@coreplane/switchboard 1.218.0 → 1.220.0

package/dist/assets/Dockerfile

@@ -71,6 +71,22 @@ RUN apt-get update \
 COPY --from=meat /go/bin/meat /usr/local/bin/meat
 RUN command -v meat >/dev/null && meat -h >/dev/null 2>&1
 
+# pi (docs/reference/specs/harness-pi.md item 12): the harness for a preset
+# without a workspace (general, research, conductor) runs pi as a child of
+# the bot itself, since such a run has no execution container. The same layer
+# as the execution images', on this image's Node (the same 24 tag), installed
+# globally as root before the switch to the bot's user, so `switchboard` finds
+# it on PATH and cannot alter it; at an EXACT pin held equal across the three
+# images by src/deploy/imagePiHarness.test.ts (the pnpm lesson: a floating tag
+# would move the harness's protocol with the build date). Proven by the layer,
+# so the BUILD fails, not a run. Dark until a deployment sets
+# `harness: { general: pi }`: nothing here starts pi on its own, and a run's
+# pi on this host holds none of pi's own shell or file tools.
+RUN npm install -g @earendil-works/pi-coding-agent@0.85.1 \
+  && npm cache clean --force \
+  && pi --version | grep -qx '0.85.1' \
+  && pi --help | grep -q -- '--mode <mode>'
+
 # Non-root user; agents run bash with this user's (container-scoped) permissions.
 RUN useradd -m -u 1001 switchboard
 WORKDIR /app

package/dist/assets/config/config.example.yaml

@@ -320,8 +320,10 @@ workspaceDir: ./workspaces
 # with the run's model-proxy bearer as its only key, every event bridged onto
 # the run's own record, our tools and tool rules as its extension. Every
 # preset declares `native` until the pi series moves it; a deployment moves a
-# preset early here. Only a preset with a workspace (coding, review, explore)
-# can run on pi. A preset the registry does not know, or a value that is
+# preset early here. A preset with a workspace (coding, review, explore) runs
+# pi in the run's container; a preset without one (general, research,
+# conductor) runs it as a child of the bot, with none of pi's own shell or
+# file tools. A preset the registry does not know, or a value that is
 # neither harness, fails the load by name.
 # harness:
 #   coding: pi

package/dist/assets/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "switchboard",
-  "version": "1.218.0",
+  "version": "1.220.0",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "switchboard",
-      "version": "1.218.0",
+      "version": "1.220.0",
       "license": "Apache-2.0",
       "workspaces": [
         "web",
@@ -16710,7 +16710,7 @@
     },
     "packages/switchboard": {
       "name": "@coreplane/switchboard",
-      "version": "1.218.0",
+      "version": "1.220.0",
       "license": "Apache-2.0",
       "dependencies": {
         "@anthropic-ai/sdk": "^0.124.0",

package/dist/assets/package.json

@@ -1,6 +1,6 @@
 {
   "name": "switchboard",
-  "version": "1.218.0",
+  "version": "1.220.0",
   "private": true,
   "description": "Mention it in Slack and an agent reviews the PR, ships the fix, or answers the question — on the model you choose, with its tools running where you decide.",
   "license": "Apache-2.0",

package/dist/assets/source.json

@@ -1,5 +1,5 @@
 {
-  "version": "1.218.0",
-  "commit": "2931369158bcc02d8e2cfd80eb661bae017188bd",
-  "builtAt": "2026-09-15T00:34:28.652Z"
+  "version": "1.220.0",
+  "commit": "cf86e5df3d3256037bf4925a9e13de0b61483df3",
+  "builtAt": "2026-09-15T02:39:37.781Z"
 }

package/dist/assets/src/agents/registry.ts

@@ -121,8 +121,9 @@ export interface AgentDef {
   residentSystem?: string;
   /** Which loop drives the preset's runs (`HARNESSES`): the native loop
    *  unless declared, and whatever a deployment's `harness.<preset>` says
-   *  over that. Only a preset with a workspace can run on pi — pi is a process
-   *  in the run's execution container. */
+   *  over that. A preset with a workspace runs pi in the run's execution
+   *  container; a preset without one (machine class `none`) runs it as a
+   *  child of the bot, with none of pi's own tools. */
   harness?: Harness;
 }
 
@@ -435,6 +436,8 @@ TIME. Your budget is up to two hours — less when a boundary or the request's \
 
 READ-ONLY: NEVER open a pull request, and never commit or push — no branch, no \`gh pr create\`, no PR or issue write of any kind. You hold a read credential and your job is to find out, not to change. If the investigation shows a change is needed, say exactly what and where in your write-up and point the user at \`agent:coding\`.
 
