@tangle-network/agent-runtime 0.126.0 → 0.131.0

package/dist/{improvement-cycle-IJgCbKWQ.js → improvement-cycle-Csp38cWg.js}

@@ -1,15 +1,17 @@
 import { i as ConfigError } from "./errors-DEAvWQPy.js";
-import { $ as parseExactAgentProfile, G as agentCandidateProfileAsAgentProfile, J as candidateMaterializerHarness, K as applyExactAgentProfileDiff, Q as parseAgentCandidateProfileActivation, Y as createAgentCandidateProfileActivation, Z as omitUndefinedObjectFields, ct as omitTopLevelDigest, d as verifiedResourceTextByDigest, f as verifyAgentCandidateBundle, it as canonicalCandidateDocument, n as executePreparedAgentCandidate, nt as canonicalCandidateBytes, q as assertCandidateProfileBinding, rt as canonicalCandidateDigest$1, st as immutableCandidateValue, t as prepareAgentCandidateExecution, ut as verifyCanonicalCandidateDocument } from "./prepare--8EvLqCr.js";
-import { $t as concreteProfileModel, Jt as runLocalHarness, Wt as harnessInvocation, xt as runSettledCommand } from "./supervisor-DpjO0Gmy.js";
-import { E as optimizerMethod, a as defaultStructuralRolloutPolicy } from "./structural-rollout-zY0oqQzO.js";
-import { t as runAnalystLoop } from "./analyst-loop-DvSciOfB.js";
+import { n as streamAgentTurn, t as collectAgentTurn } from "./stream-agent-turn-rYgaOLO0.js";
+import { t as detachedSnapshot } from "./snapshot-CXiiuHhL.js";
+import { $ as parseExactAgentProfile, G as agentCandidateProfileAsAgentProfile, J as createAgentCandidateProfileActivation, K as applyExactAgentProfileDiff, Q as parseAgentCandidateProfileActivation, X as materializeAgentCandidateProfilePlan, Z as omitUndefinedObjectFields, ct as omitTopLevelDigest, d as verifiedResourceTextByDigest, f as verifyAgentCandidateBundle, it as canonicalCandidateDocument, n as executePreparedAgentCandidate, nt as canonicalCandidateBytes, q as assertCandidateProfileBinding, rt as canonicalCandidateDigest$1, st as immutableCandidateValue, t as prepareAgentCandidateExecution, ut as verifyCanonicalCandidateDocument } from "./prepare-DYWjVcPx.js";
+import { a as concreteProfileModel, t as assertExecutableAgentProfile } from "./model-policy-CqziaqS1.js";
+import { Et as runSettledCommand, N as createExecutor } from "./supervisor-CV6Jh28D.js";
+import { D as optimizerMethod, a as defaultStructuralRolloutPolicy } from "./structural-rollout-3uxVGcg2.js";
+import { t as runAnalystLoop } from "./analyst-loop-BE8cDs5Q.js";
 import { CostLedger, canonicalJson, makeProposalFinding } from "@tangle-network/agent-eval";
 import { campaignScenarioIdentity, campaignSplitDigestFromIdentities, compareOptimizationMethods, gitWorktreeAdapter, verifyCodeSurface } from "@tangle-network/agent-eval/campaign";
-import { AGENT_IMPROVEMENT_SOURCE_METADATA_KEY, agentCandidateMaterializationReceiptSchema, agentCandidateRunReceiptSchema, agentImprovementActivationSchema, agentImprovementProposalSchema, agentImprovementReviewSchema, agentImprovementSourceMetadata, agentImprovementSourceSchema, agentProfileDiffSchema, agentProfileImprovementArmSchema, agentProfileImprovementExecutionRefSchema, agentProfileModelHintsSchema, agentProfileSchema, candidateExecutionEvidenceSchema, changedProfileImprovementSurfaces, defineAgentProfileDiff, numbersApproximatelyEqual, sha256DigestSchema } from "@tangle-network/agent-interface";
+import { AGENT_IMPROVEMENT_SOURCE_METADATA_KEY, agentCandidateMaterializationReceiptSchema, agentCandidateRunReceiptSchema, agentImprovementActivationSchema, agentImprovementProposalSchema, agentImprovementReviewSchema, agentImprovementSourceMetadata, agentImprovementSourceSchema, agentProfileDiffSchema, agentProfileImprovementArmSchema, agentProfileImprovementExecutionRefSchema, agentProfileModelHintsSchema, agentProfileSchema, candidateExecutionEvidenceSchema, canonicalAgentProfileDigest, changedProfileImprovementSurfaces, defineAgentProfileDiff, numbersApproximatelyEqual, sha256DigestSchema } from "@tangle-network/agent-interface";
