= SILENT_STALL_SECONDS) {
+                            classifySilentStall(gap, e.content ?? '', f.frictionSignals, f.correctionSignals);
+                            bump(f.silentStallsByModel, lastAssistantModel ?? 'unknown');
+                        }
                     }
                 }
                 lastAssistantTs = null;
@@ -195,8 +249,12 @@ export function computeInsightFacets(events, timezoneOffsetMinutes = new Date().
                     break;
                 if (hasTs)
                     lastAssistantTs = ts;
+                if (e.model)
+                    lastAssistantModel = shortenModel(e.model);
                 const args = e.args ?? {};
                 const toolName = e.tool ?? '';
+                if (/askuserquestion/i.test(toolName))
+                    classifyAskStall(args, f.correctionSignals);
                 // Keyed on the SHARED cross-harness vocabulary, not Claude's literals. Keying
                 // on 'Edit'|'MultiEdit'|'Write' meant codex (whose vocabulary is exec /
                 // exec_command / write_stdin) reported 5,197 tool calls and exactly zero lines
@@ -218,6 +276,7 @@ export function computeInsightFacets(events, timezoneOffsetMinutes = new Date().
                     f.shellCommandsSeen++;
                     f.gitCommits += countGitOp(e.command, 'commit');
                     f.gitPushes += countGitOp(e.command, 'push');
+                    classifyAutomation(e.command, f.automationSignals);
                 }
                 break;
             }
@@ -228,6 +287,77 @@ export function computeInsightFacets(events, timezoneOffsetMinutes = new Date().
     }
     return f;
 }
+const FRICTION_PATTERNS = [
+    [/blocked by|guard(?:rail)? (?:blocked|denied)|permission denied/i, 'blocked guard'],
+    [/\b(?:ci|check|workflow)\b.*\b(?:red|fail(?:ed|ure)?)\b|\b(?:red|failed)\b.*\bci\b/i, 'CI red loop'],
+    [/merge conflict|conflict in |CONFLICT \(/i, 'merge conflict'],
+];
+const CORRECTION_PATTERNS = [
+    [/\b(?:continue|keep going|don'?t stop)\b/i, 'continue / keep going'],
+    [/\b(?:yes|go ahead|do it|merge it)\b/i, 'approval repeated'],
+    [/\b(?:are we done|done end.to.end|what(?:'s| is) left)\b/i, 'done end-to-end?'],
+    [/\b(?:did you merge|merged\??)\b/i, 'did you merge?'],
+    [/\bwhat(?:'s| is) next\??\b/i, "what's next?"],
+    [/\b(?:check now|check again|try now|did it work)\b/i, 'check now'],
+    [/\b(?:don'?t ask|just do it|run what)\b/i, "don't ask / just do it"],
+];
+/** User text that is a pure resume nudge after the agent went silent. */
+const RESUME_NUDGE_PATTERN = /\b(?:continue|keep going|don'?t stop|resume|pick up|you stopped|still there|wake up|hello\??|are you (?:there|stuck)|go on)\b/i;
+/**
+ * Classify a long quiet gap after the assistant's last event as an agent silent stall.
+ * Optionally also marks an explicit resume nudge ("continue", …) after that silence.
