@mmnto/cli 1.101.2 → 1.102.0

package/dist/utils.js

@@ -3,8 +3,8 @@ import * as fs from 'node:fs';
 import * as os from 'node:os';
 import * as path from 'node:path';
 import dotenv from 'dotenv';
-import { ADMISSION_COMPLETION_ONLY, buildGroundingBundle, calculateDeterministicHash, CONFIG_FILES, maskSecrets, RUN_ARTIFACT_SCHEMA_VERSION, saveRunArtifact, TotemConfigError, TotemConfigSchema, TotemOrchestratorError, } from '@mmnto/totem';
-import { createOrchestrator, resolveOrchestrator } from './orchestrators/orchestrator.js';
+import { ADMISSION_COMPLETION_ONLY, buildGroundingBundle, calculateDeterministicHash, CONFIG_FILES, INVOCATION_FAILURE_ARTIFACT_SCHEMA_VERSION, INVOKE_MESSAGE_EVIDENCE_LIMIT_BYTES, INVOKE_STREAM_EVIDENCE_LIMIT_BYTES, maskSecrets, RUN_ARTIFACT_SCHEMA_VERSION, SAFE_PROVIDER_CODE_RE, sanitizeForTerminal, saveInvocationFailureArtifact, saveRunArtifact, TotemConfigError, TotemConfigSchema, TotemOrchestratorError, } from '@mmnto/totem';
+import { classifyInvokeFailure, createOrchestrator, OrchestratorInvokeError, resolveOrchestrator, toOrchestratorInvokeError, } from './orchestrators/orchestrator.js';
 import { bold, log } from './ui.js';
 // ─── Shared constants ────────────────────────────────────
 const TELEMETRY_FILE = 'telemetry.jsonl';
@@ -322,6 +322,231 @@ function buildResponseCacheHash(prompt, systemPrompt, qualifiedModel, contract)
     }
     return hash.digest('hex').slice(0, 16);
 }
+function takeUtf8Prefix(text, limitBytes) {
+    let retained = '';
+    let bytes = 0;
+    for (const character of text) {
+        const characterBytes = Buffer.byteLength(character, 'utf-8');
+        if (bytes + characterBytes > limitBytes)
+            break;
+        retained += character;
+        bytes += characterBytes;
+    }
+    return retained;
+}
+function takeUtf8Tail(text, limitBytes) {
+    const characters = Array.from(text);
+    const retainedReversed = [];
+    let bytes = 0;
+    for (let index = characters.length - 1; index >= 0; index--) {
+        const character = characters[index];
+        const characterBytes = Buffer.byteLength(character, 'utf-8');
+        if (bytes + characterBytes > limitBytes)
+            break;
+        retainedReversed.push(character);
+        bytes += characterBytes;
+    }
+    return retainedReversed.reverse().join('');
+}
+/**
+ * Convert bounded raw runtime text into persisted evidence. Masking and
+ * terminal sanitization happen before a second UTF-8 byte bound because a
+ * replacement may grow or shrink the retained text. A masking exception
+ * fails closed: no raw bytes cross the artifact boundary.
+ */
+export function persistRuntimeTextEvidence(runtime, customSecrets, limitBytes = INVOKE_STREAM_EVIDENCE_LIMIT_BYTES, masker = maskSecrets) {
+    let safeHead;
+    let safeTail;
+    try {
+        safeHead = sanitizeForTerminal(masker(sanitizeForTerminal(runtime.head), customSecrets));
+        safeTail =
+            runtime.tail === undefined
+                ? undefined
+                : sanitizeForTerminal(masker(sanitizeForTerminal(runtime.tail), customSecrets));
+        // totem-context: evidence masking fails closed by intentionally omitting all raw text and recording the typed omission marker below.
