sneakoscope 10.3.2 → 10.3.4

package/dist/core/init.js

@@ -245,6 +245,38 @@ export function mergeManagedHooksJson(existingContent, commandPrefix, projectRoo
     }
     return `${JSON.stringify({ ...root, hooks: nextHooks }, null, 2)}\n`;
 }
+export function pruneRetiredSksHookEvents(existingContent, installedEvents) {
+    let root;
+    try {
+        root = existingContent?.trim() ? JSON.parse(existingContent) : null;
+    }
+    catch {
+        return { text: existingContent, removed: [] };
+    }
+    if (!root || typeof root !== 'object' || Array.isArray(root))
+        return { text: existingContent, removed: [] };
+    const hooks = root.hooks && typeof root.hooks === 'object' && !Array.isArray(root.hooks) ? root.hooks : null;
+    if (!hooks)
+        return { text: existingContent, removed: [] };
+    const installed = new Set(installedEvents);
+    const removed = [];
+    const nextHooks = {};
+    for (const [eventName, entries] of Object.entries(hooks)) {
+        if (installed.has(eventName) || !Array.isArray(entries)) {
+            nextHooks[eventName] = entries;
+            continue;
+        }
+        const mapped = entries.map((entry) => stripSksManagedHookEntry(entry));
+        if (mapped.some((entry, index) => entry !== entries[index]))
+            removed.push(eventName);
+        const preserved = mapped.filter(Boolean);
+        if (preserved.length)
+            nextHooks[eventName] = preserved;
+    }
+    if (!removed.length)
+        return { text: existingContent, removed };
+    return { text: `${JSON.stringify({ ...root, hooks: nextHooks }, null, 2)}\n`, removed };
+}
 function stripSksManagedHookEntry(entry) {
     if (!entry || typeof entry !== 'object' || Array.isArray(entry) || !Array.isArray(entry.hooks))
         return entry;
@@ -274,11 +306,11 @@ const AGENTS_BLOCK = [
     '## Execution',
     '',
     '- Codex native `/goal` is the only persisted goal owner. Goal objectives must state the outcome, scope, constraints, verification, done-when conditions, stop conditions, and non-goals.',
-    '- General work stays parent-owned. Use `$Naruto` for explicitly requested parallel work or concrete independent slices that benefit from delegation; Answer and tiny DFix work stay lightweight.',
-    '- The parent owns decomposition, integration, verification, and the final answer. Delegate only independent slices with disjoint write scopes, reuse capacity across root-owned waves, and never nest subagents.',
-    '- Use gpt-6-astra for every subagent; vary reasoning effort by the slice: Astra Low for tiny mechanical work and instructed ordinary coding execution, Astra Medium for read-heavy context or direct tool operation, and Astra Max for planning, analysis, review, risk, or other focused judgment. Explicit Astra effort preferences, including High, remain supported.',
-    '- Preserve the user-selected parent model, reasoning effort, and service tier. SKS-launched Naruto defaults to GPT-6 Astra; parent settings and role preferences never override the sealed Astra child profiles.',
-    '- SKS child spawns must explicitly set model="gpt-6-astra" and the role reasoning effort with `fork_turns="none"` or a positive bounded turn count; carry the complete bounded slice contract in `message`. Never use full-history inheritance for children.',
+    '- Answer, Help, Goal, and tiny DFix stay parent-owned. Implementation routes to Naruto (`$sks-naruto`, alias `$sks-work`; standalone `sks naruto run`), which is parent orchestration only: decompose, assign disjoint slices, spawn, integrate, and verify. Do not implement those slices in the parent thread. The SKS PreToolUse hook denies parent-thread source edits until the first child thread starts and while children are still running; .sneakoscope artifacts stay parent-writable, and Jev may release a confirmed orchestration-scaffolding edit.',
+    '- The parent owns decomposition, integration, verification, and the final answer. Delegate implementation slices with disjoint write scopes, reuse capacity across root-owned waves, and never nest subagents.',
+    '- Every child runs the newest model of the tier its work needs (fast, balanced, context, or deep); no model family is pinned. When Jev mode is on, Jev picks the tier for each new Naruto child spawn and SKS seals it. A user role preference stays authoritative. Do not pick a child model yourself.',
+    '- Preserve the user-selected parent model, reasoning effort, and service tier. Parent settings stay on the parent thread. A role preference wins over the Jev seal for that role.',
+    '- SKS child spawns must pass the sealed model and reasoning effort with `fork_turns="none"` or a positive bounded turn count; carry the complete bounded slice contract in `message`. Never use full-history inheritance for children. Full-history forks (`fork_turns="all"`, including the omitted default) inherit the parent agent type, model, and reasoning effort and must not be combined with `agent_type`, `model`, or `reasoning_effort`.',
     '- Route-specific skills own route-specific details. Do not inject unrelated Design, PPT, image, browser, research, DB, or release policy into ordinary work.',
     '- Do not stop at a plan when implementation was requested. Finish the requested outcome, including verification of the changed path, before stopping for review. Stop early only on a hard blocker or when the stated done-when conditions are met.',
     '- User instructions outrank skill guidance. Infer routine details, honor authorization already provided, and finish authorized preparation before asking for a decision that changes scope or has irreversible effects.',
@@ -1119,8 +1151,8 @@ export function codexAppQuickReference(scope, commandPrefix) {
         coreEngineeringDirectiveReferenceText(),
         'dollar-commands:',
         ...currentDollarCommands().map((c) => `- \`${sksPrefixedDollarCommand(c.command)}\`: ${c.route}`),
-        'Routing: Answer is read-only, DFix handles tiny edits, and general work stays parent-owned. Use Naruto for explicit parallel work or independent slices that benefit from delegation.',
-        'Subagent context: explicitly set model="gpt-6-astra" and the role reasoning effort with `fork_turns="none"` or a positive bounded turn count. Pass the complete bounded slice contract in `message`; never inherit the parent model or apply a role-model preference.',
+        'Routing: Answer is read-only, DFix handles tiny edits, and Naruto implementation is parent orchestration with child slices enforced by the PreToolUse gate (spawn before any source edit, and no parent edits beside running children). Answer and tiny DFix stay on the parent.',
+        'Subagent context: pass the sealed model and reasoning effort with `fork_turns="none"` or a positive bounded turn count. Each child uses the newest model of its tier; Jev mode picks that tier at each new spawn. Pass the complete bounded slice contract in `message`. Never use `fork_turns="all"` together with a custom model.',
         'Goal: Codex native /goal is the only persisted goal owner; no SKS Goal mission, bridge, compatibility loop, or fallback state is allowed.',
         'Context: use bounded TriWiki recall when a claim needs project memory; refresh after material changes; validate before handoff/final; use Context7 or official vendor docs when external contracts or versions matter.',
         'Completion: report the result, actual verification, and remaining gaps once. Reflection and Honest Mode are optional unless explicitly requested or required by the strict profile.',

