agent-relay-runner 0.28.0 → 0.29.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-relay-runner",
-  "version": "0.28.0",
+  "version": "0.29.0",
   "description": "Unified provider lifecycle runner for Agent Relay",
   "type": "module",
   "bin": {
@@ -20,7 +20,7 @@
     "directory": "runner"
   },
   "dependencies": {
-    "agent-relay-sdk": "0.2.17"
+    "agent-relay-sdk": "0.2.18"
   },
   "devDependencies": {
     "@types/bun": "latest",

package/plugins/claude/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "agent-relay-runner",
   "description": "Thin Agent Relay runner bridge for Claude Code",
-  "version": "0.28.0",
+  "version": "0.29.0",
   "agentRelayContracts": {
     "providerPluginProtocol": 1
   }

package/src/adapters/claude-transcript.ts

@@ -21,7 +21,7 @@ interface TranscriptBlock {
   is_error?: boolean;
 }
 
-export interface TurnStep {
+interface TurnStep {
   type: "narration" | "reasoning" | "tool";
   text: string;
   label?: string;
@@ -36,6 +36,16 @@ interface TranscriptMessage {
 interface TranscriptEntry {
   type?: string;
   message?: TranscriptMessage;
+  // Claude Code stamps every transcript entry with `isSidechain`: true for
+  // entries belonging to a Task (subagent) run, false for the root session.
+  // Current CC writes sidechains to a separate subagents/*.jsonl so they don't
+  // reach the root transcript the runner tails — but older CC inlined them, and
+  // the behavior can revert, so the chat-mirror parsers below defensively skip
+  // sidechain entries to keep a subagent's reasoning/tools/responses from
+  // leaking into the parent agent's chat. Insights parsers (collectClaudeSession-
+  // Events/countSubstantiveTurns) intentionally do NOT filter — changing them
+  // would shift the #184/#185 baselines, a separate concern.
+  isSidechain?: boolean;
 }
 
 function blocks(message: TranscriptMessage | undefined): TranscriptBlock[] {
@@ -43,6 +53,11 @@ function blocks(message: TranscriptMessage | undefined): TranscriptBlock[] {
   return message.content.filter((b): b is TranscriptBlock => Boolean(b) && typeof b === "object");
 }
 
+/** True for a subagent (Task) transcript entry — see the note on TranscriptEntry.isSidechain. */
+function isSidechainEntry(entry: TranscriptEntry): boolean {
+  return entry.isSidechain === true;
+}
+
 function isRealUserPrompt(entry: TranscriptEntry): boolean {
   if (entry.type !== "user") return false;
   const content = entry.message?.content;
@@ -75,6 +90,7 @@ export function transcriptLooksComplete(jsonl: string): boolean {
     if (!trimmed) continue;
     try {
       const entry = JSON.parse(trimmed) as TranscriptEntry;
+      if (isSidechainEntry(entry)) continue;
       if (entry.type === "assistant") lastAssistantStopReason = entry.message?.stop_reason;
     } catch {
       continue;
@@ -99,6 +115,7 @@ export function extractLastAssistantTurn(jsonl: string): string {
     } catch {
       continue;
     }
+    if (isSidechainEntry(entry)) continue;
     if (isRealUserPrompt(entry)) {
       collected = [];
       continue;
@@ -128,6 +145,7 @@ export function extractFinalAssistantMessage(jsonl: string): string {
     } catch {
       continue;
     }
+    if (isSidechainEntry(entry)) continue;
     if (isRealUserPrompt(entry)) {
       pastLastUserPrompt = true;
       lastText = "";
@@ -166,6 +184,7 @@ export function extractLatestTurnSteps(jsonl: string): TurnStep[] {
     } catch {
       continue;
     }
+    if (isSidechainEntry(entry)) continue;
     if (isRealUserPrompt(entry)) {
       steps = [];
       continue;
@@ -184,6 +203,25 @@ export function extractLatestTurnSteps(jsonl: string): TurnStep[] {
   return steps;
 }
 