+ */
+export function classifySilentStall(gapSeconds, userText, friction, corrections) {
+    for (const [label, min, max] of SILENT_STALL_BUCKETS) {
+        if (gapSeconds >= min && gapSeconds < max) {
+            bump(friction, label);
+            break;
+        }
+    }
+    const normalized = userText.replace(/<[^>]+>/g, ' ').replace(/\s+/g, ' ').trim();
+    if (normalized && RESUME_NUDGE_PATTERN.test(normalized)) {
+        bump(corrections, 'resume after silent stall');
+    }
+}
+const AUTOMATION_PATTERNS = [
+    [/\bgh pr (?:checks|view|merge)\b/i, 'PR babysitting'],
+    [/\bagents secrets (?:list|exec|unlock|export)\b/i, 'secrets unlock dance'],
+    [/\bgit (?:fetch|rebase|merge|push)\b/i, 'git reconcile recipe'],
+    [/\b(?:scp|rsync|agents ssh)\b/i, 'fleet file transfer'],
+    [/\b(?:release\.sh|deploy\.sh|npm publish)\b/i, 'release / deploy recipe'],
+];
+function classifyFriction(text, counts) {
+    for (const [pattern, label] of FRICTION_PATTERNS)
+        if (pattern.test(text))
+            bump(counts, label);
+}
+function classifyCorrection(text, counts) {
+    const normalized = text.replace(/<[^>]+>/g, ' ').replace(/\s+/g, ' ').trim();
+    for (const [pattern, label] of CORRECTION_PATTERNS)
+        if (pattern.test(normalized))
+            bump(counts, label);
+}
+function classifyAskStall(args, counts) {
+    const text = JSON.stringify(args).toLowerCase();
+    const categories = [
+        [/release|ship|deploy|publish/, 'Ask stall: release / ship / deploy'],
+        [/what'?s next|next step|next move/, "Ask stall: what's next?"],
+        [/merge|reconcile|rebase/, 'Ask stall: merge / reconcile'],
+        [/direction|approach|implementation/, 'Ask stall: direction / approach'],
+    ];
+    for (const [pattern, label] of categories) {
+        if (pattern.test(text)) {
+            bump(counts, label);
+            return;
+        }
+    }
+    bump(counts, 'AskUserQuestion');
+}
+function classifyAutomation(command, counts) {
+    for (const [pattern, label] of AUTOMATION_PATTERNS)
+        if (pattern.test(command))
+            bump(counts, label);
+}
 /** Percentile of a numeric sample, nearest-rank. Returns 0 for an empty sample. */
 export function percentile(values, p) {
     if (values.length === 0)
@@ -292,12 +422,20 @@ export function mergeFacets(into, add) {
         bump(into.toolCounts, k, v);
     for (const [k, v] of Object.entries(add.models))
         bump(into.models, k, v);
+    for (const [k, v] of Object.entries(add.silentStallsByModel ?? {}))
+        bump(into.silentStallsByModel, k, v);
     for (const [k, v] of Object.entries(add.languages))
         bump(into.languages, k, v);
     for (const [k, v] of Object.entries(add.slashCommands))
         bump(into.slashCommands, k, v);
     for (const [k, v] of Object.entries(add.errorCategories))
         bump(into.errorCategories, k, v);
+    for (const [k, v] of Object.entries(add.frictionSignals ?? {}))
+        bump(into.frictionSignals, k, v);
+    for (const [k, v] of Object.entries(add.correctionSignals ?? {}))
+        bump(into.correctionSignals, k, v);
+    for (const [k, v] of Object.entries(add.automationSignals ?? {}))
+        bump(into.automationSignals, k, v);
     into.interruptions += add.interruptions;
     into.responseGaps.push(...add.responseGaps);
     into.gapsOverCeiling += add.gapsOverCeiling;
@@ -328,3 +466,75 @@ export function topEntries(counts, limit) {
         .sort((a, b) => b.count - a.count || a.name.localeCompare(b.name))
         .slice(0, limit);
 }
+/** Build a stable, evidence-backed action list without exposing transcript text. */
+export function buildInsightActions(sessions) {
+    const specs = [
+        { source: 'correctionSignals', label: 'continue / keep going', category: 'rule', action: 'Keep working through the delivery chain without waiting for another “continue”.' },
