bosun 0.42.2 → 0.42.5

package/.env.example

@@ -802,6 +802,11 @@ VK_RECOVERY_PORT=54089
 # BOSUN_HOOKS_DISABLE_TASK_COMPLETE=false
 # BOSUN_HOOKS_DISABLE_HEALTH_CHECK=false
 
+# ── OpenTelemetry Tracing & Metrics ───────────────────────────────────────────
+# External orchestration-layer observability only; never affects agent context.
+# BOSUN_OTEL_ENDPOINT=http://localhost:4318/v1/traces
+# Configure tracing.sampleRate in bosun.config.json to tune sampling.
+
 # Force hooks to fire even for non-managed sessions (debug only):
 # BOSUN_HOOKS_FORCE=false
 
@@ -1175,3 +1180,7 @@ COPILOT_CLOUD_DISABLED=true
 # AGENT_STUCK_THRESHOLD_MS=300000
 # Alert if session costs more than $N (default: 1.0)
 # AGENT_COST_ANOMALY_THRESHOLD=1.0
+
+# OpenTelemetry tracing (optional)
+# BOSUN_OTEL_ENDPOINT=http://localhost:4318/v1/traces
+

package/agent/agent-custom-tools.mjs

@@ -746,12 +746,12 @@ export function getToolsPromptBlock(rootDir, opts = {}) {
     "## Custom Tools Library",
     "",
     discoveryMode
-      ? "Only eagerly-loaded tools are listed below. Use the MCP discovery tools to find the rest at runtime."
-      : "The following reusable helper scripts are available. Run them via",
+      ? "- Eager tools only below. Discover the rest at runtime."
+      : "- Run tools via `node <tool>.mjs`, `bash <tool>.sh`, or `python3 <tool>.py`.",
     discoveryMode
-      ? "Use `search`, then `get_schema`, then `execute` for tools not listed here. Use `call_discovered_tool` only for simple direct calls."
-      : "`node <tool>.mjs`, `bash <tool>.sh`, or `python3 <tool>.py`.",
-    "Built-in tools live in `bosun/tools/`; workspace tools in `.bosun/tools/`.",
+      ? "- Use `search`, then `get_schema`, then `execute` for tools not listed here."
+      : "- Check this library before writing new helper code.",
+    "- Built-in tools: `bosun/tools/`; workspace tools: `.bosun/tools/`.",
     "",
   ];
 
@@ -791,16 +791,13 @@ export function getToolsPromptBlock(rootDir, opts = {}) {
     lines.push(
       "---",
       "",
-      "**Reflect:** Before writing repetitive inline code, check if an existing",
-      "custom tool covers the need. If you encounter a pattern that future agents",
-      "(or yourself on retry) would benefit from having as a persistent script,",
-      "save it to `.bosun/tools/` and register it via the Bosun SDK so the whole",
-      "team benefits. Good candidates: analysis helpers, test generators, codemods,",
-      "build/lint wrappers that differ from what `npm run *` provides.",
+      "Reflect:",
+      "- Check existing tools before writing new helpers.",
+      "- Promote repeated analysis, test, build, transform, or search logic into `.bosun/tools/`.",
+      "- Skip one-off scripts.",
       "",
     );
   }
-
   return lines.join("\n");
 }
 
@@ -931,3 +928,4 @@ export function getAffinityTools(rootDir, opts = {}) {
     .slice(0, limit)
     .map((s) => s.tool);
 }
+

package/agent/agent-event-bus.mjs

@@ -25,6 +25,7 @@ import {
   reduceRetryQueue,
   snapshotRetryQueue,
 } from "./retry-queue.mjs";
+import { addSpanEvent, recordAgentError, recordIntervention } from "../infra/tracing.mjs";
 
 const TAG = "[agent-event-bus]";
 
@@ -232,6 +233,14 @@ export class AgentEventBus {
   emit(type, taskId, payload = {}, opts = {}) {
     const ts = Date.now();
     const event = { type, taskId, payload, ts };
+    addSpanEvent(type, { "bosun.task.id": taskId, ...payload });
+    if (type === AGENT_EVENT.AGENT_ERROR) {
+      recordAgentError(payload?.errorType || payload?.classification || "agent_error", {
+        "bosun.task.id": taskId,
+        "bosun.executor": payload?.executor,
+        "bosun.agent.sdk": payload?.sdk,
+      });
+    }
 
     // ── Dedup
     const key = `${type}:${taskId}`;
@@ -1092,3 +1101,4 @@ export function createAgentEventBus(options) {
   return new AgentEventBus(options);
 }
 