+/**
+ * Stable dedup keys for a turn's steps, in order. Each key is salted with how many
+ * identical (type,label,text) steps preceded it in the same window — so running the
+ * same tool twice with identical input within a turn yields two distinct keys and
+ * both show in the activity trace (#265). Keying on occurrence-within-window rather
+ * than raw transcript index keeps the reasoning tailer idempotent when the "latest
+ * turn" window shrinks/resets mid-poll: a surviving step recomputes to the same or a
+ * lower occurrence, so an already-emitted step never re-fires.
+ */
+export function stepDedupKeys(steps: TurnStep[]): string[] {
+  const counts = new Map<string, number>();
+  return steps.map((step) => {
+    const base = JSON.stringify([step.type, step.label ?? "", step.text]);
+    const occ = counts.get(base) ?? 0;
+    counts.set(base, occ + 1);
+    return JSON.stringify([step.type, step.label ?? "", step.text, occ]);
+  });
+}
+
 /** Compact one-line summary of a tool invocation for the discreet activity row. */
 export function summarizeToolUse(name: string, input: Record<string, unknown> | undefined): string {
   const str = (key: string): string | undefined => (input && typeof input[key] === "string" ? (input[key] as string) : undefined);

package/src/adapters/claude.ts

@@ -453,6 +453,21 @@ function captureTmuxPane(sessionName: string, socketName?: string): string {
   return result.stdout.toString();
 }
 
+// ⚠ FRAGILE PANE HEURISTICS — both functions below string-match Claude Code's TUI
+// chrome against captured tmux scrollback (~80 lines), so they break whenever CC
+// restyles its footer/banner. They are deliberately substring/regex based because
+// there's no machine-readable ready/busy signal from the TUI. Known break conditions,
+// so the next CC restyle is a fast fix rather than a hunt:
+//   readiness (claudePaneLooksReady) breaks if CC renames/removes ALL of: the
+//     "bypass permissions" / "shift+tab to cycle" / "? for shortcuts" footer hints,
+//     the "/effort" hint, or the "Welcome back" / "Claude Code" banner.
+//   busy (claudePaneIsBusy) breaks if CC drops the live "… (<elapsed>" spinner counter
+//     (the cross-version anchor; the "esc to interrupt" hint was already dropped in 2.1.x).
+//   FALSE POSITIVES: agent output that literally QUOTES any of these strings (e.g. a
+//     transcript discussing "esc to interrupt", or this very comment shown in a pane)
+//     reads as ready/busy. Tolerated because the markers are CC-specific enough to be
+//     rare in real output; if it bites, gate on the LAST N lines only (the live footer).
+// History: 18067b5 (busy counter), and the readiness footer-vs-banner fix below.
 export function claudePaneLooksReady(text: string): boolean {
   // Claude's startup banner ("Claude Code" / "Welcome back") scrolls off the pane once the
   // conversation fills it, so a mid-session delivery (e.g. the budget warning, minutes into

package/src/adapters/codex-client.ts

@@ -74,6 +74,8 @@ interface ThreadLoadedListResponse {
   nextCursor: string | null;
 }
 
+export const CODEX_APP_CLIENT_EVENT_CAP = 5_000;
+
 export class CodexAppClient {
   private ws!: WebSocket;
   private nextId = 1;
@@ -256,6 +258,9 @@ export class CodexAppClient {
 
   private record(event: ClientEvent): void {
     this.events.push(event);
+    if (this.events.length > CODEX_APP_CLIENT_EVENT_CAP) {
+      this.events.splice(0, this.events.length - CODEX_APP_CLIENT_EVENT_CAP);
+    }
     for (const listener of this.listeners) listener(event);
   }
 

package/src/adapters/codex.ts

@@ -41,6 +41,7 @@ export class CodexAdapter implements ProviderAdapter {
   // flushed as one session response on turn/completed (mirrors Claude's chatCaptureMode).
   private turnMessages: string[] = [];
   private readonly itemTextBuffers = new Map<string, string>();
+  private readonly itemTextBufferTypes = new Map<string, string>();
   private captureMode: "final" | "full" = "final";
   // #183/#184: the normalized session-event log for the current process lifetime, fed
   // from the same completed-item stream that drives the chat mirror. The runner slices
@@ -58,6 +59,20 @@ export class CodexAdapter implements ProviderAdapter {
     this.sessionEventCb = cb;
   }
 