 import { createHash, randomUUID } from "node:crypto";
-import { applyWorkspacePlan, materializeCandidateProfile, materializeProfile } from "@tangle-network/agent-profile-materialize";
-import { existsSync, readFileSync, readdirSync, rmSync } from "node:fs";
-import { basename, join, resolve, sep } from "node:path";
+import { basename, join, resolve } from "node:path";
+import { existsSync, readFileSync, readdirSync } from "node:fs";
 import { spawnSync } from "node:child_process";
 import { assertProposalFindings } from "@tangle-network/agent-eval/analyst";
 import { measuredComparisonFromAgentProfileImprovementExperiment, measuredComparisonFromCandidateExperiment, runAgentProfileImprovementExperiment, runCandidateExperiment, sealAgentProfileImprovementExperiment, sealAgentProfileImprovementSuite, sealAgentProfileImprovementTask, sealCandidateExperiment, selfImprove, verifyAgentProfileImprovementExperimentComparison, verifyCandidateExperiment, verifyCandidateExperimentComparison } from "@tangle-network/agent-eval/contract";
@@ -17,18 +19,16 @@ import { measuredComparisonFromAgentProfileImprovementExperiment, measuredCompar
 /**
 *
 * `agenticGenerator` — the full-agentic `CandidateGenerator`. It runs a real
-* coding harness (claude / codex / opencode) inside the candidate worktree the
-* driver already created, letting the agent read the codebase + the research
-* proposal findings and make the change in place. The driver then commits the worktree
-* into a `CodeSurface`.
+* authored profile inside the candidate worktree the driver already created,
+* letting the agent read the codebase + the research proposal findings and
+* make the change in place. The driver then commits the worktree into a
+* `CodeSurface`.
 *
-* Mechanism: identical to the proven Phase-2.8 in-process executor — spawn the
-* harness as a subprocess with `cwd` = the worktree, on the same filesystem,
-* so edits land in place (no sandbox-mount round-trip). `runLocalHarness` is
-* the verified primitive. The OUTER sandbox is the improvement loop's own
-* execution context; the generator does not nest a second sandbox per
-* candidate (which would reintroduce a host↔sandbox worktree-transport
-* problem that does not need solving here).
+* Every paid shot enters Runtime through `createExecutor` + `streamAgentTurn`.
+* The caller supplies the exact `AgentProfile` and a worktree-aware executor
+* placement; Runtime validates and records the profile that actually ran. A
+* cli-bridge placement sets `cwd` to the supplied worktree so Pi and other
+* supported harnesses edit the existing candidate directly.
 *
 * `maxShots` is the DEPTH dial — a multi-shot verify-in-session loop, NOT the
 * kernel `runAgentRounds`. Each shot runs one full harness session in the (persistent)
@@ -36,32 +36,25 @@ import { measuredComparisonFromAgentProfileImprovementExperiment, measuredCompar
 *   - empty tree   → "you changed nothing, make the edits" → retry
 *   - dirty + `verify` fails → feed the verifier's failure into the next shot
 *       (the worktree persists, so the harness RESUMES atop its own failing
-*       edits with the error in hand — no `--resume` session plumbing needed,
-*       and harness-agnostic across claude/codex/opencode)
+*       edits with the error in hand — no session-specific retry path needed)
 *   - dirty + `verify` ok (or no verifier configured) → return the candidate
 * A candidate that never verifies within `maxShots` is discarded (`applied:
 * false`), never shipped — if you configured a verifier, a non-passing tree is
-* not a candidate. With no verifier the legacy behavior holds: first dirty shot
-* is the candidate.
+* not a candidate. With no verifier, the first dirty shot is the candidate.
