agent-relay-runner 0.127.9 → 0.127.11

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-relay-runner",
-  "version": "0.127.9",
+  "version": "0.127.11",
   "description": "Unified provider lifecycle runner for Agent Relay",
   "type": "module",
   "bin": {
@@ -23,7 +23,7 @@
   },
   "dependencies": {
     "agent-relay-providers": "0.104.4",
-    "agent-relay-sdk": "0.2.122",
+    "agent-relay-sdk": "0.2.124",
     "callmux": "0.23.0"
   },
   "devDependencies": {

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": "0.127.9",
+  "version": "0.127.11",
   "agentRelayContracts": {
     "providerPluginProtocol": 1
   }

package/plugins/claude/monitors/relay-monitor.provisioned.mjs

@@ -47,8 +47,8 @@ var RUNNER_PROMPT_TEMPLATE_DEFAULTS = [
     category: "runner",
     title: "Codex Relay Context",
     description: "Relay context note injected into Codex sessions.",
-    template: "[agent-relay] You are connected to Agent Relay, a real-time message bus between agents and users. When you receive a relay message: read it, do what it asks, and reply through the relay when a text response is needed. Use agent-relay /react <messageId> <emoji> for lightweight acknowledgement or approval. If Relay MCP tools are available, prefer relay_reply, relay_get_message, relay_get_thread, relay_send_message, relay_upload_artifact, relay_attach_artifact, relay_agent_status, relay_find_agents, relay_compact_and_resume, relay_recall, relay_spawn_agent, and relay_shutdown_agent. You never need to know or pass your own agent id \u2014 relay fills it from your token; use relay_whoami only if you need to reason about yourself. relay_compact_and_resume is for clean-seam self-resume after a context advisory: pass workingState and optional ruledOut; Relay owns the objective envelope. relay_recall searches your own archived pre-compaction segments by keyword when a discarded detail is needed. relay_spawn_targets / relay_spawn_agent / relay_shutdown_agent only appear if your profile grants spawning (a live-children quota); when present, call relay_spawn_targets FIRST for the live host/provider/model matrix + your quota, then stand up long-living child agents and shut down your own \u2014 find them later with relay_find_agents spawnedBy:me. For replies, prefer relay_reply or relay_send_message; CLI fallback: agent-relay /reply <messageId> \"<your reply>\". If a delivered message says it was truncated, fetch the full body with: agent-relay get-message <messageId>. For command details, run: agent-relay /guide",
-    defaultTemplate: "[agent-relay] You are connected to Agent Relay, a real-time message bus between agents and users. When you receive a relay message: read it, do what it asks, and reply through the relay when a text response is needed. Use agent-relay /react <messageId> <emoji> for lightweight acknowledgement or approval. If Relay MCP tools are available, prefer relay_reply, relay_get_message, relay_get_thread, relay_send_message, relay_upload_artifact, relay_attach_artifact, relay_agent_status, relay_find_agents, relay_compact_and_resume, relay_recall, relay_spawn_agent, and relay_shutdown_agent. You never need to know or pass your own agent id \u2014 relay fills it from your token; use relay_whoami only if you need to reason about yourself. relay_compact_and_resume is for clean-seam self-resume after a context advisory: pass workingState and optional ruledOut; Relay owns the objective envelope. relay_recall searches your own archived pre-compaction segments by keyword when a discarded detail is needed. relay_spawn_targets / relay_spawn_agent / relay_shutdown_agent only appear if your profile grants spawning (a live-children quota); when present, call relay_spawn_targets FIRST for the live host/provider/model matrix + your quota, then stand up long-living child agents and shut down your own \u2014 find them later with relay_find_agents spawnedBy:me. For replies, prefer relay_reply or relay_send_message; CLI fallback: agent-relay /reply <messageId> \"<your reply>\". If a delivered message says it was truncated, fetch the full body with: agent-relay get-message <messageId>. For command details, run: agent-relay /guide",
+    template: "[agent-relay] You are connected to Agent Relay, a real-time message bus between agents and users. When you receive a relay message: read it, do what it asks, and reply through the relay when a text response is needed. Use agent-relay /react <messageId> <emoji> for lightweight acknowledgement or approval. If Relay MCP tools are available, prefer relay_reply, relay_get_message, relay_get_thread, relay_send_message, relay_upload_artifact, relay_attach_artifact, relay_agent_status, relay_find_agents, relay_compact_and_resume, relay_recall, relay_spawn_agent, and relay_shutdown_agent. You never need to know or pass your own agent id \u2014 relay fills it from your token. Your identity is injected at startup as the `## Agent Relay Identity` block; use relay_whoami only for the full verbose identity dump. relay_compact_and_resume is for clean-seam self-resume after a context advisory: pass workingState and optional ruledOut; Relay owns the objective envelope. relay_recall searches your own archived pre-compaction segments by keyword when a discarded detail is needed. relay_spawn_targets / relay_spawn_agent / relay_shutdown_agent only appear if your profile grants spawning (a live-children quota); when present, call relay_spawn_targets FIRST for the live host/provider/model matrix + your quota, then stand up long-living child agents and shut down your own \u2014 find them later with relay_find_agents spawnedBy:me. For replies, prefer relay_reply or relay_send_message; CLI fallback: agent-relay /reply <messageId> \"<your reply>\". If a delivered message says it was truncated, fetch the full body with: agent-relay get-message <messageId>. For command details, run: agent-relay /guide",
+    defaultTemplate: "[agent-relay] You are connected to Agent Relay, a real-time message bus between agents and users. When you receive a relay message: read it, do what it asks, and reply through the relay when a text response is needed. Use agent-relay /react <messageId> <emoji> for lightweight acknowledgement or approval. If Relay MCP tools are available, prefer relay_reply, relay_get_message, relay_get_thread, relay_send_message, relay_upload_artifact, relay_attach_artifact, relay_agent_status, relay_find_agents, relay_compact_and_resume, relay_recall, relay_spawn_agent, and relay_shutdown_agent. You never need to know or pass your own agent id \u2014 relay fills it from your token. Your identity is injected at startup as the `## Agent Relay Identity` block; use relay_whoami only for the full verbose identity dump. relay_compact_and_resume is for clean-seam self-resume after a context advisory: pass workingState and optional ruledOut; Relay owns the objective envelope. relay_recall searches your own archived pre-compaction segments by keyword when a discarded detail is needed. relay_spawn_targets / relay_spawn_agent / relay_shutdown_agent only appear if your profile grants spawning (a live-children quota); when present, call relay_spawn_targets FIRST for the live host/provider/model matrix + your quota, then stand up long-living child agents and shut down your own \u2014 find them later with relay_find_agents spawnedBy:me. For replies, prefer relay_reply or relay_send_message; CLI fallback: agent-relay /reply <messageId> \"<your reply>\". If a delivered message says it was truncated, fetch the full body with: agent-relay get-message <messageId>. For command details, run: agent-relay /guide",
     variables: []
   },
   {