+  private resetProcessState(): void {
+    this.resetThreadState();
+    this.sessionEvents = []; // fresh process -> fresh segment cursor (#184)
+  }
+
+  private resetThreadState(): void {
+    this.subagentThreads.clear();
+    this.pendingApprovals.clear();
+    this.activeTurnId = undefined;
+    this.turnMessages = [];
+    this.itemTextBuffers.clear();
+    this.itemTextBufferTypes.clear();
+  }
+
   async interrupt(process: ManagedProcess): Promise<Record<string, unknown>> {
     const client = process.meta?.client as CodexAppClient | undefined;
     if (!client) throw new Error("Codex App Server client is unavailable");
@@ -68,11 +83,33 @@ export class CodexAdapter implements ProviderAdapter {
     return { method: "turn-interrupt", turnId: this.activeTurnId };
   }
 
-  // Codex streams thread/status continuously, so the runner's claim state never
-  // goes stale the way Claude's can after an out-of-band interrupt. No cheap probe
-  // is needed — defer to the live status stream.
-  async probeActivity(): Promise<"busy" | "idle" | "unknown"> {
-    return "unknown";
+  async probeActivity(process: ManagedProcess): Promise<"busy" | "idle" | "unknown"> {
+    const client = process.meta?.client as CodexAppClient | undefined;
+    if (!client?.isConnected()) return "unknown";
+    const threadId = typeof process.meta?.threadId === "string" ? process.meta.threadId : "";
+    if (!this.activeTurnId) return "idle";
+    if (!threadId) return "busy";
+    try {
+      const read = await client.threadRead(threadId, true);
+      const thread = isRecord(read.thread) ? read.thread : undefined;
+      const turns = Array.isArray(thread?.turns) ? thread.turns : [];
+      const activeTurn = turns.find((turn) => isRecord(turn) && stringValue(turn.id) === this.activeTurnId);
+      const turnStatus = isRecord(activeTurn) ? stringValue(activeTurn.status) : undefined;
+      if (turnStatus === "inProgress") return "busy";
+      if (turnStatus === "completed" || turnStatus === "interrupted" || turnStatus === "failed") {
+        this.finishMainTurn();
+        return "idle";
+      }
+      const threadStatus = statusType(thread?.status);
+      if (threadStatus === "active") return "busy";
+      if (threadStatus === "idle" || threadStatus === "notLoaded" || threadStatus === "systemError") {
+        this.finishMainTurn();
+        return "idle";
+      }
+    } catch {
+      return "unknown";
+    }
+    return "busy";
   }
 
   // The Codex app-server is headless and has no tmux session, but an unexpected
@@ -82,8 +119,8 @@ export class CodexAdapter implements ProviderAdapter {
   }
 
   async spawn(config: RunnerSpawnConfig): Promise<ManagedProcess> {
+    this.resetProcessState();
     this.captureMode = (config.providerConfig as ProviderConfig).chatCaptureMode ?? "final";
-    this.sessionEvents = []; // fresh process → fresh segment cursor (#184)
     const args = this.buildSpawnArgs(config, config.providerConfig as ProviderConfig);
     const appServer = Bun.spawn([args.command, ...args.args], {
       cwd: args.cwd,
@@ -150,7 +187,7 @@ export class CodexAdapter implements ProviderAdapter {
     if (!client) throw new Error("Codex App Server client is unavailable");
     const threadId = typeof process.meta?.threadId === "string" ? process.meta.threadId : "";
     if (!threadId) throw new Error("Codex thread is not ready");
-    await client.threadCompactStart(threadId);
+    this.statusCb({ status: "busy", reason: "provider-turn", timeline: { status: "compacting", timestamp: Date.now() } });
     const currentContext = isContextState(process.meta?.context) ? process.meta.context : undefined;
     if (currentContext) {
       process.meta = {
@@ -158,6 +195,27 @@ export class CodexAdapter implements ProviderAdapter {
         context: { ...currentContext, lifecycleState: "compacting", lastUpdatedAt: Date.now() },
       };
     }
+    try {
+      await client.threadCompactStart(threadId);
+    } catch (error) {
+      this.statusCb({ status: "idle", reason: "provider-turn" });
+      throw error;
+    }
+    const compactedAt = Date.now();
+    const compactingContext = isContextState(process.meta?.context) ? process.meta.context : currentContext;
+    if (compactingContext) {
+      process.meta = {
+        ...(process.meta ?? {}),
+        context: {
+          ...compactingContext,
+          lifecycleState: "cooling",
+          tasksSinceCompact: 0,
+          lastCompactedAt: compactedAt,
+          lastUpdatedAt: compactedAt,
+        },
+      };
+    }
+    this.statusCb({ status: "idle", reason: "provider-turn", timeline: { status: "compacted", timestamp: compactedAt } });
     return { threadId };
   }
 
