fix(miner): wire real dollar-cost tracking into the loop's budgetSpent

packages/gittensory-engine/src/miner/agent-sdk-driver.ts

@@ -144,6 +144,11 @@ export function createAgentSdkCodingAgentDriver(
 
       const turnsUsed =
         typeof resultMessage?.num_turns === "number" ? resultMessage.num_turns : undefined;
+      // Real dollar cost: the SDK's own SDKResultSuccess/SDKResultError message types both declare
+      // `total_cost_usd: number` unconditionally -- present whenever a result message arrived at all, success
+      // or not (the session was billed either way), absent only when the stream produced no result message.
+      const costUsd =
+        typeof resultMessage?.total_cost_usd === "number" ? resultMessage.total_cost_usd : undefined;
       const resultText =
         typeof resultMessage?.result === "string" ? redactSecrets(resultMessage.result) : "";
       const transcript = redactSecrets(
@@ -169,6 +174,7 @@ export function createAgentSdkCodingAgentDriver(
           summary: "agent sdk session did not complete successfully",
           transcript,
           turnsUsed,
+          costUsd,
           error: `agent_sdk_${subtype === "success" ? "errored" : subtype}`,
         };
       }
@@ -179,6 +185,7 @@ export function createAgentSdkCodingAgentDriver(
         summary: resultText.slice(0, MAX_REDACTED_TEXT_LENGTH) || `coding agent completed with ${changedFiles.size} changed file(s)`,
         transcript,
         turnsUsed,
+        costUsd,
       };
     },
   };

packages/gittensory-engine/src/miner/coding-agent-driver.ts

@@ -20,6 +20,9 @@ export type CodingAgentDriverResult = {
   /** Opaque provider transcript for operator inspection; absent when the driver did not run. */
   transcript?: string | undefined;
   turnsUsed?: number | undefined;
+  /** Real dollar cost of this driver run, when the provider reports one. Absent (not zero) when the provider
+   *  never got far enough to have a cost, or reports no cost signal at all -- never fabricated. */
+  costUsd?: number | undefined;
   error?: string | undefined;
 };
 

packages/gittensory-engine/src/miner/iterate-loop.ts

@@ -102,6 +102,10 @@ export type IterateLoopResult = {
   iterationsUsed: number;
   /** Cumulative `turnsUsed` summed across every iteration that ran. */
   totalTurnsUsed: number;
+  /** Cumulative real dollar cost summed across every iteration that ran, from each iteration's
+   *  `CodingAgentDriverResult.costUsd`. Only the `agent-sdk` provider reports this today (the CLI-subprocess
+   *  providers report no cost signal) -- always `0` for a provider that never reports one, never fabricated. */
+  totalCostUsd: number;
   iterations: readonly IterateLoopIterationRecord[];
   /** Populated only when `outcome === "handoff"`. */
   handoffPacket?: HandoffPacket | undefined;
@@ -228,7 +232,7 @@ function immediateAbandonNoIterationsPermitted(input: IterateLoopInput, deps: It
     reason: decision.reason,
     payload: { iterationNumber: 0, action: decision.action, abandonReason: decision.abandonReason },
   });
-  return { outcome: "abandon", finalDecision: decision, iterationsUsed: 0, totalTurnsUsed: 0, iterations: [] };
+  return { outcome: "abandon", finalDecision: decision, iterationsUsed: 0, totalTurnsUsed: 0, totalCostUsd: 0, iterations: [] };
 }
 
 /**
@@ -262,6 +266,7 @@ export async function runIterateLoop(input: IterateLoopInput, deps: IterateLoopD
   const iterations: IterateLoopIterationRecord[] = [];
   let previousBlockerCodes: readonly string[] | null = null;
   let totalTurnsUsed = 0;
+  let totalCostUsd = 0;
 
   for (let iterationNumber = 1; iterationNumber <= maxIterations; iterationNumber += 1) {
     const driverResult = await runDriverSafely(input, deps, {
@@ -272,6 +277,7 @@ export async function runIterateLoop(input: IterateLoopInput, deps: IterateLoopD
       maxTurns: input.maxTurnsPerIteration,
     });
     totalTurnsUsed += driverResult.turnsUsed ?? 0;
+    totalCostUsd += driverResult.costUsd ?? 0;
 
     const { outcome: selfReview, verdict } = evaluateSelfReviewOutcome(input, driverResult, deps);
 