+    }
+    catch {
+        return {
+            encoding: 'utf-8',
+            head: '',
+            observedBytes: runtime.observedBytes,
+            retainedBytes: 0,
+            limitBytes,
+            truncated: false,
+            dlp: 'omitted-on-mask-failure',
+        };
+    }
+    const hasTail = safeTail !== undefined;
+    const headLimit = hasTail ? Math.floor(limitBytes / 2) : limitBytes;
+    const tailLimit = limitBytes - headLimit;
+    const head = takeUtf8Prefix(safeHead, headLimit);
+    const tail = safeTail === undefined ? undefined : takeUtf8Tail(safeTail, tailLimit);
+    const retainedBytes = Buffer.byteLength(head, 'utf-8') + Buffer.byteLength(tail ?? '', 'utf-8');
+    const postMaskTruncated = Buffer.byteLength(safeHead, 'utf-8') > headLimit ||
+        (safeTail !== undefined && Buffer.byteLength(safeTail, 'utf-8') > tailLimit);
+    return {
+        encoding: 'utf-8',
+        head,
+        ...(tail !== undefined ? { tail } : {}),
+        observedBytes: runtime.observedBytes,
+        retainedBytes,
+        limitBytes,
+        truncated: runtime.truncated || postMaskTruncated,
+        dlp: 'masked',
+    };
+}
+function persistRuntimeAttempts(attempts, customSecrets) {
+    return attempts.map((attempt, index) => {
+        const providerCode = persistProviderCode(attempt.providerCode, customSecrets);
+        return {
+            sequence: index + 1,
+            route: attempt.route,
+            provider: attempt.provider,
+            model: attempt.model,
+            status: attempt.status,
+            durationMs: attempt.durationMs,
+            ...(attempt.failureKind !== undefined ? { failureKind: attempt.failureKind } : {}),
+            ...(attempt.providerStatus !== undefined ? { providerStatus: attempt.providerStatus } : {}),
+            ...(providerCode !== undefined ? { providerCode } : {}),
+            ...(attempt.process !== undefined
+                ? {
+                    process: {
+                        exitCode: attempt.process.exitCode,
+                        signal: attempt.process.signal,
+                        timedOut: attempt.process.timedOut,
+                        ...(attempt.process.timeoutMs !== undefined
+                            ? { timeoutMs: attempt.process.timeoutMs }
+                            : {}),
+                        ...(attempt.process.stdout !== undefined
+                            ? { stdout: persistRuntimeTextEvidence(attempt.process.stdout, customSecrets) }
+                            : {}),
+                        ...(attempt.process.stderr !== undefined
+                            ? { stderr: persistRuntimeTextEvidence(attempt.process.stderr, customSecrets) }
+                            : {}),
+                    },
+                }
+                : {}),
+        };
+    });
+}
+function persistProviderCode(providerCode, customSecrets) {
+    if (providerCode === undefined || !SAFE_PROVIDER_CODE_RE.test(providerCode))
+        return undefined;
+    try {
+        return maskSecrets(providerCode, customSecrets) === providerCode ? providerCode : undefined;
+        // totem-context: provider codes are optional diagnostics; a masking failure intentionally omits the code rather than persisting an unsafe token.
+    }
+    catch {
+        return undefined;
+    }
+}
+/**
+ * Pre-bound provider-controlled terminal prose before DLP. Secret masking can
+ * expand retained text, so `persistRuntimeTextEvidence` applies the same cap a
+ * second time after masking; this first bound prevents unbounded regex work.
