@desplega.ai/agent-swarm 1.76.2 → 1.77.0

package/src/providers/codex-adapter.ts

@@ -65,6 +65,8 @@ import {
   type Usage,
   type WebSearchItem,
 } from "@openai/codex-sdk";
+import { buildRatingsFromLlm, fetchRetrievalsForTask, postRatings } from "../be/memory/raters/llm";
+import { summarizeSession as runSummarize } from "../utils/internal-ai";
 import { scrubSecrets } from "../utils/secret-scrubber";
 import { type CodexAgentsMdHandle, writeCodexAgentsMd } from "./codex-agents-md";
 import { computeCodexCostUsd, getCodexContextWindow, resolveCodexModel } from "./codex-models";
@@ -340,6 +342,26 @@ export async function buildCodexConfig(
   };
 }
 
+/**
+ * Test-injection points for the codex session-end summarization path.
+ *
+ * Production callers omit `deps` entirely — the `CodexSession.summarizeAtEnd`
+ * helper falls back to the symbols imported at the top of this file. Tests
+ * override each function so we can exercise the summarize/index/rate flow
+ * without standing up a real API server or LLM.
+ *
+ * Why this exists (mirrors `SummarizeSessionForPiDeps`): `bun:test`'s
+ * `mock.module()` is process-wide and leaks across test files in the same
+ * `bun test` run, breaking siblings that import the real symbols. Explicit DI
+ * keeps the boundary local to this adapter.
+ */
+export interface SummarizeSessionForCodexDeps {
+  runSummarize?: typeof runSummarize;
+  fetchRetrievalsForTask?: typeof fetchRetrievalsForTask;
+  postRatings?: typeof postRatings;
+  buildRatingsFromLlm?: typeof buildRatingsFromLlm;
+}
+
 /** Running session backed by a Codex `Thread`. */
 class CodexSession implements ProviderSession {
   private readonly thread: Thread;
@@ -348,6 +370,7 @@ class CodexSession implements ProviderSession {
   private readonly resolvedModel: string;
   private readonly contextWindow: number;
   private readonly skillsDir: string;
+  private readonly summarizeDeps: SummarizeSessionForCodexDeps;
   private readonly listeners: Array<(event: ProviderEvent) => void> = [];
   private readonly eventQueue: ProviderEvent[] = [];
   private readonly logFileHandle: ReturnType<ReturnType<typeof Bun.file>["writer"]>;
@@ -355,6 +378,11 @@ class CodexSession implements ProviderSession {
   private readonly completionPromise: Promise<ProviderResult>;
   private resolveCompletion!: (result: ProviderResult) => void;
   private abortController: AbortController | null = null;
+  /**
+   * Per-session transcript buffer used to feed the session-end summarizer.
+   * Reset at the start of `runSession` and appended in `handleEvent`.
+   */
+  private transcript: string[] = [];
   /**
    * Mutable holder for the current turn's `AbortController`. Shared with the
    * swarm event handler so it can trigger an abort from outside `runSession`
@@ -366,6 +394,14 @@ class CodexSession implements ProviderSession {
   private lastUsage: Usage | null = null;
   private aborted = false;
   private settled = false;
+  /**
+   * Result captured by `settle` but held back from `resolveCompletion` until
+   * `runSession`'s `finally` block has fully cleaned up (log writer flush,
+   * AGENTS.md cleanup, session summary). Without this, callers awaiting
+   * `waitForCompletion` would race the cleanup. Phase 3 added the session
+   * summary path which materially relies on this ordering for testability.
+   */
+  private pendingResult: ProviderResult | null = null;
 
