clementine-agent 1.18.153 → 1.18.155

package/dist/agent/run-agent-cron.d.ts

@@ -175,6 +175,18 @@ export interface RunAgentCronOptions {
      *  prior progress. The fix for fire-time memory drift. Undefined =
      *  legacy behavior (inject everything). */
     predictable?: boolean;
+    /** Lean mode — strictest possible context envelope, used by meta-jobs
+     *  (insight-check, outcome-grader, route-classifier, failure-diagnostics,
+     *  __heartbeat__) that ALSO have to stay under Haiku's prompt cap.
+     *  Implies predictable + drops progress/goal/criteria/skill-context blocks
+     *  + prunes MCP catalog to just the explicit allowed servers (no Composio
+     *  auto-discovery, no claude.ai connector inventory). When unset, behaves
+     *  like predictable. The 1.18.148 fix patched outcome-grader / route-
+     *  classifier / failure-diagnostics directly; this 1.18.154 flag covers
+     *  insight-check (which still failed 65/70 because predictable alone
+     *  wasn't strict enough) and is the canonical hook for any future
+     *  meta-job that needs a tiny prompt. */
+    lean?: boolean;
     /** Extra read+execute scope for the agent's Read/Bash/Glob tools. Maps
      *  directly to the CronJobDefinition.addDirs YAML field. Combined with
      *  every pinned-skill folder so `Bash python3 scripts/render.py` works