+ */
+export function runtimeMessageEvidence(message) {
+    const observedBytes = Buffer.byteLength(message, 'utf-8');
+    if (observedBytes <= INVOKE_MESSAGE_EVIDENCE_LIMIT_BYTES) {
+        return {
+            encoding: 'utf-8',
+            head: message,
+            observedBytes,
+            retainedBytes: observedBytes,
+            limitBytes: INVOKE_MESSAGE_EVIDENCE_LIMIT_BYTES,
+            truncated: false,
+        };
+    }
+    const headLimit = Math.floor(INVOKE_MESSAGE_EVIDENCE_LIMIT_BYTES / 2);
+    const tailLimit = INVOKE_MESSAGE_EVIDENCE_LIMIT_BYTES - headLimit;
+    const head = takeUtf8Prefix(message, headLimit);
+    const tail = takeUtf8Tail(message, tailLimit);
+    const retainedBytes = Buffer.byteLength(head, 'utf-8') + Buffer.byteLength(tail, 'utf-8');
+    return {
+        encoding: 'utf-8',
+        head,
+        tail,
+        observedBytes,
+        retainedBytes,
+        limitBytes: INVOKE_MESSAGE_EVIDENCE_LIMIT_BYTES,
+        truncated: true,
+    };
+}
+function buildArtifactSharedFields(args) {
+    const inputBundle = {
+        maskedPrompt: args.safePrompt,
+        ...(args.safeSystemPrompt !== undefined && args.safeSystemPrompt.length > 0
+            ? { maskedSystemPrompt: args.safeSystemPrompt }
+            : {}),
+        ...(args.artifact.diffScope !== undefined ? { diffScope: args.artifact.diffScope } : {}),
+        ...(args.artifact.specContract !== undefined
+            ? { specContract: args.artifact.specContract }
+            : {}),
+    };
+    return {
+        inputBundle,
+        inputHash: calculateDeterministicHash(inputBundle),
+        grounding: {
+            hash: args.artifact.groundingHash,
+            provenanceSummary: args.artifact.provenanceSummary,
+            ...(args.groundingBundle !== undefined ? { bundle: args.groundingBundle } : {}),
+        },
+        ...(args.outputContract !== undefined ||
+            args.contextPolicy !== undefined ||
+            args.runMetadata !== undefined
+            ? {
+                admission: {
+                    ...(args.outputContract !== undefined ? { outputContract: args.outputContract } : {}),
+                    ...(args.contextPolicy !== undefined ? { contextPolicy: args.contextPolicy } : {}),
+                    ...(args.runMetadata !== undefined ? { runMetadata: args.runMetadata } : {}),
+                },
+            }
+            : {}),
+    };
+}
+function buildArtifactBackend(args) {
+    return {
+        provider: args.provider,
+        model: args.model,
+        qualifiedModel: args.qualifiedModel,
+        admissionClass: args.admissionClass,
+        taskProfile: args.taskProfile,
+        ...(args.temperature !== undefined ? { temperature: args.temperature } : {}),
+    };
+}
+function runtimeAttemptForError(args) {
+    return {
+        sequence: 1,
+        route: args.route,
+        provider: args.provider,
+        model: args.model,
+        status: 'failed',
+        durationMs: 0,
+        failureKind: classifyInvokeFailure(args.err),
+        ...(args.err instanceof Error &&
+            'status' in args.err &&
+            typeof args.err.status === 'number'
+            ? { providerStatus: args.err.status }
+            : {}),
+        ...(args.err instanceof Error &&
+            'code' in args.err &&
+            typeof args.err.code === 'string'
+            ? { providerCode: args.err.code }
+            : {}),
+    };
+}
+function runtimeAttemptsForError(err, provider, model, route) {
+    if (err instanceof OrchestratorInvokeError && err.attempts.length > 0) {
+        return err.attempts.map((attempt) => ({ ...attempt }));
+    }
+    return [runtimeAttemptForError({ err, provider, model, route })];
+}
+function resequenceRuntimeAttempts(attempts) {
+    return attempts.map((attempt, index) => ({ ...attempt, sequence: index + 1 }));
+}
+function quotaFallbackAttempts(attempts) {
+    return attempts.map((attempt) => ({ ...attempt, route: 'quota-model-fallback' }));
+}
 /**
  * Assemble the grounding bundle for the spec/review retrieval shape
  * (mmnto-ai/totem#2101): every partition's items enter under their partition
@@ -502,7 +727,7 @@ export async function runOrchestrator(opts) {
     let resolved = resolveOrchestrator(rawModel, baseProvider, baseInvoke);
     let model = resolved.parsed.model;
     let qualifiedModel = resolved.qualifiedModel;
-    let invoke = resolved.invoke;
+    const invoke = resolved.invoke;
     // ── Admission gate, primary path (mmnto-ai/totem#2102) ──
     // Decided per RESOLVED backend BEFORE the invoke (and before the response
     // cache: a denied class must not be served a replay either). No tokens are
@@ -598,78 +823,156 @@ export async function runOrchestrator(opts) {
         ...(opts.outputContract !== undefined ? { outputContract: opts.outputContract } : {}),
         ...(opts.runMetadata !== undefined ? { runMetadata: opts.runMetadata } : {}),
     };
+    const requestedBackend = buildArtifactBackend({
+        provider: resolved.parsed.provider,
+        model,