package/dist/core/managed-assets/managed-assets-manifest.js

@@ -23,7 +23,7 @@ export const MANAGED_OFFICIAL_SUBAGENT_ROLES = Object.freeze([
         filename: 'worker.toml',
         aliases: ['worker'],
         codexName: 'worker',
-        description: 'Astra Low execution subagent only for tiny, short-context, mechanical work with an explicit done condition.',
+        description: 'Fast-tier execution subagent only for tiny, short-context, mechanical work with an explicit done condition.',
         policy: 'luna_max_mechanical',
         keywords: ['tiny', 'short context', 'single file', 'mechanical', 'repeatable', 'exact rename', 'format only', 'typo', 'simple search', 'typing', '단순 검색', '타이핑'],
         nicknames: ['Kite', 'Moss', 'Pico', 'Reed', 'Vale', 'Wren'],
@@ -47,14 +47,14 @@ Do not claim success without direct evidence.`
         filename: 'implementation-specialist.toml',
         aliases: ['implementation-specialist', 'core-implementer'],
         codexName: 'implementation_specialist',
-        description: 'Astra Low implementation specialist for instructed backend, core, API, lifecycle, and cross-file coding with disjoint ownership.',
+        description: 'Balanced-tier implementation specialist for instructed backend, core, API, lifecycle, and cross-file coding with disjoint ownership.',
         policy: 'sol_high_implementation',
         keywords: ['implementation', 'backend', 'core', 'api', 'lifecycle implementation', 'cross-file coding', 'feature change', '구현', '백엔드', '핵심 로직'],
         nicknames: ['Builder', 'Forge', 'Mason', 'Rivet'],
         instructions: `You are the bounded implementation specialist.
 
 Own only the disjoint files and acceptance criteria assigned by the parent.
-Use this role for instructed backend, core, API, lifecycle, and cross-file coding after the parent defines the implementation scope. Escalate review, debugging, planning, architecture, security, release, and ambiguous work to an Astra Max specialist.
+Use this role for instructed backend, core, API, lifecycle, and cross-file coding after the parent defines the implementation scope. Escalate review, debugging, planning, architecture, security, release, and ambiguous work to a deep-tier specialist.
 Do not redesign unrelated architecture or integrate sibling work.
 Make the smallest defensible change, run focused verification, and return files, evidence, and residual risks.`
     }),
@@ -86,7 +86,7 @@ Return a concise result, evidence, risks, and next action.`
         filename: 'explorer.toml',
         aliases: ['explorer', 'code-explorer'],
         codexName: 'explorer',
-        description: 'Astra Medium read-only codebase explorer for read-heavy scans, entry points, ownership, dependencies, and distilled evidence.',
+        description: 'Context-tier read-only codebase explorer for read-heavy scans, entry points, ownership, dependencies, and distilled evidence.',
         policy: 'terra_max_context_tools',
         sandbox: 'read-only',
         keywords: ['explore', 'map', 'trace', 'inventory', 'locate', 'search', 'read-only', 'large search', 'repository-wide search', '대규모 검색'],
@@ -104,7 +104,7 @@ Return concise findings with exact paths and symbols.`
         filename: 'long-context-analyst.toml',
         aliases: ['long-context-analyst', 'large-context-analyst', 'document-analyst'],
         codexName: 'long_context_analyst',
-        description: 'Astra Medium read-only analyst for large files, long logs, multi-document context, and distilled evidence handoffs.',
+        description: 'Context-tier read-only analyst for large files, long logs, multi-document context, and distilled evidence handoffs.',
         policy: 'terra_max_context_tools',
         sandbox: 'read-only',
         keywords: ['long context', 'large file', 'large codebase', 'multi-document', 'supporting documents', 'extensive logs', 'context compression'],
@@ -112,7 +112,7 @@ Return concise findings with exact paths and symbols.`
         instructions: `You are the long-context evidence analyst.
 
 Read large files, long logs, or multiple supporting documents without turning raw context into unsupported conclusions.
-Return a compact, source-addressable summary to the parent and identify which claims still require Astra Max judgment.
+Return a compact, source-addressable summary to the parent and identify which claims still require deep-tier judgment.
 Use bounded TriWiki anchors first, hydrate only relevant sources, and do not edit files or spawn another subagent.`
     }),
     officialSubagentRole({
@@ -153,13 +153,13 @@ Run only the focused checks needed for the slice and report exact commands and o
         filename: 'ui-implementer.toml',
         aliases: ['ui-implementer', 'frontend-specialist'],
         codexName: 'ui_implementer',
-        description: 'Astra Low UI and terminal-interface implementation specialist for instructed visual behavior, interaction, accessibility, and rendered state changes.',
+        description: 'Balanced-tier UI and terminal-interface implementation specialist for instructed visual behavior, interaction, accessibility, and rendered state changes.',
         policy: 'sol_high_implementation',
         keywords: ['ui', 'ux', 'frontend', 'visual', 'terminal', 'accessibility'],
         nicknames: ['Canvas', 'Iris', 'Pixel', 'Turing'],
         instructions: `You are the UI implementation specialist.
 