package/src/adapters/claude-status-detectors.ts

@@ -1,5 +1,6 @@
 import { extractClaudeModelUnavailableMessage } from "agent-relay-sdk";
 import type { ProviderStatusUpdate } from "../adapter";
+import type { ClaudeProbeActivity } from "./claude-session-probe";
 import { buildRateLimitProviderState, parseClaudeRateLimitPane } from "../rate-limit";
 
 // LEGACY PANE-SCRAPE FALLBACK (#766/#768) — busy/idle is now driven primarily by
@@ -19,8 +20,8 @@ import { buildRateLimitProviderState, parseClaudeRateLimitPane } from "../rate-l
 //     (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).
+//     reads as ready/busy. The activity reconciler gates on the LAST N lines only (the
+//     live footer); direct raw predicates are still used by readiness/wait gates.
 // 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
@@ -43,6 +44,7 @@ export function claudePaneLooksReady(text: string): boolean {
 // (ctrl+o to expand)" truncation marker, the idle input box, or the persistent
 // "/btw … without interrupting Claude's current work" queue hint.
 const CLAUDE_BUSY_SPINNER_RE = /…\s*\((?:\d+h\s+)?(?:\d+m\s+)?\d+s\b/;
+const CLAUDE_ACTIVITY_FOOTER_LINES = 10;
 
 export function claudePaneIsBusy(text: string): boolean {
   // Claude Code <2.1.x rendered "(esc to interrupt)" in the spinner footer; 2.1.x
@@ -52,6 +54,17 @@ export function claudePaneIsBusy(text: string): boolean {
   return CLAUDE_BUSY_SPINNER_RE.test(text) || text.includes("esc to interrupt");
 }
 
+function claudeActivityFooter(text: string): string {
+  return text.split("\n").slice(-CLAUDE_ACTIVITY_FOOTER_LINES).join("\n");
+}
+
+export function claudeActivityFromProbeAndPane(probe: ClaudeProbeActivity | undefined, pane: string): ClaudeProbeActivity {
+  const footer = claudeActivityFooter(pane);
+  if (probe === "busy" || claudePaneIsBusy(footer)) return "busy";
+  if (claudePaneLooksReady(footer)) return "idle";
+  return probe === "idle" ? "unknown" : probe ?? "unknown";
+}
+
 // #286: detect the subscription usage-limit modal in the pane and turn it into the
 // provider-neutral rate-limit hold (idle + blocked) that the relay's resume sweep lifts
 // at reset. Mirrors claudeModelUnavailableStatus, but non-terminal (idle, not error).