+        { source: 'correctionSignals', label: 'resume after silent stall', category: 'rule', action: 'Never stop mid-task waiting for a human ping — finish the current goal or park with an explicit blocker; keep driving CI/review/merge without going idle.' },
+        { source: 'correctionSignals', label: 'approval repeated', category: 'rule', action: 'Treat the original build or ship request as authorization for routine follow-through.' },
+        { source: 'correctionSignals', label: 'done end-to-end?', category: 'rule', action: 'Verify the user-visible outcome before declaring the task complete.' },
+        { source: 'correctionSignals', label: 'did you merge?', category: 'automation', action: 'Automate PR review, CI watching, and merge-on-green as one durable workflow.' },
+        { source: 'correctionSignals', label: "what's next?", category: 'rule', action: 'State and execute the next in-scope step instead of asking the owner to steer implementation.' },
+        { source: 'correctionSignals', label: 'check now', category: 'automation', action: 'Add bounded status polling with a terminal success or failure signal.' },
+        { source: 'correctionSignals', label: "don't ask / just do it", category: 'rule', action: 'Reserve questions for genuine product or scope choices.' },
+        { source: 'correctionSignals', label: 'Ask stall: release / ship / deploy', category: 'skill', action: 'Teach release workflows to carry publish, tag, rollout, and live verification as one chain.' },
+        { source: 'correctionSignals', label: "Ask stall: what's next?", category: 'rule', action: 'Remove workflow-stall “what next?” prompts from agent guidance.' },
+        { source: 'correctionSignals', label: 'Ask stall: merge / reconcile', category: 'skill', action: 'Encode the safe merge and reconcile path in the git workflow skill.' },
+        { source: 'correctionSignals', label: 'Ask stall: direction / approach', category: 'rule', action: 'Let agents choose implementation details after scope is clear.' },
+        { source: 'frictionSignals', label: 'silent stall: 5-15m', category: 'rule', action: 'Stop ending turns while work remains open — after a tool batch, take the next step or schedule a real background wait that re-invokes you; do not sit idle until the user says continue.' },
+        { source: 'frictionSignals', label: 'silent stall: 15-60m', category: 'rule', action: 'Long idle after the assistant last spoke is an agent stop, not a user pause — drive open PRs/CI/todos to completion or name a true external blocker instead of going silent.' },
+        { source: 'frictionSignals', label: 'silent stall: 1h+', category: 'rule', action: 'Sessions that sit idle for an hour+ after the model stops are stranded work — use stop-gates, background watches, and queue drain so a human is not the only resume signal.' },
+        { source: 'frictionSignals', label: 'blocked guard', category: 'product', action: 'Make guard failures return the safe next command and exact blocked operation.' },
+        { source: 'frictionSignals', label: 'CI red loop', category: 'automation', action: 'Deduplicate CI watchers and turn repeated red checks into one stateful wait.' },
+        { source: 'frictionSignals', label: 'merge conflict', category: 'skill', action: 'Standardize conflict diagnosis and fix-forward reconciliation.' },
+        { source: 'automationSignals', label: 'PR babysitting', category: 'automation', action: 'Bundle PR checks, review collection, comment handling, and merge-on-green.' },
+        { source: 'automationSignals', label: 'secrets unlock dance', category: 'product', action: 'Provide one headless secrets-backed command path for repeated credential operations.' },
+        { source: 'automationSignals', label: 'git reconcile recipe', category: 'skill', action: 'Promote repeated git reconciliation commands into the canonical workflow.' },