-Execute the coding scope defined by the parent; return unresolved design, debugging, or planning decisions to an Astra Max specialist.
+Execute the coding scope defined by the parent; return unresolved design, debugging, or planning decisions to a deep-tier specialist.
 Trace the rendered user-visible behavior before editing.
 Make the smallest change that fixes interaction, layout, accessibility, or terminal presentation.
 Preserve the existing design system and unrelated behavior.
@@ -170,14 +170,14 @@ Verify the rendered result with the appropriate live or deterministic surface an
         filename: 'native-app-specialist.toml',
         aliases: ['native-app-specialist', 'macos-specialist', 'desktop-specialist'],
         codexName: 'native_app_specialist',
-        description: 'Astra Low native desktop coding specialist for instructed macOS AppKit and Swift menu-bar UI, app lifecycle, accessibility, and OS integration changes.',
+        description: 'Balanced-tier native desktop coding specialist for instructed macOS AppKit and Swift menu-bar UI, app lifecycle, accessibility, and OS integration changes.',
         policy: 'sol_high_implementation',
         keywords: ['native app', 'macos', 'appkit', 'swift', 'menu bar', 'nsstatusitem', 'nsworkspace', 'tcc', 'desktop app'],
         nicknames: ['Cocoa', 'Darwin', 'Quartz', 'Swift'],
         instructions: `You are the native desktop implementation specialist.
 
 Own only the assigned native macOS, AppKit, Swift, or menu-bar files.
-Execute the coding scope defined by the parent; return unresolved architecture, debugging, or planning decisions to an Astra Max specialist.
+Execute the coding scope defined by the parent; return unresolved architecture, debugging, or planning decisions to a deep-tier specialist.
 Preserve the project design system, accessibility semantics, app lifecycle, and OS permission boundaries.
 Do not substitute web UI or placeholder assets for required native behavior.
 Verify with the narrowest compile, deterministic template, or live native check available and report exact evidence.`
@@ -187,7 +187,7 @@ Verify with the narrowest compile, deterministic template, or live native check
         filename: 'computer-use-operator.toml',
         aliases: ['computer-use-operator', 'desktop-operator'],
         codexName: 'computer_use_operator',
-        description: 'Astra Medium Computer Use operator for scoped native macOS, desktop-app, and OS-settings interaction or evidence capture.',
+        description: 'Context-tier Computer Use operator for scoped native macOS, desktop-app, and OS-settings interaction or evidence capture.',
         policy: 'terra_max_context_tools',
         sandbox: 'read-only',
         keywords: ['computer use', 'desktop interaction', 'macos inspection', 'system settings', 'native app inspection', 'visual evidence'],
@@ -196,7 +196,7 @@ Verify with the narrowest compile, deterministic template, or live native check
 
 Use Codex Computer Use only for the explicit native macOS, desktop-app, OS-settings, or non-web visual slice assigned by the parent.
 Do not target the hosting Codex Desktop app (com.openai.codex). For Codex-linked checks, observe Codex through structured host/process evidence and operate only the external native target.
-Do not replace judgment, debugging, planning, or security review; return captured evidence to the appropriate Astra Max specialist.
+Do not replace judgment, debugging, planning, or security review; return captured evidence to the appropriate deep-tier specialist.
 Honor the parent permission scope, avoid destructive or irreversible UI actions, do not edit source files, and report exactly what was observed or changed.`
     }),
     officialSubagentRole({
@@ -204,7 +204,7 @@ Honor the parent permission scope, avoid destructive or irreversible UI actions,
         filename: 'browser-use-operator.toml',
         aliases: ['browser-use-operator', 'chrome-operator', 'web-operator'],
         codexName: 'browser_use_operator',
-        description: 'Astra Medium Browser/Chrome operator for scoped website, localhost, webapp, and browser-based evidence collection or verification.',
+        description: 'Context-tier Browser/Chrome operator for scoped website, localhost, webapp, and browser-based evidence collection or verification.',
         policy: 'terra_max_context_tools',
         sandbox: 'read-only',
         keywords: ['browser use', 'browser', 'chrome', 'website', 'webapp', 'localhost', 'playwright', 'browser evidence'],
@@ -212,7 +212,7 @@ Honor the parent permission scope, avoid destructive or irreversible UI actions,
         instructions: `You are the scoped Browser/Chrome operator.
 
 Use the Codex Chrome Extension path first for websites, localhost, webapps, and browser-based verification, and halt rapidly when the required extension is unavailable.
-Do not perform security, UX, debugging, or product judgment; collect precise browser evidence and hand it to the relevant Astra Max specialist.
+Do not perform security, UX, debugging, or product judgment; collect precise browser evidence and hand it to the relevant deep-tier specialist.
 Honor the parent permission scope, avoid destructive external actions, do not edit source files, and report URLs or sensitive values only in redacted form.`
     }),
     officialSubagentRole({
@@ -220,14 +220,14 @@ Honor the parent permission scope, avoid destructive external actions, do not ed
         filename: 'image-generation-operator.toml',
         aliases: ['image-generation-operator', 'imagegen-operator', 'image-tool-operator'],
         codexName: 'image_generation_operator',
-        description: 'Astra Medium image-generation operator for scoped imagegen and GPT Image execution after the parent seals the visual requirements.',
+        description: 'Context-tier image-generation operator for scoped imagegen and GPT Image execution after the parent seals the visual requirements.',
         policy: 'terra_max_context_tools',
         keywords: ['image generation', 'imagegen', 'gpt image', IMAGEGEN_MODEL, 'generate image', 'edit image', 'visual asset'],
         nicknames: ['Aperture', 'Frame', 'Palette', 'Render'],
         instructions: `You are the scoped image-generation operator.
 