   constructor(
     thread: Thread,
@@ -374,6 +410,7 @@ class CodexSession implements ProviderSession {
     resolvedModel: string,
     initialEvents: ProviderEvent[] = [],
     skillsDir?: string,
+    summarizeDeps: SummarizeSessionForCodexDeps = {},
   ) {
     this.thread = thread;
     this.config = config;
@@ -385,6 +422,7 @@ class CodexSession implements ProviderSession {
     // runtime default of `${HOME}/.codex/skills`.
     this.skillsDir =
       skillsDir ?? process.env.CODEX_SKILLS_DIR ?? join(os.homedir(), ".codex", "skills");
+    this.summarizeDeps = summarizeDeps;
     this.logFileHandle = Bun.file(config.logFile).writer();
 
     this.completionPromise = new Promise<ProviderResult>((resolve) => {
@@ -474,7 +512,10 @@ class CodexSession implements ProviderSession {
   private settle(result: ProviderResult): void {
     if (this.settled) return;
     this.settled = true;
-    this.resolveCompletion(result);
+    // Resolution deferred until `runSession`'s finally-block fully cleans up
+    // (see `pendingResult` rationale on the field above). Caller-visible
+    // ordering: cleanup → resolve waitForCompletion.
+    this.pendingResult = result;
   }
 
   /** Build CostData from the most recent turn usage. */
@@ -559,6 +600,41 @@ class CodexSession implements ProviderSession {
     );
   }
 
+  /**
+   * Render a completed tool item as a short, signal-dense one-liner for the
+   * session-end summarization transcript. Picks tool-type-specific fields so
+   * the transcript doesn't get drowned in raw JSON (a single `command_execution`
+   * can carry 100KB+ of `aggregated_output`). Each branch is capped at 500
+   * chars; unknown types fall back to a trimmed `JSON.stringify`.
+   */
+  private shortenItemResult(item: ThreadItem): string {
+    switch (item.type) {
+      case "command_execution": {
+        const cmd = item as CommandExecutionItem;
+        const stdout = (cmd.aggregated_output ?? "").slice(0, 500);
+        return `exit=${cmd.exit_code ?? "?"} status=${cmd.status ?? "?"} stdout=${stdout}`;
+      }
+      case "file_change": {
+        const fc = item as FileChangeItem;
+        const summarised = (fc.changes ?? [])
+          .slice(0, 5)
+          .map((c) => `${c.kind}:${c.path}`)
+          .join(",");
+        return `changes=[${summarised}]`;
+      }
+      case "mcp_tool_call": {
+        const mcp = item as McpToolCallItem;
+        return `server=${mcp.server} tool=${mcp.tool} status=${mcp.status ?? "?"}`;
+      }
+      case "web_search": {
+        const ws = item as WebSearchItem;
+        return `query=${(ws.query ?? "").slice(0, 200)}`;
+      }
+      default:
+        return JSON.stringify(item).slice(0, 500);
+    }
+  }
+
   private handleEvent(event: ThreadEvent): void {
     // Mirror every raw SDK event into the log as raw_log for debugability —
     // parity with Claude's JSONL envelope.
@@ -582,6 +658,14 @@ class CodexSession implements ProviderSession {
             toolName: this.toolNameForItem(event.item),
             args: this.toolArgsForItem(event.item),
           });
+          // Mirror into the transcript buffer for session-end summarization.
+          // Tools are the bulk of useful signal in a codex session, so we
+          // capture both the start (args) and the completion (result digest).
+          this.transcript.push(
+            `Tool[${this.toolNameForItem(event.item)}] started: ${JSON.stringify(
+              this.toolArgsForItem(event.item),
+            ).slice(0, 500)}`,
+          );
         }
         break;
       }
@@ -615,6 +699,9 @@ class CodexSession implements ProviderSession {
             toolName: this.toolNameForItem(item),
             result: item,
           });
+          this.transcript.push(
+            `Tool[${this.toolNameForItem(item)}] completed: ${this.shortenItemResult(item)}`,
+          );
           break;
         }
         switch (item.type) {
@@ -622,6 +709,7 @@ class CodexSession implements ProviderSession {
             const msg = item as AgentMessageItem;
             if (msg.text) {
               this.emit({ type: "message", role: "assistant", content: msg.text });
+              this.transcript.push(`Assistant: ${msg.text}`);
             }
             break;
           }
@@ -764,6 +852,11 @@ class CodexSession implements ProviderSession {
         this.emit(event),
       );
 
