Context L

kernel/bin/agent

@@ -1404,6 +1404,64 @@ function bareAgentTrimMessages(msgs, maxBytes) {
   return out
 }
 
+/**
+ * Rough token estimate (chars/4). Good enough for local ctx budgeting.
+ * @param {unknown} value
+ */
+function bareAgentEstimateTokens(value) {
+  try {
+    const n = JSON.stringify(value == null ? '' : value).length
+    return Math.max(0, Math.ceil(n / 4))
+  } catch {
+    return 0
+  }
+}
+
+/**
+ * Trim history + shrink system content so prompt fits a QVAC ctx window.
+ * Reserves room for completion and optional tools JSON.
+ * @param {unknown[]} msgs
+ * @param {number} ctxSize
+ * @param {{ tools?: unknown[], reserveCompletion?: number }} [opts]
+ */
+function bareAgentTrimMessagesForCtx(msgs, ctxSize, opts) {
+  const ctx = Math.max(2048, Math.floor(Number(ctxSize) || 4096))
+  const reserve =
+    opts && Number.isFinite(Number(opts.reserveCompletion))
+      ? Math.max(256, Math.floor(Number(opts.reserveCompletion)))
+      : Math.min(1024, Math.max(256, Math.floor(ctx * 0.15)))
+  const toolsTok = bareAgentEstimateTokens(
+    opts && Array.isArray(opts.tools) && opts.tools.length ? opts.tools : []
+  )
+  const budget = Math.max(512, ctx - reserve - toolsTok)
+
+  /** @type {unknown[]} */
+  let out = Array.isArray(msgs) ? msgs.slice() : []
+  // Drop middle turns first (keep system + recent).
+  while (bareAgentEstimateTokens(out) > budget && out.length > 3) {
+    out.splice(2, 1)
+  }
+  // Shrink system blob if still over (workspace/man/skills dominate).
+  if (bareAgentEstimateTokens(out) > budget && out[0] && typeof out[0] === 'object') {
+    const sys = /** @type {Record<string, unknown>} */ (out[0])
+    if (sys.role === 'system' && typeof sys.content === 'string') {
+      let content = sys.content
+      let guard = 0
+      while (
+        bareAgentEstimateTokens(out) > budget &&
+        content.length > 800 &&
+        guard < 24
+      ) {
+        content =
+          content.slice(0, Math.floor(content.length * 0.82)) + '\n… truncated'
+        out[0] = { ...sys, content }
+        guard++
+      }
+    }
+  }
+  return out
+}
+
 /**
  * Best-effort load instruction files.
  * @param {Record<string, unknown>} ctx
@@ -1597,6 +1655,16 @@ async function bareAgentResetChatSession(ctx, paths, argv0) {
 
 /** @typedef {'lite'|'recommended'|'strong'|'tool-tiny'} BareAgentQvacProfileId */
 
+/**
+ * Native context windows from upstream model cards (not conservative laptop defaults).
+ * Qwen3 dense (0.6B/1.7B/4B): https://huggingface.co/Qwen/Qwen3-0.6B — 32,768
+ * Llama 3.2 1B (tool-calling finetune base): Meta Llama 3.2 — 128,000
+ * Absolute ceiling for overrides / host clamp.
+ */
+const BARE_AGENT_QVAC_CTX_QWEN3 = 32768
+const BARE_AGENT_QVAC_CTX_LLAMA32_1B = 131072
+const BARE_AGENT_QVAC_CTX_ABSOLUTE_MAX = 131072
+
 /**
  * @typedef {{
  *   id: BareAgentQvacProfileId,
@@ -1617,50 +1685,69 @@ const BARE_AGENT_QVAC_PROFILES = {
     id: 'lite',
     label: 'Lite',
     description:
-      'Smallest download. Good for weak machines / ~4 GB VRAM; limited tool use.',
+      'Smallest download (Qwen3-0.6B). Model-card context 32k; limited tool use.',
     chatModel: 'QWEN3_600M_INST_Q4',
     tools: false,
     minRamGb: 4,
     minDiskGb: 2,
     approxDownloadGb: 0.5,
-    ctxSize: 4096
+    ctxSize: BARE_AGENT_QVAC_CTX_QWEN3
   },
   recommended: {
     id: 'recommended',
     label: 'Recommended',
     description:
-      'Best balance for Bare OS agent tool calling (fits ~4–8 GB VRAM at ctx 8k).',
+      'Best balance for Bare OS agent tool calling (Qwen3-1.7B, model-card context 32k).',
     chatModel: 'QWEN3_1_7B_INST_Q4',
     tools: true,
     minRamGb: 8,
     minDiskGb: 5,
     approxDownloadGb: 2.5,
-    ctxSize: 8192
+    ctxSize: BARE_AGENT_QVAC_CTX_QWEN3
   },
   strong: {
     id: 'strong',
     label: 'Strong',
-    description: 'Better reasoning. Needs more RAM/VRAM/disk.',
+    description:
+      'Better reasoning (Qwen3-4B). Model-card context 32k (131k with YaRN not enabled).',
     chatModel: 'QWEN3_4B_INST_Q4_K_M',
     tools: true,
     minRamGb: 16,
     minDiskGb: 8,
     approxDownloadGb: 3.5,
-    ctxSize: 8192
+    ctxSize: BARE_AGENT_QVAC_CTX_QWEN3
   },
   'tool-tiny': {
     id: 'tool-tiny',
     label: 'Tool-tiny',
-    description: 'Llama tool-calling 1B fallback if Qwen tools misbehave.',
+    description:
+      'Llama 3.2 1B tool-calling fallback (model-card context 128k).',
     chatModel: 'LLAMA_TOOL_CALLING_1B_INST_Q4_K',
     tools: true,
     minRamGb: 6,
     minDiskGb: 3,
     approxDownloadGb: 1,
-    ctxSize: 4096
+    ctxSize: BARE_AGENT_QVAC_CTX_LLAMA32_1B
   }
 }
 
+/**
+ * Model-card / train context for a QVAC registry id (fallback 32k).
+ * @param {string} [modelId]
+ * @returns {number}
+ */
+function bareAgentQvacModelCardCtxSize(modelId) {
+  const id = String(modelId || '')
+    .trim()
+    .toUpperCase()
+  if (!id) return BARE_AGENT_QVAC_CTX_QWEN3
+  if (id.includes('LLAMA') || id.includes('LLAMA_TOOL')) {
+    return BARE_AGENT_QVAC_CTX_LLAMA32_1B
+  }
+  if (id.includes('QWEN3')) return BARE_AGENT_QVAC_CTX_QWEN3
+  return BARE_AGENT_QVAC_CTX_QWEN3
+}
+
 /** @returns {BareAgentQvacProfile[]} */
 function bareAgentQvacProfileList() {
   return Object.values(BARE_AGENT_QVAC_PROFILES)
@@ -1749,17 +1836,33 @@ function bareAgentResolveBackend(config) {
 }
 