@@ -165,7 +223,16 @@ export class CodexAdapter implements ProviderAdapter {
     const client = process.meta?.client as CodexAppClient | undefined;
     if (!client) throw new Error("Codex App Server client is unavailable");
     const previousThreadId = typeof process.meta?.threadId === "string" ? process.meta.threadId : undefined;
-    const started = await client.threadStart({ cwd: typeof process.meta?.cwd === "string" ? process.meta.cwd : globalThis.process.cwd() });
+    this.statusCb({ status: "busy", reason: "provider-turn", timeline: { status: "clearing-context", timestamp: Date.now() } });
+    let started: Awaited<ReturnType<CodexAppClient["threadStart"]>>;
+    try {
+      started = await client.threadStart({ cwd: typeof process.meta?.cwd === "string" ? process.meta.cwd : globalThis.process.cwd() });
+    } catch (error) {
+      this.statusCb({ status: "idle", reason: "provider-turn" });
+      throw error;
+    }
+    const clearedAt = Date.now();
+    this.resetThreadState();
     process.meta = {
       ...(process.meta ?? {}),
       threadId: started.thread.id,
@@ -176,11 +243,13 @@ export class CodexAdapter implements ProviderAdapter {
         warmTopics: [],
         activeMemories: [],
         tasksSinceCompact: 0,
-        lastUpdatedAt: Date.now(),
+        lastCompactedAt: clearedAt,
+        lastUpdatedAt: clearedAt,
         source: "api",
         confidence: "reported",
       } satisfies ContextState,
     };
+    this.statusCb({ status: "idle", reason: "provider-turn", clear: ["subagent"], timeline: { status: "context-cleared", timestamp: clearedAt } });
     return { previousThreadId, threadId: started.thread.id };
   }
 
@@ -350,28 +419,25 @@ export class CodexAdapter implements ProviderAdapter {
         this.statusCb({ status: "busy", reason: "provider-turn", id: this.activeTurnId });
       }
     }
-    if (method.includes("turn/completed") || method.includes("turn.completed")) {
+    if (method.includes("turn/completed") || method.includes("turn.completed") || method.includes("turn/failed") || method.includes("turn.failed") || method.includes("turn/interrupted") || method.includes("turn.interrupted")) {
       if (threadId && this.subagentThreads.has(threadId)) {
         this.statusCb({ status: "idle", reason: "subagent", id: threadId, ...this.subagentThreads.get(threadId) });
       } else {
-        this.flushTurnResponse();
-        const completedTurnId = this.activeTurnId;
-        this.activeTurnId = undefined;
-        this.statusCb({ status: "idle", reason: "provider-turn", id: completedTurnId });
+        this.finishMainTurn();
       }
     }
     if ((method.includes("item/completed") || method.includes("item.completed")) && !isSubagent) {
       this.handleCodexItem(isRecord(params?.item) ? params.item : undefined);
     }
     if (!isSubagent) this.handleCodexItemDelta(method, params);
-    if (method.includes("thread/status")) {
+    if (method.includes("thread/status") || method.includes("thread.status")) {
       const status = statusType(params?.status);
       if (threadId && this.subagentThreads.has(threadId)) {
         if (status === "active") this.statusCb({ status: "busy", reason: "subagent", id: threadId, ...this.subagentThreads.get(threadId) });
         if (status === "idle" || status === "notLoaded") this.statusCb({ status: "idle", reason: "subagent", id: threadId, ...this.subagentThreads.get(threadId) });
       } else {
         if (status === "active") this.statusCb({ status: "busy", reason: "provider-turn", providerState: this.providerStateFromThreadStatus(params?.status, params) });
-        if (status === "idle") this.statusCb({ status: "idle", reason: "provider-turn" });
+        if (status === "idle" || status === "notLoaded" || status === "systemError") this.finishMainTurn();
       }
     }
   }
@@ -391,6 +457,7 @@ export class CodexAdapter implements ProviderAdapter {
         this.recordInsightEvent({ type: "turn" }); // a substantive assistant turn
       }
       if (itemId) this.itemTextBuffers.delete(itemId);
+      if (itemId) this.itemTextBufferTypes.delete(itemId);
       return;
     }
     if (type === "userMessage") {
@@ -406,6 +473,7 @@ export class CodexAdapter implements ProviderAdapter {
       const text = (codexReasoningText(item) || buffered || "").trim();
       if (text) this.sessionEventCb({ type: "reasoning", origin: "provider", body: text, ...(turnId ? { turnId } : {}) });
       if (itemId) this.itemTextBuffers.delete(itemId);
+      if (itemId) this.itemTextBufferTypes.delete(itemId);
       return;
     }
     const tool = codexToolSummary(type, item);
@@ -415,6 +483,7 @@ export class CodexAdapter implements ProviderAdapter {
       this.sessionEventCb({ type: "tool", origin: "provider", body: tool.body, label: tool.label, status: "completed", ...(turnId ? { turnId } : {}) });
     }
     if (itemId) this.itemTextBuffers.delete(itemId);
+    if (itemId) this.itemTextBufferTypes.delete(itemId);
   }
 
