@stage-labs/metro 0.1.0-beta.327 → 0.1.0-beta.328

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stage-labs/metro",
-  "version": "0.1.0-beta.327",
+  "version": "0.1.0-beta.328",
   "description": "The metro command line: run your agent on this machine, relaying its chat channels and connectors to Claude Code.",
   "license": "MIT",
   "repository": {

package/runtime/node_modules/@metro-labs/daemon/src/claude/memory.ts

@@ -28,5 +28,5 @@ export function inSessionScope(command: [string, string[]], host: ScopeHost = re
   const asAgent = file === 'sudo' && args[0] === '-n' && args[1] === '-u' && args[3] === '--';
   if (!host.metro || !host.systemd || !asAgent) return command;
   const limit = sessionMemoryLimit(host.total);
-  return helperArgv(['as-agent-scope', String(limit), String(Math.floor(limit * 0.9)), ...args.slice(4)]);
+  return helperArgv(['as-agent-scope', String(limit), String(limit), ...args.slice(4)]);
 }

package/runtime/node_modules/@metro-labs/daemon/src/claude/setup.ts

@@ -41,7 +41,7 @@ Keep the change to what the task asks for. A pre-existing bug, a performance pro
 
 Never call AskUserQuestion or ExitPlanMode. They are blocked by policy, because nobody is watching the terminal around the clock and a blocking prompt stalls indefinitely. When you hit an ambiguity: do everything that does not depend on it, then choose the most reasonable interpretation, state the assumption explicitly in your report, and continue.
 
-Your final message is a report to the orchestrator, not a message to a human. It gets relayed onward, so make it self-contained and concise. Lead with the outcome. Include concrete evidence, command output, file paths with line numbers, test results, for anything you claim to have verified. Report failures faithfully: if a test failed, say so and show the output; if you skipped a step, say that too.
+Your final message is a report to the orchestrator, not a message to a human. It stays in the orchestrator's context for the rest of the conversation and makes every later answer slower, so keep it short: about 150 words at most, self-contained. Lead with the outcome, then the key evidence for what you claim to have verified, such as the test result or the file you changed. Put anything longer, such as full logs, tables or detailed findings, in a file under ~/work and give its path. Report failures faithfully: if a test failed, say so with its key line of output; if you skipped a step, say that too.
 `;
 
 export interface SetupDeps {
@@ -129,7 +129,10 @@ export function setSystemPrompt(text: string, agents = agentsDir()): void {
   writeAtomic(promptPath(agents), `${trimmed}\n`, 0o600);
 }
 
-const PRIOR_WORKER: ReadonlySet<string> = new Set(['08a0cd8710df285d6512245246bb8b658b7cfc7f3e89490514ee5376779eb2ef']);
+const PRIOR_WORKER: ReadonlySet<string> = new Set([
+  '08a0cd8710df285d6512245246bb8b658b7cfc7f3e89490514ee5376779eb2ef',
+  '9db999a7bb7cf1d6c12a679520e21c78d038999d5454d7abbc6b87a87c136c54',
+]);
 const PRIOR_METRO: ReadonlySet<string> = new Set([
   '0b3d122ed95beff09d579cf912cd4238e1db524c41fce4b314de57d6ff5908ad',
   '36f4fb57f231a119d717b5e3d6ccb654f94960fcca708683f376679aead8df25',
@@ -142,6 +145,7 @@ const PRIOR_METRO: ReadonlySet<string> = new Set([
   '256c711a6175266c216e32c88b7190d0f98f5e6b702abe03220b422aba426574',
   '9e164a956437d60d30ea1984115a17d303940d191f5efed04954f244fc89d48f',
   '8b6d7c5a9bc8d02d5e3ae5e5931b36aba9f3d0eb7fd38b371294c2abaa0c94ee',
+  '8b8f21cb1b7608540d8b5ed9557e1644b03ce4809a7f320ba134f28129f08ddb',
 ]);
 const PRIOR_STAGE: ReadonlySet<string> = new Set([
   'dd7d30474765de3865ddc87c1cf3220161ded09502027cdff18597e5df0aab43',

package/runtime/runtime.json

@@ -1,3 +1,3 @@
 {
-  "version": "0.1.0-beta.327"
+  "version": "0.1.0-beta.328"
 }

package/runtime/sdk-runner/runner.json

@@ -1,5 +1,5 @@
 {
-  "version": "0.1.0-beta.327",
+  "version": "0.1.0-beta.328",
   "dependencies": {
     "@anthropic-ai/claude-agent-sdk": "0.3.287",
     "@modelcontextprotocol/sdk": "^1.29.0",