package/dist/agent/run-agent-cron.js

@@ -548,13 +548,32 @@ export async function buildCronExecutionPlan(opts) {
     // fire time. Legacy tricks (predictable === undefined) preserve existing
     // behavior so we don't surprise anyone.
     const predictable = opts.predictable === true;
-    const memoryContext = predictable ? '' : buildAutonomousMemoryContext(opts.profile);
-    const progressContext = buildProgressContext(opts.jobName); // opt-in via cron_progress writes
-    const goalContext = buildGoalContext(opts.jobName); // explicit links; not auto-inferred
-    const delegationContext = predictable ? '' : buildDelegationContext(agentSlug);
-    const teamContext = predictable ? '' : buildTeamContext(agentSlug);
-    const criteriaContext = buildCriteriaContext(opts.successCriteria);
-    const skillResult = await buildSkillContext(opts.jobName, opts.jobPrompt, agentSlug, opts.pinnedSkills, opts.memoryStore, { skipAutoMatch: predictable, projectWorkDir: opts.workDir });
+    // 1.18.154 — lean mode goes one tier stricter for meta-jobs (insight-check,
+    // outcome-grader, route-classifier, failure-diagnostics, __heartbeat__).
+    // Drops progress / goal / criteria / skill-context blocks too, leaving
+    // the agent with [jobName] + [jobPrompt] + [howToRespond] only. Implies
+    // predictable. Anything that ALSO needs a tiny MCP catalog uses the
+    // pruning logic below.
+    //
+    // Auto-apply for known meta-jobs even if the CRON.md/registry entry
+    // doesn't carry `lean: true` — these names are hard-coded singletons in
+    // heartbeat-scheduler.ts and we know they need the strict envelope. Lets
+    // existing user installs benefit from the fix without requiring them to
+    // hand-edit CRON.md.
+    const KNOWN_META_JOBS = new Set([
+        'insight-check', 'outcome-grader', 'route-classifier',
+        'failure-diagnostics', '__heartbeat__',
+    ]);
+    const lean = opts.lean === true || KNOWN_META_JOBS.has(opts.jobName);
+    const memoryContext = predictable || lean ? '' : buildAutonomousMemoryContext(opts.profile);
+    const progressContext = lean ? '' : buildProgressContext(opts.jobName); // opt-in via cron_progress writes
+    const goalContext = lean ? '' : buildGoalContext(opts.jobName); // explicit links; not auto-inferred
+    const delegationContext = predictable || lean ? '' : buildDelegationContext(agentSlug);
+    const teamContext = predictable || lean ? '' : buildTeamContext(agentSlug);
+    const criteriaContext = lean ? '' : buildCriteriaContext(opts.successCriteria);
+    const skillResult = lean
+        ? { text: '', applied: [], missing: [], pinnedBodies: [], pinnedToolsRequested: [] }
+        : await buildSkillContext(opts.jobName, opts.jobPrompt, agentSlug, opts.pinnedSkills, opts.memoryStore, { skipAutoMatch: predictable, projectWorkDir: opts.workDir });
     const skillContext = skillResult.text;
     const howToRespond = `## How to respond\n` +
         `You're sending this directly to ${ownerName} as a DM. ` +
@@ -577,15 +596,24 @@ export async function buildCronExecutionPlan(opts) {
         criteriaContext +
         `${opts.jobPrompt}\n\n` +
         howToRespond;
-    const mcp = await buildExtraMcpForRunAgent({
-        scopeText: [
-            opts.jobName,
-            opts.jobPrompt,
-            opts.profile?.description,
-            opts.profile?.systemPromptBody,
-        ].filter(Boolean).join('\n\n'),
-        profile: opts.profile,
-    });
+    // 1.18.154 — lean mode skips Composio + claude.ai MCP discovery entirely.
+    // Meta-jobs (insight-check etc.) only need the always-on Clementine MCP
+    // server (which is added in run-agent.ts and has cron_list, discord_*,
+    // etc.). Loading 200+ extra MCP tool schemas just to call cron_list is
+    // what was tipping the prompt over Haiku's cap. If a lean job DOES need
+    // extras, declare them explicitly via `allowedMcpServers` and the helper
+    // narrows discovery to those names.
+    const mcp = lean && (!opts.allowedMcpServers || opts.allowedMcpServers.length === 0)
+        ? { servers: {}, composioConnected: [], externalConnected: [] }
+        : await buildExtraMcpForRunAgent({
+            scopeText: [
+                opts.jobName,
+                opts.jobPrompt,
+                opts.profile?.description,
+                opts.profile?.systemPromptBody,
+            ].filter(Boolean).join('\n\n'),
+            profile: opts.profile,
+        });
     // 1.18.125 — pinned-skill scope widening (SDK alignment).
     // A pinned skill that declares `clementine.tools.allow` or references
     // `mcp__server__tool` in its body needs those tools/servers to actually
@@ -701,19 +729,28 @@ export async function buildCronExecutionPlan(opts) {
 export async function runAgentCron(opts) {
     const plan = await buildCronExecutionPlan(opts);
     const { builtPrompt, agentSlug, effort, maxBudgetUsd: maxBudget, effectiveAllowedTools, mcpServerMap, composioConnected, externalConnected, mcpServersApplied, additionalDirectories, } = plan;
-    logger.info({
+    // 1.18.154 — surface lean + promptChars for meta-job diagnostics. The
+    // self-improve / insight-check failure mode is "small prompt, fat context"
+    // which only shows in promptChars; plain log level was previously info but
+    // anything > 50KB is on the cusp of Haiku's cap and worth a warn.
+    const promptBytes = Buffer.byteLength(builtPrompt, 'utf8');
+    const promptOversized = promptBytes > 50_000;
+    logger[promptOversized ? 'warn' : 'info']({
         job: opts.jobName,
         tier: plan.tier,
         profile: agentSlug,
+        lean: opts.lean === true || ['insight-check', 'outcome-grader', 'route-classifier', 'failure-diagnostics', '__heartbeat__'].includes(opts.jobName),
         composioConnected,
         externalConnected,
         promptChars: builtPrompt.length,
+        promptBytes,
         pinnedSkills: opts.pinnedSkills?.length ?? 0,
         skillsApplied: plan.skillsApplied.length,
         skillsMissing: plan.skillsMissing.length,
         trickAllowedTools: effectiveAllowedTools?.length,
         trickAllowedMcp: opts.allowedMcpServers?.length,
         widenedFromSkills: plan.widenedFromSkills,
+        ...(promptOversized ? { warning: 'prompt > 50KB; risk of "Prompt is too long" failure' } : {}),
     }, 'runAgentCron: dispatching to runAgent');
     const startedAt = Date.now();
     const result = await runAgent(builtPrompt, {

