@kinetica/admin-agent 0.1.2 → 0.1.3

package/README.md

@@ -161,14 +161,15 @@ wasting ~28.7 MB as raw storage. Both issues have been remediated.
 
 Set environment variables or use a `.env` file. The agent loads `.env` automatically at startup (shell-set variables always take precedence). Any missing values are prompted interactively.
 
-| Variable              | Description                                                                                 | Required                                        |
-| --------------------- | ------------------------------------------------------------------------------------------- | ----------------------------------------------- |
-| `ANTHROPIC_API_KEY`   | Anthropic API key for Claude                                                                | No — OAuth login used if unset                  |
-| `KINETICA_URL`        | Kinetica instance URL (e.g. `http://host:9191` or bare `host:9191`)                         | Prompted if unset                               |
-| `KINETICA_USER`       | Kinetica username                                                                           | Prompted if unset                               |
-| `KINETICA_PASS`       | Kinetica password                                                                           | Prompted if unset (masked, never saved to .env) |
-| `KINETICA_HTTPS_ONLY` | Set to `1` to refuse plaintext HTTP fallback entirely — strict mode for production clusters | No                                              |
-| `DEBUG`               | Set to `1` to log HTTP requests and the assembled system-prompt token size to stderr        | No                                              |
+| Variable                 | Description                                                                                      | Required                                        |
+| ------------------------ | ------------------------------------------------------------------------------------------------ | ----------------------------------------------- |
+| `ANTHROPIC_API_KEY`      | Anthropic API key for Claude                                                                     | No — OAuth login used if unset                  |
+| `ADMIN_AGENT_MAX_BUDGET` | Per-session budget cap in USD for API-key billing (overridden by `--max-budget`; default `5.00`) | No                                              |
+| `KINETICA_URL`           | Kinetica instance URL (e.g. `http://host:9191` or bare `host:9191`)                              | Prompted if unset                               |
+| `KINETICA_USER`          | Kinetica username                                                                                | Prompted if unset                               |
+| `KINETICA_PASS`          | Kinetica password                                                                                | Prompted if unset (masked, never saved to .env) |
+| `KINETICA_HTTPS_ONLY`    | Set to `1` to refuse plaintext HTTP fallback entirely — strict mode for production clusters      | No                                              |
+| `DEBUG`                  | Set to `1` to log HTTP requests and the assembled system-prompt token size to stderr             | No                                              |
 
 ```bash
 cp .env.example .env   # fill in values — or let the agent create it for you
@@ -208,7 +209,12 @@ npm run dev -- --logout
 
 ### Session Budget
 
-Each session enforces a **$5.00 maximum API cost**. The agent reports actual spend in the session summary on exit.
+Each session has a **budget guard** to prevent runaway spend. Its form depends on how you authenticate with Anthropic:
+
+- **API-key billing** — the session enforces a dollar cap (default **$5.00**). Raise it with the `--max-budget=<USD>` flag or the `ADMIN_AGENT_MAX_BUDGET` environment variable (the flag wins when both are set). When estimated spend crosses ~80% of the cap, the agent warns on stderr and is instructed to save a partial report and wind down. If the cap is reached, the session ends with a message showing how to re-run with more headroom — and any report saved up to that point remains in `reports/`.
+- **OAuth (Claude Pro/Max subscription)** — no dollar cap is imposed (you are not billed per token). The session is bounded by the **turn limit** (100 turns) instead.
+
+The active guard is printed at startup, and the session summary reports per-investigation and total spend (API-key billing only). The dollar cap is enforced precisely by the Claude Agent SDK; the ~80% warning is an estimate from per-turn token usage, so it is approximate by design.
 
 ### Degraded Mode
 
@@ -225,9 +231,12 @@ admin-agent --login-method=TYPE   # Login method: claudeai (Pro/Max) or console
 admin-agent --login-org=UUID      # Target organization UUID for OAuth
 admin-agent --logout              # Log out from Anthropic account and exit
 admin-agent --model=NAME          # Override agent model (sonnet | haiku | opus); default: sonnet
+admin-agent --max-budget=USD      # Per-session budget cap in USD (API-key billing only); default: 5.00
 ```
 
-The `--model` flag swaps the primary model for a single session. `haiku` is cheaper and faster for simple triage; `opus` is slower and more expensive but produces deeper reasoning on complex investigations. The fallback model remains `haiku` regardless of the primary choice, so availability is unchanged.
+The `--model` flag swaps the primary model for a single session. `haiku` is cheaper and faster for simple triage; `opus` is slower and more expensive but produces deeper reasoning on complex investigations. The fallback model remains `haiku` regardless of the primary choice, so availability is unchanged. When you omit `--model` in an interactive terminal, the agent shows a startup picker (defaulting to `sonnet`); non-interactive runs use the default without prompting.
+
+The `--max-budget` flag sets the per-session dollar cap for API-key billing (see [Session Budget](#session-budget)). It overrides `ADMIN_AGENT_MAX_BUDGET` and has no effect under OAuth subscription billing, which is turn-limited instead.
 
