@pellux/goodvibes-agent 1.8.1 → 1.8.2

package/src/tools/agent-harness-learning-auto-promote.ts

@@ -173,9 +173,13 @@ function tryConsolidateNonMemoryCandidate(
 // ---------------------------------------------------------------------------
 
 /**
- * Run the autonomous promotion pass. Promotes `ready-to-promote` candidates
- * and resolves `needs-consolidation` duplicates (non-memory domains) without
- * requiring confirm:true / explicitUserRequest.
+ * Run the promotion pass. Promotes `ready-to-promote` candidates and resolves
+ * `needs-consolidation` duplicates (non-memory domains).
+ *
+ * The caller (agent_harness mode:"learning_auto_promote") gates this pass behind
+ * an explicit user request plus confirm:true, because it creates and deletes
+ * memory records, personas, and routines. This function assumes that gate has
+ * already been cleared.
  *
  * Secret scanning is enforced by each registry's create() method — this
  * function does NOT bypass it.

package/src/tools/agent-harness-memory-posture.ts

@@ -1,6 +1,13 @@
-import type { MemoryDoctorReport, MemoryEmbeddingProviderStatus, MemoryVectorStats } from '@pellux/goodvibes-sdk/platform/state';
+import {
+  MemoryUsageStatsStore,
+  type MemoryDoctorReport,
+  type MemoryEmbeddingProviderStatus,
+  type MemoryVectorStats,
+  type MemoryUsageSummary,
+} from '@pellux/goodvibes-sdk/platform/state';
 import type { CommandContext } from '../input/command-registry.ts';
 import { buildAgentWorkspaceRuntimeSnapshot } from '../input/agent-workspace-snapshot.ts';
+import { GOODVIBES_AGENT_SURFACE_ROOT } from '../config/surface.ts';
 import { previewHarnessText } from './agent-harness-text.ts';
 import { EXTERNAL_MEMORY_RECEIPT_FIELDS, EXTERNAL_MEMORY_REQUIRED_CONTRACTS, aggregateExternalProviderLiveRecord, aggregateExternalProviderSetupStatus, externalMemoryLiveProviderRecords, externalMemoryProviderCatalog, externalMemoryReceiptEvidence, externalProviderLiveReady, liveRecordForProvider, receiptEvidenceForProvider, setupStatusFromLiveRecord } from './agent-harness-memory-external-providers.ts';
 import type { ExternalMemoryProviderCatalogEntry, MemoryExternalProviderContractCheck, MemoryExternalProviderContractStatus, MemoryExternalProviderLiveRecord, MemoryExternalProviderReceiptContract, MemoryExternalProviderReceiptEvidence, MemoryExternalProviderRoute, MemoryExternalProviderSetupGuide, MemoryExternalProviderStatus, MemoryPostureProvider, MemoryProviderResolution } from './agent-harness-memory-external-providers.ts';
@@ -32,6 +39,40 @@ function memoryApi(context: CommandContext): {
   return context.clients?.agentKnowledgeApi?.memory ?? null;
 }
 
