clementine-agent 1.18.34 → 1.18.36

package/dist/agent/assistant.js

@@ -33,7 +33,7 @@ import { searchSkills as searchSkillsSync } from './skill-extractor.js';
 import { classifyIntent, getStrategyGuidance } from './intent-classifier.js';
 import { getEventLog } from './session-event-log.js';
 import { applyServiceDedup, routeToolSurface, TOOL_SURFACE_HARD_LIMIT, TOOL_SURFACE_WARN_THRESHOLD } from './tool-router.js';
-import { isRestrictedToolset, toolsetAllowsLocalWrites } from './toolsets.js';
+import { isRestrictedToolset, toolsetAllowsLocalWrites, toolsetDisablesAllTools } from './toolsets.js';
 import { looksLikeApprovalPrompt } from './local-turn.js';
 import { decideTurn } from './turn-policy.js';
 import { loadClementineJson } from '../config/clementine-json.js';
@@ -2011,7 +2011,9 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
     async buildOptions(opts = {}) {
         const { isHeartbeat = false, cronTier = null, maxTurns = null, model = null, enableTeams = true, retrievalContext = '', profile = null, sessionKey = null, streaming = false, isPlanStep = false, isUnleashed = false, sourceOverride, disableAllTools = false, verboseLevel, abortController, effort, maxBudgetUsd, toolScopeText, thinking, outputFormat, stallGuard, intentClassification, turnPolicy, contextRoutingText, toolset = 'auto', } = opts;
         const isCron = cronTier !== null;
-        const toolsDisabledForCall = disableAllTools || (isHeartbeat && !isCron);
+        const toolsDisabledForCall = disableAllTools
+            || (isHeartbeat && !isCron)
+            || toolsetDisablesAllTools(toolset);
         const promptScopeText = toolScopeText ?? '';
         const profileScopeText = [profile?.description, profile?.systemPromptBody]
             .filter(Boolean)
@@ -5390,7 +5392,24 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
             }
         }
         catch { /* non-fatal — run without skills */ }
+        // ── Sub-agent fan-out directive (Vision 2) ──────────────────────
+        // Detect multi-item / broad-scope signals in the job spec and
+        // prepend a hard-line fan-out mandate when found. This is what
+        // keeps the parent context clean on long jobs: each slice of work
+        // runs in an Agent sub-agent (its own context window, big tool
+        // responses contained), and the parent only sees compact summaries.
+        const { buildAlwaysOnParallelizationHint, buildFanoutDirectiveForText } = await import('./fanout-policy.js');
+        const fanoutScope = `${jobName}\n${jobPrompt}\n${cronProfile?.description ?? ''}\n${cronProfile?.systemPromptBody ?? ''}`;
+        const { directive: fanoutDirective, report: fanoutReport } = buildFanoutDirectiveForText(fanoutScope);
+        if (fanoutReport.needsFanout) {
+            logger.info({
+                job: jobName,
+                signals: fanoutReport.signals.map(s => s.pattern),
+            }, 'Fanout policy: directive injected for cron job');
+        }
         const prompt = `[Scheduled task: ${jobName}]\n\n` +
+            (fanoutDirective ? fanoutDirective + '\n\n' : '') +
+            buildAlwaysOnParallelizationHint() + '\n\n' +
             progressContext +
             goalContext +
             skillContext +
@@ -5782,22 +5801,30 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
             }
             let prompt;
             if (phase === 1) {
+                const { buildAlwaysOnParallelizationHint, buildFanoutDirectiveForText } = await import('./fanout-policy.js');
+                const unleashedFanoutScope = `${jobName}\n${jobPrompt}\n${unleashedProfile?.description ?? ''}\n${unleashedProfile?.systemPromptBody ?? ''}`;
+                const { directive: unleashedFanoutDirective, report: unleashedFanoutReport } = buildFanoutDirectiveForText(unleashedFanoutScope);
+                if (unleashedFanoutReport.needsFanout) {
+                    logger.info({
+                        job: jobName,
+                        phase,
+                        signals: unleashedFanoutReport.signals.map(s => s.pattern),
+                    }, 'Fanout policy: directive injected for unleashed phase 1');
+                }
                 prompt =
                     `[UNLEASHED TASK: ${jobName} — Phase ${phase} — ${timestamp}]\n\n` +
                         `You are running in unleashed mode — a long-running autonomous task.\n` +
                         `Time remaining: ${remainingHours} hours. You have ${turnsPerPhase} turns per phase.\n` +
                         `After each phase completes, your session will be resumed with fresh context.\n\n` +
+                        (unleashedFanoutDirective ? unleashedFanoutDirective + '\n\n' : '') +
+                        buildAlwaysOnParallelizationHint() + '\n\n' +
                         `TASK:\n${jobPrompt}\n\n` +
                         unleashedSkillContext +
                         `${unleashedContextSafety}\n\n` +
                         `IMPORTANT:\n` +
                         `- Work methodically through the task in phases\n` +
                         `- At the end of this phase, output a STATUS SUMMARY of what you accomplished and what remains\n` +
-                        `- Save important intermediate results to files so they persist across phases\n\n` +
-                        `PARALLELIZATION: When processing multiple items (prospects, accounts, emails, analyses), ` +
-                        `use the Agent tool to spawn sub-agents that work in parallel. For example, if you need to ` +
-                        `research 10 prospects, spawn 3-5 sub-agents that each handle a batch — don't process them ` +
-                        `one at a time. Each sub-agent should receive specific items and return structured results.`;
+                        `- Save important intermediate results to files so they persist across phases`;
             }
             else {
                 // Phase 2+ — inject structured checkpoint from previous phase if available
@@ -5821,6 +5848,8 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
                     }
                 }
                 catch { /* fall back to no checkpoint */ }
+                const { buildAlwaysOnParallelizationHint: hintFn } = await import('./fanout-policy.js');
+                const phaseParallelHint = hintFn();
                 if (sessionId) {
                     // Resuming existing session — agent has conversation history + structured checkpoint
                     prompt =
@@ -5829,6 +5858,7 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
                             `Time remaining: ${remainingHours} hours. You have ${turnsPerPhase} turns this phase.\n` +
                             checkpointContext +
                             `\n${unleashedContextSafety}\n` +
+                            `\n${phaseParallelHint}\n` +
                             `\nContinue working on the task. Pick up where you left off.\n` +
                             `If the task is COMPLETE, output "TASK_COMPLETE:" followed by a final summary.\n\n` +
                             `IMPORTANT: Output a STATUS SUMMARY at the end of this phase.`;
@@ -5843,6 +5873,7 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
                             `TASK:\n${jobPrompt}\n` +
                             checkpointContext +
                             `\n${unleashedContextSafety}\n` +
+                            `\n${phaseParallelHint}\n` +
                             `\nCheck any files or progress from prior phases, then continue the work.\n` +
                             `If the task is COMPLETE, output "TASK_COMPLETE:" followed by a final summary.\n\n` +
                             `IMPORTANT: Output a STATUS SUMMARY at the end of this phase.`;