 ## Tools
 
@@ -306,7 +315,7 @@ The agent is designed with defense-in-depth for database administration:
 - **Two-step approval for batch column alter** — `kinetica_alter_table_columns` requires the operator to select columns via a checklist, then confirm the exact SQL preview
 - **Audit trail** — every mutation logs a redacted audit line to stderr (EXECUTED/FAILED + fingerprinted input summary) and appears in the report's "Mutations Applied" table with before/after state
 - **Report scrubbing** — saved reports are scrubbed of URLs, auth headers, Basic/Bearer credentials, cookies, and passwords before writing to disk
-- **Budget cap** — $5.00 max API cost per session prevents runaway spend
+- **Budget guard** — a per-session dollar cap (default $5.00, configurable via `--max-budget` or `ADMIN_AGENT_MAX_BUDGET`) prevents runaway spend on API-key billing; OAuth subscription sessions are bounded by the turn limit instead
 
 To report a security vulnerability, please see [SECURITY.md](SECURITY.md). Do not open a public GitHub issue for security issues.
 
@@ -375,7 +384,7 @@ References provide domain knowledge (not diagnostic runbooks). Create a `.md` fi
 - `sql-create-index` — column index syntax, chunk skip index, when to use which
 - `version-quirks-7.2` — endpoint/property differences between 7.2.x and earlier releases
 
-> **Heads up — prompt budget:** all playbooks and references are front-loaded into a single system prompt at startup, so its token cost grows with the knowledge corpus. A startup tripwire (`agent/prompt-budget.ts`) prints the assembled prompt size under `DEBUG` and warns on stderr once it exceeds ~15,000 estimated tokens. Current baseline is ~13.4k tokens (6 playbooks + 9 references). If you add substantial knowledge and trip that warning, treat it as the cue to switch from "load everything" to keyword-based playbook selection.
+> **Heads up — prompt budget:** all playbooks and references are front-loaded into a single system prompt at startup, so its token cost grows with the knowledge corpus. A startup tripwire (`agent/prompt-budget.ts`) prints the assembled prompt size under `DEBUG` and warns on stderr once it exceeds ~20,000 estimated tokens. Current baseline is ~13.4k tokens (6 playbooks + 9 references). If you add substantial knowledge and trip that warning, treat it as the cue to switch from "load everything" to keyword-based playbook selection.
 
 ## Development
 
@@ -472,7 +481,7 @@ admin-agent
 
 **Agent hits budget cap**
 
-- Default is $5.00 per session. For complex multi-table investigations, consider running focused sessions per table
+- Applies to API-key billing only (default $5.00 per session). Raise it for the next run with `--max-budget=10` or `export ADMIN_AGENT_MAX_BUDGET=10`. The agent warns at ~80% so it can save a partial report before the cap is reached. For complex multi-table investigations, consider running focused sessions per table. OAuth (Pro/Max) sessions are turn-limited rather than dollar-capped.
 
 **Empty or missing report**
 

package/dist/admin-agent.js

@@ -111,7 +111,7 @@ function printBanner(model) {
   const version = getVersion();
   const subtitle = `admin-agent ${import_picocolors.default.dim(`v${version}`)}`;
   const header = model ? `${subtitle}
-${import_picocolors.default.dim(`model: ${model}`)}` : subtitle;
+${import_picocolors.default.dim(`Model: ${model}`)}` : subtitle;
   process.stderr.write("\n\n" + gradientize(LOGO) + "\n\n" + header + "\n");
   return subtitle;
 }
@@ -3550,6 +3550,13 @@ Monitor your context window usage during long investigations:
 - If you detect that context is getting full (many rounds, many large tool responses), warn the operator: "The session context is getting long. Consider starting a fresh session after this report to maintain investigation quality. Your reports are saved to disk."
 - Do NOT continue investigating when context is too full \u2014 write the report with evidence gathered so far.
 