+/**
+ * Per-memory usage counters (injected/referenced) from the honest usage-outcome
+ * instrumentation. Read-only view built from the same durable sidecar the runtime
+ * writes; null when there is no active workspace to resolve the path from.
+ */
+function readMemoryUsageSummary(context: CommandContext): MemoryUsageSummary | null {
+  const shellPaths = context.workspace?.shellPaths;
+  if (!shellPaths) return null;
+  try {
+    const store = new MemoryUsageStatsStore(shellPaths.resolveUserPath(GOODVIBES_AGENT_SURFACE_ROOT, 'learning', 'memory-usage.json'));
+    return store.summary();
+  } catch {
+    return null;
+  }
+}
+
+function buildUsageSection(summary: MemoryUsageSummary | null): Record<string, unknown> {
+  if (!summary) {
+    return { status: 'unavailable', signal: 'heuristic-overlap', note: 'No usage instrumentation available in this runtime.' };
+  }
+  return {
+    status: summary.everInjected > 0 ? 'tracked' : 'empty',
+    signal: 'heuristic-overlap',
+    signalNote: summary.signalNote,
+    totalTracked: summary.totalTracked,
+    everInjected: summary.everInjected,
+    everReferenced: summary.everReferenced,
+    neverReferenced: summary.neverReferenced,
+    mostReferenced: summary.mostReferenced.map((entry) => ({ id: entry.id, injectedCount: entry.injectedCount, referencedCount: entry.referencedCount, lastReferencedAt: entry.lastReferencedAt })),
+    neverReferencedSample: summary.neverReferencedSample.map((entry) => ({ id: entry.id, injectedCount: entry.injectedCount, lastInjectedAt: entry.lastInjectedAt })),
+    policy: 'Referenced means the model output overlapped this memory\'s distinctive content (heuristic, not ground truth). Never-referenced injected memories decay first during idle consolidation.',
+  };
+}
+
 async function readMemoryDoctor(context: CommandContext): Promise<MemoryDoctorReport | null> {
   const memory = memoryApi(context);
   if (!memory || typeof memory.doctor !== 'function') return null;
@@ -532,13 +573,21 @@ export async function memoryPostureCatalogStatus(context: CommandContext): Promi
   const liveRecords = await externalMemoryLiveProviderRecords(context);
   const status = memoryStatus(snapshot, Boolean(memoryApi(context)));
   const vectorState = vectorStatus(vector);
+  const usageSummary = readMemoryUsageSummary(context);
   return {
     modes: ['memory_posture', 'memory_provider'],
+    usageTracked: usageSummary?.everInjected ?? 0,
+    usageReferenced: usageSummary?.everReferenced ?? 0,
+    usageNeverReferenced: usageSummary?.neverReferenced ?? 0,
     modelRoute: 'memory action:"status"',
     status,
     localMemories: snapshot.localMemoryCount,
     reviewQueue: snapshot.localMemoryReviewQueueCount,
     promptActive: snapshot.localMemoryPromptActiveCount,
+    // Visible expiry: still stored (counted in localMemories), never deleted,
+    // but outside their temporal validity window so not prompt-injected.
+    temporallyPending: snapshot.localMemoryTemporallyPendingCount,
+    temporallyExpired: snapshot.localMemoryTemporallyExpiredCount,
     vector: vectorState,
     embeddingProviders: doctor?.embeddings.providers.length ?? 0,
     externalProviders: externalMemoryProviderCatalog(liveRecords, receiptEvidence).length,
@@ -569,8 +618,13 @@ export async function memoryPostureSummary(context: CommandContext, args: AgentH
       total: snapshot.localMemoryCount,
       reviewQueue: snapshot.localMemoryReviewQueueCount,
       promptActive: snapshot.localMemoryPromptActiveCount,
+      // Visible expiry (temporal validity window): still stored and counted in
+      // `total`, never deleted, but outside [validFrom, validUntil) so not
+      // prompt-injected — see memoryRecordTemporalStatus.
+      temporallyPending: snapshot.localMemoryTemporallyPendingCount,
+      temporallyExpired: snapshot.localMemoryTemporallyExpiredCount,
       sessionMemory: snapshot.sessionMemoryCount,
-      policy: 'Only reviewed, high-confidence Agent-local memories are prompt-active. Fresh/stale/setup-blocked behavior stays visible for review.',
+      policy: 'Only reviewed, high-confidence, temporally-active Agent-local memories are prompt-active. Fresh/stale/setup-blocked/pending/expired behavior stays visible for review.',
       routes: {
         list: 'memory action:"list"',
         search: 'memory action:"search" query:"..."',
@@ -579,6 +633,7 @@ export async function memoryPostureSummary(context: CommandContext, args: AgentH
         curator: 'memory action:"curator"',
       },
     },
+    usage: buildUsageSection(readMemoryUsageSummary(context)),
     vector: compactVector(vector),
     embeddings: {
       activeProviderId: doctor?.embeddings.activeProviderId ?? vector?.embeddingProviderId ?? snapshot.embeddingProvider,