+      // Reset + seed the transcript buffer so the session-end summarizer has
+      // the user's prompt as anchor context. Subsequent appends happen in
+      // `handleEvent` (tool start/end, agent_message).
+      this.transcript = [`User: ${resolvedPrompt}`];
+
       const streamed = await this.thread.runStreamed(resolvedPrompt, {
         signal: this.abortController.signal,
       });
@@ -827,6 +920,18 @@ class CodexSession implements ProviderSession {
         failureReason: message,
       });
     } finally {
+      // Session-end summarization. Pure addition for codex — no behavior to
+      // preserve. Wrapped in its own try/catch so summary failure must NOT
+      // block the existing log/AGENTS.md cleanup below. Gate `SKIP_SESSION_SUMMARY=1`
+      // matches the parity convention used by the claude Stop hook + pi/opencode.
+      if (process.env.SKIP_SESSION_SUMMARY !== "1") {
+        try {
+          await this.summarizeAtEnd();
+        } catch (err) {
+          console.error("session_summary failed (codex):", err);
+        }
+      }
+
       // Detach the abort controller now that the turn has settled.
       this.abortRef.current = null;
       try {
@@ -835,6 +940,121 @@ class CodexSession implements ProviderSession {
         // Ignore log writer cleanup failures.
       }
       await this.agentsMdHandle.cleanup();
+
+      // Resolve `waitForCompletion()` only AFTER all cleanup has finished so
+      // downstream observers (tests + the runner's `.then(...)` chain) don't
+      // race the finally-block side effects. Fallback to an error result if
+      // we somehow never called `settle` (defensive — every codepath in the
+      // try/catch above calls settle exactly once).
+      const finalResult =
+        this.pendingResult ??
+        ({
+          exitCode: 1,
+          sessionId: this._sessionId,
+          cost: this.buildCostData(this.lastUsage, true),
+          isError: true,
+          failureReason: "session did not settle",
+        } as ProviderResult);
+      this.resolveCompletion(finalResult);
+    }
+  }
+
+  /**
+   * Index a session summary into agent memory at the end of a codex turn.
+   *
+   * Mirrors `summarizeSessionForPi` and the claude Stop hook:
+   *   1. Truncate the in-memory transcript buffer to the last 20 KB.
+   *   2. Bail when the transcript is too short or the swarm context is
+   *      missing (no agentId / no taskId / no apiUrl / no apiKey).
+   *   3. (Optional) Pre-fetch retrievals when `MEMORY_RATERS` includes `llm`
+   *      so the LLM can score them alongside the summary.
+   *   4. Call `runSummarize` from `src/utils/internal-ai` (Phase 0). Returns
+   *      `null` when no credential resolves — silent skip.
+   *   5. Apply length/quality gate; POST to `/api/memory/index`.
+   *   6. POST ratings (`events:` key, NOT `ratings:`) via `postRatings` when
+   *      `MEMORY_RATERS=llm` and the LLM returned per-memory scores.
+   *