 /**
- * Resolve context window for QVAC load (profile default or config override).
+ * Resolve context window for QVAC load (profile/model-card default or config override).
+ * Tool schemas alone are ~4.5k tokens, so tool-enabled profiles need ≥8192.
  * @param {Record<string, unknown>} config
  * @param {BareAgentQvacProfile} profile
  */
 function bareAgentQvacResolveCtxSize(config, profile) {
+  const modelId = String(
+    (config && (config.qvac_model || config.model)) ||
+      (profile && profile.chatModel) ||
+      ''
+  )
+  const cardCtx = bareAgentQvacModelCardCtxSize(modelId)
+  const profileCtx = Math.max(
+    2048,
+    Number(profile && profile.ctxSize) || cardCtx
+  )
   const raw = Number(config && config.qvac_ctx_size)
+  let ctx
   if (Number.isFinite(raw) && raw >= 2048) {
-    // Cap high overrides; 32k+ KV often OOMs laptop GPUs (~4 GB VRAM).
-    return Math.min(32768, Math.floor(raw))
+    ctx = Math.min(BARE_AGENT_QVAC_CTX_ABSOLUTE_MAX, Math.floor(raw))
+  } else {
+    // Prefer explicit profile ctx (already model-card sized); else card for model id.
+    ctx = Math.min(BARE_AGENT_QVAC_CTX_ABSOLUTE_MAX, Math.max(profileCtx, cardCtx))
   }
-  return Math.max(2048, Number(profile.ctxSize) || 8192)
+  // Full Bare OS tool list ≈ 4.5k tokens; leave room for system + reply.
+  if (profile && profile.tools !== false && ctx < 8192) ctx = 8192
+  return ctx
 }
 
 /**
@@ -6846,11 +6949,36 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) {
   if (needWizard) await bareAgentSyncWorkspaceFromConfig(ctx, paths, config)
 
   const backendForPrompt = bareAgentResolveBackend(config)
+  const promptProfile = bareAgentQvacGetProfile(
+    String(config.qvac_profile || config.profile || '')
+  )
+  const promptCtxSize =
+    backendForPrompt === 'qvac'
+      ? bareAgentQvacResolveCtxSize(config, promptProfile)
+      : 32768
   // Local QVAC models have a hard ctx window; keep the system prompt lean.
-  const workspaceBudget = backendForPrompt === 'qvac' ? 8000 : 24000
-  const manBudget = backendForPrompt === 'qvac' ? 4000 : 12000
-  const instructionsBudget = backendForPrompt === 'qvac' ? 3000 : 8000
-  const skillsBudget = backendForPrompt === 'qvac' ? 2000 : 4000
+  // Scale char budgets from ctx (chars ≈ tokens*4); leave room for tools/reply.
+  const toolsOn = backendForPrompt === 'qvac' && promptProfile.tools !== false
+  const promptCharBudget = Math.max(
+    4000,
+    Math.floor(promptCtxSize * 4 * (toolsOn ? 0.35 : 0.7))
+  )
+  const workspaceBudget =
+    backendForPrompt === 'qvac'
+      ? Math.min(8000, Math.floor(promptCharBudget * 0.35))
+      : 24000
+  const manBudget =
+    backendForPrompt === 'qvac'
+      ? Math.min(4000, Math.floor(promptCharBudget * 0.2))
+      : 12000
+  const instructionsBudget =
+    backendForPrompt === 'qvac'
+      ? Math.min(3000, Math.floor(promptCharBudget * 0.2))
+      : 8000
+  const skillsBudget =
+    backendForPrompt === 'qvac'
+      ? Math.min(2000, Math.floor(promptCharBudget * 0.15))
+      : 4000
 
   const workspacePromptBlock = await bareAgentLoadWorkspacePrompt(
     ctx,
@@ -7005,6 +7133,27 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) {
       const backendIter = bareAgentResolveBackend(configRef.current)
       /** @type {Record<string, unknown>} */
       const providerNow = String(configRef.current.provider || '').trim().toLowerCase()
+      let qvacToolsForTurn = /** @type {unknown[]} */ ([])
+      if (backendIter === 'qvac') {
+        const profileEarly = bareAgentQvacGetProfile(
+          String(configRef.current.qvac_profile || 'recommended')
+        )
+        const ctxEarly = bareAgentQvacResolveCtxSize(
+          configRef.current,
+          profileEarly
+        )
+        qvacToolsForTurn =
+          profileEarly.tools !== false
+            ? bareAgentFlattenToolsForQvac(tools)
+            : []
+        messages = bareAgentTrimMessagesForCtx(messages, ctxEarly, {
+          tools: qvacToolsForTurn,
+          reserveCompletion: Math.min(
+            1024,
+            Number(configRef.current.max_tokens) || 512
+          )
+        })
+      }
       if (reasoningSettings.enabled && reasoningSettings.mode === 'trace') {
         bareAgentWriteOut(
           ctx,
@@ -7095,6 +7244,29 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) {
               EDIT_ANSI_RESET +
               '\n'
           )
+          try {
+            const st =
+              typeof ctx.bareOsQvacStatus === 'function'
+                ? ctx.bareOsQvacStatus()
+                : null
+            if (st && (st.cacheDir || st.hdmsPath || st.backendsDir)) {
+              bareAgentWriteOut(
+                ctx,
+                stdout,
+                bareEditSgr('dim', useColor) +
+                  '[qvac] cache=' +
+                  String(st.cacheDir || '') +
+                  ' hdms=' +
+                  String(st.hdmsPath || '/mnt/models') +
+                  ' backends=' +
+                  String(st.backendsDir || '(unresolved)') +
+                  EDIT_ANSI_RESET +
+                  '\n'
+              )
+            }
+          } catch {
+            /* ignore */
+          }
           if (typeof ctx.bareOsQvacLoadModel === 'function') {
             const loaded = await ctx.bareOsQvacLoadModel({
               modelSrc,
@@ -7145,7 +7317,7 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) {
           }
           await ctx.bareOsQvacComplete({
             history: messages,
-            tools: bareAgentFlattenToolsForQvac(tools),
+            tools: qvacToolsForTurn,
             stream: Boolean(configRef.current.stream !== false),
             captureThinking: reasoningSettings.enabled,
             modelSrc,