package/src/tools/agent-harness-metadata.ts

@@ -204,6 +204,18 @@ export function describeCommandPolicy(commandName: string): CommandExecutionPoli
       boundary: 'Only explicit allowlisted approval and automation operator actions should be performed from the model.',
     };
   }
+  if (root === 'ci' || root === 'principals' || root === 'principal' || root === 'channel-profiles' || root === 'channel-profile') {
+    return {
+      effect: 'connected-host-state',
+      confirmation,
+      // agent_operator_method, NOT agent_operator_action: the action tool's
+      // allowlist covers only approvals.* and automation.* and cannot invoke
+      // ci.*/principals.*/channels.profiles.* — pointing the model there would
+      // dead-end in "unknown action".
+      preferredModelTool: 'agent_operator_method (methodId "ci.status", "ci.watches.*", "principals.*", "channels.profiles.*")',
+      boundary: 'CI status/watches, principal identity mappings, and per-channel profile defaults live on the connected host. Reads run through agent_operator_method without confirmation; writes require confirm + explicitUserRequest there, and mutating slash/CLI subcommands require --yes.',
+    };
+  }
   if (root === 'schedule' || root === 'remind' || root === 'reminder') {
     return {
       effect: 'connected-host-state',

package/src/tools/agent-harness-mode-catalog.ts

@@ -100,7 +100,7 @@ export const HARNESS_MODE_DESCRIPTORS: readonly HarnessModeDescriptor[] = [
   { id: 'autonomy_queue_item', kind: 'inspect', family: 'personal-ops', summary: 'Inspect one autonomy card, live records, tails, and routes.', next: 'Prefer autonomy action:"item".', parameters: ['queueItemId', 'target', 'query'] },
   { id: 'learning_curator', kind: 'discover', family: 'personal-ops', summary: 'Rank memory, notes, personas, skills, routines review/proposals.', next: 'Prefer memory action:"curator"; use action:"candidate" for one card.', parameters: ['query', 'limit', 'includeParameters'] },
   { id: 'learning_candidate', kind: 'inspect', family: 'personal-ops', summary: 'Inspect one local-learning candidate and safe existing routes.', parameters: ['candidateId', 'target', 'query'] },
-  { id: 'learning_auto_promote', kind: 'effect', family: 'personal-ops', summary: 'Auto-promote ready-to-promote; consolidate duplicates autonomously.', next: 'Use memory action:"curator" to see eligible items, then invoke.', keywords: ['auto promote', 'autonomous promotion', 'consolidate duplicates', 'learning loop', 'auto learning'] },
+  { id: 'learning_auto_promote', kind: 'effect', family: 'personal-ops', summary: 'Promote ready-to-promote; consolidate duplicates on confirmation.', next: 'Use memory action:"curator" to see eligible items, then confirm.', requiresConfirmation: true, keywords: ['auto promote', 'promotion', 'consolidate duplicates', 'learning loop', 'auto learning'], parameters: ['confirm', 'explicitUserRequest'] },
   { id: 'research_briefing', kind: 'discover', family: 'research', summary: 'Show one next-action queue for research runs, sources, reports.', next: 'Use research action:"briefing".', keywords: ['deep research', 'research cockpit', 'research next actions', 'research queue', 'source review', 'report queue'], parameters: ['query', 'target', 'limit', 'includeParameters'] },
   { id: 'research_workflow', kind: 'discover', family: 'research', summary: 'Plan research routes across run, source, report, browser, Knowledge.', next: 'Use research action:"plan"; runner checks browser readiness only.', parameters: ['query', 'target', 'runId', 'includeParameters'] },
   { id: 'research_runs', kind: 'discover', family: 'research', summary: 'List deep-research runs, phase, log tails, and controls.', next: 'Use research_run for one run.', parameters: ['query', 'limit', 'includeParameters'] },