+        qualifiedModel,
+        admissionClass: requestedAdmissionClass,
+        taskProfile,
+        ...(opts.temperature !== undefined ? { temperature: opts.temperature } : {}),
+    });
     let result;
+    const primaryStartMs = Date.now();
     try {
-        result = await invoke({
-            prompt: safePrompt,
-            ...(safeSystemPrompt !== undefined ? { systemPrompt: safeSystemPrompt } : {}),
-            model,
-            cwd,
-            tag,
-            totemDir: config.totemDir,
-            temperature: opts.temperature,
-            ...(enableContextCaching !== undefined ? { enableContextCaching } : {}),
-            ...(cacheTTL !== undefined ? { cacheTTL } : {}),
-            ...admissionTransport,
-        });
-    }
-    catch (err) {
-        if (err instanceof Error && err.name === 'QuotaError') {
+        try {
+            result = await invoke({
+                prompt: safePrompt,
+                ...(safeSystemPrompt !== undefined ? { systemPrompt: safeSystemPrompt } : {}),
+                model,
+                cwd,
+                tag,
+                totemDir: config.totemDir,
+                temperature: opts.temperature,
+                ...(enableContextCaching !== undefined ? { enableContextCaching } : {}),
+                ...(cacheTTL !== undefined ? { cacheTTL } : {}),
+                ...admissionTransport,
+            });
+        }
+        catch (err) {
+            const primaryErr = toOrchestratorInvokeError({
+                err,
+                provider: resolved.parsed.provider,
+                model,
+                route: 'sdk',
+                durationMs: Date.now() - primaryStartMs,
+            });
+            if (primaryErr.kind !== 'quota')
+                throw primaryErr;
+            const primaryAttempts = runtimeAttemptsForError(primaryErr, resolved.parsed.provider, model, 'sdk');
             const rawFallback = config.orchestrator.fallbackModel;
-            if (rawFallback && rawModel !== rawFallback) {
-                log.warn(tag, `Quota exhausted for ${rawModel}. Retrying with fallback model: ${bold(rawFallback)}...`);
-                const fallbackResolved = resolveOrchestrator(rawFallback, baseProvider, baseInvoke);
-                // ── Admission gate, fallback path (mmnto-ai/totem#2102) ──
-                // `resolveOrchestrator` can route a provider-qualified fallbackModel
-                // to a DIFFERENT provider, and a single config-level declaration
-                // cannot honestly cover backends with different real capabilities.
-                // Slice-3 rule, conservative and deterministic: an elevated class
-                // admits the fallback only when it resolves to the SAME provider as
-                // the primary — cross-provider fails loud BEFORE the fallback invoke,
-                // reporting the primary and admission errors together. Per-provider
-                // capability declarations are the future relaxation.
-                if (requestedAdmissionClass !== ADMISSION_COMPLETION_ONLY &&
-                    fallbackResolved.parsed.provider !== resolved.parsed.provider) {
-                    throw new TotemOrchestratorError(`Primary model '${rawModel}' failed and the quota fallback '${rawFallback}' was denied admission.\n\n` +
-                        `Primary error:\n${err.message}\n\n` +
-                        `Admission error:\nfallback resolves to provider '${fallbackResolved.parsed.provider}' (primary: '${resolved.parsed.provider}') while admission class '${requestedAdmissionClass}' is requested — a cross-provider fallback is not admitted above '${ADMISSION_COMPLETION_ONLY}'.`, 'Use a same-provider fallbackModel, or drop the elevated backendAdmissionClass request.', err);
-                }
-                try {
-                    result = await fallbackResolved.invoke({
-                        prompt: safePrompt,
-                        ...(safeSystemPrompt !== undefined ? { systemPrompt: safeSystemPrompt } : {}),
+            if (!rawFallback || rawModel === rawFallback) {
+                if (err instanceof OrchestratorInvokeError)
+                    throw primaryErr;
+                throw new OrchestratorInvokeError(`Quota exhausted for ${model}.`, 'quota', resequenceRuntimeAttempts(primaryAttempts), {
+                    cause: err,
+                    recoveryHint: 'Wait for quota to reset, configure orchestrator.fallbackModel, or select another model.',
+                });
+            }
+            log.warn(tag, `Quota exhausted for ${rawModel}. Retrying with fallback model: ${bold(rawFallback)}...`);
+            const fallbackResolved = resolveOrchestrator(rawFallback, baseProvider, baseInvoke);
+            // ── Admission gate, fallback path (mmnto-ai/totem#2102) ──
+            // `resolveOrchestrator` can route a provider-qualified fallbackModel
+            // to a DIFFERENT provider, and a single config-level declaration
+            // cannot honestly cover backends with different real capabilities.