package/dist/agent/fanout-policy.d.ts

@@ -0,0 +1,59 @@
+/**
+ * Sub-agent fan-out policy for autonomous tasks.
+ *
+ * Why: even with a small tool surface, a single agent context can fill
+ * within a few turns when tool responses are large (Outlook list dumps,
+ * web search results, file reads, multi-prospect research). The SDK's
+ * autocompact then has nothing to compact and aborts with
+ * `rapid_refill_breaker`. The fix matching how Claude Code is designed:
+ * spawn sub-agents that each handle a slice of work in their own
+ * isolated context and return only a compact summary back to the parent.
+ *
+ * The Agent tool already exists in the SDK. The problem is timing —
+ * agents tend to discover the need for fan-out only after thrashing.
+ * This module front-loads the directive: scan the task description for
+ * signals that fan-out will be needed, and inject a strong, explicit
+ * mandate at the top of the prompt.
+ *
+ * Two outputs:
+ *  - buildAlwaysOnParallelizationHint()
+ *      Short reminder injected into every autonomous prompt. Cheap.
+ *  - buildFanoutDirective(detectFanoutSignals(text).signals)
+ *      Stronger, explicit fan-out contract. Only injected when signals
+ *      indicate the task is genuinely multi-item or broad-scope.
+ */
+export interface FanoutSignal {
+    /** Why fan-out matters for this task. Surfaced in the directive. */
+    reason: string;
+    /** The pattern that matched. Used for telemetry. */
+    pattern: string;
+}
+export interface FanoutSignalReport {
+    needsFanout: boolean;
+    signals: FanoutSignal[];
+}
+/**
+ * Detect patterns that strongly predict fan-out is needed. Conservative
+ * by design — false positives waste a few hundred tokens per turn; false
+ * negatives let the agent thrash. Tune for false positives.
+ */
+export declare function detectFanoutSignals(text: string): FanoutSignalReport;
+/**
+ * Always-on parallelization reminder. Short, designed to ride along in
+ * every autonomous prompt without inflating token cost.
+ */
+export declare function buildAlwaysOnParallelizationHint(): string;
+/**
+ * Strong fan-out contract injected when detector matches. Designed to be
+ * unambiguous: failing to fan out on these patterns *will* crash the run.
+ */
+export declare function buildFanoutDirective(signals: FanoutSignal[]): string;
+/**
+ * Convenience: detect signals and return the directive string in one
+ * call. Returns empty string when no fan-out is indicated.
+ */
+export declare function buildFanoutDirectiveForText(text: string): {
+    directive: string;
+    report: FanoutSignalReport;
+};
+//# sourceMappingURL=fanout-policy.d.ts.map