   // #183/#184: append to the session-event log with a soft cap. On overflow we drop the
@@ -449,7 +518,10 @@ export class CodexAdapter implements ProviderAdapter {
 
     if (type === "agentMessage" || type === "reasoning" || type === "plan") {
       const delta = codexDeltaText(params);
-      if (delta && itemId) this.itemTextBuffers.set(itemId, `${this.itemTextBuffers.get(itemId) ?? ""}${delta}`);
+      if (delta && itemId) {
+        this.itemTextBuffers.set(itemId, `${this.itemTextBuffers.get(itemId) ?? ""}${delta}`);
+        if (type) this.itemTextBufferTypes.set(itemId, type);
+      }
       return;
     }
 
@@ -459,13 +531,28 @@ export class CodexAdapter implements ProviderAdapter {
   }
 
   private flushTurnResponse(): void {
-    if (!this.turnMessages.length) return;
-    const joined = this.captureMode === "full" ? this.turnMessages.join("\n\n") : this.turnMessages[this.turnMessages.length - 1]!;
+    const pendingAgentMessages = [...this.itemTextBuffers.entries()]
+      .filter(([itemId]) => this.itemTextBufferTypes.get(itemId) === "agentMessage")
+      .map(([, text]) => text.trim())
+      .filter(Boolean);
+    const messages = [...this.turnMessages, ...pendingAgentMessages];
+    if (!messages.length) return;
+    const joined = this.captureMode === "full" ? messages.join("\n\n") : messages[messages.length - 1]!;
     this.turnMessages = [];
     const text = joined.trim();
     if (text) this.sessionEventCb({ type: "response", origin: "provider", body: text, ...(this.activeTurnId ? { turnId: this.activeTurnId } : {}) });
   }
 
+  private finishMainTurn(): void {
+    this.flushTurnResponse();
+    const turnId = this.activeTurnId;
+    this.activeTurnId = undefined;
+    this.pendingApprovals.clear();
+    this.itemTextBuffers.clear();
+    this.itemTextBufferTypes.clear();
+    this.statusCb({ status: "idle", reason: "provider-turn", id: turnId });
+  }
+
   private providerStateFromThreadStatus(status: unknown, params?: Record<string, unknown>): Record<string, unknown> | undefined {
     const state = codexProviderStateFromThreadStatus(status, params);
     if (state?.state !== "blocked" || state.reason !== "waitingOnApproval" || state.pendingApproval) return state;

package/src/attachment-cache.ts

@@ -7,11 +7,11 @@ import { sanitizeFsName } from "agent-relay-sdk/fs-name";
 
 const DEFAULT_CACHE_MAX_AGE_MS = 7 * 24 * 60 * 60 * 1000;
 
-export interface AttachmentCacheClient {
+interface AttachmentCacheClient {
   downloadArtifact(id: string): Promise<{ stream: ReadableStream<Uint8Array>; meta: Artifact }>;
 }
 