kernel/etc/bare-os/posix_utilities.json

@@ -1,7 +1,7 @@
 {
   "schema": 2,
   "profileId": "bare-os-posix-like",
-  "generatedAt": "2026-07-31T21:35:54.625Z",
+  "generatedAt": "2026-07-31T23:16:21.588Z",
   "note": "Sparse POSIX Issue 7 coverage hints; commandIndex is generated each coreutils build.",
   "commandIndex": [
     {

kernel/lib/bare/shell-completion.json

@@ -1,6 +1,6 @@
 {
   "schema": 1,
-  "atMs": 1785533754625,
+  "atMs": 1785539781587,
   "commands": [
     "agent",
     "appctl",

kernel/share/agent-workspace/README-agent.md

@@ -30,10 +30,17 @@ After **`agent --config`** / **`--setup`** (or changing **`owner_name`** / **`ag
 
 Default backend is **QVAC** (QuantumVerse Automatic Computer) — local in-process LLM via the booter **`ctx.bareOsQvac*`** bridge (`@qvac/sdk`). Weights download on first use after you pick a profile in **`agent --config`**.
 
+Models are stored for durability on HDMS label **`models`** (guest path **`/mnt/models/`**) and materialize into a host cache for llama.cpp load:
+
+- Host cache: **`$BARE_OS_HOST_DATA/qvac/models`** (default **`~/.bare-os/qvac/models`**)
+- Runtime config: **`$BARE_OS_HOST_DATA/qvac/qvac.config.json`** (`QVAC_CONFIG_PATH`, absolute `cacheDirectory`)
+
+First load downloads via the QVAC registry into the **host cache**, then mirrors into HDMS. If HDMS already has the GGUF, Bare OS materializes it to the host cache before load. If auto-create fails, run **`hdms create models`** after unlock.
+
 - `backend` — `qvac` (default) or `rest`
 - `qvac_profile` — `lite` / `recommended` / `strong` / `tool-tiny`
 - `qvac_model` — registry id (e.g. `QWEN3_1_7B_INST_Q4`)
-- `qvac_ctx_size` — optional context window override (profile defaults: lite/tool-tiny 4096, recommended/strong 8192; max 32768)
+- `qvac_ctx_size` — optional context window override. Profile defaults match model cards: Qwen3 profiles **32768**, tool-tiny (Llama 3.2 1B) **131072** (max override 131072). Tool-enabled profiles floor at 8192 (tool schemas alone are ~4.5k tokens). Cap further with host env `BARE_OS_QVAC_MAX_CTX` if VRAM is tight.
 - `qvac_device` — `gpu` (default) or `cpu` (bypass Vulkan; slower)
 - `qvac_main_gpu` — `auto` (default: pick highest-VRAM Vulkan GPU, else CPU), `dedicated`, `integrated`, or device index (`0`, `1`, …)
 - `qvac_gpu_layers` — optional layer offload count (`0` = CPU weights)

packages/bare-os-booter/lib/bare-os-qvac-host.mjs

@@ -491,13 +491,14 @@ export function createBareOsQvacBridge(opts = {}) {
     if (!key) throw new Error('bareOsQvacLoadModel: modelSrc required')
 
     const toolsEnabled = o.tools !== false
-    // Respect caller/profile ctx; clamp default max to 8k (old agents sent 16k/32k).
+    // Allow model-card windows (Qwen3 32k, Llama 3.2 1B 128k). Override with
+    // BARE_OS_QVAC_MAX_CTX to cap KV on small GPUs.
     const maxCtxRaw = Number(env.BARE_OS_QVAC_MAX_CTX)
     const maxCtx =
       Number.isFinite(maxCtxRaw) && maxCtxRaw >= 2048
-        ? Math.min(32768, Math.floor(maxCtxRaw))
-        : 8192
-    const ctxSize = Math.max(2048, Math.min(maxCtx, Number(o.ctxSize) || 4096))
+        ? Math.min(131072, Math.floor(maxCtxRaw))
+        : 131072
+    const ctxSize = Math.max(2048, Math.min(maxCtx, Number(o.ctxSize) || 32768))
     const resolved = bareOsQvacResolveDeviceConfig(o, env)
     const verbEnv = Number(env.BARE_OS_QVAC_VERBOSITY)
     const verbosity =
@@ -847,7 +848,10 @@ export function createBareOsQvacBridge(opts = {}) {
     const s = await ensureApi()
     const onEvent =
       typeof o.onEvent === 'function' ? o.onEvent : () => {}
-    const flatTools = bareOsQvacFlattenTools(o.tools || [])
+    const toolsEnabled = o.toolsEnabled !== false
+    const flatTools = toolsEnabled
+      ? bareOsQvacFlattenTools(o.tools || [])
+      : []
     const history = Array.isArray(o.history) ? o.history : []
 
     const run = s.completion({

packages/bare-os-coreutils/lib/agent-qvac.js

@@ -4,6 +4,16 @@
 
 /** @typedef {'lite'|'recommended'|'strong'|'tool-tiny'} BareAgentQvacProfileId */
 