 Execute only the sealed image-generation or image-editing instructions supplied by the parent, using the official Codex image generation surface when available.
-Do not perform UX review, art-direction judgment, or product strategy; return generated artifact paths and tool evidence to a Astra Max reviewer when judgment is required.
+Do not perform UX review, art-direction judgment, or product strategy; return generated artifact paths and tool evidence to a deep-tier reviewer when judgment is required.
 Write only assigned generated-asset paths, preserve source images, and never fabricate successful image output.`
     }),
     officialSubagentRole({
@@ -404,7 +404,7 @@ Return release blockers, exact evidence, and the minimal verification still requ
         filename: 'docs-maintainer.toml',
         aliases: ['docs-maintainer', 'documentation'],
         codexName: 'docs_maintainer',
-        description: 'Astra Medium documentation maintainer for multi-source README, changelog, migration, and reference consistency after behavior is known.',
+        description: 'Context-tier documentation maintainer for multi-source README, changelog, migration, and reference consistency after behavior is known.',
         policy: 'terra_max_context_tools',
         keywords: ['docs', 'documentation', 'readme', 'changelog', 'migration guide', 'reference'],
         nicknames: ['Ink', 'Page', 'Scribe', 'Slate'],

package/dist/core/pipeline-internals/runtime-core.js

@@ -1108,7 +1108,7 @@ async function prepareResearch(root, route, task, required, opts = {}) {
     const researchPlan = await writeResearchPlan(dir, task, {});
     const pipelinePlan = await writePipelinePlan(dir, { missionId: id, route, task, required, ambiguity: { required: false, status: 'direct_route' } });
     await setCurrent(root, routeState(id, route, 'RESEARCH_PREPARED', required, { prompt: task, ...pipelinePlanState(pipelinePlan) }), { sessionKey: opts.sessionKey });
-    return routeContext(route, id, task, required, `Run sks research run latest as a real long-running source-gathering pass, never an automatic mock fallback; do not modify repository source code. Run layered Super Search first and allow only correlated verified-content rows to support real claims. Then run exactly three independent official research_reviewer threads on GPT-6 Astra Max. Any objection requires a mission-local research_synthesizer revision and a fresh three-thread review cycle; do not launch a custom scheduler or debate pool. Keep subagent-plan.json, subagent-events.jsonl, subagent-parent-summary.json, and subagent-evidence.json current, write research-report.md and ${researchPaperArtifactForPlan(researchPlan)}, and pass the adversarial convergence, Honest Mode, and research-gate.json checks.`, null, root);
+    return routeContext(route, id, task, required, `Run sks research run latest as a real long-running source-gathering pass, never an automatic mock fallback; do not modify repository source code. Run layered Super Search first and allow only correlated verified-content rows to support real claims. Then run exactly three independent official research_reviewer threads on the latest deep-tier model at max effort. Any objection requires a mission-local research_synthesizer revision and a fresh three-thread review cycle; do not launch a custom scheduler or debate pool. Keep subagent-plan.json, subagent-events.jsonl, subagent-parent-summary.json, and subagent-evidence.json current, write research-report.md and ${researchPaperArtifactForPlan(researchPlan)}, and pass the adversarial convergence, Honest Mode, and research-gate.json checks.`, null, root);
 }
 async function prepareAutoResearch(root, route, task, required, opts = {}) {
     const { id, dir } = await createMission(root, { mode: 'autoresearch', prompt: task, sessionKey: opts.sessionKey });
@@ -1489,7 +1489,7 @@ ${intakeLine}
 Pipeline plan: .sneakoscope/missions/${id}/${PIPELINE_PLAN_ARTIFACT}
 Required skills: ${route.requiredSkills.join(', ')}
 Stop gate: ${route.stopGate}
-Official subagents: ${routeRequiresSubagents(route, visibleTask) ? 'required for this explicit Naruto/parallel task; use independent disjoint slices, official agent threads, matched SubagentStart/SubagentStop events, and a parent integration summary.' : 'not required by this task profile; keep the work parent-owned unless a concrete independent decomposition emerges.'}
+Official subagents: ${routeRequiresSubagents(route, visibleTask) ? 'required; the parent orchestrates only: spawn a child per independent disjoint slice through official agent threads, wait, then integrate with matched SubagentStart/SubagentStop events and a parent integration summary.' : 'not required by this task profile; keep the work parent-owned unless a concrete independent decomposition emerges.'}
 TriWiki: use a coordinate+voxel-overlay context pack when a claim needs project memory; hydrate low-trust claims from source; refresh after new findings or artifact changes; validate before handoffs/final claims. Coordinate-only packs are invalid and must be refreshed before pipeline decisions.
 Final closeout: every pipeline final answer must summarize what was done, what changed for the user/repo, what was verified, and any remaining gaps.
 ${stopFinalizationRitualsEnforced(root) && route.stopGate !== 'none' && reflectionRequiredForRoute(route) ? `Reflection: ${reflectionInstructionText()}` : 'Reflection: not required for this route.'}