+   * All catches log via `console.error(..., err)` — silent-fail behavior is
+   * gone. The outer try in `runSession`'s finally-block is the final safety
+   * net guaranteeing existing cleanup runs regardless.
+   */
+  private async summarizeAtEnd(): Promise<void> {
+    const transcriptStr = this.transcript.join("\n").slice(-20_000);
+    const { agentId, taskId, apiUrl, apiKey } = this.config;
+    if (!agentId || !taskId || !apiUrl || !apiKey) return;
+    if (transcriptStr.length <= 100) return;
+
+    const _runSummarize = this.summarizeDeps.runSummarize ?? runSummarize;
+    const _fetchRetrievals = this.summarizeDeps.fetchRetrievalsForTask ?? fetchRetrievalsForTask;
+    const _postRatings = this.summarizeDeps.postRatings ?? postRatings;
+    const _buildRatings = this.summarizeDeps.buildRatingsFromLlm ?? buildRatingsFromLlm;
+
+    const memoryRaters = (process.env.MEMORY_RATERS ?? "")
+      .split(",")
+      .map((s) => s.trim())
+      .filter(Boolean);
+    const wantRatings = memoryRaters.includes("llm");
+    const retrievals = wantRatings
+      ? await _fetchRetrievals({ apiUrl, apiKey, agentId, taskId }).catch(() => [])
+      : [];
+
+    const result = await _runSummarize({
+      harness: "codex",
+      transcript: transcriptStr,
+      retrievals,
+      taskContext: {
+        sourceTaskId: taskId,
+        agentId,
+        prompt: this.config.prompt,
+      },
+      apiUrl,
+      apiKey,
+    });
+    // null = no auth resolved or wrapper exhausted retries (already logged inside)
+    if (!result) return;
+
+    const summary = result.summary.trim();
+    if (summary.length <= 20 || summary.toLowerCase().includes("no significant learnings")) {
+      return;
+    }
+
+    const indexResp = await fetch(`${apiUrl}/api/memory/index`, {
+      method: "POST",
+      headers: {
+        "Content-Type": "application/json",
+        Authorization: `Bearer ${apiKey}`,
+        "X-Agent-ID": agentId,
+      },
+      body: JSON.stringify({
+        scope: "agent",
+        source: "session_summary",
+        sourceTaskId: taskId,
+        content: summary,
+        name: "session-summary",
+        agentId,
+      }),
+    });
+    if (!indexResp.ok) {
+      const text = await indexResp.text().catch(() => "");
+      console.error(
+        "session_summary: /api/memory/index POST failed (codex):",
+        indexResp.status,
+        text,
+      );
+      return;
+    }
+
+    if (wantRatings && result.ratings && result.ratings.length > 0) {
+      const ratingEvents = _buildRatings(result.ratings, retrievals);
+      if (ratingEvents.length > 0) {
+        await _postRatings({
+          apiUrl,
+          apiKey,
+          agentId,
+          taskId,
+          events: ratingEvents,
+        }).catch((err) => console.error("session_summary: postRatings failed (codex):", err));
+      }
     }
   }
 }
@@ -851,8 +1071,17 @@ export class CodexAdapter implements ProviderAdapter {
    */
   private readonly skillsDir?: string;
 
