pi-bedrouter 0.1.0 → 0.2.1

package/README.md

@@ -39,6 +39,7 @@ Bedrouter needs AWS credentials that can call Bedrock; see its README for the `a
 | `debug` | `false` | Start the server with `BEDROUTER_DEBUG=1` (per-request trace in `server.log`) |
 | `footer` | `true` | Show the routing status line in Pi's footer |
 | `providerName` | `"bedrouter"` | Provider name registered in Pi |
+| `stopOnExit` | `"if-started-here"` | What happens to the server when Pi **quits** (`/reload` and session switches never stop it). `if-started-here`: stop it if this session started it and no other client sent a request in the last 5 minutes; `always`: stop it whenever this session started it; `never`: leave it running. Pi tears down its UI before extensions are told about the quit, so this cannot be a prompt; the policy is shown when the server is started and in `/bedrouter status` |
 | `healthPollS` | `15` | Seconds between background health checks. A dead server flips the footer to `bedrouter: DOWN` and, with `autoStart`, is restarted (at most once a minute); `0` disables the poll |
 
 Example for a developer with a checkout:
@@ -52,7 +53,7 @@ Example for a developer with a checkout:
 | Command | Does |
 | --- | --- |
 | `/bedrouter` or `/bedrouter status` | Install location, server health (pid, version, region, classifier), registered models, current model, last decision |
-| `/bedrouter start` / `stop` / `restart` | Manage the server. It is shared by every Pi session, so `stop` affects all of them |
+| `/bedrouter start` / `stop` / `restart` | Manage the server. It is shared by every Pi session, so `stop` affects all of them; see `stopOnExit` for what happens when Pi quits |
 | `/bedrouter install` | `npm install` the dependency, or `npm run build` a checkout that has no `dist/` |
 | `/bedrouter doctor` | Which credential source resolved, expiry, the loaded ladder |
 | `/bedrouter probe` | One 1-token request per rung: which models this AWS account can actually invoke |
@@ -64,7 +65,7 @@ Example for a developer with a checkout:
 
 ## How the footer works
 
-Bedrouter echoes every routing decision in response headers (`x-bedrouter-model`, `-requested`, `-class`, `-reason`, `-conversation`, `-classifier`). Pi hands extensions those headers in the `after_provider_response` event, so the status line updates the moment a response starts, before any tokens stream. When the turn ends, the extension asks bedrouter for the conversation's running totals (`GET /v1/conversations/:key`) and appends cost: spend so far (classifier calls included), what the same tokens would have cost on the model the client asked for, and the difference as a percentage. `↑n` counts escalations in this conversation. The line clears when you switch to a non-bedrouter model. A background health check (every `healthPollS` seconds) keeps it honest between turns: if the server dies the line reads `bedrouter: DOWN` and, with `autoStart` on, the extension restarts it and says so.
+Bedrouter echoes every routing decision in response headers (`x-bedrouter-model`, `-requested`, `-class`, `-reason`, `-conversation`, `-classifier`). Pi hands extensions those headers in the `after_provider_response` event, so the status line updates the moment a response starts, before any tokens stream. When the turn ends, the extension asks bedrouter for the conversation's running totals (`GET /v1/conversations/:key`) and appends cost: spend so far, then the routing effect against what the same tokens would have cost on the model the client asked for: `saved $x (n%)`, `same as asked-for`, or `+$x over asked-for (routed up)` when the router chose a stronger rung. The classifier's one-off call is shown separately (`classifier $0.0004`) rather than counted against routing. `↑n` counts escalations in this conversation. The line clears when you switch to a non-bedrouter model. A background health check (every `healthPollS` seconds) keeps it honest between turns: if the server dies the line reads `bedrouter: DOWN` and, with `autoStart` on, the extension restarts it and says so.
 
 Under a coding agent every request carries tools and a large system prompt, so you will see `execute` and `explore` decided by keywords or the classifier, then `sticky` for the rest of the session, `up:kw:explore` when an explicit design question moves the conversation up, and escalations after failures. `trivial` shows up for bare chat clients, not for Pi.
 

package/extensions/index.ts

@@ -23,6 +23,8 @@ export default async function (pi: ExtensionAPI) {
   let lastCtx: ExtensionContext | null = null; // most recent context, for the background poll to update the footer
   let serverUp: boolean | null = null;         // last known health; null = never checked
   let restartAttemptAt = 0;                    // throttle auto-restarts to one per 60 s
+  let startedHere = false;                     // this session launched the server (so it may stop it on quit)
+  const ourConversations = new Set<string>();  // bedrouter conversation keys seen from this session
 
   const isOurs = (ctx: ExtensionContext) => ctx.model?.provider === settings.providerName;
   const setStatus = (ctx: ExtensionContext, text: string | undefined) => { lastCtx = ctx; if (ctx.hasUI && settings.footer) ctx.ui.setStatus("bedrouter", text); };
@@ -57,6 +59,8 @@ export default async function (pi: ExtensionAPI) {
     return `registered ${models.length} models from ${source}: ${registeredModelIds.join(", ")}`;
   }
 