package/dist/core/provider/model-router.js

@@ -1,4 +1,5 @@
-import { ASTRA_SUBAGENT_MODEL, NARUTO_LUNA_MODEL, NARUTO_SOL_MODEL, NARUTO_TERRA_MODEL, decideSubagentModel, subagentModelProfile } from '../subagents/model-policy.js';
+import { decideSubagentModel, subagentModelProfile } from '../subagents/model-policy.js';
+import { latestTierModelSet } from '../subagents/model-tiers.js';
 const CATEGORY_POLICY = {
     quick: { reasoning: 'low', serviceTier: 'fast' },
     standard: { reasoning: 'medium', serviceTier: 'fast' },
@@ -10,7 +11,9 @@ const CATEGORY_POLICY = {
     refactor: { reasoning: 'max', serviceTier: 'fast' },
     strategy: { reasoning: 'max', serviceTier: 'fast' }
 };
-export const NARUTO_MODELS = [NARUTO_LUNA_MODEL, NARUTO_SOL_MODEL, NARUTO_TERRA_MODEL, ASTRA_SUBAGENT_MODEL];
+export function narutoModels() {
+    return [...latestTierModelSet()];
+}
 const E2E_WORK_RE = /(e2e|end[-\s]?to[-\s]?end|test_execution|browser|chrome|computer[-\s]?use|computer\s+use|cross[-\s]?app|playwright|selenium|puppeteer|브라우저|컴퓨터\s*유즈)/i;
 export async function routeModel(category, opts = {}) {
     if (opts.narutoOnly) {
@@ -47,9 +50,9 @@ export function routeNarutoGpt56Model(input = {}) {
             || category === 'ultrabrain'
             || explicitHighRisk
     });
-    const preferred = explicit || ASTRA_SUBAGENT_MODEL;
+    const preferred = explicit || automatic.model;
     const available = input.availableModels == null
-        ? [...NARUTO_MODELS]
+        ? narutoModels()
         : input.availableModels.map(normalizeNarutoGpt56Model).filter((model) => Boolean(model));
     const degraded = new Set((input.degradedModels || []).map((model) => String(model).toLowerCase()));
     const usable = available.filter((model) => !degraded.has(model));
@@ -65,7 +68,7 @@ export function isNarutoGpt56Model(value) {
 }
 export function normalizeNarutoGpt56Model(value) {
     const model = String(value || '').trim().toLowerCase();
-    return NARUTO_MODELS.includes(model) ? model : null;
+    return narutoModels().includes(model) ? model : null;
 }
 export function childInheritsActiveMainModel(_value) {
     return false;

package/dist/core/recallpulse/policy.js

@@ -1,4 +1,4 @@
-import { ASTRA_SUBAGENT_MODEL, SUBAGENT_EFFORT } from '../subagents/model-policy.js';
+import { thinkingSubagentModel, SUBAGENT_EFFORT } from '../subagents/model-policy.js';
 export const RECALLPULSE_DECISION_ARTIFACT = 'recallpulse-decision.json';
 export const RECALLPULSE_HISTORY_ARTIFACT = 'recallpulse-history.jsonl';
 export const MISSION_STATUS_LEDGER_ARTIFACT = 'mission-status-ledger.json';
@@ -201,7 +201,7 @@ export const RESEARCH_REVIEWER_CONTRACT = Object.freeze([
     ...agent,
     persona_boundary: 'Apply only the assigned review dimension and report evidence-bound findings.',
     custom_agent: 'research_reviewer',
-    model: ASTRA_SUBAGENT_MODEL,
+    model: thinkingSubagentModel(),
     reasoning_effort: SUBAGENT_EFFORT,
     service_tier: 'fast'
 })));

package/dist/core/recallpulse.js

@@ -200,7 +200,7 @@ export async function evaluateRecallPulseFixtures(root, opts = {}) {
             fixture('repeated-stop-hook-blocker', true, 'Duplicate suppression keys collapse repeated blocker text into one durable status row.'),
             fixture('hook-only-status-visibility', true, 'mission-status-ledger.json preserves recoverable user-visible status.'),
             fixture('research-persona-missing', true, 'Research validation blocks missing agent display_name/persona/persona_boundary.'),
-            fixture('research-model-policy-not-sol-max', true, 'Research validation blocks reviewer rows that are not bound to the research_reviewer GPT-6 Astra Max policy.'),
+            fixture('research-model-policy-not-sol-max', true, 'Research validation blocks reviewer rows that are not bound to the research_reviewer latest deep-tier max policy.'),
             fixture('research-review-evidence-missing', true, 'Research validation blocks reviewer outcomes without source evidence, falsifiers, or probes.'),
             fixture('research-impersonation', true, 'Research validation blocks persona-boundary violations.'),
             fixture('oversized-l1', true, 'L1 token and item limits reject oversized active recall.'),
@@ -369,7 +369,7 @@ export async function buildRecallPulseGovernanceReport(root, opts = {}) {
             ],
             migration_paths: {
                 existing_missions: 'Run sks recallpulse run <mission-id> and sks recallpulse governance <mission-id> to add report-only artifacts.',
-                existing_research_artifacts: 'Research gates require agent display_name/persona/persona_boundary fields and the research_reviewer GPT-6 Astra Max binding; old ledgers must be migrated before claiming pass.',
+                existing_research_artifacts: 'Research gates require agent display_name/persona/persona_boundary fields and the research_reviewer latest deep-tier max binding; old ledgers must be migrated before claiming pass.',
                 generated_skills: 'Do not edit generated installed skills directly; rerun init/bootstrap from engine source when generated text needs refreshing.'
             },
             release_gate: 'RecallPulse remains report-only unless packcheck, selftest, sizecheck, registry metadata check, TriWiki validate, and RecallPulse fixture eval pass.'
@@ -882,7 +882,7 @@ function preservedRoutePersonality(routeId = '', routeName = '') {
         ImageUXReview: ("Image UX Review keeps " + IMAGEGEN_MODEL + " annotated raster review identity"),
         ComputerUse: 'Computer Use keeps maximum-speed native Mac/non-web visual lane identity',
         Goal: 'Goal uses Codex native /goal only and creates no SKS-owned persistence, artifact, loop, or fallback state',
-        Research: 'Research keeps Super Search evidence, three independent Astra Max review dimensions, bounded revision, paper, and falsification identity',
+        Research: 'Research keeps Super Search evidence, three independent deep-tier review dimensions, bounded revision, paper, and falsification identity',
         AutoResearch: 'AutoResearch keeps iterative experiment loop identity',
         DB: 'DB keeps conservative read-first destructive-operation safety identity',
         MadSKS: 'MAD-SKS keeps explicit scoped high-risk authorization identity',