 *
-* @experimental
+* @stable
 */
 const RAW_TRACE_ANALYST_ID = "raw-trace-distiller";
 const RAW_TRACE_AREA = "raw-trace-context";
 const RAW_TRACE_DIAGNOSIS_PATH = ".improve/raw-trace-diagnosis.md";
-/** Dedicated ephemeral root for generic author-profile files. Every declared
-* file must live below this root so cleanup cannot alter candidate-owned files. */
-const AGENTIC_PROFILE_RESOURCE_ROOT = ".agent-runtime-profile-resources";
-/** Full-agentic `CandidateGenerator` (the `shots=N, sandbox=on` setting): run a real coding harness inside the candidate worktree so the agent makes the change in place. */
-function agenticGenerator(opts = {}) {
-	const harness = opts.harness ?? "claude-code";
-	if (opts.codexReproducible && harness !== "codex") throw new Error("agenticGenerator: codexReproducible requires harness 'codex'");
-	if (opts.codexReproducible && !opts.profile) throw new Error("agenticGenerator: codexReproducible requires an explicit author profile");
-	if (opts.codexReadDeniedPaths && !opts.codexReproducible) throw new Error("agenticGenerator: codexReadDeniedPaths requires codexReproducible");
-	if (opts.maximumCharge && !opts.codexReproducible) throw new Error("agenticGenerator: maximumCharge requires codexReproducible");
-	const profileResourcePlan = opts.codexReproducible ? authorProfileResourcePlan(opts.profile) : null;
-	const buildPrompt = opts.buildPrompt ?? defaultBuildPrompt;
-	const run = opts.runHarness ?? runLocalHarness;
+/** Full-agentic `CandidateGenerator`: run an exact profiled author inside the existing candidate worktree. */
+function agenticGenerator(opts) {
+	const profile = detachedSnapshot(agentProfileSchema.parse(opts.profile), "agenticGenerator profile");
+	assertExecutableAgentProfile(profile, "agenticGenerator");
+	const harness = profile.harness;
+	const provider = profile.model.provider.trim();
+	const model = concreteProfileModel(profile);
+	const profileDigest = canonicalAgentProfileDigest(profile);
 	const dirty = opts.isDirty ?? worktreeDirty;
 	const verify = opts.verify;
 	return {
@@ -69,59 +62,40 @@ function agenticGenerator(opts = {}) {
 		proposesWithoutFindings: true,
 		async generate({ worktreePath, findings, maxShots, signal, generation, candidateIndex, costLedger, costPhase }) {
 			signal.throwIfAborted();
-			let reproducibleCostLedger;
-			if (opts.codexReproducible) {
-				if (!costLedger) throw new Error("agenticGenerator: reproducible Codex requires the run-wide CostLedger supplied by agent-eval");
-				reproducibleCostLedger = costLedger;
-			}
-			const basePrompt = appendProfileResourcePaths(buildPrompt({ findings }), profileResourcePlan);
+			if (opts.maximumCharge && !costLedger) throw new Error("agenticGenerator: maximumCharge requires the run-wide CostLedger supplied by agent-eval");
+			const basePrompt = opts.buildPrompt({ findings });
+			if (typeof basePrompt !== "string" || basePrompt.trim().length === 0) throw new Error("agenticGenerator: buildPrompt must return a non-empty string");
+			const executorConfig = opts.executorForWorktree(worktreePath);
+			if (executorConfig.backend === "bridge" && executorConfig.cwd !== worktreePath) throw new Error("agenticGenerator: bridge executor cwd must equal the candidate worktree path");
+			const factory = createExecutor(executorConfig);
 			const needsRawTraceEvidence = requiresRawTraceEvidence(findings);
 			const shots = Math.max(1, maxShots);
 			let attemptNote = "";
 			for (let shot = 0; shot < shots; shot++) {
 				signal.throwIfAborted();
 				const taskPrompt = attemptNote ? `${basePrompt}\n\n${attemptNote}` : basePrompt;
-				const invocation = opts.profile ? harnessInvocation(harness, opts.profile, taskPrompt, {
-					dangerouslySkipPermissions: true,
-					...opts.codexReproducible ? { codexReproducible: true } : {}
-				}) : void 0;
-				const exactPrompt = invocation?.prompt ?? taskPrompt;
-				const readDeniedPaths = typeof opts.codexReadDeniedPaths === "function" ? opts.codexReadDeniedPaths(worktreePath) : opts.codexReadDeniedPaths;
 				const startedAt = /* @__PURE__ */ new Date();
-				let harnessResult = null;
-				let profileWorkspaceReceipt = null;
+				let turn = null;
 				let costReceipt = null;
 				let costCallId = null;
 				let shotError = null;
 				try {
-					const execute = async (executionSignal) => {
-						harnessResult = await withAuthorProfileResources(profileResourcePlan, worktreePath, async (receipt) => {
-							profileWorkspaceReceipt = receipt;
-							const result = await run({
-								harness,
-								cwd: worktreePath,