package/dist/cli/dashboard.js

@@ -25418,8 +25418,18 @@ function renderScheduledTaskCard(task) {
   } else {
     badges += '<span class="badge badge-gray" title="Legacy CRON.md job. Carries its own prompt/tools/MCP. Convert to a scheduled skill when you can.">LEGACY CRON</span>';
   }
-  if (task.predictable === true) badges += '<span class="badge badge-green" title="Contract mode — runs with only the prompt + pinned skills/tools. No MEMORY.md, no auto-matched skills, no team comms injection at fire-time.">🔒 predictable</span>';
-  else if (task.predictable === false) badges += '<span class="badge badge-yellow" title="Dynamic mode — fire-time injects MEMORY.md, recent team activity, and auto-matched skills. Can drift from chat-time intent.">🔄 reads memory</span>';
+  // 1.18.155 — replace the "🔒 predictable" lock icon (which appeared on
+  // ~every row and added pure noise) with a single "Strict" pill that ONLY
+  // appears when behavior diverges from the default. predictable=true is
+  // the default for new tasks since 1.18.68, so suppress that badge — only
+  // show one when the user has explicitly opted into dynamic mode (the
+  // less safe path, worth flagging) or when lean mode is in effect for a
+  // meta-job (worth showing because it explains why the prompt is small).
+  if (task.lean === true) {
+    badges += '<span class="badge badge-purple" title="Lean envelope — drops every auto-injected context block (memory, progress, goal, criteria, skills) and prunes the MCP catalog. Used for meta-jobs that must stay under Haiku\'s prompt cap.">Lean</span>';
+  } else if (task.predictable === false) {
+    badges += '<span class="badge badge-yellow" title="Dynamic mode — fire-time injects MEMORY.md, recent team activity, and auto-matched skills. Can drift from chat-time intent.">Reads memory</span>';
+  }
   if (task.mode === 'unleashed') badges += '<span class="badge badge-purple">long-running</span>';
   if (task.after) badges += '<span class="badge badge-yellow" title="Triggered after ' + esc(task.after) + '">after ' + esc(task.after) + '</span>';
   if (task.maxRetries != null) badges += '<span class="badge badge-gray">' + esc(task.maxRetries) + ' retries</span>';
@@ -25670,7 +25680,13 @@ function renderRecentHistoryList(runs) {
       var msg = String(entry.error).slice(0, 120);
       errorPreview = '<div style="font-size:11px;color:var(--red);margin-top:2px;word-break:break-word">' + esc(msg) + '</div>';
     } else if (entry.outputPreview) {
-      var preview = String(entry.outputPreview).slice(0, 140);
+      // 1.18.155 — strip the __NOTHING__ sentinel agents emit when they
+      // intentionally stay silent (it's the "no signal" contract from
+      // run-agent-cron.ts howToRespond block, not output the user should see).
+      var rawPreview = String(entry.outputPreview);
+      var preview = (rawPreview === '__NOTHING__' || rawPreview.trim() === '__NOTHING__')
+        ? '(no output — agent reported nothing worth saying)'
+        : rawPreview.slice(0, 140);
       errorPreview = '<div style="font-size:11px;color:var(--text-muted);margin-top:2px;word-break:break-word">' + esc(preview) + '</div>';
     }
     // PRD Phase 1.1: goal cell. Empty cell when no goal configured (status='skipped'
@@ -25775,16 +25791,31 @@ async function refreshHealthStrip() {
     var ops = await apiFetch('/api/build/operations?hours=1&limit=10').then(function(rr) { return rr.json(); });
     activeRuns = ((ops && ops.runningNow) || []).length;
   } catch (e) { /* fall back to 0 */ }
-  // Top failure category for the day.
+  // Top failure category for the day. 1.18.155 — when no failureCategory
+  // is set on the failed runs (which is common for raw SDK errors like
+  // "Prompt is too long"), fall back to grouping by jobName so the user
+  // sees WHICH job is failing instead of a useless dash. Categories still
+  // win when present (more semantic than a job name).
   var catCounts = {};
+  var jobCounts = {};
   for (var i = 0; i < last24.length; i++) {
     var c = last24[i].failureCategory;
     if (c) catCounts[c] = (catCounts[c] || 0) + 1;
+    if (last24[i].status === 'error' || last24[i].status === 'failed') {