package/dist/core/release/gate-affected-globs.js

@@ -31,7 +31,6 @@ export function affectedGlobsFor(id) {
                 'src/core/routes/design-policy.ts',
                 'src/scripts/installed-package-smoke-check.ts',
                 'src/scripts/postinstall-safe-side-effects-check.ts',
-                'test/blackbox/postinstall-safe-side-effects-packed.test.mjs',
                 'test/unit/postinstall-command.test.mjs',
                 'test/unit/publish-workflow-safety.test.mjs'
             ];

package/dist/core/research/mock-result.js

@@ -1,5 +1,5 @@
 import path from 'node:path';
-import { ASTRA_SUBAGENT_MODEL, SUBAGENT_EFFORT } from '../subagents/model-policy.js';
+import { thinkingSubagentModel, SUBAGENT_EFFORT } from '../subagents/model-policy.js';
 import { nowIso, writeJsonAtomic, writeTextAtomic } from '../fsx.js';
 import { CLAIM_EVIDENCE_MATRIX_ARTIFACT, buildClaimEvidenceMatrixFromLedgers, writeClaimEvidenceMatrix } from './claim-evidence-matrix.js';
 import { DEFAULT_RESEARCH_QUALITY_CONTRACT, writeResearchQualityContract } from './research-quality-contract.js';
@@ -174,7 +174,7 @@ export async function writeMockResearchResult(dir, plan) {
             mandate: agent.mandate,
             model_policy: {
                 custom_agent: RESEARCH_REVIEWER_CUSTOM_AGENT,
-                model: ASTRA_SUBAGENT_MODEL,
+                model: thinkingSubagentModel(),
                 reasoning_effort: SUBAGENT_EFFORT,
                 enforcement_source: 'mock_fixture'
             },

package/dist/core/research/research-adversarial-review.js

@@ -6,7 +6,7 @@ import { buildOfficialSubagentPrompt } from '../subagents/official-subagent-prom
 import { codexAppSessionKey, detectCodexAppSession, runOfficialSubagentWorkflow } from '../subagents/official-subagent-runner.js';
 import { readOfficialSubagentConfig } from '../subagents/official-subagent-config.js';
 import { SUBAGENT_EVENT_LOG_FILENAME, SUBAGENT_EVIDENCE_FILENAME, SUBAGENT_PARENT_SUMMARY_FILENAME, bindTrustworthySubagentParentSummaryToRun, normalizeSubagentParentSummary, persistOrReuseTrustworthySubagentParentSummary, readSubagentEvents, writeSubagentEvidence } from '../subagents/subagent-evidence.js';
-import { THINKING_SUBAGENT_MODEL, SUBAGENT_EFFORT } from '../subagents/model-policy.js';
+import { thinkingSubagentModel, SUBAGENT_EFFORT } from '../subagents/model-policy.js';
 import { RESEARCH_AGENT_COUNCIL, RESEARCH_REVIEWER_CONFIG_ARTIFACT, RESEARCH_REVIEWER_CUSTOM_AGENT, researchAgentAgentName, researchPaperArtifactForPlan } from '../research.js';
 import { normalizeResearchSynthesisOutput } from './research-synthesis-writer.js';
 import { buildResearchReviewArtifactDigest, validateResearchReviewArtifactDigest } from './research-review-artifact-digest.js';
@@ -301,7 +301,7 @@ export function buildResearchAdversarialPlan(plan, maxCycles = 3, maxThreads = R
             persona: agent.persona,
             persona_boundary: agent.persona_boundary,
             custom_agent: RESEARCH_REVIEWER_CUSTOM_AGENT,
-            model_policy: `${THINKING_SUBAGENT_MODEL} ${SUBAGENT_EFFORT}`,
+            model_policy: `${thinkingSubagentModel()} ${SUBAGENT_EFFORT}`,
             model_policy_source: modelPolicyEvidence?.source || RESEARCH_REVIEWER_CONFIG_ARTIFACT,
             model_policy_sha256: modelPolicyEvidence?.sha256 || null
         })),
@@ -533,7 +533,7 @@ async function finalizeResearchAdversarialArtifacts(input, plan, reviewCycles, r
         publication_acceptance_guaranteed: false,
         reviewer_model_policy: {
             custom_agent: RESEARCH_REVIEWER_CUSTOM_AGENT,
-            model: THINKING_SUBAGENT_MODEL,
+            model: thinkingSubagentModel(),
             reasoning_effort: SUBAGENT_EFFORT,
             enforcement_source: plan?.model_policy_evidence?.source || RESEARCH_REVIEWER_CONFIG_ARTIFACT,
             config_sha256: plan?.model_policy_evidence?.sha256 || null,
@@ -581,7 +581,7 @@ async function writeCompatibilityCouncilArtifacts(dir, plan, finalReview, gate)
             mandate: agent.mandate,
             model_policy: {
                 custom_agent: RESEARCH_REVIEWER_CUSTOM_AGENT,
-                model: THINKING_SUBAGENT_MODEL,
+                model: thinkingSubagentModel(),
                 reasoning_effort: SUBAGENT_EFFORT,
                 enforcement_source: gate?.reviewer_model_policy?.enforcement_source || RESEARCH_REVIEWER_CONFIG_ARTIFACT,
                 config_sha256: gate?.reviewer_model_policy?.config_sha256 || null
@@ -951,7 +951,7 @@ async function prepareResearchSubagentRun(input, opts) {
         })),
         model_policy: {
             custom_agent: plan.phase === 'review' ? RESEARCH_REVIEWER_CUSTOM_AGENT : 'research_synthesizer',
-            model: THINKING_SUBAGENT_MODEL,
+            model: thinkingSubagentModel(),
             reasoning_effort: SUBAGENT_EFFORT,
             config: plan.phase === 'review' ? RESEARCH_REVIEWER_CONFIG_ARTIFACT : '.codex/agents/research-synthesizer.toml'
         }
@@ -997,7 +997,7 @@ async function verifyResearchReviewerRoleConfig(root) {
     const blockers = [
         ...(text.trim() ? [] : ['research_reviewer_agent_config_missing']),
         ...(name === RESEARCH_REVIEWER_CUSTOM_AGENT ? [] : [`research_reviewer_name_mismatch:${name || 'missing'}`]),
-        ...(model === THINKING_SUBAGENT_MODEL ? [] : [`research_reviewer_model_mismatch:${model || 'missing'}`]),
+        ...(model === thinkingSubagentModel() ? [] : [`research_reviewer_model_mismatch:${model || 'missing'}`]),
         ...(effort === SUBAGENT_EFFORT ? [] : [`research_reviewer_effort_mismatch:${effort || 'missing'}`]),
         ...(sandbox === 'read-only' ? [] : [`research_reviewer_sandbox_mismatch:${sandbox || 'missing'}`])
     ];
@@ -1019,7 +1019,7 @@ function mockResearchModelPolicyEvidence() {
         ok: true,
         source: 'mock_fixture:research_reviewer',
         name: RESEARCH_REVIEWER_CUSTOM_AGENT,
-        model: THINKING_SUBAGENT_MODEL,
+        model: thinkingSubagentModel(),
         reasoning_effort: SUBAGENT_EFFORT,
         sandbox_mode: 'read-only',
         sha256: null,