-  constructor(opts: { skillsDir?: string } = {}) {
+  /**
+   * Optional dependency-injection points for session-end summarization. Tests
+   * pass stubs in here to exercise the summarize → index → rate flow without
+   * standing up a real API server or LLM. Production callers omit this and the
+   * `CodexSession` falls back to the module-level imports.
+   */
+  private readonly summarizeDeps: SummarizeSessionForCodexDeps;
+
+  constructor(opts: { skillsDir?: string; summarizeDeps?: SummarizeSessionForCodexDeps } = {}) {
     this.skillsDir = opts.skillsDir;
+    this.summarizeDeps = opts.summarizeDeps ?? {};
   }
 
   async createSession(config: ProviderSessionConfig): Promise<ProviderSession> {
@@ -937,6 +1166,7 @@ export class CodexAdapter implements ProviderAdapter {
         resolvedModel,
         preSessionEvents,
         this.skillsDir,
+        this.summarizeDeps,
       );
     } catch (err) {
       // If we failed to construct the thread, clean up the managed AGENTS.md

package/src/providers/codex-oauth/storage.ts

@@ -85,6 +85,27 @@ export async function deleteCodexOAuth(apiUrl: string, apiKey: string): Promise<
   });
 }
 
+/**
+ * Best-effort persistence of refreshed OAuth credentials back to the config
+ * store. Wraps {@link storeCodexOAuth} with a try/catch + `console.error` —
+ * a write failure MUST NOT block the current caller from using the refreshed
+ * `apiKey` (the rotation will just have to retry on the next call). This is
+ * called from `src/utils/internal-ai/credentials.ts` after pi-ai's
+ * `getOAuthApiKey` returns `{ newCredentials, apiKey }` so the rotated refresh
+ * token isn't lost in-memory.
+ */
+export async function persistCodexOAuth(
+  apiUrl: string,
+  apiKey: string,
+  creds: CodexOAuthCredentials,
+): Promise<void> {
+  try {
+    await storeCodexOAuth(apiUrl, apiKey, creds);
+  } catch (err) {
+    console.error("[codex-oauth] persistCodexOAuth failed (non-fatal):", err);
+  }
+}
+
 export async function getValidCodexOAuth(
   apiUrl: string,
   apiKey: string,

package/src/providers/pi-mono-extension.ts

@@ -7,9 +7,11 @@
  */
 
 import type { ExtensionFactory } from "@mariozechner/pi-coding-agent";
+import { buildRatingsFromLlm, fetchRetrievalsForTask, postRatings } from "../be/memory/raters/llm";
 import { checkToolLoop, clearToolHistory } from "../hooks/tool-loop-detection";
+import { summarizeSession as runSummarize } from "../utils/internal-ai";
 
-interface SwarmHooksConfig {
+export interface SwarmHooksConfig {
   apiUrl: string;
   apiKey: string;
   agentId: string;
@@ -276,13 +278,54 @@ async function fetchConcurrentContext(config: SwarmHooksConfig): Promise<string
   }
 }
 
-/** Run session summarization via Claude Haiku on shutdown */
-async function summarizeSession(
+/**
+ * Test-injection points for `summarizeSessionForPi`. Production callers omit
+ * `deps` entirely — pi-mono-extension's `session_shutdown` handler uses the
+ * default implementations bound via `import` at the top of this file.
+ *
+ * Why this exists: `bun:test`'s `mock.module()` is process-wide and leaks
+ * across test files, so the pi-mono-extension test cannot stub out
+ * `runSummarize`/`postRatings` via module mocking without breaking siblings
+ * that import the real symbols (e.g. `buildRatingsFromLlm` step-6 tests,
+ * `summarize-session.test.ts`). Explicit DI is the safer pattern.
+ */
+export interface SummarizeSessionForPiDeps {
+  runSummarize?: typeof runSummarize;
+  fetchRetrievalsForTask?: typeof fetchRetrievalsForTask;
+  postRatings?: typeof postRatings;
+  buildRatingsFromLlm?: typeof buildRatingsFromLlm;
+}
+
+/**
+ * Run session summarization via the shared `internal-ai` abstraction on
+ * shutdown. Replaces the previous `Bun.spawn(claude -p ...)` shellout
+ * (which silently failed in production because pi sessions typically don't
+ * have Anthropic CLI auth).
+ *
+ * Flow:
+ *   1. Read tail of session transcript file.
+ *   2. Fetch task details + (optionally) memory retrievals for ratings.
+ *   3. Call `runSummarize` from `src/utils/internal-ai` — picks credentials
+ *      out of env / codex OAuth, returns structured `{summary, ratings}`.
+ *   4. Apply length/quality gate; POST summary to `/api/memory/index`.
+ *   5. If `MEMORY_RATERS` includes `llm` AND ratings came back, POST them
+ *      via `postRatings` (events-based; mirrors the claude Stop hook).
+ *
+ * All catches log via `console.error(..., err)` — silent-fail behavior is
+ * gone.
+ */
+export async function summarizeSessionForPi(
   config: SwarmHooksConfig,
   sessionFile: string | undefined,
+  deps: SummarizeSessionForPiDeps = {},
 ): Promise<void> {
   if (!sessionFile) return;
 