+## Budget & Length Awareness
+
+The session has a per-session budget guard that can end the run before you finish \u2014 and the operator may see a warning that you are "approaching the budget guard". To make sure a diagnostic always survives an early cutoff:
+- During a long or expensive investigation (many rounds or many large tool responses), proactively call \`save_report\` with \`partial: true\` to checkpoint your findings so far. A partial report is far better than none.
+- If the operator warns you that the budget guard is approaching, STOP gathering new evidence: immediately save a \`partial: true\` report with the evidence you have, state your best current hypothesis, and wind down the turn.
+- Treat the guard as a normal limit, not an error \u2014 never apologize for it; just preserve the work.
+
 ---
 
 ## Output Formatting
@@ -3678,6 +3685,69 @@ function checkPromptBudget(prompt, opts) {
   };
 }
 
+// src/agent/session-budget.ts
+var DEFAULT_MAX_BUDGET_USD = 5;
+var DEFAULT_WARN_FRACTION = 0.8;
+var BUDGET_ENV_VAR = "ADMIN_AGENT_MAX_BUDGET";
+var MODEL_PRICING = {
+  sonnet: { inputPerMTok: 3, outputPerMTok: 15, cacheReadPerMTok: 0.3, cacheCreationPerMTok: 3.75 },
+  haiku: { inputPerMTok: 1, outputPerMTok: 5, cacheReadPerMTok: 0.1, cacheCreationPerMTok: 1.25 },
+  opus: { inputPerMTok: 15, outputPerMTok: 75, cacheReadPerMTok: 1.5, cacheCreationPerMTok: 18.75 }
+};
+function fromSdkUsage(raw) {
+  const u = raw ?? {};
+  return {
+    inputTokens: u.input_tokens,
+    outputTokens: u.output_tokens,
+    cacheReadInputTokens: u.cache_read_input_tokens,
+    cacheCreationInputTokens: u.cache_creation_input_tokens
+  };
+}
+function safeCount(value) {
+  return typeof value === "number" && Number.isFinite(value) && value > 0 ? value : 0;
+}
+function isValidBudget(value) {
+  return typeof value === "number" && Number.isFinite(value) && value > 0;
+}
+function estimateTurnCostUsd(usage, model) {
+  if (!usage) return 0;
+  const price = MODEL_PRICING[model];
+  const input5 = safeCount(usage.inputTokens);
+  const output = safeCount(usage.outputTokens);
+  const cacheRead = safeCount(usage.cacheReadInputTokens);
+  const cacheCreation = safeCount(usage.cacheCreationInputTokens);
+  return (input5 * price.inputPerMTok + output * price.outputPerMTok + cacheRead * price.cacheReadPerMTok + cacheCreation * price.cacheCreationPerMTok) / 1e6;
+}
+function resolveMaxBudgetUsd(flagValue, env = process.env) {
+  if (isValidBudget(flagValue)) return flagValue;
+  const raw = env[BUDGET_ENV_VAR];
+  if (raw !== void 0 && raw !== "") {
+    const parsed = Number(raw);
+    if (isValidBudget(parsed)) return parsed;
+  }
+  return DEFAULT_MAX_BUDGET_USD;
+}
+function createBudgetTracker(opts) {
+  const warnFraction = opts.warnFraction ?? DEFAULT_WARN_FRACTION;
+  const warnAt = opts.maxUsd * warnFraction;
+  let spent = 0;
+  let warned = false;
+  return {
+    add(usage, model) {
+      spent += estimateTurnCostUsd(usage, model);
+    },
+    spentUsd() {
+      return spent;
+    },
+    shouldWarn() {
+      return !warned && spent > warnAt;
+    },
+    markWarned() {
+      warned = true;
+    }
+  };
+}
+
 // src/report/save-report.ts
 var import_promises3 = require("fs/promises");
 var import_node_path4 = require("path");
@@ -3970,13 +4040,29 @@ function createSpinner() {
 