-export interface AttachmentCacheOptions {
+interface AttachmentCacheOptions {
   agentId: string;
   rootDir?: string;
   maxAgeMs?: number;
@@ -34,7 +34,7 @@ function attachmentRefs(message: Message): Record<string, unknown>[] {
   return refs.filter(isRecord);
 }
 
-export function attachmentCacheRoot(agentId: string, rootDir = process.env.AGENT_RELAY_ATTACHMENT_CACHE_DIR): string {
+function attachmentCacheRoot(agentId: string, rootDir = process.env.AGENT_RELAY_ATTACHMENT_CACHE_DIR): string {
   return join(attachmentCacheBase(rootDir), safePathPart(agentId));
 }
 

package/src/control-server.ts

@@ -237,7 +237,7 @@ async function handlePermissionRequest(
   return Response.json(claudePermissionHookResponse(decision, body));
 }
 
-export function claudePermissionApprovalView(id: string, body: Record<string, unknown>): Record<string, unknown> {
+function claudePermissionApprovalView(id: string, body: Record<string, unknown>): Record<string, unknown> {
   const toolName = typeof body.tool_name === "string" ? body.tool_name : "Tool";
   const toolInput = isRecord(body.tool_input) ? body.tool_input : {};
   // AskUserQuestion is not a yes/no gate — it asks the user to pick answers.
@@ -299,7 +299,7 @@ export function claudePermissionApprovalView(id: string, body: Record<string, un
   };
 }
 
-export function claudePermissionHookResponse(decision: ProviderPermissionDecisionInput, body: Record<string, unknown>): Record<string, unknown> {
+function claudePermissionHookResponse(decision: ProviderPermissionDecisionInput, body: Record<string, unknown>): Record<string, unknown> {
   // AskUserQuestion comes through a PreToolUse hook. The only way to satisfy it
   // headlessly is permissionDecision "allow" + updatedInput carrying the answers
   // (echoing back the original questions). A bare "allow" is not sufficient, so

package/src/logger.ts

@@ -16,7 +16,7 @@ import { sanitizeFsName } from "agent-relay-sdk/fs-name";
 // flipped at runtime via the control port (no restart) — so a phase refactor can
 // be watched at debug without bouncing the agent.
 
-export type LogLevel = "debug" | "info" | "warn" | "error" | "fatal";
+type LogLevel = "debug" | "info" | "warn" | "error" | "fatal";
 
 const ORDER: Record<LogLevel, number> = { debug: 10, info: 20, warn: 30, error: 40, fatal: 50 };
 export const LOG_LEVELS = Object.keys(ORDER) as LogLevel[];
@@ -33,7 +33,7 @@ function safeLogName(value: string): string {
   return sanitizeFsName(value, { replacement: "_", maxLen: 180 });
 }
 
-export interface LoggerConfig {
+interface LoggerConfig {
   agentId?: string;
   level?: LogLevel;
   headless?: boolean;

package/src/outbox.ts

@@ -21,9 +21,9 @@ import { logger } from "./logger";
 // last-wins and self-heals on reconnect (so it already satisfies "coalesce, don't replay
 // stale busyes"). The coalesce mode below exists so a future state event could migrate here.
 
-export type OutboxMode = "append" | "coalesce";
+type OutboxMode = "append" | "coalesce";
 
-export interface OutboxEventInput {
+interface OutboxEventInput {
   kind: string;
   payload: unknown;
   mode?: OutboxMode;
@@ -46,9 +46,9 @@ export interface OutboxRecord {
 }
 
 // The transport. Resolve = delivered (row deleted). Reject = failed (retried with backoff).
-export type OutboxSend = (record: OutboxRecord) => Promise<void>;
+type OutboxSend = (record: OutboxRecord) => Promise<void>;
 
-export interface OutboxOptions {
+interface OutboxOptions {
   agentId: string;
   send: OutboxSend;
   // Storage directory. Defaults to AGENT_RELAY_RUNNER_OUTBOX_DIR, else a per-host temp dir.