-								taskPrompt,
-								...invocation ? { invocation: {
-									command: invocation.command,
-									args: invocation.args
-								} } : {},
-								dangerouslySkipPermissions: true,
-								...opts.codexReproducible ? { codexReproducible: true } : {},
-								...readDeniedPaths ? { codexReadDeniedPaths: readDeniedPaths } : {},
-								timeoutMs: opts.timeoutMs,
-								signal: executionSignal
-							});
-							harnessResult = result;
-							return result;
-						});
-						const failure = shotFailure(harnessResult, exactPrompt, opts.codexReproducible === true);
+					const execute = async (executionSignal, callId) => {
+						turn = await collectAgentTurn(streamAgentTurn({
+							kind: "executor",
+							profile,
+							factory
+						}, taskPrompt, {
+							signal: executionSignal,
+							...opts.timeoutMs !== void 0 ? { timeoutMs: opts.timeoutMs } : {},
+							...callId ? { callId } : {}
+						}));
+						const failure = shotFailure(turn);
 						if (failure) throw failure;
-						return harnessResult;
+						return turn;
 					};
-					if (reproducibleCostLedger) {
-						const model = opts.profile ? concreteProfileModel(opts.profile) : void 0;
-						if (!model) throw new Error("agenticGenerator: reproducible Codex requires profile.model.default");
-						const paid = await reproducibleCostLedger.runPaidCall({
+					if (costLedger) {
+						const paid = await costLedger.runPaidCall({
 							channel: "driver",
 							phase: costPhase ?? "search.proposal",
 							actor: `agentic-generator:${harness}`,
@@ -133,38 +107,40 @@ function agenticGenerator(opts = {}) {
 							},
 							signal,
 							...opts.maximumCharge ? { maximumCharge: opts.maximumCharge } : {},
-							execute,
-							receipt: (result) => costReceiptFromHarness(result, model),
-							receiptFromError: () => harnessResult?.usage ? costReceiptFromHarness(harnessResult, model) : void 0
+							execute: (executionSignal, callId) => execute(executionSignal, callId),
+							receipt: (result) => costReceiptFromTurn(result, model),
+							receiptFromError: () => turn ? costReceiptFromTurn(turn, model) : void 0
 						});
 						costCallId = paid.callId ?? null;
 						costReceipt = paid.receipt ?? null;
 						if (!paid.succeeded) throw paid.error;
-						harnessResult = paid.value;
-					} else harnessResult = await execute(signal);
+						turn = paid.value;
+					} else turn = await execute(signal);
 				} catch (cause) {
 					shotError = cause instanceof Error ? cause : new Error(String(cause));
 				}
-				const execution = shotExecutionSnapshot(harnessResult);
+				const execution = turn ? detachedSnapshot(turn, "agenticGenerator shot execution") : null;
 				const receipt = shotReceipt({
 					generation,
 					candidateIndex,
 					shot,
 					maxShots: shots,
 					harness,
-					profile: opts.profile,
-					prompt: exactPrompt,
+					provider,
+					model,
+					profile,
+					profileDigest,
+					prompt: taskPrompt,
 					startedAt,
 					completedAt: /* @__PURE__ */ new Date(),
 					result: execution,
-					profileWorkspaceReceipt,
 					costCallId,
 					costReceipt,
 					error: shotError
 				});
 				await emitShotReceipt(opts.onShotCompleted, receipt, execution, shotError);
 				signal.throwIfAborted();
-				if (!execution) throw new Error("agenticGenerator: author shot completed without a harness result");
+				if (!execution) throw new Error("agenticGenerator: author shot completed without a Runtime turn");
 				let worktreeChanged;
 				try {
 					worktreeChanged = dirty(worktreePath);
@@ -249,67 +225,6 @@ function agenticGenerator(opts = {}) {
 		}
 	};
 }
-function authorProfileResourcePlan(profile) {
-	const resources = profile.resources;
-	if (!resources) return null;
-	const unsupportedKinds = [
-		resources.tools?.length ? "tools" : null,
-		resources.skills?.length ? "skills" : null,
-		resources.agents?.length ? "agents" : null
-	].filter((kind) => kind !== null);
-	if (unsupportedKinds.length > 0) throw new Error(`agenticGenerator: reproducible Codex author resources support files only; unsupported: ${unsupportedKinds.join(", ")}`);
-	if (!resources.files || resources.files.length === 0) return null;
-	const plan = materializeProfile({
-		name: profile.name,
-		resources: { files: resources.files }
-	}, "codex");
-	if (plan.unsupported.length > 0) throw new Error(`agenticGenerator: author profile files could not be materialized: ${plan.unsupported.map((item) => item.reason).join("; ")}`);