+You cannot attach or post files: your whole answer is text. Never say a file is attached or below — name its path in the workspace and describe it (what it shows, its size) instead; a person who needs the file itself asks \`agent:coding\`, which can attach.
+
 Maintain the user-facing status card with the update_status tool: post your plan as a checklist (○ pending) once you have it, and update items as they start (✱) and finish (✓ — only after they actually happened). Items are short outcomes ("Clone and install", "Time the full suite"), never commands.
 
 ${NOTEPAD}
@@ -600,7 +603,7 @@ const WORK_PRESETS = {
   explore: {
     name: "explore",
     description:
-      "Long, read-only investigation of a repository in a cold sandbox: runs builds, suites and pipelines, searches the web, and reports a claim table with commands and numbers. Never opens a PR. Up to two hours.",
+      "Read-only investigation of a repository in a cold sandbox: runs builds, suites and pipelines, searches the web, and reports a claim table with commands and numbers. Never opens a PR; cannot attach or post files.",
     system: EXPLORE_SYSTEM,
     toolset: "explore",
     // Always a cold per-thread sandbox with the checkout, never the resident a

package/dist/assets/src/core/runEvents.ts

@@ -71,6 +71,11 @@ export type RunNoteKind =
    *  executor's bounded wait (docs/reference/specs/execution.md item 14). Capacity, not a
    *  dead sandbox: the run goes on and the model is told to retry or finish. */
   | "fleet_busy"
+  /** The sandbox restarted under the run and came back (docs/reference/specs/
+   *  resident-repos.md item 65): the executor waited for the resident's wake
+   *  and re-attached; the interrupted call was settled with a synthetic
+   *  result (run-loop.md item 19) and the run goes on. */
+  | "sandbox_restarted"
   | "stop_requested"
   | "stopped"
   /** Setup spans the request's stream sink had to drop before this run was
@@ -154,6 +159,7 @@ export const RUN_NOTE_KINDS = [
   "turn_budget_exhausted",
   "sandbox_dead",
   "fleet_busy",
+  "sandbox_restarted",
   "stop_requested",
   "stopped",
   "spans_dropped",

package/dist/assets/src/core/runFriction.ts

@@ -38,7 +38,8 @@ export type FrictionCategory =
   | "setup_install"
   | "wrap_up"
   | "budget_hit"
-  | "infra_failure";
+  | "infra_failure"
+  | "unkept_promise";
 
 export const FRICTION_CATEGORIES: readonly FrictionCategory[] = [
   "slow_tool",
@@ -49,6 +50,7 @@ export const FRICTION_CATEGORIES: readonly FrictionCategory[] = [
   "wrap_up",
   "budget_hit",
   "infra_failure",
+  "unkept_promise",
 ];
 
 /** Human labels — the verdict line, the report's table and its finding lines. */
@@ -61,6 +63,7 @@ export const CATEGORY_LABEL: Record<FrictionCategory, string> = {
   wrap_up: "agent wind-down",
   budget_hit: "budget hits",
   infra_failure: "infra failures",
+  unkept_promise: "unkept promises",
 };
 
 /** What a category's time is a share OF: tool time for the categories whose
@@ -75,8 +78,17 @@ export const DENOMINATOR_OF: Record<FrictionCategory, "tool" | "run"> = {
   wrap_up: "run",
   budget_hit: "run",
   infra_failure: "run",
+  unkept_promise: "run",
 };
 
+/** The ways a reply tells the person a file is attached — the promise `unkept_promise` holds a run to.
+ *  One deliberate expression, case-insensitive: "attached below/here/above", "see (the) attached",
+ *  "I've/I have attached", "is/are attached", "attachment(s) (is/are) below/here". A path such as
+ *  `./attachments/1-clip.mp4` or the bare word "attachments/" names a place, not a promise, and
+ *  does not match; nor does prose about attaching in an `assistant` turn — only the `answer` counts. */
+export const ATTACHMENT_PROMISE =
+  /\b(?:attached (?:below|here|above)|see (?:the )?attached|i(?:'ve| have) attached|(?:is|are) attached\b|attachments? (?:(?:is|are) )?(?:below|here))/i;
+
 export type FrictionSeverity = "low" | "medium" | "high";
 