package/src/profile-home.ts

@@ -44,7 +44,7 @@ const CLAUDE_AUTH_ITEMS = [".credentials.json", "statsig"];
 // Shared skeleton for both providers: gate on isolated-profile, make the
 // instance-keyed home, run the provider-specific first-run bootstrap. The
 // bootstrap step is the only genuinely provider-specific part.
-export function prepareProviderHome(provider: "claude" | "codex", config: RunnerSpawnConfig): ProviderHome | undefined {
+function prepareProviderHome(provider: "claude" | "codex", config: RunnerSpawnConfig): ProviderHome | undefined {
   if (!profileRequiresIsolatedHome(config)) return undefined;
   const target = providerHomePath(provider, config);
   mkdirSync(target, { recursive: true });

package/src/relay-mcp-proxy.ts

@@ -40,7 +40,7 @@ const SSE_KEEPALIVE_MS = 25_000;
 // The write tools whose loss during a relay outage is unacceptable and whose result the agent
 // does not need synchronously — safe to queue durably and replay on reconnect. Reads, claims
 // (409 contention), spawn/shutdown (need a real ack) are deliberately NOT bufferable.
-export const DEFAULT_BUFFERABLE_TOOLS = new Set<string>([
+const DEFAULT_BUFFERABLE_TOOLS = new Set<string>([
   "relay_send_message",
   "relay_reply",
   "relay_workspace_ready",
@@ -59,7 +59,7 @@ const WORKTREE_ONLY_TOOLS = new Set<string>([
   "relay_workspace_land",
 ]);
 
-export interface ProxyContext {
+interface ProxyContext {
   // The agent owns a live (non-terminal) isolated git worktree → workspace tools apply.
   isolatedWorktree: boolean;
 }

package/src/reply-obligation-cache.ts

@@ -16,9 +16,9 @@ import { logger } from "./logger";
 // - A background interval keeps the snapshot warm; `markDirty()` requests an extra,
 //   debounced refresh when state likely just changed (a message arrived, a turn ended).
 
-export type ReplyObligationFetch = () => Promise<ReplyObligation[]>;
+type ReplyObligationFetch = () => Promise<ReplyObligation[]>;
 
-export interface ReplyObligationCacheOptions {
+interface ReplyObligationCacheOptions {
   fetch: ReplyObligationFetch;
   // Background freshness backstop. Default 10s — well under any turn cadence, cheap.
   intervalMs?: number;

package/src/runner.ts

@@ -11,7 +11,7 @@ import { ClaimTracker } from "./claim-tracker";
 import { startControlServer, type ControlServer } from "./control-server";
 import { ReplyObligationCache } from "./reply-obligation-cache";
 import { Outbox, type OutboxRecord } from "./outbox";
-import { extractLastAssistantTurn, extractFinalAssistantMessage, extractHookAssistantMessage, extractLatestTurnSteps, transcriptLooksComplete } from "./adapters/claude-transcript";
+import { extractLastAssistantTurn, extractFinalAssistantMessage, extractHookAssistantMessage, extractLatestTurnSteps, stepDedupKeys, transcriptLooksComplete } from "./adapters/claude-transcript";
 import { computeContextRatio } from "./session-insights";
 import { agentProfileProjectionReport } from "./profile-projection";
 import { profileUsesHostProviderGlobals } from "./profile-home";
@@ -1256,6 +1256,12 @@ export class AgentRunner {
   // the same pre-destroy seam the bus commands use. `clear`/`compact` continue the session;
   // anything else (logout, prompt_input_exit, other) is a real termination.
   private async handleSessionBoundary(input: { reason?: string; transcriptPath?: string }): Promise<void> {
+    // Reason mapping is fail-safe-toward-termination: only the two known session-
+    // CONTINUING reasons are special-cased; everything else (logout, prompt_input_exit,
+    // other, AND any future reason) maps to "shutdown" → full pre-destroy capture.
+    // ⚠ If Claude Code adds a new BENIGN/continuing boundary reason, add it here — until
+    // then it will trigger a (harmless but wasteful) full context capture on a session
+    // that isn't actually ending.
     const reason = input.reason === "compact" ? "compact"
       : input.reason === "clear" ? "clear"
       : "shutdown";
@@ -1494,6 +1500,8 @@ export class AgentRunner {
     // turn" window in the transcript can shrink/reset (a tool_result entry, a
     // mid-turn user line), and an index cursor would then either re-emit or stall
     // and drop the rest of the turn. A seen-set is idempotent under any reshuffle.
+    // The signature is salted with each step's occurrence-within-window (stepDedupKeys)
+    // so two identical steps in one turn — same tool, same input — both surface (#265).
     const seen = new Set<string>();
     const turnIdAtStart = this.currentTurnId;
     // On the first poll the new prompt usually hasn't landed in the transcript yet,
@@ -1509,16 +1517,16 @@ export class AgentRunner {
       try { jsonl = await readFile(transcriptPath, "utf8"); } catch { return; }
       let steps: ReturnType<typeof extractLatestTurnSteps>;
       try { steps = extractLatestTurnSteps(jsonl); } catch { return; }
+      const keyed = stepDedupKeys(steps).map((sig, i) => ({ sig, step: steps[i]! }));
       if (!seeded) {
         seeded = true;
         if (transcriptLooksComplete(jsonl)) {
-          for (const s of steps) seen.add(JSON.stringify([s.type, s.label ?? "", s.text]));
+          for (const { sig } of keyed) seen.add(sig);
         }
       }
       const turnId = this.currentTurnId ?? turnIdAtStart;
       let emitted = 0;
-      for (const step of steps) {
-        const sig = JSON.stringify([step.type, step.label ?? "", step.text]);
+      for (const { sig, step } of keyed) {
         if (seen.has(sig)) continue;
         seen.add(sig);
         emitted += 1;
@@ -2056,7 +2064,7 @@ export function latestClaudeResumeIdFromText(text: string): string | undefined {
   return latest;
 }
 