-	if (Object.keys(plan.env).length > 0 || plan.flags.length > 0) throw new Error("agenticGenerator: generic author profile files unexpectedly changed spawn values");
-	const virtualRoot = resolve("/", AGENTIC_PROFILE_RESOURCE_ROOT);
-	const seen = /* @__PURE__ */ new Set();
-	for (const file of plan.files) {
-		const target = resolve("/", file.relPath);
-		if (!target.startsWith(`${virtualRoot}${sep}`)) throw new Error(`agenticGenerator: author profile file must be below ${AGENTIC_PROFILE_RESOURCE_ROOT}: ${file.relPath}`);
-		if (seen.has(target)) throw new Error(`agenticGenerator: duplicate author profile file path: ${file.relPath}`);
-		seen.add(target);
-	}
-	return plan;
-}
-function appendProfileResourcePaths(prompt, plan) {
-	if (!plan) return prompt;
-	return [
-		prompt,
-		"",
-		"Profile resource files available for this shot:",
-		...plan.files.map((file) => `- ${file.relPath}`)
-	].join("\n");
-}
-async function withAuthorProfileResources(plan, worktreePath, run) {
-	if (!plan) return run(null);
-	const rootPath = resolve(worktreePath, AGENTIC_PROFILE_RESOURCE_ROOT);
-	if (existsSync(rootPath)) throw new Error(`agenticGenerator: ephemeral author profile root already exists: ${AGENTIC_PROFILE_RESOURCE_ROOT}`);
-	let value;
-	let primaryError;
-	try {
-		value = await run(applyWorkspacePlan(plan, worktreePath));
-	} catch (cause) {
-		primaryError = cause;
-	}
-	let cleanupError;
-	try {
-		rmSync(rootPath, {
-			recursive: true,
-			force: true
-		});
-		if (existsSync(rootPath)) throw new Error(`agenticGenerator: ephemeral author profile root survived cleanup: ${AGENTIC_PROFILE_RESOURCE_ROOT}`);
-	} catch (cause) {
-		cleanupError = cause;
-	}
-	if (primaryError !== void 0 && cleanupError !== void 0) throw new AggregateError([primaryError, cleanupError], "agenticGenerator: author shot and profile resource cleanup both failed");
-	if (primaryError !== void 0) throw primaryError;
-	if (cleanupError !== void 0) throw cleanupError;
-	return value;
-}
 async function emitShotReceipt(callback, receipt, execution, primaryError) {
 	try {
 		await callback?.(receipt, execution);
@@ -339,92 +254,77 @@ async function rethrowShotSetupError(callback, receipt, worktreePath, stage, cau
 	}
 	throw cause;
 }
-function shotExecutionSnapshot(result) {
-	if (result === null) return null;
-	const usage = result.usage ? Object.freeze({ ...result.usage }) : void 0;
-	const evidence = result.evidence ? Object.freeze({
-		...result.evidence,
-		readDeniedPaths: Object.freeze([...result.evidence.readDeniedPaths]),
-		policy: Object.freeze({ ...result.evidence.policy })
-	}) : void 0;
-	return Object.freeze({
-		exitCode: result.exitCode,
-		stdout: result.stdout,
-		stderr: result.stderr,
-		killedBySignal: result.killedBySignal,
-		durationMs: result.durationMs,
-		timedOut: result.timedOut,
-		...result.aborted !== void 0 ? { aborted: result.aborted } : {},
-		...usage ? { usage } : {},
-		...evidence ? { evidence } : {}
-	});
-}
 function shotReceipt(input) {
 	const error = input.error ? {
 		name: input.error instanceof Error ? input.error.name : "Error",
 		message: input.error instanceof Error ? input.error.message : String(input.error)
 	} : null;
 	const result = input.result;
-	const costBasis = costBasisFor(input.costReceipt);
+	const costBasis = costBasisFor(input.costReceipt, result);
+	const costUsd = input.costReceipt?.costUsd ?? result?.usage.costUsd ?? result?.usage.estimatedCostUsd ?? null;
 	return {
 		generation: input.generation ?? null,
 		candidateIndex: input.candidateIndex ?? null,
 		shot: input.shot + 1,
 		maxShots: input.maxShots,
+		profileDigest: input.profileDigest,
 		harness: input.harness,
-		model: input.profile ? concreteProfileModel(input.profile) ?? null : null,
-		reasoningEffort: input.profile?.model?.reasoningEffort ?? null,
+		provider: input.provider,
+		model: input.model,
+		reasoningEffort: input.profile.model?.reasoningEffort ?? null,
 		promptSha256: sha256(input.prompt),
 		startedAt: input.startedAt.toISOString(),
 		completedAt: input.completedAt.toISOString(),
-		durationMs: result?.durationMs ?? input.completedAt.getTime() - input.startedAt.getTime(),
-		exitCode: result?.exitCode ?? null,
-		timedOut: result?.timedOut ?? false,
-		aborted: result?.aborted ?? false,
-		killedBySignal: result?.killedBySignal ?? null,
-		stdoutBytes: result ? Buffer.byteLength(result.stdout) : null,
-		stdoutSha256: result ? sha256(result.stdout) : null,
-		stderrBytes: result ? Buffer.byteLength(result.stderr) : null,
-		stderrSha256: result ? sha256(result.stderr) : null,
-		usage: result?.usage ? { ...result.usage } : null,