package/src/tools/agent-harness-personal-ops-lanes.ts

@@ -5,7 +5,13 @@ import { providerBackedQueueRecords } from './agent-harness-personal-ops-provide
 import { providerBackedReminderRecords, providerBackedTaskRecords } from './agent-harness-personal-ops-provider-task-records.ts';
 import { channelRecords, connectorRecords, localRecord, refreshableSavedRecordCount, reminderOperationRecords, routineReceiptRecord, savedProviderEffectReceiptRecords, savedReviewArtifactRecords, savedReviewQueueRecords, taskOperationRecords } from './agent-harness-personal-ops-records.ts';
 import type { McpToolRecord, McpToolSchema, PersonalOpsLane, PersonalOpsLiveRecord } from './agent-harness-personal-ops-types.ts';
-import { buildStyleReplyLaneAdditions } from '../agent/email/style-reply-lane.ts';
+// Capability-honesty note: the writing-style-matched draft-reply composer
+// (src/agent/email/style-reply.ts + style-reply-lane.ts) is NOT wired into the
+// advertised inbox lane. The Agent has no reader for the user's own sent-message
+// corpus, so it cannot honestly claim a style match, and competitive-feature-
+// inventory.ts records writing-style-matched draft replies as "not yet shipped".
+// The composer stays as tested internal code; do not re-advertise it here until
+// the sent-corpus input and a real compose route ship together.
 
 export function buildLanes(
   context: CommandContext,
@@ -51,9 +57,6 @@ export function buildLanes(
     && routine.reviewState === 'reviewed'
     && (routine.missingRequirementCount ?? 0) === 0
   )).length;
-  const hasEmailCapability = emailMethods.length > 0 || emailConnectors.length > 0;
-  const styleReplyAdditions = buildStyleReplyLaneAdditions(hasEmailCapability);
-
   return [
     {
       id: 'inbox',
@@ -101,14 +104,13 @@ export function buildLanes(
       ],
       methodIds: emailMethods,
       connectorSignals: emailConnectors,
-      workflows: [...inboxWorkflows(emailMethods, emailConnectors, unavailableEmailMethods), styleReplyAdditions.workflow],
+      workflows: inboxWorkflows(emailMethods, emailConnectors, unavailableEmailMethods),
       liveRecords: [
         ...inboxProviderRecords,
         ...inboxReviewQueueRecords,
         ...inboxArtifactRecords,
         ...inboxEffectReceiptRecords,
         ...connectorRecords(emailConnectors, 'Inbox', 'inbox'),
-        styleReplyAdditions.liveRecord,
       ],
     },
     {

package/src/tools/agent-harness-prompt-context.ts

@@ -190,7 +190,9 @@ function formatMemoryLine(record: MemoryRecord): string {
 
 function promptMemoryRecords(memory: PromptMemoryApi | undefined): readonly MemoryRecord[] {
   if (!memory) return [];
-  return memory.getAll().filter(isPromptActiveMemory).sort(memorySort).slice(0, 10);
+  // Bound to one arg — see prompt-context-receipts.ts for why a bare
+  // `.filter(isPromptActiveMemory)` leaks the array index in as `now`.
+  return memory.getAll().filter((record) => isPromptActiveMemory(record)).sort(memorySort).slice(0, 10);
 }
 
 function buildMemoryPromptFromRecords(records: readonly MemoryRecord[]): string {

package/src/tools/agent-harness-tool.ts

@@ -493,6 +493,7 @@ export function createAgentHarnessTool(deps: AgentHarnessToolDeps): Tool {
         }
         if (args.mode === 'learning_curator') return output(learningCuratorSummary(deps.commandContext, args));
         if (args.mode === 'learning_auto_promote') {
+          const confirmationError = requireConfirmedAction(args, 'Learning promotion'); if (confirmationError) return error(confirmationError);
           const shellPaths = deps.commandContext.workspace?.shellPaths;
           if (!shellPaths) return error('learning_auto_promote requires an active workspace.');
           const memoryApi = deps.commandContext.clients?.agentKnowledgeApi?.memory;
@@ -503,8 +504,8 @@ export function createAgentHarnessTool(deps: AgentHarnessToolDeps): Tool {
             ...result,
             message: result.promoted > 0
               ? `Promoted ${result.promoted} item(s); consolidated ${result.consolidated} duplicate(s).`
-              : 'No items eligible for autonomous promotion this pass.',
-            policy: 'Autonomous promotion. Skills are created via the skill-draft runner. Memory, persona, routine creates call registry directly. Consolidation runs the full merge-stale-delete pipeline. Secret scanning enforced by each registry create().',
+              : 'No items eligible for promotion this pass.',
+            policy: 'Confirmed promotion. Skills are created via the skill-draft runner. Memory, persona, routine creates call registry directly. Consolidation runs the full merge-stale-delete pipeline. Secret scanning enforced by each registry create().',
           });
         }
         if (args.mode === 'learning_candidate') {