package/runtime/sdk-runner/src/app.ts

@@ -63,7 +63,7 @@ async function connected(cfg: RunnerConfig, hooks: AppHooks, release: () => Prom
     tools = metroTools(link);
     const resume = store.resumable();
     log.info({ resume, model: cfg.model, mode: cfg.permissionMode }, 'sdk-runner: starting the Agent SDK session');
-    const approvals = approvalsThrough(link, (id, ask) => { hooks.activity?.approval(id, ask); }, (tool, input) => runner.calls.approval(tool, input));
+    const approvals = approvalsThrough(link, (id, ask) => { hooks.activity?.approval(id, ask); runner.approval(id, ask); }, (tool, input) => runner.calls.approval(tool, input));
     runner.start(runnerOptions(cfg, tools, approvals, resume, hooks.env));
     return { runner, link, done: runner.run(hooks.observe), stop };
   } catch (err) {

package/runtime/sdk-runner/src/runner.ts

@@ -7,7 +7,6 @@ import { isRecord } from '@metro-labs/core/is-record';
 import type { AutomationStore } from '@metro-labs/core/automation-store';
 import type { Activity } from './activity.js';
 import { Automation } from './automation.js';
-import { ActivityTasks } from './activity-tasks.js';
 import { RunnerCalls } from './calls.js';
 import { channelText, type ChannelEvent } from './channel-text.js';
 import type { RunnerConfig } from './config.js';
@@ -21,7 +20,8 @@ import { METRO_SERVER, type MetroTools } from './tool-proxy.js';
 import { debugFailure, failureSummary, SessionError, StartupDiagnostics } from './errors.js';
 
 const STARTUP_WAIT = 'CLAUDE_CODE_MCP_STARTUP_WAIT_MS';
-export const COMPACT_AT = 120_000;
+export const COMPACT_AT = 80_000;
+const URGENT_FACTOR = 1.5;
 export const COMPACT_IDLE_MS = 60_000;
 
 const MODEL_PINS: ReadonlySet<string> = new Set(['ANTHROPIC_MODEL', 'CLAUDE_CODE_SUBAGENT_MODEL']);
@@ -97,7 +97,6 @@ export class Runner {
   private readonly recovery: TaskRecovery;
   private readonly watch: SessionWatch;
   private readonly diagnostics = new StartupDiagnostics();
-  private readonly tasks = new ActivityTasks(() => undefined);
   private readonly skillWork = new SkillWork();
   private compactTimer: ReturnType<typeof setTimeout> | null = null;
   private session: Query | null = null;
@@ -108,11 +107,12 @@ export class Runner {
   private model: string | null = null;
   private switching: Promise<void> = Promise.resolve();
   private ended = false;
+  private readonly mainAsks = new Set<string>();
   private readonly interrupted: number;
   private interruptionReported = false;
 
   constructor(private readonly parts: RunnerParts) {
-    this.watch = new SessionWatch(parts.readOnly);
+    this.watch = new SessionWatch(parts.readOnly, () => this.mainAsks.size > 0);
     this.inbox = new Inbox((unanswered) => { this.keep(unanswered); }, {
       ready: () => { parts.skills?.check(); return this.watch.admitting && parts.skills?.busy !== true; },
       dispatch: (message) => { this.automation?.dispatched(message.uuid); this.watch.dispatched(message.uuid); },
@@ -167,6 +167,12 @@ export class Runner {
     return this.session;
   }
 
+  approval(id: string, ask: { worker: string | null } | null): void {
+    if (ask === null) this.mainAsks.delete(id);
+    else if (ask.worker === null) this.mainAsks.add(id);
+    this.inbox.notify();
+  }
+
   switchModel(model: string | null): Promise<void> {
     this.queueModel(model);
     return this.switching;
@@ -178,7 +184,6 @@ export class Runner {
       for await (const message of this.session) {
         if (this.ended) break;
         this.cancelCompact();
-        if (message.type === 'system') this.tasks.observe(message);
         const m: Record<string, unknown> = { ...message };
         this.recovery.observe(m);
         this.automation?.observe(m);
@@ -186,8 +191,8 @@ export class Runner {
         this.observeWork(m);
         this.note(m);
         this.parts.activity?.observe(message);
-        this.inbox.notify();
         this.maybeCompact();
+        this.inbox.notify();
         observe?.(message);
       }
     } catch (err) {
@@ -318,14 +323,19 @@ export class Runner {
     return this.ended || this.parts.skills?.busy === true || this.compactAsked !== null || !this.watch.safe || !this.watch.admitting;
   }
 
+  private get compactUrgent(): boolean {
+    return this.watch.context >= (this.parts.compactAt ?? COMPACT_AT) * URGENT_FACTOR;
+  }
+
   private compactReady(resuming = false): boolean {
-    const busy = this.maintenanceBusy || (!resuming && (this.calls.live || this.inbox.pending > 0 || this.tasks.running > 0));
+    const waiting = this.inbox.pending > 0 && !this.compactUrgent;
+    const busy = this.maintenanceBusy || (!resuming && (this.calls.live || waiting));
     return compactDue({ context: this.watch.context, floor: this.compactFloor, limit: this.parts.compactAt ?? COMPACT_AT, busy });
   }
 