 // src/agent/run-agent.ts
 var MCP_SERVER_NAME = "kinetica-diagnostics";
+var SAVE_REPORT_TOOL_NAME = `mcp__${MCP_SERVER_NAME}__save_report`;
+function contentCallsSaveReport(content) {
+  if (!Array.isArray(content)) return false;
+  return content.some((block) => {
+    if (typeof block !== "object" || block === null) return false;
+    const { type, name } = block;
+    return type === "tool_use" && name === SAVE_REPORT_TOOL_NAME;
+  });
+}
+function formatCostSuffix(costUsd) {
+  return costUsd !== void 0 && costUsd > 0 ? ` Cost: $${costUsd.toFixed(4)}.` : "";
+}
+function formatMetricsLine(turns, durationMs, durationApiMs, costUsd) {
+  const durationSec = Math.round(durationMs / 1e3);
+  const apiPct = durationMs > 0 ? Math.round(durationApiMs / durationMs * 100) : 0;
+  return `Turns: ${turns}. Duration: ${durationSec}s (${apiPct}% API).${formatCostSuffix(costUsd)}`;
+}
 var EXIT_COMMANDS = /* @__PURE__ */ new Set(["exit", "quit", "end", "q"]);
 var SUPPORTED_MODELS = ["sonnet", "haiku", "opus"];
 var DEFAULT_AGENT_MODEL = "sonnet";
-var DEFAULT_MAX_BUDGET_USD = 5;
 var ALLOWED_TOOL_NAMES = [
   ...DIAGNOSTIC_TOOL_NAMES.map((name) => `mcp__${MCP_SERVER_NAME}__${name}`),
-  `mcp__${MCP_SERVER_NAME}__save_report`,
+  SAVE_REPORT_TOOL_NAME,
   `mcp__${MCP_SERVER_NAME}__${ALTER_TABLE_COLUMNS_TOOL_NAME}`
 ];
 var DISALLOWED_TOOLS = ["Bash", "Edit", "Write", "MultiEdit"];
@@ -4070,7 +4156,10 @@ async function displayDegradedStatus(session2) {
   }
   process.stderr.write("\n");
 }
-async function runAgent(session2, kineticaVersion, degraded, model) {
+async function runAgent(session2, kineticaVersion, degraded, model, runOptions) {
+  const authMethod = runOptions?.authMethod ?? "api_key";
+  const dollarCapped = authMethod === "api_key";
+  const resolvedBudgetUsd = runOptions?.maxBudgetUsd ?? DEFAULT_MAX_BUDGET_USD;
   const [catalogSchemas, playbooks, references] = await Promise.all([
     degraded ? Promise.resolve(void 0) : discoverCatalogSchemas(session2),
     loadPlaybooks(),
@@ -4086,7 +4175,7 @@ async function runAgent(session2, kineticaVersion, degraded, model) {
   const budget = checkPromptBudget(systemPrompt);
   if (process.env.DEBUG) {
     process.stderr.write(
-      import_picocolors8.default.dim(`system prompt: ~${budget.tokens} tokens (${budget.chars} chars)
+      import_picocolors8.default.dim(`System prompt: ~${budget.tokens} tokens (${budget.chars} chars)
 `)
     );
   }
@@ -4126,23 +4215,28 @@ async function runAgent(session2, kineticaVersion, degraded, model) {
     fallbackModel: "haiku",
     thinking: { type: "adaptive" },
     maxTurns: 100,
-    maxBudgetUsd: DEFAULT_MAX_BUDGET_USD,
+    // Only impose a dollar cap for per-token billing. For OAuth subscription users
+    // the SDK would otherwise cut them off at a notional dollar figure they never pay;
+    // omitting it leaves the turn limit (maxTurns) as their guard.
+    ...dollarCapped ? { maxBudgetUsd: resolvedBudgetUsd } : {},
     persistSession: false,
     includePartialMessages: true,
     abortController,
     env: { ...process.env, CLAUDE_AGENT_SDK_CLIENT_APP: "admin-agent" }
   };
+  const guardLine = dollarCapped ? import_picocolors8.default.dim(`Budget guard: $${resolvedBudgetUsd.toFixed(2)} (raise with --max-budget)
+`) : import_picocolors8.default.dim("Budget guard: subscription (Pro/Max) \u2014 turn-limited\n");
   if (degraded) {
     process.stderr.write("\nKinetica Diagnostic Session Ready (DEGRADED MODE)\n");
     process.stderr.write(
       "DB engine (port 9191) is unreachable. Only host manager tools are available.\n\n"
     );
     await displayDegradedStatus(session2);
-    process.stderr.write("Type 'exit' to end the session.\n\n");
   } else {
     process.stderr.write("\nKinetica Diagnostic Session Ready\n");
-    process.stderr.write("Type 'exit' to end the session.\n\n");
   }
+  process.stderr.write(guardLine);
+  process.stderr.write("Type 'exit' to end the session.\n\n");
   const turnGate = createTurnGate();
   const agentQuery = (0, import_claude_agent_sdk4.query)({