package/src/tools/agent-policy-explanation.ts

@@ -233,6 +233,46 @@ function predictPermission(context: CommandContext, toolName: string, category:
     };
   }
 
+  if (mode === 'plan') {
+    if (category === 'read') {
+      return {
+        outcome: 'allowed',
+        sourceLayer: 'runtime_mode',
+        reasonCode: 'mode_allow_all',
+        mode,
+        reason: 'Plan mode auto-allows read-only actions.',
+      };
+    }
+    return {
+      outcome: 'denied',
+      sourceLayer: 'runtime_mode',
+      reasonCode: 'plan_mode',
+      mode,
+      reason: `Plan mode refuses ${category} actions outright — it never asks; the model presents a plan instead of acting.`,
+    };
+  }
+
+  if (mode === 'accept-edits') {
+    if (category === 'read' || category === 'write') {
+      return {
+        outcome: 'allowed',
+        sourceLayer: 'runtime_mode',
+        reasonCode: category === 'write' ? 'mode_accept_edits' : 'mode_allow_all',
+        mode,
+        reason: category === 'write'
+          ? 'Accept-edits mode auto-approves file write/edit actions.'
+          : 'Accept-edits mode auto-allows read-only actions.',
+      };
+    }
+    return {
+      outcome: 'prompt',
+      sourceLayer: 'user_prompt_pending',
+      reasonCode: 'prompt_required',
+      mode,
+      reason: `Accept-edits mode still asks before ${category} actions.`,
+    };
+  }
+
   if (mode === 'custom') {
     const toolConfigKey = TOOL_CONFIG_KEYS[toolName];
     if (!toolConfigKey) {

package/src/tools/tool-definition-compaction.ts

@@ -13,6 +13,7 @@ const TOOL_DESCRIPTION_OVERRIDES: Readonly<Record<string, string>> = {
   computer: 'Inspect/open browser, desktop, and computer routes.',
   context: 'Inspect project instructions and prompt context.',
   delegation: 'Inspect build delegation routes and policy.',
+  context_accounting: 'Report live session context: usage and compaction.',
   control: 'Inspect commands, panels, and runtime posture.',
   device: 'Inspect/open device, voice, and browser routes.',
   execution: 'Inspect local work routes, history, and recovery.',
@@ -36,6 +37,7 @@ const TOOL_DESCRIPTION_OVERRIDES: Readonly<Record<string, string>> = {
   registry: 'Discover local skills, agents, tools, and templates.',
   remote: 'Manage remote runner pools and artifacts.',
   repl: 'Evaluate bounded JS, TS, Python, SQL, or GraphQL.',
+  repo_map: 'Return a token-budgeted map of the repository.',
   route: 'Choose the best visible route for a user task.',
   schedule: 'List, create, edit, run, pause, resume schedules.',
   security: 'Read security posture, findings, and policy decisions.',

package/src/version.ts

@@ -6,7 +6,7 @@ import { join } from 'node:path';
 // The prebuild script updates the fallback value before compilation.
 // Uses import.meta.dir (Bun) to locate package.json relative to this file,
 // which is correct regardless of the process working directory.
-let _version = '1.8.1';
+let _version = '1.8.2';
 try {
   const pkg = JSON.parse(readFileSync(join(import.meta.dir, '..', 'package.json'), 'utf-8')) as {
     readonly version?: unknown;