+        { source: 'automationSignals', label: 'fleet file transfer', category: 'product', action: 'Add a first-class fleet file transfer command with host/path validation.' },
+        { source: 'automationSignals', label: 'release / deploy recipe', category: 'automation', action: 'Turn repeated release shell recipes into a checked-in release script.' },
+    ];
+    const actions = [];
+    for (const spec of specs) {
+        let evidenceCount = 0;
+        const ids = [];
+        for (const session of sessions) {
+            const count = session.facets[spec.source]?.[spec.label] ?? 0;
+            if (count <= 0)
+                continue;
+            evidenceCount += count;
+            if (ids.length < 3)
+                ids.push(session.id.slice(0, 8));
+        }
+        if (evidenceCount === 0)
+            continue;
+        actions.push({
+            priority: evidenceCount >= 10 ? 'high' : evidenceCount >= 3 ? 'medium' : 'low',
+            category: spec.category,
+            action: spec.action,
+            evidenceCount,
+            sampleSessionIds: ids,
+        });
+    }
+    let failedLoopCount = 0;
+    const failedLoopIds = [];
+    for (const session of sessions) {
+        const count = Object.entries(session.facets.frictionSignals ?? {})
+            .filter(([label]) => label.startsWith('failed tool loop:'))
+            .reduce((sum, [, value]) => sum + value, 0);
+        if (count <= 0)
+            continue;
+        failedLoopCount += count;
+        if (failedLoopIds.length < 3)
+            failedLoopIds.push(session.id.slice(0, 8));
+    }
+    if (failedLoopCount > 0) {
+        actions.push({
+            priority: failedLoopCount >= 10 ? 'high' : failedLoopCount >= 3 ? 'medium' : 'low',
+            category: 'automation',
+            action: 'Detect repeated failures of the same tool and stop the retry loop with a different recovery path.',
+            evidenceCount: failedLoopCount,
+            sampleSessionIds: failedLoopIds,
+        });
+    }
+    return actions.sort((a, b) => b.evidenceCount - a.evidenceCount || a.action.localeCompare(b.action));
+}

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

@@ -5,6 +5,7 @@ export type SessionRecoveryTarget = {
     mode: 'native';
     agent: AgentId;
     version: string;
+    cwd?: string;
     reason: string;
 } | {
     mode: 'continue';
@@ -12,6 +13,13 @@ export type SessionRecoveryTarget = {
     version: string;
     reason: string;
 };
+export type NativeResumeInspection = {
+    available: true;
+    cwd?: string;
+} | {
+    available: false;
+    reason: string;
+};
 export declare class SessionRecoveryError extends Error {
     constructor(message: string);
 }
@@ -21,6 +29,14 @@ export declare function sessionOriginDevice(session: Pick<SessionMeta, 'machine'
 export declare function sessionRecoveryPeer(session: Pick<SessionMeta, 'machine'>, selfCheck?: (host: string) => boolean): string | undefined;
 /** Whether an explicit placement names the session's origin device. */
 export declare function sessionRecoveryDestinationMatches(session: Pick<SessionMeta, 'machine'>, requestedHost: string, self?: string): boolean;
+/**
+ * Prove that the indexed transcript is reachable from the exact active version
+ * home that would receive native resume. Retained trash/backup transcripts are
+ * intentionally rejected here: they remain readable by `/continue`, but a new
+ * installation with the same version number must not native-resume an empty
+ * isolated home.
+ */
+export declare function inspectNativeResumeSession(session: SessionMeta, versionHome: string): NativeResumeInspection;
 /**
  * Decide how a durable session resumes on the device that owns it.
  *
@@ -29,7 +45,7 @@ export declare function sessionRecoveryDestinationMatches(session: Pick<SessionM
  * same harness and uses `/continue`, whose indexed transcript reader can reach
  * retained version trash. No healthy same-harness account is a loud failure.