+

package/agent/agent-pool.mjs

@@ -791,6 +791,9 @@ function buildCodexSdkOptions(envInput = process.env) {
             wire_api: "responses",
           },
         },
+        features: {
+          remote_models: false,
+        },
         ...(azureModel ? { model: azureModel } : {}),
       },
     };
@@ -1222,6 +1225,7 @@ async function launchCodexThread(prompt, cwd, timeoutMs, extra = {}) {
   codexOpts.config = {
     ...(codexOpts.config || {}),
     features: {
+      ...(codexOpts.config?.features || {}),
       child_agents_md: true,
       multi_agent: true,
       memories: true,
@@ -3158,6 +3162,17 @@ async function resumeCodexThread(threadId, prompt, cwd, timeoutMs, extra = {}) {
     codexOpts.env = { ...(codexOpts.env || {}), CODEX_MODEL: modelOverride };
     codexOpts.config = { ...(codexOpts.config || {}), model: modelOverride };
   }
+  codexOpts.config = {
+    ...(codexOpts.config || {}),
+    features: {
+      ...(codexOpts.config?.features || {}),
+      child_agents_md: true,
+      multi_agent: true,
+      memories: true,
+      undo: true,
+      steer: true,
+    },
+  };
   const codex = new CodexClass(codexOpts);
 
   let thread;

package/agent/agent-prompt-catalog.mjs

@@ -909,7 +909,7 @@ Examples:
 - For numbers, say them naturally: "You have 12 tasks in the backlog."
 - When tasks or agents are busy, keep the user informed.
 - For long outputs (code, logs), summarize the key points vocally.
-- When delegating to an agent, let the user know: "I'm sending that to Codex now."
+- When delegating to an agent, say: "Sending that to Codex now."
 
 ## Error Handling
 
@@ -938,7 +938,7 @@ Be concise. Lead with answers. Summarize long outputs.
 
   customToolReflect: `## Reflect: Custom Tool Extraction
 
-Before closing this task, reflect on the work you just completed:
+Before closing the task, check for reusable tooling:
 
 1. **Did you write any utility code (≥ 10 lines) that you'd write again?**
    If yes — extract it into a persistent custom tool in \`.bosun/tools/\`.
@@ -948,7 +948,7 @@ Before closing this task, reflect on the work you just completed:
    If yes — package it as a custom tool so future agents skip the re-derivation.
 
 3. **Did an existing custom tool help you?**
-   Consider whether it should be promoted to global scope (\`promoteToGlobal\`).
+   Consider promoting it to global scope (\`promoteToGlobal\`).
 
 4. **What category does the extracted logic fall into?**
    analysis | testing | git | build | transform | search | validation | utility
@@ -962,7 +962,7 @@ registerCustomTool(rootDir, {
 });
 \`\`\`
 
-Only extract if the tool has clear reuse value. Skip one-off logic.
+Extract only when reuse is clear. Skip one-off logic.
 `,
 
   customToolsContext: `{{CUSTOM_TOOLS_BLOCK}}

package/agent/agent-prompts.mjs

@@ -174,6 +174,26 @@ export function resolvePromptTemplate(template, values, fallback) {
   return rendered && rendered.trim() ? rendered : base;
 }
 
+export async function buildCustomToolsContextPrompt(rootDir, opts = {}) {
+  const { getToolsPromptBlock, listCustomTools } = await import("./agent-custom-tools.mjs");
+  const registeredTools = listCustomTools(rootDir, {
+    includeBuiltins: false,
+  });
+  if (registeredTools.length === 0) return "";
+
+  const promptTemplate = DEFAULT_PROMPTS.customToolsContext || "{{CUSTOM_TOOLS_BLOCK}}";
+  const toolsBlock = getToolsPromptBlock(rootDir, opts);
+  if (!toolsBlock.trim()) return "";
+
+  return renderPromptTemplate(
+    promptTemplate,
+    {
+      CUSTOM_TOOLS_BLOCK: toolsBlock,
+    },
+    rootDir,
+  ).trim();
+}
+
 export function ensureAgentPromptWorkspace(repoRoot) {
   const root = resolve(repoRoot || process.cwd());
   let workspaceDir = getDefaultPromptWorkspace(root);

package/agent/agent-supervisor.mjs

@@ -29,6 +29,8 @@
  * @module agent-supervisor
  */
 
+import { addSpanEvent, recordIntervention } from "../infra/tracing.mjs";
+
 const TAG = "[agent-supervisor]";
 const API_ERROR_CONTINUE_COOLDOWNS_MS = Object.freeze([
   3 * 60_000,
@@ -473,6 +475,12 @@ export class AgentSupervisor {
     }
 
     this._lastDecision.set(taskId, { situation, intervention, ts: Date.now() });
+    addSpanEvent("bosun.supervisor.assess", {
+      "bosun.task.id": taskId,
+      "bosun.health.score": healthScore,
+      "bosun.situation": situation,
+      "bosun.intervention.type": intervention,
+    });
 
     return { situation, healthScore, intervention, prompt, reason };
   }
@@ -489,6 +497,18 @@ export class AgentSupervisor {
     );
 
     try {
+      if (intervention !== INTERVENTION.NONE) {
+        recordIntervention(intervention, {
+          "bosun.task.id": taskId,
+          "bosun.situation": situation,
+        });
+      }
+      addSpanEvent("bosun.supervisor.intervention", {
+        "bosun.task.id": taskId,
+        "bosun.intervention.type": intervention,
+        "bosun.situation": situation,
+        "bosun.reason": reason,
+      });
       switch (intervention) {
         case INTERVENTION.NONE:
           break;