package/dist/agent/fanout-policy.js

@@ -0,0 +1,133 @@
+/**
+ * Sub-agent fan-out policy for autonomous tasks.
+ *
+ * Why: even with a small tool surface, a single agent context can fill
+ * within a few turns when tool responses are large (Outlook list dumps,
+ * web search results, file reads, multi-prospect research). The SDK's
+ * autocompact then has nothing to compact and aborts with
+ * `rapid_refill_breaker`. The fix matching how Claude Code is designed:
+ * spawn sub-agents that each handle a slice of work in their own
+ * isolated context and return only a compact summary back to the parent.
+ *
+ * The Agent tool already exists in the SDK. The problem is timing —
+ * agents tend to discover the need for fan-out only after thrashing.
+ * This module front-loads the directive: scan the task description for
+ * signals that fan-out will be needed, and inject a strong, explicit
+ * mandate at the top of the prompt.
+ *
+ * Two outputs:
+ *  - buildAlwaysOnParallelizationHint()
+ *      Short reminder injected into every autonomous prompt. Cheap.
+ *  - buildFanoutDirective(detectFanoutSignals(text).signals)
+ *      Stronger, explicit fan-out contract. Only injected when signals
+ *      indicate the task is genuinely multi-item or broad-scope.
+ */
+/**
+ * Detect patterns that strongly predict fan-out is needed. Conservative
+ * by design — false positives waste a few hundred tokens per turn; false
+ * negatives let the agent thrash. Tune for false positives.
+ */
+export function detectFanoutSignals(text) {
+    const signals = [];
+    const lower = text.toLowerCase();
+    const checks = [
+        {
+            pattern: 'multi_item_iteration',
+            re: /\b(for each|for every|process each|iterate over|loop through|across all|across each)\b/,
+            reason: 'task explicitly iterates over multiple items — process them in parallel sub-agents, not one at a time in this conversation',
+        },
+        {
+            pattern: 'collective_with_quantifier',
+            re: /\b(all|every|each)\s+(prospects?|accounts?|leads?|contacts?|customers?|deals?|emails?|messages?|threads?|files?|records?|rows?|tasks?|items?|results?|pages?|articles?|posts?|repos?|repositories|projects?)\b/,
+            reason: 'task spans every item in a collection — fan out by batching items across sub-agents',
+        },
+        {
+            pattern: 'numeric_collection',
+            re: /\b\d{2,}\s+(prospects?|accounts?|leads?|contacts?|customers?|deals?|emails?|messages?|threads?|files?|records?|rows?|items?|results?|pages?|articles?)\b/,
+            reason: 'task names a numeric count of items (10+) — split into batches of 3-5 per sub-agent',
+        },
+        {
+            pattern: 'comprehensive_research',
+            re: /\b(comprehensive|exhaustive|deep[- ]dive|deep dive|full audit|competitive intel|market map|content intel|brief|landscape|panorama)\b/,
+            reason: 'broad-scope research task — each step (news, search, brand, competitor, social) should run in its own sub-agent so the parent context stays clean',
+        },
+        {
+            pattern: 'broad_scan_or_crawl',
+            re: /\b(scan all|crawl|backfill|inventory|migrate|refactor)\b.{0,80}\b(all|entire|every|full)\b/s,
+            reason: 'broad scan/crawl — partition by directory, date range, or ID range and fan out per partition',
+        },
+        {
+            pattern: 'long_history_pull',
+            re: /\b(last|past)\s+\d+\s+(days|weeks|months)|\bsince\s+(yesterday|last week|last month)\b/,
+            reason: 'pulling a history range that is likely to return many items — sub-agents per day/week chunk',