+/**
+ * Native context windows from upstream model cards (not conservative laptop defaults).
+ * Qwen3 dense (0.6B/1.7B/4B): https://huggingface.co/Qwen/Qwen3-0.6B — 32,768
+ * Llama 3.2 1B (tool-calling finetune base): Meta Llama 3.2 — 128,000
+ * Absolute ceiling for overrides / host clamp.
+ */
+const BARE_AGENT_QVAC_CTX_QWEN3 = 32768
+const BARE_AGENT_QVAC_CTX_LLAMA32_1B = 131072
+const BARE_AGENT_QVAC_CTX_ABSOLUTE_MAX = 131072
+
 /**
  * @typedef {{
  *   id: BareAgentQvacProfileId,
@@ -24,50 +34,69 @@ const BARE_AGENT_QVAC_PROFILES = {
     id: 'lite',
     label: 'Lite',
     description:
-      'Smallest download. Good for weak machines / ~4 GB VRAM; limited tool use.',
+      'Smallest download (Qwen3-0.6B). Model-card context 32k; limited tool use.',
     chatModel: 'QWEN3_600M_INST_Q4',
     tools: false,
     minRamGb: 4,
     minDiskGb: 2,
     approxDownloadGb: 0.5,
-    ctxSize: 4096
+    ctxSize: BARE_AGENT_QVAC_CTX_QWEN3
   },
   recommended: {
     id: 'recommended',
     label: 'Recommended',
     description:
-      'Best balance for Bare OS agent tool calling (fits ~4–8 GB VRAM at ctx 8k).',
+      'Best balance for Bare OS agent tool calling (Qwen3-1.7B, model-card context 32k).',
     chatModel: 'QWEN3_1_7B_INST_Q4',
     tools: true,
     minRamGb: 8,
     minDiskGb: 5,
     approxDownloadGb: 2.5,
-    ctxSize: 8192
+    ctxSize: BARE_AGENT_QVAC_CTX_QWEN3
   },
   strong: {
     id: 'strong',
     label: 'Strong',
-    description: 'Better reasoning. Needs more RAM/VRAM/disk.',
+    description:
+      'Better reasoning (Qwen3-4B). Model-card context 32k (131k with YaRN not enabled).',
     chatModel: 'QWEN3_4B_INST_Q4_K_M',
     tools: true,
     minRamGb: 16,
     minDiskGb: 8,
     approxDownloadGb: 3.5,
-    ctxSize: 8192
+    ctxSize: BARE_AGENT_QVAC_CTX_QWEN3
   },
   'tool-tiny': {
     id: 'tool-tiny',
     label: 'Tool-tiny',
-    description: 'Llama tool-calling 1B fallback if Qwen tools misbehave.',
+    description:
+      'Llama 3.2 1B tool-calling fallback (model-card context 128k).',
     chatModel: 'LLAMA_TOOL_CALLING_1B_INST_Q4_K',
     tools: true,
     minRamGb: 6,
     minDiskGb: 3,
     approxDownloadGb: 1,
-    ctxSize: 4096
+    ctxSize: BARE_AGENT_QVAC_CTX_LLAMA32_1B
   }
 }
 
+/**
+ * Model-card / train context for a QVAC registry id (fallback 32k).
+ * @param {string} [modelId]
+ * @returns {number}
+ */
+function bareAgentQvacModelCardCtxSize(modelId) {
+  const id = String(modelId || '')
+    .trim()
+    .toUpperCase()
+  if (!id) return BARE_AGENT_QVAC_CTX_QWEN3
+  if (id.includes('LLAMA') || id.includes('LLAMA_TOOL')) {
+    return BARE_AGENT_QVAC_CTX_LLAMA32_1B
+  }
+  if (id.includes('QWEN3')) return BARE_AGENT_QVAC_CTX_QWEN3
+  return BARE_AGENT_QVAC_CTX_QWEN3
+}
+
 /** @returns {BareAgentQvacProfile[]} */
 function bareAgentQvacProfileList() {
   return Object.values(BARE_AGENT_QVAC_PROFILES)
@@ -156,17 +185,33 @@ function bareAgentResolveBackend(config) {
 }
 
 /**
- * Resolve context window for QVAC load (profile default or config override).
+ * Resolve context window for QVAC load (profile/model-card default or config override).
+ * Tool schemas alone are ~4.5k tokens, so tool-enabled profiles need ≥8192.
  * @param {Record<string, unknown>} config
  * @param {BareAgentQvacProfile} profile
  */
 function bareAgentQvacResolveCtxSize(config, profile) {
+  const modelId = String(
+    (config && (config.qvac_model || config.model)) ||
+      (profile && profile.chatModel) ||
+      ''
+  )
+  const cardCtx = bareAgentQvacModelCardCtxSize(modelId)
+  const profileCtx = Math.max(
+    2048,
+    Number(profile && profile.ctxSize) || cardCtx
+  )
   const raw = Number(config && config.qvac_ctx_size)
+  let ctx
   if (Number.isFinite(raw) && raw >= 2048) {
-    // Cap high overrides; 32k+ KV often OOMs laptop GPUs (~4 GB VRAM).
-    return Math.min(32768, Math.floor(raw))
+    ctx = Math.min(BARE_AGENT_QVAC_CTX_ABSOLUTE_MAX, Math.floor(raw))
+  } else {
+    // Prefer explicit profile ctx (already model-card sized); else card for model id.
+    ctx = Math.min(BARE_AGENT_QVAC_CTX_ABSOLUTE_MAX, Math.max(profileCtx, cardCtx))
   }
-  return Math.max(2048, Number(profile.ctxSize) || 8192)
+  // Full Bare OS tool list ≈ 4.5k tokens; leave room for system + reply.
+  if (profile && profile.tools !== false && ctx < 8192) ctx = 8192
+  return ctx
 }
 
 /**

packages/bare-os-coreutils/lib/agent-state.js

@@ -369,6 +369,64 @@ function bareAgentTrimMessages(msgs, maxBytes) {
   return out
 }
 
+/**
+ * Rough token estimate (chars/4). Good enough for local ctx budgeting.
+ * @param {unknown} value
+ */
+function bareAgentEstimateTokens(value) {
+  try {
+    const n = JSON.stringify(value == null ? '' : value).length
+    return Math.max(0, Math.ceil(n / 4))
+  } catch {
+    return 0
+  }
+}
+
+/**
+ * Trim history + shrink system content so prompt fits a QVAC ctx window.
+ * Reserves room for completion and optional tools JSON.
+ * @param {unknown[]} msgs
+ * @param {number} ctxSize
+ * @param {{ tools?: unknown[], reserveCompletion?: number }} [opts]
+ */
+function bareAgentTrimMessagesForCtx(msgs, ctxSize, opts) {
+  const ctx = Math.max(2048, Math.floor(Number(ctxSize) || 4096))
+  const reserve =
+    opts && Number.isFinite(Number(opts.reserveCompletion))
+      ? Math.max(256, Math.floor(Number(opts.reserveCompletion)))
+      : Math.min(1024, Math.max(256, Math.floor(ctx * 0.15)))
+  const toolsTok = bareAgentEstimateTokens(
+    opts && Array.isArray(opts.tools) && opts.tools.length ? opts.tools : []
+  )
+  const budget = Math.max(512, ctx - reserve - toolsTok)
+
+  /** @type {unknown[]} */
+  let out = Array.isArray(msgs) ? msgs.slice() : []
+  // Drop middle turns first (keep system + recent).
+  while (bareAgentEstimateTokens(out) > budget && out.length > 3) {
+    out.splice(2, 1)
+  }
+  // Shrink system blob if still over (workspace/man/skills dominate).
+  if (bareAgentEstimateTokens(out) > budget && out[0] && typeof out[0] === 'object') {
+    const sys = /** @type {Record<string, unknown>} */ (out[0])
+    if (sys.role === 'system' && typeof sys.content === 'string') {
+      let content = sys.content
+      let guard = 0
+      while (
+        bareAgentEstimateTokens(out) > budget &&
+        content.length > 800 &&
+        guard < 24
+      ) {
+        content =
+          content.slice(0, Math.floor(content.length * 0.82)) + '\n… truncated'
+        out[0] = { ...sys, content }
+        guard++
+      }
+    }
+  }
+  return out
+}
+
 /**
  * Best-effort load instruction files.
  * @param {Record<string, unknown>} ctx