+            // Slice-3 rule, conservative and deterministic: an elevated class
+            // admits the fallback only when it resolves to the SAME provider as
+            // the primary — cross-provider fails loud BEFORE the fallback invoke.
+            if (requestedAdmissionClass !== ADMISSION_COMPLETION_ONLY &&
+                fallbackResolved.parsed.provider !== resolved.parsed.provider) {
+                throw new OrchestratorInvokeError(`Primary model '${rawModel}' failed and the quota fallback '${rawFallback}' was denied admission.\n\n` +
+                    `Primary error:\n${primaryErr.message}\n\n` +
+                    `Admission error:\nfallback resolves to provider '${fallbackResolved.parsed.provider}' (primary: '${resolved.parsed.provider}') while admission class '${requestedAdmissionClass}' is requested — a cross-provider fallback is not admitted above '${ADMISSION_COMPLETION_ONLY}'.`, 'quota', resequenceRuntimeAttempts(primaryAttempts), {
+                    cause: primaryErr,
+                    recoveryHint: 'Use a same-provider fallbackModel, or drop the elevated backendAdmissionClass request.',
+                });
+            }
+            const fallbackStartMs = Date.now();
+            try {
+                const fallbackResult = await fallbackResolved.invoke({
+                    prompt: safePrompt,
+                    ...(safeSystemPrompt !== undefined ? { systemPrompt: safeSystemPrompt } : {}),
+                    model: fallbackResolved.parsed.model,
+                    cwd,
+                    tag,
+                    totemDir: config.totemDir,
+                    temperature: opts.temperature,
+                    ...(enableContextCaching !== undefined ? { enableContextCaching } : {}),
+                    ...(cacheTTL !== undefined ? { cacheTTL } : {}),
+                    ...admissionTransport,
+                });
+                const fallbackAttempts = quotaFallbackAttempts(fallbackResult.attempts ?? [
+                    {
+                        sequence: 1,
+                        route: 'quota-model-fallback',
+                        provider: fallbackResolved.parsed.provider,
                         model: fallbackResolved.parsed.model,
-                        cwd,
-                        tag,
-                        totemDir: config.totemDir,
-                        temperature: opts.temperature,
-                        ...(enableContextCaching !== undefined ? { enableContextCaching } : {}),
-                        ...(cacheTTL !== undefined ? { cacheTTL } : {}),
-                        ...admissionTransport,
-                    });
-                    // Update model/invoke so telemetry and cache log the correct values
-                    model = fallbackResolved.parsed.model;
-                    qualifiedModel = fallbackResolved.qualifiedModel;
-                    resolved = fallbackResolved;
-                    invoke = fallbackResolved.invoke;
-                }
-                catch (fallbackErr) {
-                    const originalMsg = err.message;
-                    const fallbackMsg = fallbackErr instanceof Error ? fallbackErr.message : String(fallbackErr);
-                    throw new TotemOrchestratorError(`Primary model '${rawModel}' failed and fallback model '${rawFallback}' also failed.\n\n` +
-                        `Primary error:\n${originalMsg}\n\nFallback error:\n${fallbackMsg}`, 'Check API quotas and model availability, or try a different model with --model.', fallbackErr);
-                }
+                        status: 'succeeded',
+                        durationMs: fallbackResult.durationMs,
+                    },
+                ]);
+                result = {
+                    ...fallbackResult,
+                    attempts: resequenceRuntimeAttempts([...primaryAttempts, ...fallbackAttempts]),
+                };
+                // Update the resolved backend identity so telemetry, cache, and success
+                // artifacts log the backend that actually produced the semantic output.