+  const _runSummarize = deps.runSummarize ?? runSummarize;
+  const _fetchRetrievals = deps.fetchRetrievalsForTask ?? fetchRetrievalsForTask;
+  const _postRatings = deps.postRatings ?? postRatings;
+  const _buildRatings = deps.buildRatingsFromLlm ?? buildRatingsFromLlm;
+
   try {
     let transcript = "";
     try {
@@ -294,84 +337,78 @@ async function summarizeSession(
 
     if (transcript.length <= 100) return;
 
-    let taskContext = "";
-    try {
-      const taskDetails = await fetchTaskDetails(config);
-      if (taskDetails) {
-        taskContext = `Task: ${taskDetails.task}`;
-      }
-    } catch {
-      /* no task context */
-    }
-
-    const summarizePrompt = `You are summarizing an AI agent's work session. Extract ONLY high-value learnings.
-
-DO NOT include:
-- Generic descriptions of what was done ("worked on task X")
-- Tool calls or file reads
-- Routine progress updates
-
-DO include (if present):
-- **Mistakes made and corrections** — what went wrong and what fixed it
-- **Discovered patterns** — reusable approaches, APIs, or codebase conventions
-- **Codebase knowledge** — important file paths, architecture decisions, gotchas
-- **Environment knowledge** — service URLs, config details, tool quirks
-- **Failed approaches** — what was tried and didn't work (and why)
-
-Format as a bulleted list of concrete, reusable facts. If the session was routine with no significant learnings, respond with exactly: "No significant learnings."
-${taskContext ? `\nTask context: ${taskContext}` : ""}
-Transcript:
-${transcript}`;
-
-    const tmpFile = `/tmp/session-summary-${Date.now()}.txt`;
-    await Bun.write(tmpFile, summarizePrompt);
-    const proc = Bun.spawn(
-      [
-        "bash",
-        "-c",
-        `cat "${tmpFile}" | ${process.env.CLAUDE_BINARY || "claude"} -p --model haiku --output-format json`,
-      ],
-      {
-        stdout: "pipe",
-        stderr: "pipe",
-        env: { ...process.env, SKIP_SESSION_SUMMARY: "1" },
+    const sourceTaskId = config.taskId;
+    const agentId = config.agentId;
+    if (!sourceTaskId || !agentId) return;
+
+    const taskDetails = await fetchTaskDetails(config).catch(() => null);
+
+    const memoryRaters = (process.env.MEMORY_RATERS ?? "")
+      .split(",")
+      .map((s) => s.trim())
+      .filter(Boolean);
+    const wantRatings = memoryRaters.includes("llm");
+    const retrievals = wantRatings
+      ? await _fetchRetrievals({
+          apiUrl: config.apiUrl,
+          apiKey: config.apiKey,
+          agentId,
+          taskId: sourceTaskId,
+        }).catch(() => [])
+      : [];
+
+    const result = await _runSummarize({
+      harness: "pi",
+      transcript,
+      retrievals,
+      taskContext: {
+        sourceTaskId,
+        agentId,
+        prompt: taskDetails?.task,
       },
-    );
-    const timeoutId = setTimeout(() => proc.kill(), 30000);
-    const result = { stdout: await new Response(proc.stdout).text() };
-    clearTimeout(timeoutId);
-    await Bun.$`rm -f ${tmpFile}`.quiet();
+      apiUrl: config.apiUrl,
+      apiKey: config.apiKey,
+    });
+    // null = no auth resolved or wrapper exhausted retries (already logged inside)
+    if (!result) return;
 
-    let summary: string;
-    try {
-      const summaryOutput = JSON.parse(result.stdout);
-      summary = summaryOutput.result ?? result.stdout;
-    } catch {
-      summary = result.stdout;
+    const summary = result.summary.trim();
+    if (summary.length <= 20 || summary.toLowerCase().includes("no significant learnings")) {
+      return;
     }
 