+        },
+        {
+            pattern: 'multiple_steps',
+            re: /\b(steps?|phases?|stages?)\s*[:0-9]/,
+            reason: 'task has explicit multi-step structure — each step in its own sub-agent, parent only sees the step summaries',
+        },
+    ];
+    for (const check of checks) {
+        if (check.re.test(lower)) {
+            signals.push({ pattern: check.pattern, reason: check.reason });
+        }
+    }
+    return {
+        needsFanout: signals.length > 0,
+        signals,
+    };
+}
+/**
+ * Always-on parallelization reminder. Short, designed to ride along in
+ * every autonomous prompt without inflating token cost.
+ */
+export function buildAlwaysOnParallelizationHint() {
+    return [
+        '## Sub-agent fan-out',
+        'When you process multiple items, spawn ONE Agent sub-agent per batch of 3–5 items. Sub-agents return ONE-LINE summaries (no raw tool output). Do not iterate sequentially in this conversation — that fills your context and aborts the run.',
+        'Cost: pass `model: "haiku"` to Agent for routine extraction, summarization, or per-item lookups. Use Sonnet only when the sub-agent must reason across many sources or write something durable.',
+    ].join('\n');
+}
+/**
+ * Strong fan-out contract injected when detector matches. Designed to be
+ * unambiguous: failing to fan out on these patterns *will* crash the run.
+ */
+export function buildFanoutDirective(signals) {
+    if (signals.length === 0)
+        return '';
+    const reasonLines = signals
+        .map((s, i) => `${i + 1}. ${s.reason}`)
+        .join('\n');
+    return [
+        '## Sub-agent fan-out is MANDATORY for this task',
+        '',
+        'Preflight detected patterns that will fill the context window if you run them sequentially in this conversation:',
+        '',
+        reasonLines,
+        '',
+        '### Required pattern',
+        '',
+        'Use the `Agent` tool to spawn parallel sub-agents. Each sub-agent runs in its own isolated context, so big tool responses live and die there — your context only sees the summary.',
+        '',
+        '- **Batch size**: 3–5 items per sub-agent (or one slice of work per sub-agent for research tasks)',
+        '- **Sub-agent model**: pass `model: "haiku"` to the Agent tool by default — sub-agents that just extract fields, summarize a single email, or pull a single record do not need Sonnet. Reserve Sonnet for sub-agents that must reason across multiple sources or write something durable.',
+        '- **Sub-agent prompt MUST include**: the narrow task, the exact return format (e.g. `Return ONE LINE: <id> | <status> | <next-action>`), and an explicit "do not include raw tool output" directive',
+        '- **Parent context keeps**: only the sub-agent return strings, not their tool transcripts',
+        '',
+        'If you anticipate a single tool call returning more than ~5 KB of text (full email lists, web search result pages, large database queries, file dumps), wrap THAT call in an Agent invocation too. The sub-agent runs the tool, extracts only the fields you need, and returns those.',
+        '',
+        'Failing to fan out on this task will cause the SDK to abort with `rapid_refill_breaker` and the run will be lost.',
+    ].join('\n');
+}
+/**
+ * Convenience: detect signals and return the directive string in one
+ * call. Returns empty string when no fan-out is indicated.
+ */
+export function buildFanoutDirectiveForText(text) {
+    const report = detectFanoutSignals(text);
+    return {
+        directive: buildFanoutDirective(report.signals),
+        report,
+    };
+}
+//# sourceMappingURL=fanout-policy.js.map