packages/bare-os-coreutils/lib/agent-tui.js

@@ -900,11 +900,36 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) {
   if (needWizard) await bareAgentSyncWorkspaceFromConfig(ctx, paths, config)
 
   const backendForPrompt = bareAgentResolveBackend(config)
+  const promptProfile = bareAgentQvacGetProfile(
+    String(config.qvac_profile || config.profile || '')
+  )
+  const promptCtxSize =
+    backendForPrompt === 'qvac'
+      ? bareAgentQvacResolveCtxSize(config, promptProfile)
+      : 32768
   // Local QVAC models have a hard ctx window; keep the system prompt lean.
-  const workspaceBudget = backendForPrompt === 'qvac' ? 8000 : 24000
-  const manBudget = backendForPrompt === 'qvac' ? 4000 : 12000
-  const instructionsBudget = backendForPrompt === 'qvac' ? 3000 : 8000
-  const skillsBudget = backendForPrompt === 'qvac' ? 2000 : 4000
+  // Scale char budgets from ctx (chars ≈ tokens*4); leave room for tools/reply.
+  const toolsOn = backendForPrompt === 'qvac' && promptProfile.tools !== false
+  const promptCharBudget = Math.max(
+    4000,
+    Math.floor(promptCtxSize * 4 * (toolsOn ? 0.35 : 0.7))
+  )
+  const workspaceBudget =
+    backendForPrompt === 'qvac'
+      ? Math.min(8000, Math.floor(promptCharBudget * 0.35))
+      : 24000
+  const manBudget =
+    backendForPrompt === 'qvac'
+      ? Math.min(4000, Math.floor(promptCharBudget * 0.2))
+      : 12000
+  const instructionsBudget =
+    backendForPrompt === 'qvac'
+      ? Math.min(3000, Math.floor(promptCharBudget * 0.2))
+      : 8000
+  const skillsBudget =
+    backendForPrompt === 'qvac'
+      ? Math.min(2000, Math.floor(promptCharBudget * 0.15))
+      : 4000
 
   const workspacePromptBlock = await bareAgentLoadWorkspacePrompt(
     ctx,
@@ -1059,6 +1084,27 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) {
       const backendIter = bareAgentResolveBackend(configRef.current)
       /** @type {Record<string, unknown>} */
       const providerNow = String(configRef.current.provider || '').trim().toLowerCase()
+      let qvacToolsForTurn = /** @type {unknown[]} */ ([])
+      if (backendIter === 'qvac') {
+        const profileEarly = bareAgentQvacGetProfile(
+          String(configRef.current.qvac_profile || 'recommended')
+        )
+        const ctxEarly = bareAgentQvacResolveCtxSize(
+          configRef.current,
+          profileEarly
+        )
+        qvacToolsForTurn =
+          profileEarly.tools !== false
+            ? bareAgentFlattenToolsForQvac(tools)
+            : []
+        messages = bareAgentTrimMessagesForCtx(messages, ctxEarly, {
+          tools: qvacToolsForTurn,
+          reserveCompletion: Math.min(
+            1024,
+            Number(configRef.current.max_tokens) || 512
+          )
+        })
+      }
       if (reasoningSettings.enabled && reasoningSettings.mode === 'trace') {
         bareAgentWriteOut(
           ctx,
@@ -1222,7 +1268,7 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) {
           }
           await ctx.bareOsQvacComplete({
             history: messages,
-            tools: bareAgentFlattenToolsForQvac(tools),
+            tools: qvacToolsForTurn,
             stream: Boolean(configRef.current.stream !== false),
             captureThinking: reasoningSettings.enabled,
             modelSrc,

packages/bare-os-coreutils/share/agent-workspace/README-agent.md

@@ -40,7 +40,7 @@ First load downloads via the QVAC registry into the **host cache**, then mirrors
 - `backend` — `qvac` (default) or `rest`
 - `qvac_profile` — `lite` / `recommended` / `strong` / `tool-tiny`
 - `qvac_model` — registry id (e.g. `QWEN3_1_7B_INST_Q4`)
-- `qvac_ctx_size` — optional context window override (profile defaults: lite/tool-tiny 4096, recommended/strong 8192; max 32768)
+- `qvac_ctx_size` — optional context window override. Profile defaults match model cards: Qwen3 profiles **32768**, tool-tiny (Llama 3.2 1B) **131072** (max override 131072). Tool-enabled profiles floor at 8192 (tool schemas alone are ~4.5k tokens). Cap further with host env `BARE_OS_QVAC_MAX_CTX` if VRAM is tight.
 - `qvac_device` — `gpu` (default) or `cpu` (bypass Vulkan; slower)
 - `qvac_main_gpu` — `auto` (default: pick highest-VRAM Vulkan GPU, else CPU), `dedicated`, `integrated`, or device index (`0`, `1`, …)
 - `qvac_gpu_layers` — optional layer offload count (`0` = CPU weights)