+      var jn = last24[i].jobName || last24[i].name;
+      if (jn) jobCounts[jn] = (jobCounts[jn] || 0) + 1;
+    }
   }
   var topCat = null, topCount = 0;
   Object.keys(catCounts).forEach(function(k) {
     if (catCounts[k] > topCount) { topCat = k; topCount = catCounts[k]; }
   });
+  var topJob = null, topJobCount = 0;
+  Object.keys(jobCounts).forEach(function(k) {
+    if (jobCounts[k] > topJobCount) { topJob = k; topJobCount = jobCounts[k]; }
+  });
+  // Prefer the categorized one (more semantic), fall back to job name.
+  if (!topCat && topJob) { topCat = topJob; topCount = topJobCount; }
   // Render six tiles.
   function tile(label, value, sub, color) {
     return '<div class="health-tile">'
@@ -26450,7 +26481,12 @@ function renderRunListBody(allRuns) {
     if (status === 'error' && entry.error) {
       preview = '<div style="font-size:11px;color:var(--red);margin-top:2px;word-break:break-word">' + esc(String(entry.error).slice(0, 140)) + '</div>';
     } else if (entry.outputPreview) {
-      preview = '<div style="font-size:11px;color:var(--text-muted);margin-top:2px;word-break:break-word">' + esc(String(entry.outputPreview).slice(0, 120)) + '</div>';
+      // 1.18.155 — see renderRecentHistoryList for the same __NOTHING__ strip.
+      var rawOp = String(entry.outputPreview);
+      var opText = (rawOp === '__NOTHING__' || rawOp.trim() === '__NOTHING__')
+        ? '(no output — agent reported nothing worth saying)'
+        : rawOp.slice(0, 120);
+      preview = '<div style="font-size:11px;color:var(--text-muted);margin-top:2px;word-break:break-word">' + esc(opText) + '</div>';
     }
     // 1.18.89: cost label. Showing 4 decimals for sub-penny costs (Haiku
     // runs land in fractions of a cent), 2 decimals when ≥ $0.01.
@@ -27079,7 +27115,11 @@ async function refreshCron() {
     var bannerHtml = '';
     var dismissed = localStorage.getItem('clem-skill-migrate-banner-dismissed') === '1';
     if (legacyCount > 0 && !dismissed) {
-      bannerHtml = '<div style="background:rgba(124,58,237,0.08);border:1px solid var(--purple);border-radius:8px;padding:12px 14px;margin-bottom:14px;display:flex;align-items:center;gap:12px;flex-wrap:wrap">'
+      // 1.18.155 — data-banner-kind tags this as the legacy-cron soft-
+      // deprecation banner so refreshCronMigrateBanner can suppress its
+      // secondary "clean up preambles" banner when this one is showing
+      // (full migration is a superset; showing both is confusing noise).
+      bannerHtml = '<div data-banner-kind="legacy-cron-soft-deprecation" style="background:rgba(124,58,237,0.08);border:1px solid var(--purple);border-radius:8px;padding:12px 14px;margin-bottom:14px;display:flex;align-items:center;gap:12px;flex-wrap:wrap">'
         + '<span style="font-size:18px">⚡</span>'
         + '<div style="flex:1;min-width:200px">'
         +   '<div style="font-size:13px;font-weight:500;color:var(--text-primary)">' + legacyCount + ' legacy cron task' + (legacyCount === 1 ? '' : 's') + ' can become scheduled skills</div>'
@@ -29049,6 +29089,19 @@ async function refreshCronMigrateBanner() {
       return;
     }
     _cronMigratePreview = d;
+    // 1.18.155 — suppress this banner when the upstream "X legacy crons
+    // can become scheduled skills" banner is already showing. The full
+    // migration there is a superset of this cleanup (the destination skill
+    // gets a clean prompt automatically), so showing both at once is just
+    // confusing noise. If the upstream banner is dismissed via localStorage
+    // OR there are no legacy crons to migrate, this orange banner shows up
+    // as the secondary action.
+    var dismissed = localStorage.getItem('clem-skill-migrate-banner-dismissed') === '1';
+    var hasLegacyCronBanner = !dismissed && document.querySelector('[data-banner-kind="legacy-cron-soft-deprecation"]');
+    if (hasLegacyCronBanner) {
+      host.innerHTML = '';
+      return;
+    }
     var n = d.eligible.length;
     host.innerHTML =
       '<div style="margin:0 0 14px;padding:12px 16px;border:1px solid var(--accent);background:rgba(255,141,0,0.06);border-radius:8px;display:flex;align-items:center;gap:14px;flex-wrap:wrap">'