-		profileWorkspacePlanDigest: input.profileWorkspaceReceipt?.workspacePlanDigest ?? null,
-		profileWorkspaceFileCount: input.profileWorkspaceReceipt?.written.length ?? 0,
+		durationMs: input.completedAt.getTime() - input.startedAt.getTime(),
+		status: result?.status ?? null,
+		usage: result ? { ...result.usage } : null,
+		transportAttempts: result?.transportAttempts ?? null,
 		costCallId: input.costCallId,
 		costBasis,
-		costUsd: costBasis === "unknown" ? null : input.costReceipt?.costUsd ?? null,
+		costUsd: costBasis === "unknown" ? null : costUsd,
 		costUsdKnown: costBasis === "provider-reported",
-		evidence: result?.evidence ? {
-			...result.evidence,
-			readDeniedPaths: [...result.evidence.readDeniedPaths],
-			policy: { ...result.evidence.policy }
-		} : null,
 		error
 	};
 }
-function costBasisFor(receipt) {
-	if (receipt === null || receipt.costUnknown) return "unknown";
-	return receipt.actualCostUsd === void 0 ? "estimated-pricing" : "provider-reported";
-}
-function shotFailure(result, exactPrompt, codexReproducible) {
-	if (result.timedOut) return /* @__PURE__ */ new Error("agenticGenerator: author shot timed out");
-	if (result.aborted) return /* @__PURE__ */ new Error("agenticGenerator: author shot was cancelled by the caller");
-	if (result.killedBySignal) return /* @__PURE__ */ new Error(`agenticGenerator: author shot was killed by ${result.killedBySignal}`);
-	if (result.exitCode !== 0) return /* @__PURE__ */ new Error(`agenticGenerator: author shot exited with code ${String(result.exitCode)}`);
-	if (!codexReproducible) return null;
-	if (!result.usage || !result.evidence) return /* @__PURE__ */ new Error("agenticGenerator: reproducible Codex shot completed without usage or execution evidence");
-	const expectedPromptSha256 = sha256(exactPrompt).slice(7);
-	if (result.evidence.requestedPromptSha256 !== expectedPromptSha256) return /* @__PURE__ */ new Error("agenticGenerator: reproducible Codex prompt evidence does not match the exact authored prompt");
-	return null;
-}
-function costReceiptFromHarness(result, model) {
-	if (!result.usage) throw new Error("agenticGenerator: author shot did not report terminal token usage");
+function costBasisFor(receipt, turn) {
+	if (receipt) {
+		if (receipt.costUnknown) return "unknown";
+		return receipt.actualCostUsd === void 0 ? "estimated-pricing" : "provider-reported";
+	}
+	if (turn?.usage.usdKnown !== false && turn?.usage.costUsd !== void 0) return "provider-reported";
+	return turn?.usage.estimatedCostUsd !== void 0 ? "estimated-pricing" : "unknown";
+}
+function shotFailure(result) {
+	if (result.status === "completed") return null;
+	if (result.status === "aborted") return /* @__PURE__ */ new Error("agenticGenerator: author shot was cancelled by the caller");
+	return /* @__PURE__ */ new Error(`agenticGenerator: author shot failed${result.error?.message ? `: ${result.error.message}` : ""}`);
+}
+function costReceiptFromTurn(result, model) {
+	const tokensKnown = result.usage.tokensKnown !== false;
+	const actualCostUsd = result.usage.usdKnown === false ? void 0 : result.usage.costUsd;
+	const cachedTokens = promptCacheReadTokens(result.usage);
 	return {
 		model,
-		inputTokens: result.usage.inputTokens - result.usage.cachedInputTokens,
-		outputTokens: result.usage.outputTokens,
-		...result.usage.cachedInputTokens > 0 ? { cachedTokens: result.usage.cachedInputTokens } : {}
+		inputTokens: tokensKnown ? Math.max(0, result.usage.input - (cachedTokens ?? 0)) : 0,
+		outputTokens: tokensKnown ? result.usage.output : 0,
+		...tokensKnown ? {} : { usageUnknown: true },
+		...cachedTokens !== void 0 ? { cachedTokens } : {},
+		...result.usage.reasoningTokens !== void 0 ? { reasoningTokens: result.usage.reasoningTokens } : {},
+		...actualCostUsd !== void 0 ? { actualCostUsd } : result.usage.estimatedCostUsd !== void 0 ? { estimatedCostUsd: result.usage.estimatedCostUsd } : { costUnknown: true }
 	};
 }
+function promptCacheReadTokens(usage) {
+	const cache = usage.promptCache;
+	if (!cache) return void 0;
+	for (const field of [
+		"readTokens",
+		"cachedInputTokens",
+		"cacheReadInputTokens"
+	]) {
+		const value = cache[field];
+		if (typeof value === "number" && Number.isFinite(value) && value >= 0) return value;
+	}
+}
 function sha256(value) {
 	return `sha256:${createHash("sha256").update(value).digest("hex")}`;
 }
@@ -618,7 +518,7 @@ function copyImproveCost(cost) {
 * A `CandidateGenerator` edits an isolated checkout. This driver finalizes each
 * accepted edit as a `CodeSurface` and disposes rejected worktrees.