package/src/adapters/claude.ts

@@ -12,7 +12,7 @@ import { assembleLaunch, materializeLaunchAssembly, sessionStatusLineSettingsArg
 import { claudeProviderMessageText } from "./claude-delivery";
 import { claudeProbeActivity, readManagedClaudeStatus, resolveClaudePid, type ClaudeProbeActivity } from "./claude-session-probe";
 import { evaluateClaudePromptGatePane, initialClaudePromptGatePaneState, type ClaudePromptGatePaneState } from "./claude-prompt-gates";
-import { claudePaneLooksReady, claudePaneIsBusy, claudeRateLimitStatus, claudeConnectionRetryStatus, claudeModelUnavailableStatus } from "./claude-status-detectors";
+import { claudeActivityFromProbeAndPane, claudePaneLooksReady, claudePaneIsBusy, claudeRateLimitStatus, claudeConnectionRetryStatus, claudeModelUnavailableStatus } from "./claude-status-detectors";
 import { ClaudeSessionCapture, collectClaudeTranscriptArchiveSegment, collectClaudeTranscriptSessionEvents } from "./claude-session-capture";
 import { CLAUDE_TMUX_READY_TIMEOUT_MS, captureTmuxPane, submitTextToTmux, tmuxEnvKeys, tmuxSessionName, tmuxSocketName, waitForClaudeInputReady, writeLauncherScript } from "./claude-tmux";
 import { claudePermissionPromptHandler } from "./claude-permission";
@@ -138,12 +138,11 @@ export class ClaudeAdapter implements ProviderAdapter {
       tmuxSocket: socket,
       configHomeEnv: (process.meta?.env as Record<string, string | undefined> | undefined)?.CLAUDE_CONFIG_DIR,
     });
-    if (managed?.reading.available) return claudeProbeActivity(managed.reading.status);
-    let pane: string;
-    try { pane = captureTmuxPane(session, socket); } catch { return "unknown"; }
-    if (claudePaneIsBusy(pane)) return "busy";
-    if (claudePaneLooksReady(pane)) return "idle";
-    return "unknown";
+    const probe = managed?.reading.available ? claudeProbeActivity(managed.reading.status) : undefined;
+    if (probe === "busy") return "busy";
+    let pane: string | undefined;
+    try { pane = captureTmuxPane(session, socket); } catch { return probe ?? "unknown"; }
+    return claudeActivityFromProbeAndPane(probe, pane);
   }
 
   // #769 P1-A: an independent pane-busy read, consumed ONLY as a veto on the busy
@@ -532,7 +531,7 @@ export { tmuxHasSession };
 
 // Re-export pane-status detectors (moved to ./claude-status-detectors) so existing
 // import paths (including claude.test.ts) continue to resolve via this module.
-export { claudePaneLooksReady, claudePaneIsBusy, claudeRateLimitStatus, claudeConnectionRetryStatus, claudeModelUnavailableStatus } from "./claude-status-detectors";
+export { claudeActivityFromProbeAndPane, claudePaneLooksReady, claudePaneIsBusy, claudeRateLimitStatus, claudeConnectionRetryStatus, claudeModelUnavailableStatus } from "./claude-status-detectors";
 
 // Shared shell-quoting; re-exported so `./claude` consumers + tests resolve it.
 export { shellQuote };

package/src/control-server.ts

@@ -323,34 +323,19 @@ async function handlePermissionRequest(
   const toolInput = isRecord(body.tool_input) ? body.tool_input : undefined;
   const approvalId = crypto.randomUUID();
 