package/dist/dashboard/build-operations.d.ts

@@ -101,6 +101,10 @@ export interface ScheduledTaskCard {
      *  team comms / auto-matched skills. The visibility-on-card flag for
      *  "this trick will run with only what you see here." */
     predictable?: boolean;
+    /** 1.18.154/155 — Lean envelope (meta-jobs). Surfaced in the row badge
+     *  so the user understands why insight-check / outcome-grader / etc.
+     *  show a tiny prompt + restricted MCP catalog. */
+    lean?: boolean;
 }
 export interface ScheduledWorkflowCard {
     type: 'scheduled_workflow';

package/dist/dashboard/build-operations.js

@@ -216,6 +216,7 @@ export function buildOperationsSnapshot(input) {
             tags: asStringArray(job.tags),
             category: typeof job.category === 'string' && job.category.trim() ? job.category.trim() : undefined,
             predictable: typeof job.predictable === 'boolean' ? job.predictable : undefined,
+            lean: typeof job.lean === 'boolean' ? job.lean : undefined,
         };
     }).sort((a, b) => a.owner.localeCompare(b.owner) || a.displayName.localeCompare(b.displayName));
     const scheduledWorkflows = input.workflowSummaries

package/dist/gateway/cron-scheduler.js

@@ -169,6 +169,8 @@ function parseJobYaml(job) {
         ? categoryRaw.trim().slice(0, 64)
         : undefined;
     const predictable = typeof job.predictable === 'boolean' ? job.predictable : undefined;
+    // 1.18.154 — strictest envelope for meta-jobs (insight-check, heartbeat).
+    const lean = typeof job.lean === 'boolean' ? job.lean : undefined;
     const description = typeof job.description === 'string' && job.description.trim()
         ? job.description.trim().slice(0, 500)
         : undefined;
@@ -177,7 +179,7 @@ function parseJobYaml(job) {
         maxHours, maxRetries, after, successCriteria, successCriteriaText, successSchema, addDirs,
         alwaysDeliver, context, preCheck, attachments, requiresConfirmation, confirmationTimeoutMin,
         agentSlug,
-        skills, allowedTools, allowedMcpServers, tags, category, predictable,
+        skills, allowedTools, allowedMcpServers, tags, category, predictable, lean,
     };
 }
 /**
@@ -1224,12 +1226,12 @@ export class CronScheduler {
                 const startedAt = new Date();
                 try {
                     // Standard cron jobs get a timeout via SDK AbortController (advisor may override)
-                    let response = await this.gateway.handleCronJob(job.name, jobPrompt, job.tier, job.maxTurns, job.model, job.workDir, job.mode, job.maxHours, effectiveTimeoutMs, job.successCriteria, job.agentSlug, job.skills, job.allowedTools, job.allowedMcpServers, job.predictable, job.addDirs);
+                    let response = await this.gateway.handleCronJob(job.name, jobPrompt, job.tier, job.maxTurns, job.model, job.workDir, job.mode, job.maxHours, effectiveTimeoutMs, job.successCriteria, job.agentSlug, job.skills, job.allowedTools, job.allowedMcpServers, job.predictable, job.lean, job.addDirs);
                     // alwaysDeliver: retry once if the response is empty/noise
                     if (job.alwaysDeliver && (!response || CronScheduler.isCronNoise(response))) {
                         logger.info({ job: job.name }, 'alwaysDeliver: empty/noise response — retrying once');
                         try {
-                            const retryResponse = await this.gateway.handleCronJob(job.name, jobPrompt + '\n\nYou MUST produce a brief status update. Do NOT return __NOTHING__.', job.tier, job.maxTurns, job.model, job.workDir, job.mode, job.maxHours, effectiveTimeoutMs, job.successCriteria, job.agentSlug, job.skills, job.allowedTools, job.allowedMcpServers, job.predictable, job.addDirs);
+                            const retryResponse = await this.gateway.handleCronJob(job.name, jobPrompt + '\n\nYou MUST produce a brief status update. Do NOT return __NOTHING__.', job.tier, job.maxTurns, job.model, job.workDir, job.mode, job.maxHours, effectiveTimeoutMs, job.successCriteria, job.agentSlug, job.skills, job.allowedTools, job.allowedMcpServers, job.predictable, job.lean, job.addDirs);
                             if (retryResponse && !CronScheduler.isCronNoise(retryResponse)) {
                                 response = retryResponse;
                             }