+  const exitPolicyText = () => settings.stopOnExit === "never" ? "left running" : settings.stopOnExit === "always" ? "stopped" : "stopped unless another client is using it";
+
   /** Locate → (install) → start → register. Returns a human summary. Never throws. */
   async function bringUp(ctx: ExtensionContext | null, opts: { install?: boolean; start?: boolean } = {}): Promise<{ ok: boolean; lines: string[] }> {
     const lines: string[] = [];
@@ -76,10 +80,13 @@ export default async function (pi: ExtensionAPI) {
     }
     lines.push(await registerProvider(loc as br.Found));
     if (opts.start) {
+      const before = await br.health(settings);
       const r = await br.start(settings, loc as br.Found);
       for (const f of r.created) lines.push(`created ${f} from the example; edit it for this machine (AWS_PROFILE, ladder)`);
-      if (r.ok) lines.push(`bedrouter ${r.health.version} up on ${br.baseUrl(settings)} (pid ${r.health.pid}, region ${r.health.region}, classifier ${r.health.classifier ?? "off"})`);
-      else { lines.push(r.error); return { ok: false, lines }; }
+      if (r.ok) {
+        if (!before?.ok) startedHere = true;
+        lines.push(`bedrouter ${r.health.version} up on ${br.baseUrl(settings)} (pid ${r.health.pid}, region ${r.health.region}, classifier ${r.health.classifier ?? "off"})${!before?.ok ? `; started by this session, on quit: ${exitPolicyText()}` : "; was already running (not started here, left alone on quit)"}`);
+      } else { lines.push(r.error); return { ok: false, lines }; }
     }
     return { ok: true, lines };
   }
@@ -132,6 +139,7 @@ export default async function (pi: ExtensionAPI) {
     if (!d) return;
     last = d;
     serverUp = true;
+    if (d.conversation) ourConversations.add(d.conversation);
     setStatus(ctx, statusLine(last, stats));
   });
 
@@ -163,7 +171,14 @@ export default async function (pi: ExtensionAPI) {
   }
   if (settings.healthPollS > 0) { poll = setInterval(() => void checkHealth(), settings.healthPollS * 1000); poll.unref(); }
 
-  pi.on("session_shutdown", async () => { if (poll) clearInterval(poll); /* the server is left running on purpose: other Pi sessions share it */ });
+  pi.on("session_shutdown", async (ev) => {
+    if (poll) clearInterval(poll);
+    // Only a real quit ends the server; /reload and session switches keep it (the next session picks it straight up).
+    // The TUI is already gone at this point, so the policy is a setting, not a prompt (see stopOnExit).
+    if (ev.reason !== "quit" || !startedHere || settings.stopOnExit === "never") return;
+    if (settings.stopOnExit === "if-started-here" && (await br.othersActive(settings, ourConversations))) return;
+    await br.stop(settings);
+  });
 
   // ---- /bedrouter ------------------------------------------------------------------------------------------------
   const SUB = ["status", "start", "stop", "restart", "install", "doctor", "probe", "report", "log", "models", "fitnotes", "config", "help"];
@@ -186,6 +201,7 @@ export default async function (pi: ExtensionAPI) {
             `provider: ${settings.providerName} → ${registeredModelIds.length ? registeredModelIds.join(", ") : "(not registered)"}`,
             `session:  ${isOurs(ctx) ? `using ${ctx.model?.id}` : `not using bedrouter (model ${ctx.model?.provider}/${ctx.model?.id})`}`,
             `last:     ${last ? `${last.requested} → ${last.model}  ${last.cls} · ${last.reason}${last.classifier ? `  classifier: ${last.classifier}` : ""}` : "-"}`,
+            `on quit:  ${startedHere ? `server ${exitPolicyText()} (stopOnExit: ${settings.stopOnExit})` : "server was not started by this session; left alone"}`,
             `settings: ${settingsPath()}${fs.existsSync(settingsPath()) ? "" : " (defaults; /bedrouter config to create)"}`,
           ];
           show("bedrouter status", lines.join("\n"));
@@ -246,7 +262,7 @@ export default async function (pi: ExtensionAPI) {
         }
         case "config": {
           if (!fs.existsSync(settingsPath())) saveSettings(settings);
-          show("bedrouter settings", `${settingsPath()}\n\n${fs.readFileSync(settingsPath(), "utf8")}\nKeys: path, home, port, autoStart, autoSelect (model id or false), debug, footer, providerName. Edit the file, then /reload.`);
+          show("bedrouter settings", `${settingsPath()}\n\n${fs.readFileSync(settingsPath(), "utf8")}\nKeys: path, home, port, autoStart, autoSelect (model id or false), debug, footer, providerName, healthPollS, stopOnExit (if-started-here | always | never). Edit the file, then /reload.`);
           break;
         }
         default:

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-bedrouter",
-  "version": "0.1.0",
+  "version": "0.2.1",
   "description": "Pi extension for bedrouter: starts the cost-aware Bedrock model router, registers it as a provider, and shows which model served each request",
   "keywords": [
     "pi-package",

package/src/bedrouter.ts

@@ -124,6 +124,13 @@ export const FALLBACK_CONFIG: BedrouterConfig = {
   routing: { classes: { anthropic: { trivial: "haiku", execute: "sonnet", explore: "opus" }, openai: { execute: "gpt-oss-20b", explore: "gpt-oss-120b" } } },
 };
 export const conversation = (s: Settings, key: string) => getJson<ConversationStats>(`${baseUrl(s)}/v1/conversations/${key}`);
+export const recentConversations = async (s: Settings) => (await getJson<{ data: ConversationStats[] }>(`${baseUrl(s)}/v1/conversations`))?.data ?? [];
+
+/** True when a conversation other than `ours` sent a request within `windowMs`: another client is using the server. */
+export async function othersActive(s: Settings, ours: Set<string>, windowMs = 5 * 60_000, now = Date.now()): Promise<boolean> {
+  const recent = await recentConversations(s);
+  return recent.some((c) => !ours.has(c.key) && now - Date.parse(c.lastTs) < windowMs);
+}
 
 export const serverLog = (s: Settings) => path.join(homeDir(s), "server.log");
 export const decisionLog = (s: Settings) => path.join(homeDir(s), "bedrouter.log.jsonl");

package/src/footer.ts

@@ -19,9 +19,16 @@ export function statusLine(d: LastDecision | null, c: ConversationStats | null):
   let s = `⇄ ${d.model} ${arrow} ${d.requested}  ${d.cls}·${short(d.reason)}`;
   if (c && c.requests > 0) {
     const spend = c.costUsd + c.classifierCostUsd;
-    const diff = c.requestedCostUsd - spend;
+    // Routing effect = model cost vs the same tokens on the asked-for model. The classifier's one-off call is overhead
+    // and shown as such, never folded into "over asked-for" (which would flag every break-even session as a loss).
+    const diff = c.requestedCostUsd - c.costUsd;
     const pct = c.requestedCostUsd > 0 ? Math.round((diff / c.requestedCostUsd) * 100) : 0;
-    s += `  ${usd(spend)}${diff >= 0 ? ` saved ${usd(diff)} (${pct}%)` : ` +${usd(-diff)} over asked-for`}${c.escalations ? `  ↑${c.escalations}` : ""}`;
+    const clf = c.classifierCostUsd > 0 ? `, classifier ${usd(c.classifierCostUsd)}` : "";
+    let verdict: string;
+    if (Math.abs(diff) < 1e-6) verdict = `same as asked-for${clf}`;
+    else if (diff > 0) verdict = `saved ${usd(diff)} (${pct}%)${clf}`;
+    else verdict = `+${usd(-diff)} over asked-for (routed up)${clf}`;
+    s += `  ${usd(spend)} · ${verdict}${c.escalations ? `  ↑${c.escalations}` : ""}`;
   }
   return s;
 }

package/src/settings.ts

@@ -22,9 +22,15 @@ export type Settings = {
   providerName: string;
   /** Seconds between background health checks that keep the footer honest and restart a dead server (0 disables). */
   healthPollS: number;
+  /**
+   * What happens to the server when Pi quits (not on /reload or session switches):
+   * "if-started-here": stop it if this session started it and no other client used it in the last few minutes;
+   * "always": stop it whenever this session started it; "never": leave it running.
+   */
+  stopOnExit: "if-started-here" | "always" | "never";
 };
 
-export const DEFAULTS: Settings = { port: 20129, autoStart: true, autoSelect: "auto", debug: false, footer: true, providerName: "bedrouter", healthPollS: 15 };
+export const DEFAULTS: Settings = { port: 20129, autoStart: true, autoSelect: "auto", debug: false, footer: true, providerName: "bedrouter", healthPollS: 15, stopOnExit: "if-started-here" };
 
 export const agentDir = () => process.env.PI_CODING_AGENT_DIR ?? path.join(os.homedir(), ".pi", "agent");
 export const settingsPath = () => path.join(agentDir(), "pi-bedrouter.json");