@@ -296,12 +302,13 @@ export async function runIterateLoop(input: IterateLoopInput, deps: IterateLoopD
         finalDecision: decision,
         iterationsUsed: iterationNumber,
         totalTurnsUsed,
+        totalCostUsd,
         iterations,
         handoffPacket: buildHandoffPacket(input, verdict as SelfReviewVerdict, driverResult),
       };
     }
     if (decision.action === "abandon") {
-      return { outcome: "abandon", finalDecision: decision, iterationsUsed: iterationNumber, totalTurnsUsed, iterations };
+      return { outcome: "abandon", finalDecision: decision, iterationsUsed: iterationNumber, totalTurnsUsed, totalCostUsd, iterations };
     }
     previousBlockerCodes = blockerCodesFromContinuingOutcome(selfReview);
   }
@@ -313,5 +320,5 @@ export async function runIterateLoop(input: IterateLoopInput, deps: IterateLoopD
    * this package's fail-closed discipline, in case a future edit to the precedence ladder ever removes that
    * guarantee. */
   const fallbackDecision: IterateLoopDecision = { action: "abandon", abandonReason: "max_iterations_reached", reason: "Iterate loop exhausted its iteration budget." };
-  return { outcome: "abandon", finalDecision: fallbackDecision, iterationsUsed: maxIterations, totalTurnsUsed, iterations };
+  return { outcome: "abandon", finalDecision: fallbackDecision, iterationsUsed: maxIterations, totalTurnsUsed, totalCostUsd, iterations };
 }

packages/gittensory-miner/lib/attempt-cli.d.ts

@@ -38,6 +38,7 @@ export type AttemptCliResult =
       outcome: `attempt_${RunMinerAttemptResult["outcome"]}`;
       submissionMode: "observe" | "enforce";
       totalTurnsUsed: number;
+      totalCostUsd: number;
       iterationsUsed: number;
       reason?: string;
       decision?: unknown;

packages/gittensory-miner/lib/attempt-cli.js

@@ -374,10 +374,14 @@ export async function runAttempt(args, options = {}) {
       mode,
       attemptId,
       submissionMode: amsPolicy.spec.submissionMode,
-      // Every runMinerAttempt outcome carries a real loopResult (#5135's loop needs its genuine turn-usage to
-      // save real GovernorCapUsage via governor-state.js's saveCapUsage -- nothing else in the codebase calls
-      // it yet). Surfaced flat rather than the whole loopResult object, matching this result's own shallow shape.
+      // Every runMinerAttempt outcome carries a real loopResult (#5135's loop needs its genuine turn-usage and
+      // cost to save real GovernorCapUsage via governor-state.js's saveCapUsage -- nothing else in the codebase
+      // calls it yet). Surfaced flat rather than the whole loopResult object, matching this result's own
+      // shallow shape. costUsd is real only for the agent-sdk provider (its own SDK result message reports
+      // total_cost_usd); CLI-subprocess providers (claude-cli/codex-cli) report no cost signal today, so this
+      // is 0 for those -- an honest absence, not a fabricated number.
       totalTurnsUsed: result.loopResult.totalTurnsUsed,
+      totalCostUsd: result.loopResult.totalCostUsd,
       iterationsUsed: result.loopResult.iterationsUsed,
       ...("reason" in result ? { reason: result.reason } : {}),
       ...("decision" in result ? { decision: result.decision } : {}),

packages/gittensory-miner/lib/loop-cli.js

@@ -349,7 +349,11 @@ export async function runLoop(args, options = {}) {
       const cycleElapsedMs = nowMsFn() - cycleStartMs;
 
       usage = {
-        budgetSpent: usage.budgetSpent,
+        // Real for the agent-sdk provider (its own SDK result message reports total_cost_usd, wired through
+        // runMinerAttempt's real loopResult.totalCostUsd); the CLI-subprocess providers (claude-cli/codex-cli)
+        // report no cost signal today, so this contributes 0 for those runs -- an honest absence, not a
+        // fabricated number. A capLimits.budget dimension only ever meaningfully trips against agent-sdk spend.
+        budgetSpent: usage.budgetSpent + (lastResult?.totalCostUsd ?? 0),
         turnsTaken: usage.turnsTaken + (lastResult?.totalTurnsUsed ?? 0),
         elapsedMs: usage.elapsedMs + cycleElapsedMs,
       };