package/dist/agent/toolsets.d.ts

@@ -1,4 +1,4 @@
-export type ToolsetName = 'auto' | 'safe' | 'diagnostic' | 'communications' | 'memory' | 'full';
+export type ToolsetName = 'auto' | 'safe' | 'diagnostic' | 'communications' | 'memory' | 'none' | 'full';
 export interface ToolsetPreset {
     name: ToolsetName;
     label: string;
@@ -11,4 +11,10 @@ export declare function getToolsetPreset(name: ToolsetName): ToolsetPreset;
 export declare function formatToolsetChoices(): string;
 export declare function isRestrictedToolset(name: ToolsetName): boolean;
 export declare function toolsetAllowsLocalWrites(name: ToolsetName): boolean;
+/**
+ * "none" toolset: not just restricted, but actively disables ALL tools
+ * including the core Clementine MCP server. Used by builder/JSON-gen
+ * chats where tool schemas in the system prompt are pure cost overhead.
+ */
+export declare function toolsetDisablesAllTools(name: ToolsetName): boolean;
 //# sourceMappingURL=toolsets.d.ts.map

package/dist/agent/toolsets.js

@@ -29,6 +29,12 @@ export const TOOLSET_PRESETS = [
         description: 'Memory, transcript, and relationship tools only unless explicitly changed.',
         directive: 'Toolset memory: use memory_read, memory_search, memory_recall, transcript_search, working_memory, and user_model. Avoid external integrations and local shell/file writes.',
     },
+    {
+        name: 'none',
+        label: 'None',
+        description: 'No tools at all — pure-LLM conversation. Used by builders and JSON-generating chats where tool schemas are dead weight in the system prompt.',
+        directive: 'Toolset none: do not call any tools. Respond from the prompt context only. Generate JSON, summaries, or text directly.',
+    },
     {
         name: 'full',
         label: 'Full',
@@ -60,9 +66,17 @@ export function formatToolsetChoices() {
         .join('\n');
 }
 export function isRestrictedToolset(name) {
-    return name === 'safe' || name === 'diagnostic' || name === 'memory';
+    return name === 'safe' || name === 'diagnostic' || name === 'memory' || name === 'none';
 }
 export function toolsetAllowsLocalWrites(name) {
     return name === 'auto' || name === 'full';
 }
+/**
+ * "none" toolset: not just restricted, but actively disables ALL tools
+ * including the core Clementine MCP server. Used by builder/JSON-gen
+ * chats where tool schemas in the system prompt are pure cost overhead.
+ */
+export function toolsetDisablesAllTools(name) {
+    return name === 'none';
+}
 //# sourceMappingURL=toolsets.js.map