 *
-* @experimental
+* @stable
 */
 /** Build the code-only proposer used internally by `improve({ surface: 'code' })`. */
 function improvementDriver(opts) {
@@ -756,8 +656,7 @@ function normalizeRolloutPolicy(raw) {
 		k,
 		repairRounds,
 		testgen,
-		...typeof bag.diverse === "boolean" ? { diverse: bag.diverse } : {},
-		...typeof bag.temperature === "number" ? { temperature: bag.temperature } : {}
+		...typeof bag.diverse === "boolean" ? { diverse: bag.diverse } : {}
 	};
 }
 /** Stable serialization with fixed field order. */
@@ -766,8 +665,7 @@ function serializeRolloutPolicy(policy) {
 		k: policy.k,
 		repairRounds: policy.repairRounds,
 		testgen: policy.testgen,
-		...policy.diverse !== void 0 ? { diverse: policy.diverse } : {},
-		...policy.temperature !== void 0 ? { temperature: policy.temperature } : {}
+		...policy.diverse !== void 0 ? { diverse: policy.diverse } : {}
 	});
 }
 /** Read the persisted policy off the profile. `undefined` when the profile does
@@ -784,8 +682,7 @@ function applyRolloutPolicyToProfile(profile, policy) {
 		k: policy.k,
 		repairRounds: policy.repairRounds,
 		testgen: policy.testgen,
-		...policy.diverse !== void 0 ? { diverse: policy.diverse } : {},
-		...policy.temperature !== void 0 ? { temperature: policy.temperature } : {}
+		...policy.diverse !== void 0 ? { diverse: policy.diverse } : {}
 	};
 	return {
 		...candidate,
@@ -1075,7 +972,7 @@ function assertFailClosedResources(profile, surface) {
 *         cached-result.json              ← the cell's score + artifact ref
 *         <artifacts…>                    ← whatever the dispatch wrote
 *
-* @experimental
+* @stable
 */
 const ANALYST_ID = "raw-trace-distiller";
 /** A cell counts as "failing" below this mean composite (matches the default
@@ -1091,7 +988,7 @@ const PASS_THRESHOLD = .999;
 *   await improve({
 *     surface: 'code',
 *     findings: seedFindings,
-*     code: { repoRoot },
+*     code: { repoRoot, profile, executorForWorktree, buildPrompt },
 *     runDir: '/abs/run',                 // MUST be a real path — the traces live here
 *     analyzeGeneration: rawTraceDistiller(),
 *     scenarios, judge, agent,
@@ -1350,6 +1247,8 @@ async function discardPreparedBaseline(worktree, baselineWorktree, cause) {
 }
 /** Create a clean incumbent checkout and the candidate producer for a code run. */
 async function prepareCodeRun(code) {
+	const authorProfile = agentProfileSchema.parse(code.profile);
+	assertExecutableAgentProfile(authorProfile, "improve(code) author");
 	const baseRef = code.baseRef ?? "main";
 	const worktree = code.worktree ?? gitWorktreeAdapter({
 		repoRoot: code.repoRoot,
@@ -1364,10 +1263,13 @@ async function prepareCodeRun(code) {
 		let baselineDiscarded = false;
 		const managed = improvementDriver({
 			worktree,
-			generator: code.generator ?? agenticGenerator({
-				...code.harness ? { harness: code.harness } : {},
+			generator: code.generator ? code.generator : agenticGenerator({
+				profile: authorProfile,
+				executorForWorktree: code.executorForWorktree,
+				buildPrompt: code.buildPrompt,
 				...code.verify ? { verify: code.verify } : {},
-				...code.timeoutMs ? { timeoutMs: code.timeoutMs } : {}
+				...code.timeoutMs !== void 0 ? { timeoutMs: code.timeoutMs } : {},
+				...code.maximumCharge ? { maximumCharge: code.maximumCharge } : {}
 			}),
 			baseRef
 		});
@@ -2960,7 +2862,14 @@ function verifyCandidateExecutionEvidence(input, options) {
 	assertCapturedInput(materialization.benchmark.task, task, "benchmark task");
 	assertEvidenceMaterialDigest(plan, "candidate execution plan");
 	assertEvidenceMaterialDigest(materialization.profileActivation.profilePlan, "candidate profile plan");
-	const expectedProfilePlan = materializeCandidateProfile(bundle.profile, candidateMaterializerHarness(materialization.harness), { resolvedResources: options.resolvedResources });
+	const expectedProfilePlan = materializeAgentCandidateProfilePlan({
+		profile: bundle.profile,
+		harness: bundle.execution.harness,