 export interface FrictionFinding {
@@ -382,6 +394,9 @@ export function analyzeRunFriction(events: readonly RunEvent[], opts: FrictionOp
   let sideFactEvents = 0; // skill_use / review_artifact / pr_description / pr_opened / review_posted / ship_round / route: facts about the run, not steps
   let spanEvents = 0; // span_start / span_end (docs/reference/specs/tracing.md): timing records, not steps
   let wrapUp: { index: number; at?: number } | undefined;
+  // The reply and the files the run sent out: what `unkept_promise` compares after the loop.
+  let answer: { index: number; text: string } | undefined;
+  let outboundArtifacts = 0;
   events.forEach((ev, index) => {
     flushTurns(index);
     if (isSpanRecord(ev)) {
@@ -390,8 +405,10 @@ export function analyzeRunFriction(events: readonly RunEvent[], opts: FrictionOp
     }
     if (isNarrative(ev)) {
       narrativeEvents++; // the narrative and `run_meta` are neither steps nor findings
+      if (ev.type === "answer") answer = { index, text: ev.text };
       return;
     }
+    if (ev.type === "artifact" && ev.direction === "out") outboundArtifacts++;
     // Side facts about the run, not steps: skill_use rides beside a use_skill
     // call that already produced its own tool pair, and artifact beside the
     // attach_file call (or the dispatcher's staging) that moved the file;
@@ -545,6 +562,17 @@ export function analyzeRunFriction(events: readonly RunEvent[], opts: FrictionOp
           eventIndex: index,
         });
         return;
+      case "sandbox_restarted":
+        // The container rolled under the run and came back: the run went on,
+        // but the minutes spent waiting for the wake are friction the deploy
+        // window owns.
+        findings.push({
+          category: "infra_failure",
+          severity: "medium",
+          summary: `sandbox restarted: ${ev.summary}`,
+          eventIndex: index,
+        });
+        return;
     }
   });
   // A slow turn whose produced event is past the stream (a turn that ended the run).
@@ -582,6 +610,24 @@ export function analyzeRunFriction(events: readonly RunEvent[], opts: FrictionOp
       f.interval = { start: at, end: Math.max(at, windowEnd) };
     }
   }
+  // The reply promised a file the run never produced: no tool failed (a preset without
+  // attach_file makes no call), so the answer's own words are the only witness. Once per
+  // run, anchored to the answer, no extent; the line that made the promise is quoted.
+  if (answer !== undefined && outboundArtifacts === 0) {
+    const promised = ATTACHMENT_PROMISE.exec(answer.text);
+    if (promised) {
+      const line =
+        answer.text.slice(0, promised.index).split("\n").pop()! + answer.text.slice(promised.index).split("\n")[0]!;
+      const trimmed = line.trim();
+      const quoted = trimmed.length > 120 ? `…${trimmed.slice(trimmed.length - 119)}` : trimmed;
+      findings.push({
+        category: "unkept_promise",
+        severity: "low",
+        summary: `unkept promise: the reply says a file is attached but the run produced none — "${quoted}"`,
+        eventIndex: answer.index,
+      });
+    }
+  }
 
   // ---- totals ----------------------------------------------------------------
   let toolTimeMs: number | undefined;

package/dist/assets/src/core/runRecord.ts

@@ -158,6 +158,39 @@ export interface RunRecord {
    *  on records written before the session log existed and on runs without
    *  a conversation of their own. */
   session?: RunSession;
+  /** The pull request the run's post-step opened or edited (item 2): the last
+   *  `pr_opened` event, folded at the assembly. Present only on a run that
+   *  reached one. A follow-up in the thread continues on its head branch
+   *  (docs/reference/specs/resident-repos.md item 29) — read off the record,
+   *  never off the reply's text. */
+  pr?: RunPullRequest;
+}
+
+/** A pull request as the record names it (item 2): its number and its GitHub
+ *  URL, both from the `pr_opened` event the coding post-step published. */
+export interface RunPullRequest {
+  number: number;
+  url: string;
+}
+
+/** The pull request a run's events say it opened or edited — the last
+ *  `pr_opened` wins, as an edit after an open names the same PR — or nothing. */
+export function prOfEvents(events: readonly RunEvent[]): RunPullRequest | undefined {
+  let pr: RunPullRequest | undefined;
+  for (const e of events) if (e.type === "pr_opened") pr = { number: e.number, url: e.url };
+  return pr;
+}
+
+function isRunPullRequestShape(v: unknown): v is RunPullRequest {
+  if (typeof v !== "object" || v === null) return false;
+  const p = v as Record<string, unknown>;
+  return (
+    typeof p.number === "number" &&
+    Number.isInteger(p.number) &&
+    p.number > 0 &&
+    typeof p.url === "string" &&
+    p.url.length > 0
+  );
 }
 
 /** The three places a run's conversation can start (item 52): the thread's
@@ -574,6 +607,8 @@ export function isRunRecord(v: unknown): v is RunRecord {
   if (r.dispositions !== undefined && !isFindingDispositionsShape(r.dispositions)) return false;
   if (r.reviewPost !== undefined && !isReviewPostShape(r.reviewPost)) return false;
   if (r.profile !== undefined && !isRunProfileRecord(r.profile)) return false;
+  // The pull request the run reached (item 2): a positive integer and a URL, or absent.
+  if (r.pr !== undefined && !isRunPullRequestShape(r.pr)) return false;
   // A parent is named by a run id (item 46): the same shape as the record's own.
   if (r.parentRunId !== undefined && (typeof r.parentRunId !== "string" || !RUN_ID_PATTERN.test(r.parentRunId)))
     return false;