-export function latestClaudeResumeIdFromLogFile(path: string): string | undefined {
+function latestClaudeResumeIdFromLogFile(path: string): string | undefined {
   let fd: number | undefined;
   try {
     const stat = statSync(path);

package/src/session-insights.ts

@@ -37,7 +37,7 @@ export function isGatheringTool(name: string): boolean {
   return GATHERING_NAME.test(name);
 }
 
-export interface ContextRatioMetric {
+interface ContextRatioMetric {
   /** Session-wide gathering fraction: gatheringCalls / totalToolCalls. The headline metric. */
   ratio: number;
   gatheringCalls: number;
@@ -49,7 +49,7 @@ export interface ContextRatioMetric {
   turns: number;
 }
 
-export interface SessionOutcomeProxy {
+interface SessionOutcomeProxy {
   /** Real user prompts in the session — more back-and-forth ~ more clarification/correction. */
   userPrompts: number;
   /** tool_result blocks flagged is_error — failures/workarounds the agent hit. */

package/src/session-scratch.ts

@@ -2,7 +2,7 @@ import { execFileSync } from "node:child_process";
 import { accessSync, appendFileSync, constants, mkdirSync, readdirSync, readFileSync, rmSync, statSync } from "node:fs";
 import { dirname, isAbsolute, join, resolve } from "node:path";
 
-export const SCRATCH_DIR_NAME = ".agent-relay";
+const SCRATCH_DIR_NAME = ".agent-relay";
 // The local-ignore entry. Leading + trailing slash scopes it to the dir at the
 // base, matching git's gitignore semantics.
 const EXCLUDE_ENTRY = "/.agent-relay/";
@@ -16,7 +16,7 @@ export interface SessionScratchLayout {
   replyFile: string; // <tmp>/reply.md
 }
 
-export interface SessionScratchTarget {
+interface SessionScratchTarget {
   agentId: string;
   cwd: string;
   // Orchestrator base dir, used only when cwd is not writable. NEVER home — a
@@ -44,7 +44,7 @@ export function resolveScratchBase(cwd: string, fallbackBaseDir?: string): strin
   return cwd;
 }
 
-export function sessionScratchLayout(baseDir: string, agentId: string): SessionScratchLayout {
+function sessionScratchLayout(baseDir: string, agentId: string): SessionScratchLayout {
   const rootDir = join(baseDir, SCRATCH_DIR_NAME);
   const sessionsDir = join(rootDir, "sessions");
   const sessionDir = join(sessionsDir, agentId);
@@ -131,7 +131,7 @@ export function reapSessionScratch(target: SessionScratchTarget): void {
   }
 }
 
-export interface SweepOptions {
+interface SweepOptions {
   cwd: string;
   fallbackBaseDir?: string;
   // Agent ids to keep (currently-known agents + self). Any session dir whose id