-    if (
-      summary &&
-      summary.length > 20 &&
-      !summary.trim().toLowerCase().includes("no significant learnings")
-    ) {
-      await fetch(`${config.apiUrl}/api/memory/index`, {
-        method: "POST",
-        headers: apiHeaders(config),
-        body: JSON.stringify({
-          agentId: config.agentId,
-          content: summary,
-          name: taskContext
-            ? `Session: ${taskContext.slice(0, 80)}`
-            : `Session: ${new Date().toISOString().slice(0, 16)}`,
-          scope: "agent",
-          source: "session_summary",
-          sourceTaskId: config.taskId,
-        }),
-      });
+    const indexResp = await fetch(`${config.apiUrl}/api/memory/index`, {
+      method: "POST",
+      headers: apiHeaders(config),
+      body: JSON.stringify({
+        scope: "agent",
+        source: "session_summary",
+        sourceTaskId,
+        content: summary,
+        name: "session-summary",
+        agentId,
+      }),
+    });
+    if (!indexResp.ok) {
+      const text = await indexResp.text().catch(() => "");
+      console.error("session_summary: /api/memory/index POST failed (pi):", indexResp.status, text);
+      return;
     }
-  } catch {
-    /* non-blocking */
+
+    if (wantRatings && result.ratings && result.ratings.length > 0) {
+      const ratingEvents = _buildRatings(result.ratings, retrievals);
+      if (ratingEvents.length > 0) {
+        await _postRatings({
+          apiUrl: config.apiUrl,
+          apiKey: config.apiKey,
+          agentId,
+          taskId: sourceTaskId,
+          events: ratingEvents,
+        }).catch((err) => console.error("session_summary: postRatings failed (pi):", err));
+      }
+    }
+  } catch (err) {
+    console.error("session_summary failed (pi):", err);
   }
 }
 
@@ -661,7 +698,7 @@ export function createSwarmHooksExtension(config: SwarmHooksConfig): ExtensionFa
       // Session summarization — get session file from context's session manager
       const sessionFile = ctx.sessionManager.getSessionFile?.();
       if (!process.env.SKIP_SESSION_SUMMARY) {
-        await summarizeSession(config, sessionFile);
+        await summarizeSessionForPi(config, sessionFile);
       }
 
       // Mark agent offline

package/src/telemetry.ts

@@ -73,6 +73,32 @@ interface TrackOptions {
   metadata?: Record<string, unknown>;
 }
 
+/**
+ * Read SWARM_ORG_ID / SWARM_ORG_NAME from process.env at call time. Reading
+ * fresh each track() lets reloaded swarm_config values land in telemetry
+ * without restarting (loadGlobalConfigsIntoEnv mutates process.env on
+ * `POST /api/config/reload` with override=true). Returns only the keys that
+ * are set, so the spread below stays a clean noop on self-host.
+ */
+function getOrgIdentity(): { organization_id?: string; organization_name?: string } {
+  const out: { organization_id?: string; organization_name?: string } = {};
+  const orgId = process.env.SWARM_ORG_ID?.trim();
+  if (orgId) out.organization_id = orgId;
+  const orgName = process.env.SWARM_ORG_NAME?.trim();
+  if (orgName) out.organization_name = orgName;
+  return out;
+}
+
+/**
+ * Mirror of `buildIdentity()`'s SWARM_CLOUD parsing — accepts "true" or "1".
+ * Always emitted (not optional) so consumers can split cloud vs self-host
+ * cohorts without ambiguity between "false" and "unset".
+ */
+function isCloudDeployment(): boolean {
+  const raw = process.env.SWARM_CLOUD;
+  return raw === "true" || raw === "1";
+}
+
 /** Fire-and-forget telemetry event. Never throws, never blocks. */
 export function track(options: TrackOptions): void {
   if (!isEnabled() || !installationId) return;
@@ -89,6 +115,8 @@ export function track(options: TrackOptions): void {
         transport: "https",
         schema_version: 1,
         environment: process.env.NODE_ENV ?? "production",
+        is_cloud: isCloudDeployment(),
+        ...getOrgIdentity(),
         ...options.metadata,
       },
     };