  */
-export declare function resolveSessionRecoveryFromCandidates(session: SessionMeta, candidates: RotateCandidate[], supportsNative?: (agent: AgentId, version?: string) => boolean): SessionRecoveryTarget;
+export declare function resolveSessionRecoveryFromCandidates(session: SessionMeta, candidates: RotateCandidate[], supportsNative?: (agent: AgentId, version?: string) => boolean, nativeInspection?: NativeResumeInspection): SessionRecoveryTarget;
 export declare function resolveSessionRecovery(session: SessionMeta): Promise<SessionRecoveryTarget>;
 /** Stable self-command used by focus, resume, and attach. The owning device runs
  * the recovery resolver above; callers must not native-resume another version's

package/dist/lib/session/recovery.js

@@ -1,8 +1,11 @@
-import { AGENTS } from '../agents.js';
+import * as fs from 'fs';
+import * as path from 'path';
+import { AGENTS, agentConfigDirName } from '../agents.js';
 import { isSelfHost } from '../devices/self-host.js';
 import { nativeResume } from '../exec.js';
 import { machineId, normalizeHost } from '../machine-id.js';
 import { collectRunCandidates, formatNoHealthyAccountError, pickBalancedCandidate, readinessFromCandidate, } from '../rotate.js';
+import { getVersionHomePath } from '../versions.js';
 export class SessionRecoveryError extends Error {
     constructor(message) {
         super(message);
@@ -41,6 +44,99 @@ function sourceReason(session, candidates) {
         ? `origin ${session.agent}@${session.version} has no native resume form`
         : `origin ${session.agent}@${session.version} is ${readiness.reason}`;
 }
+function isPathInside(candidate, dir) {
+    const rel = path.relative(dir, candidate);
+    return rel === '' || (!!rel && !rel.startsWith('..') && !path.isAbsolute(rel));
+}
+function existingDirectory(dir) {
+    if (!dir)
+        return undefined;
+    try {
+        return fs.statSync(dir, { throwIfNoEntry: false })?.isDirectory() ? dir : undefined;
+    }
+    catch {
+        return undefined;
+    }
+}
+/** Read the launch cwd Claude used to choose its projects/<cwd-key> directory.
+ * Claude can record attachment envelopes before the first user turn, and those
+ * envelopes retain the actual launch cwd even after the session changes dirs. */
+function readClaudeLaunchCwd(filePath) {
+    const maxBytes = 2 * 1024 * 1024;
+    let fd;
+    try {
+        fd = fs.openSync(filePath, 'r');
+    }
+    catch {
+        return undefined;
+    }
+    try {
+        const chunk = Buffer.alloc(maxBytes);
+        const bytesRead = fs.readSync(fd, chunk, 0, maxBytes, 0);
+        const lines = chunk.toString('utf8', 0, bytesRead).split('\n');
+        for (const line of lines) {
+            if (!line.trim())
+                continue;
+            try {
+                const parsed = JSON.parse(line);
+                if (typeof parsed?.cwd !== 'string' || !path.isAbsolute(parsed.cwd))
+                    continue;
+                if (existingDirectory(parsed.cwd))
+                    return parsed.cwd;
+            }
+            catch {
+                // A malformed line or vanished cwd cannot identify a usable native home.
+            }
+        }
+    }
+    finally {
+        fs.closeSync(fd);
+    }
+    return undefined;
+}
+/**
+ * Prove that the indexed transcript is reachable from the exact active version
+ * home that would receive native resume. Retained trash/backup transcripts are
+ * intentionally rejected here: they remain readable by `/continue`, but a new
+ * installation with the same version number must not native-resume an empty
+ * isolated home.
+ */
+export function inspectNativeResumeSession(session, versionHome) {
+    let realFile;
+    try {
+        realFile = fs.realpathSync(session.filePath);
+    }
+    catch {
+        return { available: false, reason: 'the indexed transcript is no longer present in the origin home' };
+    }
+    const roots = [versionHome, path.join(versionHome, agentConfigDirName(session.agent))];
+    const owned = roots.some((root) => {
+        try {
+            return isPathInside(realFile, fs.realpathSync(root));
+        }
+        catch {
+            return false;
+        }
+    });
+    if (!owned) {
+        return {
+            available: false,
+            reason: `the indexed transcript is retained outside the active ${session.agent}@${session.version ?? 'unknown'} home`,
+        };
+    }
+    if (session.agent === 'claude') {
+        const cwd = readClaudeLaunchCwd(realFile);
+        if (!cwd) {
+            return {
+                available: false,
+                reason: 'the Claude transcript does not identify an existing original project directory',
+            };
+        }
+        return { available: true, cwd };
+    }
+    const cwd = existingDirectory(session.cwd);
+    return { available: true, cwd };
+}
 /**
  * Decide how a durable session resumes on the device that owns it.