package/dist/cli/dashboard.js

@@ -8053,6 +8053,12 @@ If the tool returns nothing or errors, return an empty array \`[]\`.`,
         }
         try {
             const gateway = await getGateway();
+            // Builder generates JSON artifacts — no tool calls. Pin the session
+            // toolset to 'none' so buildOptions strips all MCP servers and tool
+            // schemas from the system prompt. Without this, every tiny builder
+            // turn writes 60–280 KB of cache_creation for tool schemas the
+            // model never uses.
+            gateway.setSessionToolset(sessionKey, 'none');
             const response = await gateway.handleMessage(sessionKey, enrichedMessage);
             // Parse any json-artifact blocks from the response
             let artifact = null;
@@ -8180,6 +8186,9 @@ If the tool returns nothing or errors, return an empty array \`[]\`.`,
         try {
             writeEvent('progress', { status: 'thinking…' });
             const gateway = await getGateway();
+            // Builder generates JSON artifacts — no tool calls. Pin to 'none'
+            // toolset so the SDK system prompt drops the tool inventory.
+            gateway.setSessionToolset(sessionKey, 'none');
             let lastText = '';
             const response = await gateway.handleMessage(sessionKey, enrichedMessage, async (text) => {
                 lastText = text ?? '';

package/dist/gateway/agent-heartbeat-scheduler.js

@@ -171,7 +171,15 @@ export class AgentHeartbeatScheduler {
         catch {
             signals.latestGoalUpdate = '';
         }
-        // 3. Latest cron run for any of this agent's crons (file mtime is enough)
+        // 3. Latest cron run for any of this agent's crons (file mtime is enough).
+        //
+        // INFO ONLY — kept in `signals` for the LLM prompt context, but NOT
+        // included in the fingerprint hash below. A cron firing is expected
+        // background activity, not a "wake up Sonnet" signal. Including it
+        // caused every cron run to bump the fingerprint → fire a $1+ Sonnet
+        // pass that just confirmed "yep, the cron ran, nothing else to do."
+        // Actionable wake-ups are pendingTasks growing and goal-state changes;
+        // those still trip the fingerprint below.
         try {
             const runsDir = path.join(this.baseDir, 'cron', 'runs');
             let latestMs = 0;
@@ -195,8 +203,15 @@ export class AgentHeartbeatScheduler {
         catch {
             signals.latestCronRunMs = 0;
         }
+        // Fingerprint only includes ACTIONABLE signals. latestCronRunMs is
+        // info-only and explicitly excluded.
+        const fingerprintSource = {
+            slug: signals.slug,
+            pendingTasks: signals.pendingTasks,
+            latestGoalUpdate: signals.latestGoalUpdate,
+        };
         const fingerprint = createHash('sha1')
-            .update(JSON.stringify(signals))
+            .update(JSON.stringify(fingerprintSource))
             .digest('hex')
             .slice(0, 16);
         return { fingerprint, signals };