package/dist/core/research/research-claim-synthesizer.js

@@ -2,7 +2,7 @@ import path from 'node:path';
 import { readJson } from '../fsx.js';
 import { uniqueValues as unique } from '../text/strings.js';
 import { runCodexTask } from '../codex-control/codex-task-runner.js';
-import { THINKING_SUBAGENT_MODEL, SUBAGENT_EFFORT } from '../subagents/model-policy.js';
+import { thinkingSubagentModel, SUBAGENT_EFFORT } from '../subagents/model-policy.js';
 import { normalizeClaimEvidenceMatrix, validateClaimEvidenceMatrix } from './claim-evidence-matrix.js';
 export async function synthesizeResearchClaimEvidenceMatrix(input) {
     const result = await runCodexTask({
@@ -33,7 +33,7 @@ export async function synthesizeResearchClaimEvidenceMatrix(input) {
             hardTimeoutMs: input.timeoutMs,
             ...(input.deadlineMs === undefined ? {} : { deadlineEpochMs: input.deadlineMs })
         },
-        model: THINKING_SUBAGENT_MODEL,
+        model: thinkingSubagentModel(),
         reasoningEffort: SUBAGENT_EFFORT,
         modelReasoningEffort: SUBAGENT_EFFORT,
         serviceTier: 'fast'
@@ -178,7 +178,7 @@ function buildResearchClaimSynthesisPrompt(input) {
     const contract = input.plan?.quality_contract || {};
     return [
         'Build a semantic claim-evidence matrix for this Research mission.',
-        `This is a judgment-heavy task: use ${THINKING_SUBAGENT_MODEL} with ${SUBAGENT_EFFORT} reasoning.`,
+        `This is a judgment-heavy task: use ${thinkingSubagentModel()} with ${SUBAGENT_EFFORT} reasoning.`,
         'Return exactly one JSON object matching sks.claim-evidence-matrix.v1.',
         'Never reuse or merge discovery claim IDs merely because their strings match.',
         'Group sources only when their hydrated notes/content actually support the same written claim.',

package/dist/core/research/research-falsification-runner.js

@@ -2,7 +2,7 @@ import path from 'node:path';
 import { nowIso, readJson } from '../fsx.js';
 import { uniqueValues as unique } from '../text/strings.js';
 import { runCodexTask } from '../codex-control/codex-task-runner.js';
-import { THINKING_SUBAGENT_MODEL, SUBAGENT_EFFORT } from '../subagents/model-policy.js';
+import { thinkingSubagentModel, SUBAGENT_EFFORT } from '../subagents/model-policy.js';
 export async function runResearchFalsification(input) {
     const result = await runCodexTask({
         route: '$Research',
@@ -32,7 +32,7 @@ export async function runResearchFalsification(input) {
             hardTimeoutMs: input.timeoutMs,
             ...(input.deadlineMs === undefined ? {} : { deadlineEpochMs: input.deadlineMs })
         },
-        model: THINKING_SUBAGENT_MODEL,
+        model: thinkingSubagentModel(),
         reasoningEffort: SUBAGENT_EFFORT,
         modelReasoningEffort: SUBAGENT_EFFORT,
         serviceTier: 'fast'
@@ -107,7 +107,7 @@ export function normalizeResearchFalsification(value, claimMatrix, sourceLedger)
 function buildResearchFalsificationPrompt(input) {
     return [
         'Attempt to falsify the key claims in this Research mission before manuscript synthesis.',
-        `This is a judgment-heavy task: use ${THINKING_SUBAGENT_MODEL} with ${SUBAGENT_EFFORT} reasoning.`,
+        `This is a judgment-heavy task: use ${thinkingSubagentModel()} with ${SUBAGENT_EFFORT} reasoning.`,
         'Return exactly one JSON object matching sks.falsification-ledger.v1.',
         'Do not mark a claim as surviving by default. Compare the written claim with actual source notes/content and counterevidence.',
         'Use only known claim IDs and source IDs. A generic attack with no source-linked reasoning is invalid.',

package/dist/core/research/research-super-search.js

@@ -2,10 +2,13 @@ import path from 'node:path';
 import { readJson, sha256 } from '../fsx.js';
 import { runCodexTask } from '../codex-control/codex-task-runner.js';
 import { runSuperSearch } from '../super-search/index.js';
-import { TERRA_SUBAGENT_EFFORT, TERRA_SUBAGENT_MODEL } from '../subagents/model-policy.js';
+import { TERRA_SUBAGENT_EFFORT } from '../subagents/model-policy.js';
+import { latestModelForTier } from '../subagents/model-tiers.js';
 import {} from './research-source-shards.js';
 export const RESEARCH_SOURCE_ACQUISITION_MODEL_POLICY = Object.freeze({
-    model: TERRA_SUBAGENT_MODEL,
+    get model() {
+        return latestModelForTier('context');
+    },
     model_reasoning_effort: TERRA_SUBAGENT_EFFORT
 });
 export async function runResearchSuperSearchShard(input) {