  *
@@ -49,7 +145,7 @@ function sourceReason(session, candidates) {
  * same harness and uses `/continue`, whose indexed transcript reader can reach
  * retained version trash. No healthy same-harness account is a loud failure.
  */
-export function resolveSessionRecoveryFromCandidates(session, candidates, supportsNative = nativeResume) {
+export function resolveSessionRecoveryFromCandidates(session, candidates, supportsNative = nativeResume, nativeInspection) {
     const agent = runnableSessionAgent(session);
     const device = sessionOriginDevice(session);
     const source = session.version
@@ -69,11 +165,22 @@ export function resolveSessionRecoveryFromCandidates(session, candidates, suppor
     }
     const version = selection.picked.version;
     if (session.version === version && supportsNative(agent, version)) {
+        const inspection = nativeInspection
+            ?? inspectNativeResumeSession(session, getVersionHomePath(agent, version));
+        if (inspection.available) {
+            return {
+                mode: 'native',
+                agent,
+                version,
+                cwd: inspection.cwd,
+                reason: `origin ${agent}@${version} is installed, healthy, and owns the indexed transcript`,
+            };
+        }
         return {
-            mode: 'native',
+            mode: 'continue',
             agent,
             version,
-            reason: `origin ${agent}@${version} is installed and healthy`,
+            reason: `${inspection.reason}; continuing with healthy ${agent}@${version}`,
         };
     }
     return {

package/dist/lib/session/remote-list.d.ts

@@ -2,20 +2,12 @@ import { type DeviceProfile } from '../devices/registry.js';
 import { type RemoteAgentsJsonParseResult } from '../remote-agents-json.js';
 import type { SessionMeta } from './types.js';
 import { type ToolProgramCountEnvelope, type ToolSearchEnvelope } from './tool-index.js';
-export declare const REMOTE_STDOUT_MAX_BYTES: number;
+export { REMOTE_STDOUT_MAX_BYTES, RemoteUtf8Accumulator } from '../ssh-exec.js';
 export declare const REMOTE_TOOL_AGGREGATE_MAX_BYTES: number;
 export interface RemoteToolByteBudget {
     remainingBytes: number;
     exhausted: boolean;
 }
-/** Preserve UTF-8 code points when SSH splits them across stdout chunks. */
-export declare class RemoteUtf8Accumulator {
-    private readonly decoder;
-    private value;
-    write(chunk: Buffer): void;
-    end(): string;
-    current(): string;
-}
 /** Claim received bytes against one fleet-query budget before retaining them. */
 export declare function consumeRemoteToolByteBudget(budget: RemoteToolByteBudget, bytes: number): boolean;
 /** Charge sanitized, machine-stamped evidence because redaction may expand it. */
@@ -130,7 +122,16 @@ export declare function resolvePeerTarget(machine: string): Promise<{
  * not via a local `--host` hop, which would discover locally and dead-end for a
  * session that exists only on the peer. Resolves 'no-target' when the machine
  * isn't a dialable registered device; the caller surfaces a clear message.
+ *
+ * `opts.env` adds variables to the remote command. It deliberately does NOT
+ * carry `AGENTS_FLEET_REMOTE` the way the `--host` passthrough does: that marker
+ * gates consent-sensitive actions on the far side
+ * (lib/browser/remote-control.ts), and a resumed agent is a long-lived session
+ * that would inherit it for its whole life — `agents browser start` inside it
+ * would then be refused as a cross-machine drive. A one-shot `--host` command
+ * can carry the marker; a session cannot.
  */
 export declare function runOnPeer(args: string[], machine: string, opts?: {
     tty?: boolean;
+    env?: Record<string, string>;
 }): Promise<'ok' | 'no-target'>;