test/contract/coding-agent-driver-parity.test.ts

@@ -112,7 +112,7 @@ const DRIVER_HARNESSES: DriverHarness[] = [
               return;
             }
             yield { type: "assistant", message: { content: [{ type: "tool_use", name: "Edit", input: { file_path: "src/a.ts" } }] } };
-            yield { type: "result", subtype: "success", is_error: false, num_turns: 3, result: "edited src/a.ts" };
+            yield { type: "result", subtype: "success", is_error: false, num_turns: 3, total_cost_usd: 0.15, result: "edited src/a.ts" };
           })();
         },
       });
@@ -130,6 +130,7 @@ function expectDriverResultShape(result: CodingAgentDriverResult): void {
   expect(result.summary.length).toBeGreaterThan(0);
   if (result.transcript !== undefined) expect(typeof result.transcript).toBe("string");
   if (result.turnsUsed !== undefined) expect(typeof result.turnsUsed).toBe("number");
+  if (result.costUsd !== undefined) expect(typeof result.costUsd).toBe("number");
   if (result.error !== undefined) {
     expect(typeof result.error).toBe("string");
     expect(result.error.length).toBeGreaterThan(0);
@@ -229,4 +230,16 @@ describe("documented divergences, locked in explicitly", () => {
     const result = await driver.run(task);
     expect(result.turnsUsed).toBe(3);
   });
+
+  it("CLI driver reports no cost signal (a subprocess exposes none without --output-format json) — costUsd stays undefined", async () => {
+    const { driver } = DRIVER_HARNESSES[0]!.make("success");
+    const result = await driver.run(task);
+    expect(result.costUsd).toBeUndefined();
+  });
+
+  it("Agent-SDK driver reports the result frame's real total_cost_usd", async () => {
+    const { driver } = DRIVER_HARNESSES[1]!.make("success");
+    const result = await driver.run(task);
+    expect(result.costUsd).toBe(0.15);
+  });
 });

test/unit/miner-attempt-cli.test.ts

@@ -243,7 +243,7 @@ describe("runAttempt (#5132)", () => {
       outcome: "submitted",
       spec: { command: "gh pr create", cwd: worktreeResult.worktreePath, timeoutMs: 1000 },
       execResult: { code: 0 },
-      loopResult: { outcome: "handoff", totalTurnsUsed: 3, iterationsUsed: 2 },
+      loopResult: { outcome: "handoff", totalTurnsUsed: 3, totalCostUsd: 0.42, iterationsUsed: 2 },
     });
 
     const exitCode = await runAttempt(["acme/widgets", "7", "--miner-login", "alice", "--json"], {
@@ -270,6 +270,7 @@ describe("runAttempt (#5132)", () => {
       attemptId: "fixed-attempt-id",
       submissionMode: "observe",
       totalTurnsUsed: 3,
+      totalCostUsd: 0.42,
       iterationsUsed: 2,
       spec: { command: "gh pr create", cwd: worktreeResult.worktreePath, timeoutMs: 1000 },
       execResult: { code: 0 },

test/unit/miner-attempt-runner.test.ts

@@ -92,8 +92,8 @@ function driverReturning(result: CodingAgentDriverResult): CodingAgentDriver {
   return { async run() { return result; } };
 }
 
-function okDriverResult(changedFiles: string[] = ["src/upload.ts"], turnsUsed = 5): CodingAgentDriverResult {
-  return { ok: true, changedFiles, summary: "added retry logic", turnsUsed };
+function okDriverResult(changedFiles: string[] = ["src/upload.ts"], turnsUsed = 5, costUsd = 0.42): CodingAgentDriverResult {
+  return { ok: true, changedFiles, summary: "added retry logic", turnsUsed, costUsd };
 }
 