packages/bare-os-coreutils/test/agent-qvac.test.mjs

@@ -10,7 +10,7 @@ function loadAgentQvacHelpers() {
   vm.createContext(sandbox)
   vm.runInContext(
     QVAC_SRC +
-      '\n;this.__exports = { bareAgentFlattenToolsForQvac, bareAgentQvacGetProfile, bareAgentQvacProfileList, bareAgentResolveBackend, bareAgentQvacBridgeAvailable, bareAgentQvacResolveCtxSize, bareAgentQvacResolveDeviceOpts }',
+      '\n;this.__exports = { bareAgentFlattenToolsForQvac, bareAgentQvacGetProfile, bareAgentQvacProfileList, bareAgentResolveBackend, bareAgentQvacBridgeAvailable, bareAgentQvacResolveCtxSize, bareAgentQvacResolveDeviceOpts, bareAgentQvacModelCardCtxSize, BARE_AGENT_QVAC_CTX_QWEN3, BARE_AGENT_QVAC_CTX_LLAMA32_1B }',
     sandbox
   )
   return sandbox.__exports
@@ -40,18 +40,33 @@ test('qvac profiles include recommended + lite', async (t) => {
   const list = bareAgentQvacProfileList()
   t.ok(list.length >= 4)
   t.is(bareAgentQvacGetProfile('recommended').chatModel, 'QWEN3_1_7B_INST_Q4')
-  t.is(bareAgentQvacGetProfile('recommended').ctxSize, 8192)
+  t.is(bareAgentQvacGetProfile('recommended').ctxSize, 32768)
   t.is(bareAgentQvacGetProfile('lite').id, 'lite')
-  t.is(bareAgentQvacGetProfile('lite').ctxSize, 4096)
+  t.is(bareAgentQvacGetProfile('lite').ctxSize, 32768)
+  t.is(bareAgentQvacGetProfile('tool-tiny').ctxSize, 131072)
+  t.is(bareAgentQvacGetProfile('strong').ctxSize, 32768)
   t.is(bareAgentQvacGetProfile('nope').id, 'recommended')
 })
 
+test('model-card ctx sizes match upstream cards', async (t) => {
+  const { bareAgentQvacModelCardCtxSize } = loadAgentQvacHelpers()
+  t.is(bareAgentQvacModelCardCtxSize('QWEN3_600M_INST_Q4'), 32768)
+  t.is(bareAgentQvacModelCardCtxSize('QWEN3_1_7B_INST_Q4'), 32768)
+  t.is(bareAgentQvacModelCardCtxSize('QWEN3_4B_INST_Q4_K_M'), 32768)
+  t.is(bareAgentQvacModelCardCtxSize('LLAMA_TOOL_CALLING_1B_INST_Q4_K'), 131072)
+})
+
 test('resolve qvac ctx size from profile or config override', async (t) => {
   const { bareAgentQvacGetProfile, bareAgentQvacResolveCtxSize } = loadAgentQvacHelpers()
   const rec = bareAgentQvacGetProfile('recommended')
-  t.is(bareAgentQvacResolveCtxSize({}, rec), 8192)
-  t.is(bareAgentQvacResolveCtxSize({ qvac_ctx_size: 4096 }, rec), 4096)
-  t.is(bareAgentQvacResolveCtxSize({ qvac_ctx_size: 999999 }, rec), 32768)
+  const lite = bareAgentQvacGetProfile('lite')
+  const tiny = bareAgentQvacGetProfile('tool-tiny')
+  t.is(bareAgentQvacResolveCtxSize({}, rec), 32768)
+  t.is(bareAgentQvacResolveCtxSize({ qvac_ctx_size: 4096 }, rec), 8192) // tools need ≥8k
+  t.is(bareAgentQvacResolveCtxSize({ qvac_ctx_size: 4096 }, lite), 4096) // lite tools=false
+  t.is(bareAgentQvacResolveCtxSize({ qvac_ctx_size: 999999 }, rec), 131072)
+  t.is(bareAgentQvacResolveCtxSize({}, lite), 32768)
+  t.is(bareAgentQvacResolveCtxSize({}, tiny), 131072)
 })
 
 test('resolve qvac device opts', async (t) => {

packages/bare-os-coreutils/test/agent-trim-ctx.test.mjs

@@ -0,0 +1,35 @@
+import test from 'brittle'
+import { readFileSync } from 'node:fs'
+import vm from 'node:vm'
+
+const STATE_SRC = readFileSync(new URL('../lib/agent-state.js', import.meta.url), 'utf8')
+
+function loadTrimHelpers() {
+  const sandbox = { console, TextEncoder, TextDecoder }
+  vm.createContext(sandbox)
+  vm.runInContext(
+    STATE_SRC +
+      '\n;this.__exports = { bareAgentTrimMessages, bareAgentTrimMessagesForCtx, bareAgentEstimateTokens }',
+    sandbox
+  )
+  return sandbox.__exports
+}
+
+test('trim for ctx shrinks system when tools consume budget', async (t) => {
+  const { bareAgentTrimMessagesForCtx, bareAgentEstimateTokens } = loadTrimHelpers()
+  const tools = [{ name: 'x', description: 'y'.repeat(8000), parameters: {} }]
+  const msgs = [
+    { role: 'system', content: 'S'.repeat(20_000) },
+    { role: 'user', content: 'hello' }
+  ]
+  const out = bareAgentTrimMessagesForCtx(msgs, 4096, {
+    tools,
+    reserveCompletion: 512
+  })
+  t.is(out.length, 2)
+  t.ok(String(/** @type {any} */ (out[0]).content).length < 20_000)
+  t.ok(
+    bareAgentEstimateTokens(out) + bareAgentEstimateTokens(tools) + 512 <=
+      4096 + 200
+  )
+})

packages/bare-os-seeder/kernel/bin/agent

@@ -1404,6 +1404,64 @@ function bareAgentTrimMessages(msgs, maxBytes) {
   return out
 }
 
+/**
+ * Rough token estimate (chars/4). Good enough for local ctx budgeting.
+ * @param {unknown} value
+ */
+function bareAgentEstimateTokens(value) {
+  try {
+    const n = JSON.stringify(value == null ? '' : value).length
+    return Math.max(0, Math.ceil(n / 4))
+  } catch {
+    return 0
+  }
+}
+
+/**
+ * Trim history + shrink system content so prompt fits a QVAC ctx window.
+ * Reserves room for completion and optional tools JSON.
+ * @param {unknown[]} msgs
+ * @param {number} ctxSize
+ * @param {{ tools?: unknown[], reserveCompletion?: number }} [opts]
+ */
+function bareAgentTrimMessagesForCtx(msgs, ctxSize, opts) {
+  const ctx = Math.max(2048, Math.floor(Number(ctxSize) || 4096))
+  const reserve =
+    opts && Number.isFinite(Number(opts.reserveCompletion))
+      ? Math.max(256, Math.floor(Number(opts.reserveCompletion)))
+      : Math.min(1024, Math.max(256, Math.floor(ctx * 0.15)))
+  const toolsTok = bareAgentEstimateTokens(
+    opts && Array.isArray(opts.tools) && opts.tools.length ? opts.tools : []
+  )
+  const budget = Math.max(512, ctx - reserve - toolsTok)
+
+  /** @type {unknown[]} */
+  let out = Array.isArray(msgs) ? msgs.slice() : []
+  // Drop middle turns first (keep system + recent).
+  while (bareAgentEstimateTokens(out) > budget && out.length > 3) {
+    out.splice(2, 1)
+  }
+  // Shrink system blob if still over (workspace/man/skills dominate).
+  if (bareAgentEstimateTokens(out) > budget && out[0] && typeof out[0] === 'object') {
+    const sys = /** @type {Record<string, unknown>} */ (out[0])
+    if (sys.role === 'system' && typeof sys.content === 'string') {
+      let content = sys.content
+      let guard = 0
+      while (
+        bareAgentEstimateTokens(out) > budget &&
+        content.length > 800 &&
+        guard < 24
+      ) {
+        content =
+          content.slice(0, Math.floor(content.length * 0.82)) + '\n… truncated'
+        out[0] = { ...sys, content }
+        guard++
+      }
+    }
+  }
+  return out
+}
+
 /**
  * Best-effort load instruction files.
  * @param {Record<string, unknown>} ctx
@@ -1597,6 +1655,16 @@ async function bareAgentResetChatSession(ctx, paths, argv0) {
 