     prompt: makeInteractivePrompt(abortController, turnGate, spinner),
@@ -4163,6 +4257,9 @@ async function runAgent(session2, kineticaVersion, degraded, model) {
   let lastStreamCharWasNewline = true;
   const tableAligner = createStreamingTableAligner();
   let hadNonAbortError = false;
+  let reportSavedThisRun = false;
+  let invBase = { turns: 0, duration: 0, api: 0, cost: 0 };
+  const budgetTracker = dollarCapped ? createBudgetTracker({ maxUsd: resolvedBudgetUsd }) : void 0;
   try {
     for await (const message of agentQuery) {
       if (message.type === "stream_event") {
@@ -4189,6 +4286,23 @@ async function runAgent(session2, kineticaVersion, degraded, model) {
           process.stderr.write("\n");
           lastStreamCharWasNewline = true;
         }
+        if (budgetTracker) {
+          budgetTracker.add(fromSdkUsage(assistantMsg.message.usage), effectiveModel);
+          if (budgetTracker.shouldWarn()) {
+            spinner.stop();
+            process.stderr.write(
+              import_picocolors8.default.yellow(
+                `
+\u26A0 Approaching budget guard (~$${budgetTracker.spentUsd().toFixed(2)} / $${resolvedBudgetUsd.toFixed(2)}) \u2014 wrapping up soon. Save a partial report now if you want to preserve findings.
+`
+              )
+            );
+            budgetTracker.markWarned();
+          }
+        }
+        if (contentCallsSaveReport(assistantMsg.message.content)) {
+          reportSavedThisRun = true;
+        }
         if (assistantMsg.message.stop_reason === "end_turn") {
           spinner.stop();
           turnGate.open();
@@ -4217,10 +4331,21 @@ API error: ${label}
         cacheCreationTokens = usages.reduce((sum, u) => sum + (u.cacheCreationInputTokens ?? 0), 0);
         if (resultMsg.subtype === "error_max_turns") {
           process.stderr.write(
-            "\nInvestigation hit turn limit. Partial report may be available.\n"
+            import_picocolors8.default.yellow(
+              `
+Reached the turn limit (${numTurns} turns) \u2014 a safety guard, not an error. Any report the agent saved is in reports/. Start a fresh session to continue.
+`
+            )
           );
         } else if (resultMsg.subtype === "error_max_budget_usd") {
-          process.stderr.write("\nBudget limit reached.\n");
+          const spentStr = totalCostUsd > 0 ? ` ($${totalCostUsd.toFixed(2)} spent)` : "";
+          process.stderr.write(
+            import_picocolors8.default.yellow(
+              `
+Reached the $${resolvedBudgetUsd.toFixed(2)} budget guard${spentStr} \u2014 a safety limit, not an error. Re-run with --max-budget=<amount> (or set ADMIN_AGENT_MAX_BUDGET) for more headroom. Any report the agent saved is in reports/.
+`
+            )
+          );
         } else if (resultMsg.subtype === "error_during_execution") {
           process.stderr.write(
             "\nExecution error \u2014 the agent encountered an unrecoverable failure.\n"
@@ -4236,6 +4361,24 @@ Agent session ended with error: ${resultMsg.subtype}
 Permission denials: ${denied}
 `);
         }
+        if (reportSavedThisRun) {
+          const line = formatMetricsLine(
+            numTurns - invBase.turns,
+            durationMs - invBase.duration,
+            durationApiMs - invBase.api,
+            dollarCapped ? totalCostUsd - invBase.cost : void 0
+          );
+          process.stderr.write(`
+Investigation complete \u2014 ${line}
+`);
+          invBase = {
+            turns: numTurns,
+            duration: durationMs,
+            api: durationApiMs,
+            cost: totalCostUsd
+          };
+          reportSavedThisRun = false;
+        }
         turnGate.open();
       } else if (message.type === "system") {
         const sysMsg = message;
@@ -4307,27 +4450,26 @@ Agent error: ${message}
       process.stderr.write(remaining);
     }
     turnGate.open();
-    const durationSec = Math.round(durationMs / 1e3);
-    const apiPct = durationMs > 0 ? Math.round(durationApiMs / durationMs * 100) : 0;
-    const costStr = totalCostUsd > 0 ? ` Cost: $${totalCostUsd.toFixed(4)}.` : "";