-  // #1173: a PreToolUse "questions" prompt (the provider's AskUserQuestion) used to block this hook
-  // response for up to ~15 minutes waiting on the human — which meant the turn's reasoning that
-  // led to the question stayed buffered in the provider's process (never flushed to the transcript
-  // JSONL) for exactly as long as the human took to answer, so the dashboard could only ever
-  // show the question AFTER the reasoning that motivated it. Denying the tool call immediately
-  // is the ordinary, already-supported "rejected tool call" hook contract: the provider persists the
-  // buffered turn + a rejected tool_result right away, on its own, same as any other denied tool
-  // — so the settle-poll harvest below (previously doomed to time out for this one case) now
-  // finds the freshly-flushed anchor within moments. The human's eventual answer is delivered
-  // afterward as a normal injected prompt (see the `.then` below), not as this hook's response.
   const probeView = handler.buildView(approvalId, body, true);
-  if (body.hook_event_name === "PreToolUse" && probeView.kind === "questions" && handler.buildAnswerInjectionPrompt) {
-    const response = Response.json(handler.buildHookResponse(
-      { approvalId, decision: "deny", reason: PENDING_QUESTION_DENY_REASON },
-      body,
-    ));
-    void awaitPermissionDecision({
+  const answerInput = permissionRequestQuestionAnswers(body);
+  if (body.hook_event_name === "PermissionRequest" && probeView.kind === "questions" && answerInput) {
+    return Response.json(handler.buildHookResponse({ approvalId, decision: "answer", answers: answerInput }, body));
+  }
+  if (
+    (body.hook_event_name === "PreToolUse" || body.hook_event_name === "PermissionRequest")
+    && probeView.kind === "questions"
+    && handler.buildAnswerInjectionPrompt
+  ) {
+    return holdQuestionWithImmediateDeny({
       options, pendingPermissionRequests, handler, approvalId, body, occurredAt, transcriptPath, toolName, toolInput,
-      blockedReason: "pendingQuestion",
-      recommendedAction: "Answer the question in Agent Relay.",
-      suppressResolvedStatus: true,
-    }).then(({ decision, view }) => {
-      const text = decision.decision === "answer" && decision.answers && Object.keys(decision.answers).length
-        ? handler.buildAnswerInjectionPrompt!(view, decision, body)
-        : pendingQuestionDismissalPrompt(decision);
-      return options.onAnswerInject?.({ approvalId, text });
-    }).catch(() => {});
-    return response;
+    });
   }
 
   const { decision } = await awaitPermissionDecision({
@@ -361,6 +346,51 @@ async function handlePermissionRequest(
   return Response.json(handler.buildHookResponse(decision, body));
 }
 
+function permissionRequestQuestionAnswers(body: Record<string, unknown>): Record<string, string> | undefined {
+  if (body.hook_event_name !== "PermissionRequest") return undefined;
+  const toolInput = isRecord(body.tool_input) ? body.tool_input : {};
+  if (!isRecord(toolInput.answers)) return undefined;
+  const answers = Object.fromEntries(
+    Object.entries(toolInput.answers)
+      .filter((entry): entry is [string, string] => typeof entry[1] === "string" && entry[1].trim().length > 0),
+  );
+  return Object.keys(answers).length ? answers : undefined;
+}
+
+function holdQuestionWithImmediateDeny(params: {
+  options: ControlServerOptions;
+  pendingPermissionRequests: Map<string, PendingPermissionWaiter>;
+  handler: ProviderPermissionPromptHandler;
+  approvalId: string;
+  body: Record<string, unknown>;
+  occurredAt: number;
+  transcriptPath: string;
+  toolName?: string;
+  toolInput?: unknown;
+}): Response {
+  const { options, pendingPermissionRequests, handler, approvalId, body, occurredAt, transcriptPath, toolName, toolInput } = params;
+  // #1173/#1194: a fresh "questions" prompt (the provider's AskUserQuestion) must not block
+  // its own hook response while the human answers. Denying immediately makes the provider persist
+  // the buffered turn plus rejected tool_result, so the pre-flush harvest below can publish the
+  // reasoning before the dashboard form appears. The eventual answer is injected as a normal prompt.
+  const response = Response.json(handler.buildHookResponse(
+    { approvalId, decision: "deny", reason: PENDING_QUESTION_DENY_REASON },
+    body,
+  ));
+  void awaitPermissionDecision({
+    options, pendingPermissionRequests, handler, approvalId, body, occurredAt, transcriptPath, toolName, toolInput,
+    blockedReason: "pendingQuestion",
+    recommendedAction: "Answer the question in Agent Relay.",
+    suppressResolvedStatus: true,
+  }).then(({ decision, view }) => {
+    const text = decision.decision === "answer" && decision.answers && Object.keys(decision.answers).length
+      ? handler.buildAnswerInjectionPrompt!(view, decision, body)
+      : pendingQuestionDismissalPrompt(decision);
+    return options.onAnswerInject?.({ approvalId, text });
+  }).catch(() => {});
+  return response;
+}
+
 // Shared by both the traditional block-until-decided flow and the #1173 deny-then-hold flow:
 // pre-flushes narrative (#435/#1116), builds+persists the interactive view, publishes the
 // blocked providerState, and resolves once a decision arrives (or the wait times out). Callers

package/src/profile-home.ts

@@ -140,7 +140,7 @@ function seedClaudeConfigIfMissing(claudeHome: string, config: RunnerSpawnConfig
 
 function readHostClaudeConfig(): Record<string, unknown> | undefined {
   try {
-    const hostPath = join(homedir(), ".claude.json");
+    const hostPath = join(providerSourceHome("claude"), ".claude.json");
     if (!existsSync(hostPath)) return undefined;
     return JSON.parse(readFileSync(hostPath, "utf8")) as Record<string, unknown>;
   } catch {