package/dist/gateway/cron-scheduler.js

@@ -23,7 +23,7 @@ import { listBackgroundTasks, markDone as markBgTaskDone, markFailed as markBgTa
 import { outcomeStatusFromGoalDisposition, recentDecisions, recordDecisionOutcome, } from '../agent/proactive-ledger.js';
 import { formatCreditBlock, getBackgroundCreditBlock, isCreditBalanceError, markBackgroundCreditBlocked, } from './credit-guard.js';
 import { isRunHealthFailure } from './job-health.js';
-import { analyzeLongTaskPreflight, compactLongTaskPreflight, formatLongTaskPromptPrefix, } from './long-task-preflight.js';
+import { analyzeLongTaskPreflight, compactLongTaskPreflight, shouldDowngradeUnleashed, formatLongTaskPromptPrefix, } from './long-task-preflight.js';
 const logger = pino({ name: 'clementine.cron' });
 /** Default timeout for standard cron jobs (10 minutes). */
 const CRON_STANDARD_TIMEOUT_MS = 10 * 60 * 1000;
@@ -1061,6 +1061,34 @@ export class CronScheduler {
             // Sonnet runs every job by default. Opus 1M is opt-in: set
             // `model: claude-opus-4-7[1m]` in CRON.md per-job, or flip
             // CLEMENTINE_1M_CONTEXT_MODE=on for global enable.
+            // ── Auto-downgrade unleashed → standard ────────────────────────
+            // CRON.md `mode: unleashed` is a CEILING, not a floor. If the
+            // job's history shows it's a quiet probe that completes in 1
+            // phase with __NOTHING__ or short output, the multi-phase
+            // wrapper is wasteful overhead — each phase is a fresh SDK
+            // query with full system prompt + tool schemas in cache_creation.
+            // For a "did anything new come in?" cron firing every 2 hours,
+            // that's 12+ unleashed runs/day at ~$1/each instead of standard
+            // mode at ~$0.05/each.
+            if (job.mode === 'unleashed') {
+                const downgrade = shouldDowngradeUnleashed(this.runLog.readRecent(job.name, 5));
+                if (downgrade.downgrade) {
+                    job = { ...job, mode: 'standard' };
+                    logger.info({
+                        job: job.name,
+                        reason: downgrade.reason,
+                        quietRatio: downgrade.quietRatio,
+                        avgDurationMs: downgrade.avgDurationMs,
+                    }, 'Cron mode downgraded unleashed → standard based on run history');
+                    this.logAutonomy('mode_downgrade', job, {
+                        from: 'unleashed',
+                        to: 'standard',
+                        reason: downgrade.reason,
+                        quietRatio: downgrade.quietRatio,
+                        avgDurationMs: downgrade.avgDurationMs,
+                    });
+                }
+            }
             let longTaskPreflight;
             const preflight = analyzeLongTaskPreflight(job, jobPrompt, this.runLog.readRecent(job.name, 5));
             if (preflight.risk !== 'normal') {

package/dist/gateway/long-task-preflight.d.ts

@@ -15,6 +15,15 @@ export interface LongTaskPreflightOptions {
     opusModel?: string;
     sonnetModel?: string;
 }
+export interface UnleashedDowngradeDecision {
+    downgrade: boolean;
+    reason: string;
+    /** Quiet ratio observed in the sample (for telemetry). */
+    quietRatio?: number;
+    /** Average duration of recent runs in ms (for telemetry). */
+    avgDurationMs?: number;
+}
+export declare function shouldDowngradeUnleashed(recentRuns: CronRunEntry[], now?: number): UnleashedDowngradeDecision;
 export declare function analyzeLongTaskPreflight(job: CronJobDefinition, prompt: string, recentRuns?: CronRunEntry[], opts?: LongTaskPreflightOptions): LongTaskPreflightDecision;
 export declare function formatLongTaskPromptPrefix(decision: LongTaskPreflightDecision): string;
 export declare function compactLongTaskPreflight(decision: LongTaskPreflightDecision): LongTaskPreflightSnapshot;