package/dist/gateway/router.d.ts

@@ -183,6 +183,9 @@ export declare class Gateway {
     _mode?: 'standard' | 'unleashed', maxHours?: number, timeoutMs?: number, successCriteria?: string[], agentSlug?: string, pinnedSkills?: string[], allowedTools?: string[], allowedMcpServers?: string[], 
     /** Predictable (contract) mode — runner skips memory/team/auto-skills. */
     predictable?: boolean, 
+    /** Lean (meta-job) mode — strictest envelope, drops every auto-injected
+     *  context block. For insight-check, the four meta-routers, heartbeat. */
+    lean?: boolean, 
     /** Extra read+exec scope for the SDK's Read/Bash/Glob tools. From the
      *  CronJobDefinition.addDirs YAML field. Combined inside runAgentCron
      *  with each pinned-skill folder. (1.18.121) */

package/dist/gateway/router.js

@@ -1977,6 +1977,9 @@ export class Gateway {
     pinnedSkills, allowedTools, allowedMcpServers, 
     /** Predictable (contract) mode — runner skips memory/team/auto-skills. */
     predictable, 
+    /** Lean (meta-job) mode — strictest envelope, drops every auto-injected
+     *  context block. For insight-check, the four meta-routers, heartbeat. */
+    lean, 
     /** Extra read+exec scope for the SDK's Read/Bash/Glob tools. From the
      *  CronJobDefinition.addDirs YAML field. Combined inside runAgentCron
      *  with each pinned-skill folder. (1.18.121) */
@@ -2026,6 +2029,7 @@ export class Gateway {
                     allowedTools,
                     allowedMcpServers,
                     predictable,
+                    lean,
                     addDirs,
                 });
                 scanner.refreshIntegrity();

package/dist/types.d.ts

@@ -550,6 +550,15 @@ export interface CronJobDefinition {
      * Existing tricks (no field set) keep current behavior — backward compat.
      */
     predictable?: boolean;
+    /** 1.18.154 — strictest possible context envelope, for meta-jobs that
+     *  need to stay under Haiku's prompt cap (insight-check, the four
+     *  meta-routers, __heartbeat__). Implies predictable + drops progress/
+     *  goal/criteria/skill blocks entirely. Use only for jobs whose prompt
+     *  is self-contained and whose tools are explicitly allow-listed. The
+     *  fix for the 1.18.148 regression where insight-check kept failing
+     *  with "Prompt is too long" because predictable alone wasn't strict
+     *  enough. */
+    lean?: boolean;
     /** 1.18.129 — where this cron job came from. 'cron-md' is the legacy
      *  fat-cron format in vault/00-System/CRON.md (or per-agent CRON.md).
      *  'scheduled-skill' is the Anthropic-pure registry — a thin entry in

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "clementine-agent",
-  "version": "1.18.153",
+  "version": "1.18.155",
   "description": "Clementine — Personal AI Assistant (TypeScript)",
   "type": "module",
   "main": "dist/index.js",