 /** @typedef {'lite'|'recommended'|'strong'|'tool-tiny'} BareAgentQvacProfileId */
 
+/**
+ * Native context windows from upstream model cards (not conservative laptop defaults).
+ * Qwen3 dense (0.6B/1.7B/4B): https://huggingface.co/Qwen/Qwen3-0.6B — 32,768
+ * Llama 3.2 1B (tool-calling finetune base): Meta Llama 3.2 — 128,000
+ * Absolute ceiling for overrides / host clamp.
+ */
+const BARE_AGENT_QVAC_CTX_QWEN3 = 32768
+const BARE_AGENT_QVAC_CTX_LLAMA32_1B = 131072
+const BARE_AGENT_QVAC_CTX_ABSOLUTE_MAX = 131072
+
 /**
  * @typedef {{
  *   id: BareAgentQvacProfileId,
@@ -1617,50 +1685,69 @@ const BARE_AGENT_QVAC_PROFILES = {
     id: 'lite',
     label: 'Lite',
     description:
-      'Smallest download. Good for weak machines / ~4 GB VRAM; limited tool use.',
+      'Smallest download (Qwen3-0.6B). Model-card context 32k; limited tool use.',
     chatModel: 'QWEN3_600M_INST_Q4',
     tools: false,
     minRamGb: 4,
     minDiskGb: 2,
     approxDownloadGb: 0.5,
-    ctxSize: 4096
+    ctxSize: BARE_AGENT_QVAC_CTX_QWEN3
   },
   recommended: {
     id: 'recommended',
     label: 'Recommended',
     description:
-      'Best balance for Bare OS agent tool calling (fits ~4–8 GB VRAM at ctx 8k).',
+      'Best balance for Bare OS agent tool calling (Qwen3-1.7B, model-card context 32k).',
     chatModel: 'QWEN3_1_7B_INST_Q4',
     tools: true,
     minRamGb: 8,
     minDiskGb: 5,
     approxDownloadGb: 2.5,
-    ctxSize: 8192
+    ctxSize: BARE_AGENT_QVAC_CTX_QWEN3
   },
   strong: {
     id: 'strong',
     label: 'Strong',
-    description: 'Better reasoning. Needs more RAM/VRAM/disk.',
+    description:
+      'Better reasoning (Qwen3-4B). Model-card context 32k (131k with YaRN not enabled).',
     chatModel: 'QWEN3_4B_INST_Q4_K_M',
     tools: true,
     minRamGb: 16,
     minDiskGb: 8,
     approxDownloadGb: 3.5,
-    ctxSize: 8192
+    ctxSize: BARE_AGENT_QVAC_CTX_QWEN3
   },
   'tool-tiny': {
     id: 'tool-tiny',
     label: 'Tool-tiny',
-    description: 'Llama tool-calling 1B fallback if Qwen tools misbehave.',
+    description:
+      'Llama 3.2 1B tool-calling fallback (model-card context 128k).',
     chatModel: 'LLAMA_TOOL_CALLING_1B_INST_Q4_K',
     tools: true,
     minRamGb: 6,
     minDiskGb: 3,
     approxDownloadGb: 1,
-    ctxSize: 4096
+    ctxSize: BARE_AGENT_QVAC_CTX_LLAMA32_1B
   }
 }
 
+/**
+ * Model-card / train context for a QVAC registry id (fallback 32k).
+ * @param {string} [modelId]
+ * @returns {number}
+ */
+function bareAgentQvacModelCardCtxSize(modelId) {
+  const id = String(modelId || '')
+    .trim()
+    .toUpperCase()
+  if (!id) return BARE_AGENT_QVAC_CTX_QWEN3
+  if (id.includes('LLAMA') || id.includes('LLAMA_TOOL')) {
+    return BARE_AGENT_QVAC_CTX_LLAMA32_1B
+  }
+  if (id.includes('QWEN3')) return BARE_AGENT_QVAC_CTX_QWEN3
+  return BARE_AGENT_QVAC_CTX_QWEN3
+}
+
 /** @returns {BareAgentQvacProfile[]} */
 function bareAgentQvacProfileList() {
   return Object.values(BARE_AGENT_QVAC_PROFILES)
@@ -1749,17 +1836,33 @@ function bareAgentResolveBackend(config) {
 }
 
 /**
- * Resolve context window for QVAC load (profile default or config override).
+ * Resolve context window for QVAC load (profile/model-card default or config override).
+ * Tool schemas alone are ~4.5k tokens, so tool-enabled profiles need ≥8192.
  * @param {Record<string, unknown>} config
  * @param {BareAgentQvacProfile} profile
  */
 function bareAgentQvacResolveCtxSize(config, profile) {
+  const modelId = String(
+    (config && (config.qvac_model || config.model)) ||
+      (profile && profile.chatModel) ||
+      ''
+  )
+  const cardCtx = bareAgentQvacModelCardCtxSize(modelId)
+  const profileCtx = Math.max(
+    2048,
+    Number(profile && profile.ctxSize) || cardCtx
+  )
   const raw = Number(config && config.qvac_ctx_size)
+  let ctx
   if (Number.isFinite(raw) && raw >= 2048) {
-    // Cap high overrides; 32k+ KV often OOMs laptop GPUs (~4 GB VRAM).
-    return Math.min(32768, Math.floor(raw))
+    ctx = Math.min(BARE_AGENT_QVAC_CTX_ABSOLUTE_MAX, Math.floor(raw))
+  } else {
+    // Prefer explicit profile ctx (already model-card sized); else card for model id.
+    ctx = Math.min(BARE_AGENT_QVAC_CTX_ABSOLUTE_MAX, Math.max(profileCtx, cardCtx))
   }
-  return Math.max(2048, Number(profile.ctxSize) || 8192)
+  // Full Bare OS tool list ≈ 4.5k tokens; leave room for system + reply.
+  if (profile && profile.tools !== false && ctx < 8192) ctx = 8192
+  return ctx
 }
 