+		launch: bundle.execution.launch,
+		workspace: bundle.execution.cwd.workspace,
+		workspaces: plan.material.workspaces,
+		resolvedResources: options.resolvedResources
+	});
 	const activation = parseAgentCandidateProfileActivation(materialization.profileActivation, materialization.profileActivation.profilePlan.digest);
 	const regeneratedActivation = createAgentCandidateProfileActivation(expectedProfilePlan, materialization.profileActivation.profilePlan);
 	if (activation.digest !== regeneratedActivation.digest) throw new Error("candidate profile activation does not match the experiment bundle");
@@ -2987,6 +2896,6 @@ function sameOrderedValues(left, right) {
 	return left.length === right.length && left.every((value, index) => value === right[index]);
 }
 //#endregion
-export { improve as A, agenticGenerator as B, agentImprovementTargetInput as C, buildAgentImprovementActivationTargets as D, assertProfileImprovementTargetsShareIdentity as E, normalizeRolloutPolicy as F, summarizeFindings as G, defaultBuildPrompt as H, parseRolloutPolicy as I, worktreeChangedPaths as K, serializeRolloutPolicy as L, rawTraceDistiller as M, ROLLOUT_POLICY_EXTENSION as N, isAgentImprovementProfileSurface as O, applyRolloutPolicyToProfile as P, structuralRolloutPolicyFromProfile as R, agentImprovementTargetDigest as S, agentProfileImprovementStateDigest as T, rawTraceEvidenceProblem as U, commandVerifier as V, requiresRawTraceEvidence as W, AGENT_IMPROVEMENT_PROFILE_SURFACES as _, executeAgentCandidateExperimentCell as a, agentImprovementProfileSurfaceDigest as b, requireSealedCandidateExperiment as c, verifyAgentImprovementActivation as d, verifyAgentImprovementProposal as f, optimizationActivationReceiptFromMetadata as g, createOptimizationActivationReceipt as h, createAgentImprovementProposal as i, withMethodRuntimeControls as j, isAgentProfileMeasuredSurface as k, reviewAgentImprovementProposal as l, verifyCandidateExecutionEvidence as m, createAgentImprovementActivation as n, proposeAgentImprovement as o, verifyAgentImprovementReview as p, createAgentImprovementMeasuredComparison as r, proposeAgentProfileImprovement as s, AgentCandidateExperimentCellExecutionError as t, runAgentCandidateExperiment as u, AGENT_PROFILE_MEASURED_SURFACES as v, agentImprovementTargetProfileDiffs as w, agentImprovementProfileSurfaceInput as x, agentImprovementProfileDiffs as y, AGENTIC_PROFILE_RESOURCE_ROOT as z };
+export { improve as A, commandVerifier as B, agentImprovementTargetInput as C, buildAgentImprovementActivationTargets as D, assertProfileImprovementTargetsShareIdentity as E, normalizeRolloutPolicy as F, parseRolloutPolicy as I, serializeRolloutPolicy as L, rawTraceDistiller as M, ROLLOUT_POLICY_EXTENSION as N, isAgentImprovementProfileSurface as O, applyRolloutPolicyToProfile as P, structuralRolloutPolicyFromProfile as R, agentImprovementTargetDigest as S, agentProfileImprovementStateDigest as T, defaultBuildPrompt as V, AGENT_IMPROVEMENT_PROFILE_SURFACES as _, executeAgentCandidateExperimentCell as a, agentImprovementProfileSurfaceDigest as b, requireSealedCandidateExperiment as c, verifyAgentImprovementActivation as d, verifyAgentImprovementProposal as f, optimizationActivationReceiptFromMetadata as g, createOptimizationActivationReceipt as h, createAgentImprovementProposal as i, withMethodRuntimeControls as j, isAgentProfileMeasuredSurface as k, reviewAgentImprovementProposal as l, verifyCandidateExecutionEvidence as m, createAgentImprovementActivation as n, proposeAgentImprovement as o, verifyAgentImprovementReview as p, createAgentImprovementMeasuredComparison as r, proposeAgentProfileImprovement as s, AgentCandidateExperimentCellExecutionError as t, runAgentCandidateExperiment as u, AGENT_PROFILE_MEASURED_SURFACES as v, agentImprovementTargetProfileDiffs as w, agentImprovementProfileSurfaceInput as x, agentImprovementProfileDiffs as y, agenticGenerator as z };
 
-//# sourceMappingURL=improvement-cycle-IJgCbKWQ.js.map
+//# sourceMappingURL=improvement-cycle-Csp38cWg.js.map