package/dist/core/research/research-synthesis-writer.js

@@ -2,7 +2,7 @@ import path from 'node:path';
 import { readJson, writeJsonAtomic, writeTextAtomic, nowIso } from '../fsx.js';
 import { runCodexTask } from '../codex-control/codex-task-runner.js';
 import { researchPaperArtifactForPlan } from '../research.js';
-import { THINKING_SUBAGENT_MODEL, SUBAGENT_EFFORT } from '../subagents/model-policy.js';
+import { thinkingSubagentModel, SUBAGENT_EFFORT } from '../subagents/model-policy.js';
 import { analyzeResearchReportQuality, countWords } from './research-report-quality.js';
 import { analyzeResearchRepetition } from './research-repetition-detector.js';
 import { buildRealisticResearchPaper, buildRealisticResearchReport } from './research-realistic-report.js';
@@ -78,7 +78,7 @@ export async function runResearchCodexSynthesisWriter(input) {
             hardTimeoutMs: input.timeoutMs || 120000,
             ...(input.deadlineMs === undefined ? {} : { deadlineEpochMs: input.deadlineMs })
         },
-        model: THINKING_SUBAGENT_MODEL,
+        model: thinkingSubagentModel(),
         reasoningEffort: SUBAGENT_EFFORT,
         modelReasoningEffort: SUBAGENT_EFFORT,
         serviceTier: 'fast'

package/dist/core/research.js

@@ -1,5 +1,5 @@
 import path from 'node:path';
-import { ASTRA_SUBAGENT_MODEL, SUBAGENT_EFFORT } from './subagents/model-policy.js';
+import { thinkingSubagentModel, SUBAGENT_EFFORT } from './subagents/model-policy.js';
 import { appendJsonlBounded, nowIso, readJson, readText, writeJsonAtomic, writeTextAtomic, exists } from './fsx.js';
 import { OUTCOME_RUBRIC } from './proof-field.js';
 import { RESEARCH_REVIEWER_CONTRACT } from './recallpulse.js';
@@ -100,7 +100,7 @@ export function researchNativeAgentPlan(prompt = '', opts = {}) {
         role: persona.role,
         mandate: persona.mandate,
         custom_agent: RESEARCH_REVIEWER_CUSTOM_AGENT,
-        model: ASTRA_SUBAGENT_MODEL,
+        model: thinkingSubagentModel(),
         reasoning_effort: SUBAGENT_EFFORT,
         read_only: true
     }));
@@ -202,10 +202,10 @@ export function createResearchPlan(prompt, opts = {}) {
             policy: 'Assign distinct evidence, method, and falsification review dimensions.',
             effort_policy: {
                 custom_agent: RESEARCH_REVIEWER_CUSTOM_AGENT,
-                required_model: ASTRA_SUBAGENT_MODEL,
+                required_model: thinkingSubagentModel(),
                 required_effort: SUBAGENT_EFFORT,
                 applies_to: 'every_official_adversarial_reviewer',
-                rule: 'Every adversarial reviewer uses the verified research_reviewer custom agent configuration with GPT-6 Astra Max. Long-context and source-tool acquisition uses Astra Medium; synthesis, falsification, and review use Astra Max.'
+                rule: 'Every adversarial reviewer uses the verified research_reviewer custom agent configuration on the latest deep-tier model at max effort. Long-context and source-tool acquisition uses the context tier; synthesis, falsification, and review use the deep tier.'
             },
             debate_policy: {
                 mode: 'independent_adversarial_reviews_with_bounded_revision',
@@ -473,7 +473,7 @@ export function defaultAgentLedger(plan = null) {
             mandate: agent.mandate,
             model_policy: {
                 custom_agent: RESEARCH_REVIEWER_CUSTOM_AGENT,
-                model: ASTRA_SUBAGENT_MODEL,
+                model: thinkingSubagentModel(),
                 reasoning_effort: SUBAGENT_EFFORT,
                 enforcement_source: RESEARCH_REVIEWER_CONFIG_ARTIFACT
             },

package/dist/core/routes/dollar-manifest-lite.js

@@ -1,6 +1,6 @@
 import { IMAGEGEN_MODEL } from '../imagegen/imagegen-model-policy.js';
 import { normalizeDollarSkillName, prefixKnownSksDollarReferences, sksPrefixedDollarCommand } from './dollar-prefix.js';
-const NARUTO_DESCRIPTION = '$Naruto runs explicit parallel work through Codex official subagents. The selected parent owns decomposition, integration, and verification; standalone launches default to GPT-6 Astra. Delegate independent slices with disjoint writes, use Astra for every child and vary only effort by task, and honor explicit counts and measured host limits. max_threads is a cap, never a target.';
+const NARUTO_DESCRIPTION = '$Naruto runs implementation work through Codex official subagents. The parent orchestrates: it owns decomposition, integration, and verification, spawns a child per disjoint slice, and does not implement slices itself; standalone launches default to the latest deep-tier model. Each child runs the newest model of its tier, Jev mode picks the tier on spawn, and explicit counts and measured host limits are honored. max_threads is a cap, never a target.';
 const COMPUTER_USE_DESCRIPTION = 'Maximum-speed Codex Computer Use lane for native macOS, desktop-app, OS-settings, and non-web visual tasks only. Browser, localhost, website, webapp, and web-based app verification must route through Codex Chrome Extension readiness first.';
 const DOLLAR_COMMANDS_LITE_BASE = [
     { command: '$DFix', route: 'fast direct fix', description: 'Tiny simple direct edits such as copy, labels, typos, wording, spacing, colors, or clearly scoped one-line changes. Bypasses the general SKS pipeline and runs an ultralight, no-record task-list path.' },