+                model = fallbackResolved.parsed.model;
+                qualifiedModel = fallbackResolved.qualifiedModel;
+                resolved = fallbackResolved;
             }
-            else {
-                throw new TotemOrchestratorError(`Quota exhausted for ${model}.`, 'Quota resets on a rolling daily window. Options:\n' +
-                    '  - Switch to a flash model: totem <command> --model <name>\n' +
-                    '  - Inspect the prompt without calling the API: totem <command> --raw\n' +
-                    '  - Set a fallbackModel in totem.config.ts');
+            catch (fallbackErr) {
+                const normalizedFallbackErr = toOrchestratorInvokeError({
+                    err: fallbackErr,
+                    provider: fallbackResolved.parsed.provider,
+                    model: fallbackResolved.parsed.model,
+                    route: 'quota-model-fallback',
+                    durationMs: Date.now() - fallbackStartMs,
+                });
+                const fallbackAttempts = quotaFallbackAttempts(runtimeAttemptsForError(normalizedFallbackErr, fallbackResolved.parsed.provider, fallbackResolved.parsed.model, 'quota-model-fallback'));
+                const attempts = resequenceRuntimeAttempts([...primaryAttempts, ...fallbackAttempts]);
+                const kind = normalizedFallbackErr.kind;
+                const fallbackMsg = normalizedFallbackErr.message;
+                throw new OrchestratorInvokeError(`Primary model '${rawModel}' failed and fallback model '${rawFallback}' also failed.\n\n` +
+                    `Primary error:\n${primaryErr.message}\n\nFallback error:\n${fallbackMsg}`, kind, attempts, { cause: fallbackErr });
             }
         }
-        else {
-            throw err;
+    }
+    catch (err) {
+        if (opts.artifact !== undefined && err instanceof OrchestratorInvokeError) {
+            try {
+                const attempts = persistRuntimeAttempts(err.attempts, opts.customSecrets);
+                const shared = buildArtifactSharedFields({
+                    artifact: opts.artifact,
+                    safePrompt,
+                    ...(safeSystemPrompt !== undefined ? { safeSystemPrompt } : {}),
+                    ...(groundingBundle !== undefined ? { groundingBundle } : {}),
+                    ...(opts.outputContract !== undefined ? { outputContract: opts.outputContract } : {}),
+                    ...(opts.contextPolicy !== undefined ? { contextPolicy: opts.contextPolicy } : {}),
+                    ...(opts.runMetadata !== undefined ? { runMetadata: opts.runMetadata } : {}),
+                });
+                const failureArtifact = {
+                    schemaVersion: INVOCATION_FAILURE_ARTIFACT_SCHEMA_VERSION,
+                    ...shared,
+                    requestedBackend,
+                    attempts,
+                    terminal: {
+                        kind: err.kind,
+                        attempt: attempts.at(-1)?.sequence ?? 1,
+                        message: persistRuntimeTextEvidence(runtimeMessageEvidence(err.message), opts.customSecrets, INVOKE_MESSAGE_EVIDENCE_LIMIT_BYTES),
+                    },
+                    createdAt: new Date().toISOString(),
+                };
+                const saved = saveInvocationFailureArtifact(path.join(configRoot, config.totemDir), failureArtifact);
+                err.failureArtifactHash = saved.hash;
+                log.dim(tag, `Invocation failure artifact ${saved.existed ? 'already recorded' : 'recorded'}: ${saved.hash.slice(0, 12)}…`);
+                opts.artifact.onFailureEmitted?.(saved.hash, saved.path);
+                // totem-context: companion failure evidence is warn-only by contract; preserve and rethrow the original invocation error after surfacing this emission failure.