package/dist/gateway/long-task-preflight.js

@@ -48,6 +48,88 @@ function recentContextFailures(recentRuns, now = Date.now()) {
     }
     return [...new Set(reasons)];
 }
+// ── Auto-downgrade unleashed → standard for quiet/probe jobs ──────────
+//
+// `mode: unleashed` wraps a job in multi-phase machinery: each phase is a
+// fresh SDK query with the full system prompt + tool schemas, and the
+// orchestrator chains phases until TASK_COMPLETE or max-phases. That
+// machinery is essential for genuinely-long tasks (sasha briefs, market
+// outreach), but it's pure overhead on quiet probe jobs that finish in
+// 1 phase with `__NOTHING__` or a short output.
+//
+// Detect that pattern from history and downgrade the next run to
+// standard mode. Single SDK call, single cache write, fraction of the
+// cost. The user's CRON.md `mode: unleashed` becomes a "ceiling" rather
+// than a forced floor — actual mode chosen dynamically per-run.
+//
+// Conservative by design: requires 3+ prior runs of evidence, refuses
+// to downgrade if any recent run hit context overflow (the unleashed
+// wrapper might be actively saving us), and only triggers on jobs that
+// historically complete fast with short or empty output.
+const UNLEASHED_DOWNGRADE_SAMPLE_SIZE = 5;
+const UNLEASHED_DOWNGRADE_MIN_HISTORY = 3;
+const UNLEASHED_DOWNGRADE_QUIET_RATIO = 0.6;
+const UNLEASHED_DOWNGRADE_MAX_DURATION_MS = 90_000;
+const UNLEASHED_DOWNGRADE_AVG_DURATION_MS = 60_000;
+const UNLEASHED_DOWNGRADE_QUIET_PREVIEW_CHARS = 200;
+export function shouldDowngradeUnleashed(recentRuns, now = Date.now()) {
+    const sample = recentRuns
+        .slice(0, UNLEASHED_DOWNGRADE_SAMPLE_SIZE)
+        .filter(r => r.status === 'ok' || r.status === 'error');
+    if (sample.length < UNLEASHED_DOWNGRADE_MIN_HISTORY) {
+        return { downgrade: false, reason: 'insufficient_history' };
+    }
+    // Refuse to downgrade if any recent run hit a context-window failure —
+    // the unleashed multi-phase wrapper might be the only thing keeping
+    // this job from thrashing on a single huge SDK query. Pair this guard
+    // with the existing fanout-policy directive (1.18.35) so by the next
+    // few runs the agent has learned to fan out and the wrapper can be
+    // shed safely.
+    const cutoff = now - RECENT_CONTEXT_FAILURE_WINDOW_MS;
+    const hadOverflow = sample.some(r => {
+        const startedMs = Date.parse(r.startedAt);
+        if (!Number.isFinite(startedMs) || startedMs < cutoff)
+            return false;
+        return r.terminalReason === 'rapid_refill_breaker'
+            || r.terminalReason === 'prompt_too_long';
+    });
+    if (hadOverflow) {
+        return { downgrade: false, reason: 'recent_context_overflow_protect_unleashed' };
+    }
+    // Quiet pattern: most recent runs returned __NOTHING__ or a short
+    // output. These jobs don't need multi-phase orchestration.
+    const quietCount = sample.filter(r => {
+        const preview = (r.outputPreview ?? '').trim();
+        if (!preview)
+            return false;
+        if (/__nothing__/i.test(preview))
+            return true;
+        return preview.length < UNLEASHED_DOWNGRADE_QUIET_PREVIEW_CHARS;
+    }).length;
+    const quietRatio = quietCount / sample.length;
+    if (quietRatio >= UNLEASHED_DOWNGRADE_QUIET_RATIO) {
+        return {
+            downgrade: true,
+            reason: `quiet_pattern_${Math.round(quietRatio * 100)}pct`,
+            quietRatio,
+        };
+    }
+    // Fast-completion pattern: every run finishes well under the standard
+    // cron timeout, average is short. Multi-phase wrapper is overhead.
+    const durations = sample.map(r => r.durationMs || 0).filter(d => d > 0);
+    if (durations.length === sample.length) {
+        const avgDuration = durations.reduce((a, b) => a + b, 0) / durations.length;
+        const allFast = durations.every(d => d < UNLEASHED_DOWNGRADE_MAX_DURATION_MS);
+        if (allFast && avgDuration < UNLEASHED_DOWNGRADE_AVG_DURATION_MS) {
+            return {
+                downgrade: true,
+                reason: `fast_completion_avg_${Math.round(avgDuration / 1000)}s`,
+                avgDurationMs: Math.round(avgDuration),
+            };
+        }
+    }
+    return { downgrade: false, reason: 'workload_warrants_unleashed' };
+}
 function classifyRisk(args) {
     const { inputTokens, projectedTokens, signalCount, recentContextIssue, job, oneMillionAvailable } = args;
     if (inputTokens >= 185_000)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "clementine-agent",
-  "version": "1.18.34",
+  "version": "1.18.36",
   "description": "Clementine — Personal AI Assistant (TypeScript)",
   "type": "module",
   "main": "dist/index.js",