 // ── Governor "everything allows" fixture, mirroring test/unit/miner-governor-chokepoint.test.ts's own ────────
@@ -160,6 +160,8 @@ describe("runMinerAttempt (#2337) — the real create->review->gate->submit pipe
     expect(result.spec.command).toContain("'miner/attempt-1'");
     expect(result.execResult).toEqual({ ranAt: 10_000 });
     expect(result.loopResult.outcome).toBe("handoff");
+    // Real per-iteration driver costUsd summed into the loop result (#5135's loop needs this for budgetSpent).
+    expect(result.loopResult.totalCostUsd).toBe(0.42);
   });
 
   it("defaults the open_pr body to an empty string when the loop input never set one", async () => {
@@ -180,6 +182,20 @@ describe("runMinerAttempt (#2337) — the real create->review->gate->submit pipe
     expect(claimLedgerListClaims).not.toHaveBeenCalled();
   });
 
+  it("abandon mid-loop: a real iteration ran (and was billed) before the ceiling forced abandon", async () => {
+    // maxIterations: 1 with a driver that never produces a passing self-review runs ONE real iteration, then
+    // decideNextActionWithReason's own iterationNumber>=maxIterations check abandons -- a genuinely different
+    // code path from the maxIterations:0 case above (that one never invokes the driver at all).
+    const failingDriverResult: CodingAgentDriverResult = { ok: false, changedFiles: [], summary: "driver failed", turnsUsed: 2, costUsd: 0.11, error: "driver_error" };
+    const deps = baseDeps({ driver: driverReturning(failingDriverResult) });
+    const result = await runMinerAttempt(baseAttemptInput({ loopInput: passingLoopInput({ maxIterations: 1 }) }), deps);
+
+    expect(result.outcome).toBe("abandon");
+    // The failed iteration's real turnsUsed/costUsd still counted -- an abandoned attempt was still billed.
+    expect(result.loopResult.totalTurnsUsed).toBe(2);
+    expect(result.loopResult.totalCostUsd).toBe(0.11);
+  });
+
   it("stale: a superseded claim aborts before the submission-gate or governor ever run", async () => {
     const deps = baseDeps({ claimLedger: { listClaims: () => [{ repoFullName: "acme/widgets", issueNumber: 7, status: "released" }] } });
     const result = await runMinerAttempt(baseAttemptInput(), deps);

test/unit/miner-loop-cli.test.ts

@@ -209,6 +209,7 @@ describe("runLoop (#5135)", () => {
         attemptId: "loop-attempt-1",
         submissionMode: "observe",
         totalTurnsUsed: 4,
+        totalCostUsd: 0.37,
         iterationsUsed: 1,
         execResult: { action: "open_pr", stdout: "https://github.com/acme/widgets/pull/123\n", stderr: "", code: 0, timedOut: false },
       });
@@ -249,8 +250,9 @@ describe("runLoop (#5135)", () => {
     // The claimed item resolved to done (real success), not left in_progress or requeued.
     expect(after.portfolioQueue.listQueue()).toEqual([expect.objectContaining({ identifier: "issue:7", status: "done" })]);
 
-    // Real governor cap usage was saved using runAttempt's own real totalTurnsUsed, not fabricated.
+    // Real governor cap usage was saved using runAttempt's own real totalTurnsUsed/totalCostUsd, not fabricated.
     expect(after.governorState.loadCapUsage().turnsTaken).toBe(4);
+    expect(after.governorState.loadCapUsage().budgetSpent).toBe(0.37);
 
     // Re-entry actually fired (a real loop_reentry_decision event, reentered on a merged outcome).
     const reentryEvents = after.eventLedger.readEvents({}).filter((e) => e.type === "loop_reentry_decision");
@@ -304,6 +306,9 @@ describe("runLoop (#5135)", () => {
     // The run-loop boundary gate released the in-flight item back to 'queued' on the fresh halt.
     expect(after.portfolioQueue.listQueue()).toHaveLength(1);
     expect(after.portfolioQueue.listQueue()[0]).toMatchObject({ status: "queued" });
+    // No totalCostUsd on any of these results (mirrors a CLI-subprocess provider, which reports no cost signal
+    // today) -- budgetSpent stays honestly at 0 across all 3 real attempts, never fabricated.
+    expect(after.governorState.loadCapUsage().budgetSpent).toBe(0);
   });
 
   it("REGRESSION: a permanent (AI-usage-policy) block marks the item done instead of re-queuing it forever", async () => {