+    const sessionCost = dollarCapped ? totalCostUsd : void 0;
     if (process.env.DEBUG && (cacheReadTokens > 0 || cacheCreationTokens > 0)) {
       process.stderr.write(
         import_picocolors8.default.dim(
-          `cache: ${cacheReadTokens} read / ${cacheCreationTokens} created input tokens (read > 0 confirms the system prompt is served from cache)
+          `Cache: ${cacheReadTokens} read / ${cacheCreationTokens} created input tokens (read > 0 confirms the system prompt is served from cache)
 `
         )
       );
     }
     if (hadNonAbortError) {
-      process.stderr.write(`
-Session ended due to error. Turns: ${numTurns}.${costStr}
-`);
-    } else {
       process.stderr.write(
         `
-Session ended. Turns: ${numTurns}. Duration: ${durationSec}s (${apiPct}% API).${costStr}
+Session ended due to error. Turns: ${numTurns}.${formatCostSuffix(sessionCost)}
 `
       );
+    } else {
+      const line = formatMetricsLine(numTurns, durationMs, durationApiMs, sessionCost);
+      process.stderr.write(`
+Session ended. ${line}
+`);
     }
   }
 }
@@ -4904,12 +5046,14 @@ function printHelp() {
     "    --login-org=UUID      Target organization UUID for OAuth",
     "    --logout              Log out from Anthropic account and exit",
     `    --model=NAME          Override agent model (${SUPPORTED_MODELS.join(" | ")}); default: sonnet`,
+    "    --max-budget=USD      Per-session budget cap in USD (API-key billing only); default: 5.00",
     "",
     "  Environment variables:",
-    "    ANTHROPIC_API_KEY  Anthropic API key (if not set, OAuth login via browser is used)",
-    "    KINETICA_URL       Kinetica endpoint URL",
-    "    KINETICA_USER      Admin username",
-    "    KINETICA_PASS      Admin password",
+    "    ANTHROPIC_API_KEY      Anthropic API key (if not set, OAuth login via browser is used)",
+    "    ADMIN_AGENT_MAX_BUDGET Per-session budget cap in USD (overridden by --max-budget)",
+    "    KINETICA_URL           Kinetica endpoint URL",
+    "    KINETICA_USER          Admin username",
+    "    KINETICA_PASS          Admin password",
     ""
   ];
   process.stdout.write(lines.join("\n") + "\n");
@@ -4954,13 +5098,30 @@ async function main() {
       return;
     }
   }
+  const budgetArg = args.find((a) => a.startsWith("--max-budget="));
+  const budgetValue = budgetArg?.split("=")[1];
+  let maxBudgetFlag;
+  if (budgetValue !== void 0) {
+    const parsed = Number(budgetValue);
+    if (!isValidBudget(parsed)) {
+      process.stderr.write(
+        import_picocolors14.default.red(
+          `Error: invalid --max-budget value "${budgetValue}". Use a positive number, e.g. --max-budget=10
+`
+        )
+      );
+      process.exitCode = 1;
+      return;
+    }
+    maxBudgetFlag = parsed;
+  }
   loadEnvFile();
   printBanner();
   if (model === void 0 && process.stdin.isTTY) {
     model = await selectModel();
   }
   const effectiveModel = model ?? DEFAULT_AGENT_MODEL;
-  process.stderr.write(import_picocolors14.default.dim(`model: ${effectiveModel}
+  process.stderr.write(import_picocolors14.default.dim(`Model: ${effectiveModel}
 `));
   const authResult = await authenticateAnthropic({ forceLogin, loginMethod, loginOrgUUID });
   if (authResult.method === "oauth") {
@@ -4970,9 +5131,13 @@ async function main() {
   } else {
     process.stderr.write(import_picocolors14.default.dim("Authenticated via API key\n"));
   }
+  const maxBudgetUsd = resolveMaxBudgetUsd(maxBudgetFlag);
   const { session: connectedSession, kineticaVersion, degraded } = await connectWithRetry();
   session = connectedSession;
-  await runAgent(session, kineticaVersion, degraded, model);
+  await runAgent(session, kineticaVersion, degraded, model, {
+    authMethod: authResult.method,
+    maxBudgetUsd
+  });
 }
 function getSession() {
   return session;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@kinetica/admin-agent",
-  "version": "0.1.2",
+  "version": "0.1.3",
   "description": "Autonomous diagnostic agent for Kinetica databases",
   "license": "Apache-2.0",
   "author": "Kinetica",