   private maybeCompact(resuming = false): void {
     if (!this.compactReady(resuming)) return;
-    if (resuming) { this.askCompact(true); return; }
+    if (resuming || this.compactUrgent) { this.askCompact(resuming); return; }
     this.compactTimer = setTimeout(() => {
       this.compactTimer = null;
       if (this.compactReady()) this.askCompact(false);
@@ -335,7 +345,7 @@ export class Runner {
 
   private askCompact(resuming: boolean): void {
     this.compactAsked = this.inbox.compact();
-    log.info({ context: this.watch.context, resuming }, 'sdk-runner: compacting at a safe idle boundary');
+    log.info({ context: this.watch.context, resuming, urgent: this.compactUrgent, waiting: this.inbox.pending }, 'sdk-runner: compacting at a safe boundary');
   }
 
   private cancelCompact(): void {

package/runtime/sdk-runner/src/session-watch.ts

@@ -35,7 +35,7 @@ export class SessionWatch {
   private lastContext = 0;
   private streamUsage: Record<string, unknown> = {};
 
-  constructor(private readonly readOnly: (tool: string) => boolean) {}
+  constructor(private readonly readOnly: (tool: string) => boolean, private readonly mainAsking: () => boolean = () => true) {}
 
   get compacting(): boolean {
     return this.compactingNow;
@@ -52,7 +52,7 @@ export class SessionWatch {
   get safe(): boolean { return !this.busy && !this.writing && !this.compacting; }
 
   get admitting(): boolean {
-    return this.awaitingInput === null && !this.requiresAction && !this.compacting;
+    return this.awaitingInput === null && !(this.requiresAction && this.mainAsking()) && !this.compacting;
   }
 
   dispatched(uuid?: string): void {

package/skills/METRO.md

@@ -22,7 +22,7 @@ You are the main thread of a Claude Code session that is connected to chat throu
 1. `tool_search` finds a tool and returns its full definition: `select:notion.search` for an exact name, words (`notion create page`) to rank, or `+linear issue` to require `linear` in the name. Each name is `<connector>.<tool>`.
 2. `tool_call` runs it, with the name exactly as found and arguments that match its schema. After `tool_search` with `select:notion.search` returned `notion.search`: `{"calls": [{"tool": "notion.search", "arguments": {"query": "roadmap"}}]}`.
 
-Put up to 10 independent calls in one `tool_call`; they run in parallel and come back in order, each under `[n] <tool>`. `tool_search` may run on this thread, but never call `tool_call` here: delegate it to a worker, as all other work. When a call answers `needs_approval`, the owner is asked in chat and on the page: tell the person what is waiting, then wait for the channel message `Approved: <tool> (<id>)` or `Denied: <tool> (<id>)`. After an approval, have a worker repeat the same call with `"approval": "<id>"` added to it.
+Put up to 10 independent calls in one `tool_call`; they run in parallel and come back in order, each under `[n] <tool>`. Both steps run in a worker, never on this thread: `tool_search` answers carry full tool definitions, and anything on this thread stays in its context and slows every later answer. When a call answers `needs_approval`, the owner is asked in chat and on the page: tell the person what is waiting, then wait for the channel message `Approved: <tool> (<id>)` or `Denied: <tool> (<id>)`. After an approval, have a worker repeat the same call with `"approval": "<id>"` added to it.
 
 **Finish what was asked.** You work on your own. Nobody is watching a screen, and a question sent to chat may go unanswered for hours, so asking "Shall I?" about work already requested simply stops it. Anything reversible that follows from the request, do. Stop only for a destructive action or a genuine change of scope. Before you end a turn, read your own last paragraph: if it is a plan, a question, a list of next steps or a promise ("I'll…", "Next…"), then the work is not done, so do it now, retries and missing information included. End the turn when the task is finished, or when you are truly blocked on something only the person can give you.
 
@@ -30,6 +30,6 @@ The exception is a person describing a problem, asking a question or thinking ou
 
 **The whole request is the deliverable.** Do not quietly make it smaller, larger or different. Read an ambiguity the way a careful colleague would, make the ordinary judgement calls yourself, and name the assumption in your reply. If one part turns out to be blocked, finish every other part and say plainly what you left out and why: scaling the work down is the person's call, not yours. Something else you notice on the way is a suggestion at the end of the reply, not extra work to do now.
 
-**Report like a relay.** A worker's report is the only evidence you have, so pass on what it verified and say plainly what it did not. Short messages on chat: lead with the outcome, one idea per sentence, no tool names, no file paths unless the person needs them. Send a file out with `create_upload` (the curl line it hands back runs in a worker) and then `send` with the upload id.
+**Report like a relay.** A worker's report is the only evidence you have, so pass on what it verified and say plainly what it did not. Keep this thread small: ask workers for short reports, and give them file paths and details rather than pasting long text into their prompts. Short messages on chat: lead with the outcome, one idea per sentence, no tool names, no file paths unless the person needs them. Send a file out with `create_upload` (the curl line it hands back runs in a worker) and then `send` with the upload id.
 
 **Keep going.** Long tasks run as background agents and loops; check in over chat when something lands rather than going quiet. The session is restarted by the daemon if it exits, so an interrupted task is resumed from chat, not from memory.