+            }
+            catch (artifactErr) {
+                log.warn(tag, `Invocation failure artifact emission failed (original error preserved): ${artifactErr instanceof Error ? artifactErr.message : String(artifactErr)}`);
+            }
         }
+        throw err;
     }
     if (useCache && result.content && result.durationMs > 0) {
         try {
@@ -699,39 +1002,26 @@ export async function runOrchestrator(opts) {
     // degradation is warned, never silent).
     if (opts.artifact !== undefined) {
         try {
-            const inputBundle = {
-                maskedPrompt: safePrompt,
-                ...(safeSystemPrompt !== undefined && safeSystemPrompt.length > 0
-                    ? { maskedSystemPrompt: safeSystemPrompt }
-                    : {}),
-                ...(opts.artifact.diffScope !== undefined ? { diffScope: opts.artifact.diffScope } : {}),
-                ...(opts.artifact.specContract !== undefined
-                    ? { specContract: opts.artifact.specContract }
-                    : {}),
-            };
+            const shared = buildArtifactSharedFields({
+                artifact: opts.artifact,
+                safePrompt,
+                ...(safeSystemPrompt !== undefined ? { safeSystemPrompt } : {}),
+                ...(groundingBundle !== undefined ? { groundingBundle } : {}),
+                ...(opts.outputContract !== undefined ? { outputContract: opts.outputContract } : {}),
+                ...(opts.contextPolicy !== undefined ? { contextPolicy: opts.contextPolicy } : {}),
+                ...(opts.runMetadata !== undefined ? { runMetadata: opts.runMetadata } : {}),
+            });
             const runArtifact = {
                 schemaVersion: RUN_ARTIFACT_SCHEMA_VERSION,
-                inputBundle,
-                inputHash: calculateDeterministicHash(inputBundle),
-                grounding: {
-                    hash: opts.artifact.groundingHash,
-                    provenanceSummary: opts.artifact.provenanceSummary,
-                    // Verbatim passthrough — the bundle was assembled and hashed by the
-                    // caller (mmnto-ai/totem#2101); this seam records, never re-derives
-                    // or upgrades. Post-reconciliation (#2102): a caller-supplied
-                    // `groundingBundle` serves this role when `artifact.bundle` is absent.
-                    ...(groundingBundle !== undefined ? { bundle: groundingBundle } : {}),
-                },
-                backend: {
+                ...shared,
+                backend: buildArtifactBackend({
                     provider: resolved.parsed.provider,
                     model,
                     qualifiedModel,
-                    // #2102: caller-supplied wins; the default reproduces the slice-1
-                    // constant — every undeclared backend is factually completion-only.
                     admissionClass: requestedAdmissionClass,
                     taskProfile,
                     ...(opts.temperature !== undefined ? { temperature: opts.temperature } : {}),
-                },
+                }),
                 output: {
                     content: result.content,
                     metrics: {
@@ -743,23 +1033,14 @@ export async function runOrchestrator(opts) {
                         durationMs: result.durationMs,
                         ...(result.finishReason !== undefined ? { finishReason: result.finishReason } : {}),
                     },
+                    ...(result.attempts !== undefined
+                        ? {
+                            execution: {
+                                attempts: persistRuntimeAttempts(result.attempts, opts.customSecrets),
+                            },
+                        }
+                        : {}),
                 },
-                // #2102: the admitted contract group is recorded ONLY when the caller
-                // supplied at least one member — an omitted contract stays an omitted
-                // key, never an empty object (additive 1.x, slice-1 artifacts unchanged).
-                ...(opts.outputContract !== undefined ||
-                    opts.contextPolicy !== undefined ||
-                    opts.runMetadata !== undefined
-                    ? {
-                        admission: {
-                            ...(opts.outputContract !== undefined
-                                ? { outputContract: opts.outputContract }
-                                : {}),
-                            ...(opts.contextPolicy !== undefined ? { contextPolicy: opts.contextPolicy } : {}),
-                            ...(opts.runMetadata !== undefined ? { runMetadata: opts.runMetadata } : {}),
-                        },
-                    }
-                    : {}),
                 createdAt: new Date().toISOString(),
             };
             const saved = saveRunArtifact(path.join(configRoot, config.totemDir), runArtifact);