agent-think: use observability preview and 250 steps

agent-think/AGENTS.md

@@ -192,6 +192,9 @@ npm run seed:r2  # push skills/** to the R2 bucket (add -- --local for dev)
   and team AI Gateway token, with a client-side fallback to
   `claude-opus-4-8` when the primary dispatch fails. Responses use
   `store: false`, so agent-think does not persist provider reasoning state.
+- The observability-preview packages are pinned to PR #1860 commit
+  `be4f7abc`. Agent-think opts into full message/tool payload spans while this
+  preview is validated in production. The turn safety cap is 250 steps.
   Production reads `CLOUDFLARE_AIG_TOKEN` from a Worker secret and
   attributes every request to the `agents-team-agent-think` project. Local
   agent turns read the same variable from `.env`.

agent-think/package.json

@@ -18,11 +18,11 @@
   "dependencies": {
     "@ai-sdk/anthropic": "^3.0.83",
     "@ai-sdk/openai": "^3.0.70",
-    "@cloudflare/ai-chat": "workspace:*",
-    "@cloudflare/think": "workspace:*",
+    "@cloudflare/ai-chat": "https://pkg.pr.new/cloudflare/agents/@cloudflare/ai-chat@be4f7abcf892a39449ec630b24e7f19e61309e92",
+    "@cloudflare/think": "https://pkg.pr.new/cloudflare/agents/@cloudflare/think@be4f7abcf892a39449ec630b24e7f19e61309e92",
     "@cloudflare/workspace": "https://pkg.pr.new/cloudflare/workspace/@cloudflare/workspace@18",
     "@platformatic/vfs": "^0.4.0",
-    "agents": "workspace:*",
+    "agents": "https://pkg.pr.new/cloudflare/agents/agents@be4f7abcf892a39449ec630b24e7f19e61309e92",
     "ai": "^6.0.202",
     "diff": "^9.0.0",
     "isomorphic-git": "^1.38.5",

agent-think/src/agent.ts

@@ -142,6 +142,10 @@ export class ThinkAgent extends ThinkBase {
   /** repro/pr can be long: clone, install, deploy or fix, verify. */
   override maxSteps = AGENT_THINK_MAX_STEPS;
 
+  /** Capture full model/tool payloads while validating the tracing preview. */
+  override storeMessages = true;
+  override storeTools = true;
+
   /**
    * We expose our own container-backed `bash` tool; skip Think's built-in
    * just-bash. Note: even if this flag is ever flipped back,

agent-think/src/turn-outcome.ts

@@ -1,4 +1,4 @@
-export const AGENT_THINK_MAX_STEPS = 100;
+export const AGENT_THINK_MAX_STEPS = 250;
 
 export type TurnOutcomeReason =
   | "completed"

agent-think/test/turn-outcome.test.ts

@@ -22,6 +22,10 @@ function result(
 }
 
 describe("classifyTurnOutcome", () => {
+  it("allows 250 steps for long review and demo runs", () => {
+    expect(AGENT_THINK_MAX_STEPS).toBe(250);
+  });
+
   it("reports max-step exhaustion when the final step ends with a tool call", () => {
     const outcome = classifyTurnOutcome(
       result(toolSteps(AGENT_THINK_MAX_STEPS)),