teapot-coding-agent 0.29.17 → 0.29.18

package/dist/agent/agent.js

@@ -10,7 +10,7 @@ import { promises as fs, existsSync, readFileSync, readdirSync, statSync } from
 import path from "node:path";
 import { EventLog, readEvents } from "../log/events.js";
 import { chat, chatStream } from "./llm.js";
-import { executeTool, toolSpecs, currentSkills, safeJoin, isContextOverflow, hasRunningBgShells } from "./tools.js";
+import { executeTool, toolSpecs, currentSkills, safeJoin, isContextOverflow, hasRunningBgShells, killBgShells } from "./tools.js";
 import { effortChoosable } from "./llm.js";
 import { skillRootsFingerprint, foreignSkillRoots } from "./skills.js";
 import { bus } from "../bus.js";
@@ -685,6 +685,15 @@ export class Agent {
             this.pendingPrompts.length > 0 ||
             this.compactPhase !== "" ||
             this.retrySleeping ||
+            // #158: a BACKGROUND job is work in flight, and `stop()` now ends it. The
+            // round that started it deliberately ended (the loop breaks rather than
+            // nagging — see hasRunningBgShells in loop()), so the status reads "idle"
+            // while a dev server / watcher / long build is still running. Without this
+            // the UI offered "▶ start" and there was no stop control for the one thing
+            // the operator most needs to stop. Note this is checked AFTER stopRequested:
+            // once stopped, the jobs are already dead, so `live` must go false or the
+            // button would offer stop forever (#133).
+            hasRunningBgShells(this.toolCtx.cwd) ||
             // ask_user: the loop is parked on the operator's answer — stoppable work
             // in flight, not idleness. Without this the UI offered "start", and
             // pressing it cleared awaitingUser, silently discarding the open question
@@ -1677,6 +1686,13 @@ export class Agent {
         // without aborting toolCtx.signal the park ran to its full timeout
         this.abort?.abort();
         this.toolAbort.abort();
+        // #158: and the BACKGROUND shells, which neither signal above reaches. A job
+        // spawned with `background: true` is `detached` and its spawning tool returned
+        // long ago, so `toolAbort` has nothing on the stack to abort: the agent parked
+        // in its "wait for the job" idle state, the job kept running, and `stop()` —
+        // the control the operator pressed for exactly this — ended nothing. MEASURED:
+        // `sleep 300` outlived both stop() and dispose().
+        killBgShells(this.toolCtx.cwd);
         this.wake?.();
         // #86: the status write is DEFERRED onto the run chain, so it can land
         // AFTER a start() that followed it. Observed: `idle -> stopped -> running ->
@@ -3666,6 +3682,29 @@ export class Agent {
         onDelta
             ? (s) => onDelta(s.text)
             : undefined);
+        // Record what the summarizer itself cost, especially `cachedInputTokens`.
+        //
+        // This is the ONLY way to observe the property the whole replay-the-prefix shape
+        // exists for: the summarizer is supposed to be served from the provider's KV
+        // cache because its input is a byte-identical prefix of the request the
+        // conversation just made. Without this line the claim was UNVERIFIABLE — a
+        // compaction event reported tokensBefore/tokensAfter (the CONVERSATION's size)
+        // and nothing at all about the call that produced the summary, so a shape that
+        // silently stopped sharing the prefix would look identical in the log.
+        //
+        // MEASURED on the session that prompted this: the 12k-line history had 2342
+        // usage events, none of them the summarizer's.
+        if (res.usage) {
+            const u = res.usage;
+            this.stats.inputTokens += u.inputTokens ?? 0;
+            this.stats.outputTokens += u.outputTokens ?? 0;
+            this.stats.cachedInputTokens += u.cachedInputTokens ?? 0;
+            await this.log.append("usage", this.currentSession, this.currentBranch, {
+                ...u,
+                // distinguishes it from a conversation turn when reading the log back
+                kind: "summarize",
+            });
+        }
         return this.formatCompactSummary(res.message.content ?? "");
     }
     /** strip the <analysis> scratchpad — it improves thinking but must not

package/dist/agent/tools.js

@@ -301,6 +301,34 @@ export function hasRunningBgShells(cwd) {
     }
     return false;
 }
+/**
+ * Kill every background shell for one workspace and forget the jobs (#158).
+ *
+ * `stop()` left these running. They are spawned `detached: true` for group-killing
+ * and are NOT covered by `toolAbort` (that signal only reaches a tool that is
+ * still on the stack), so pressing stop on an agent that had parked its round on
+ * a background job killed nothing: the loop sat in its deliberate "wait for the
+ * job" idle state while the job kept burning CPU, and no control could end it.
+ *
+ * MEASURED: `sleep 300` outlived both `stop()` and `dispose()`.
+ *
+ * Returns how many were actually killed, so a caller can report it.
+ */
+export function killBgShells(cwd) {
+    const m = bgShellsByWs.get(cwd);
+    if (!m)
+        return 0;
+    let killed = 0;
+    for (const [id, sh] of m) {
+        if (!sh.exited) {
+            describeKill(killTree(sh.child));
+            killed++;
+        }
+        bgReadCursors.delete(id);
+    }
+    m.clear();
+    return killed;
+}
 /** Run a command in its own process group; kill the whole group on timeout. */
 function runShell(cmd, ctx, timeoutMs) {
     return new Promise((resolve) => {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "teapot-coding-agent",
-  "version": "0.29.17",
+  "version": "0.29.18",
   "description": "A lightweight, always-on multi-agent harness for AI coding agents",
   "type": "module",
   "license": "AGPL-3.0-or-later",