 /**
@@ -6846,11 +6949,36 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) {
   if (needWizard) await bareAgentSyncWorkspaceFromConfig(ctx, paths, config)
 
   const backendForPrompt = bareAgentResolveBackend(config)
+  const promptProfile = bareAgentQvacGetProfile(
+    String(config.qvac_profile || config.profile || '')
+  )
+  const promptCtxSize =
+    backendForPrompt === 'qvac'
+      ? bareAgentQvacResolveCtxSize(config, promptProfile)
+      : 32768
   // Local QVAC models have a hard ctx window; keep the system prompt lean.
-  const workspaceBudget = backendForPrompt === 'qvac' ? 8000 : 24000
-  const manBudget = backendForPrompt === 'qvac' ? 4000 : 12000
-  const instructionsBudget = backendForPrompt === 'qvac' ? 3000 : 8000
-  const skillsBudget = backendForPrompt === 'qvac' ? 2000 : 4000
+  // Scale char budgets from ctx (chars ≈ tokens*4); leave room for tools/reply.
+  const toolsOn = backendForPrompt === 'qvac' && promptProfile.tools !== false
+  const promptCharBudget = Math.max(
+    4000,
+    Math.floor(promptCtxSize * 4 * (toolsOn ? 0.35 : 0.7))
+  )
+  const workspaceBudget =
+    backendForPrompt === 'qvac'
+      ? Math.min(8000, Math.floor(promptCharBudget * 0.35))
+      : 24000
+  const manBudget =
+    backendForPrompt === 'qvac'
+      ? Math.min(4000, Math.floor(promptCharBudget * 0.2))
+      : 12000
+  const instructionsBudget =
+    backendForPrompt === 'qvac'
+      ? Math.min(3000, Math.floor(promptCharBudget * 0.2))
+      : 8000
+  const skillsBudget =
+    backendForPrompt === 'qvac'
+      ? Math.min(2000, Math.floor(promptCharBudget * 0.15))
+      : 4000
 
   const workspacePromptBlock = await bareAgentLoadWorkspacePrompt(
     ctx,
@@ -7005,6 +7133,27 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) {
       const backendIter = bareAgentResolveBackend(configRef.current)
       /** @type {Record<string, unknown>} */
       const providerNow = String(configRef.current.provider || '').trim().toLowerCase()
+      let qvacToolsForTurn = /** @type {unknown[]} */ ([])
+      if (backendIter === 'qvac') {
+        const profileEarly = bareAgentQvacGetProfile(
+          String(configRef.current.qvac_profile || 'recommended')
+        )
+        const ctxEarly = bareAgentQvacResolveCtxSize(
+          configRef.current,
+          profileEarly
+        )
+        qvacToolsForTurn =
+          profileEarly.tools !== false
+            ? bareAgentFlattenToolsForQvac(tools)
+            : []
+        messages = bareAgentTrimMessagesForCtx(messages, ctxEarly, {
+          tools: qvacToolsForTurn,
+          reserveCompletion: Math.min(
+            1024,
+            Number(configRef.current.max_tokens) || 512
+          )
+        })
+      }
       if (reasoningSettings.enabled && reasoningSettings.mode === 'trace') {
         bareAgentWriteOut(
           ctx,
@@ -7095,6 +7244,29 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) {
               EDIT_ANSI_RESET +
               '\n'
           )
+          try {
+            const st =
+              typeof ctx.bareOsQvacStatus === 'function'
+                ? ctx.bareOsQvacStatus()
+                : null
+            if (st && (st.cacheDir || st.hdmsPath || st.backendsDir)) {
+              bareAgentWriteOut(
+                ctx,
+                stdout,
+                bareEditSgr('dim', useColor) +
+                  '[qvac] cache=' +
+                  String(st.cacheDir || '') +
+                  ' hdms=' +
+                  String(st.hdmsPath || '/mnt/models') +
+                  ' backends=' +
+                  String(st.backendsDir || '(unresolved)') +
+                  EDIT_ANSI_RESET +
+                  '\n'
+              )
+            }
+          } catch {
+            /* ignore */
+          }
           if (typeof ctx.bareOsQvacLoadModel === 'function') {
             const loaded = await ctx.bareOsQvacLoadModel({
               modelSrc,
@@ -7145,7 +7317,7 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) {
           }
           await ctx.bareOsQvacComplete({
             history: messages,
-            tools: bareAgentFlattenToolsForQvac(tools),
+            tools: qvacToolsForTurn,
             stream: Boolean(configRef.current.stream !== false),
             captureThinking: reasoningSettings.enabled,
             modelSrc,

packages/bare-os-seeder/kernel/etc/bare-os/posix_utilities.json

@@ -1,7 +1,7 @@
 {
   "schema": 2,
   "profileId": "bare-os-posix-like",
-  "generatedAt": "2026-07-31T21:35:54.625Z",
+  "generatedAt": "2026-07-31T23:16:21.588Z",
   "note": "Sparse POSIX Issue 7 coverage hints; commandIndex is generated each coreutils build.",
   "commandIndex": [
     {

packages/bare-os-seeder/kernel/lib/bare/shell-completion.json

@@ -1,6 +1,6 @@
 {
   "schema": 1,
-  "atMs": 1785533754625,
+  "atMs": 1785539781587,
   "commands": [
     "agent",
     "appctl",

packages/bare-os-seeder/kernel/share/agent-workspace/README-agent.md

@@ -30,10 +30,17 @@ After **`agent --config`** / **`--setup`** (or changing **`owner_name`** / **`ag
 
 Default backend is **QVAC** (QuantumVerse Automatic Computer) — local in-process LLM via the booter **`ctx.bareOsQvac*`** bridge (`@qvac/sdk`). Weights download on first use after you pick a profile in **`agent --config`**.
 
+Models are stored for durability on HDMS label **`models`** (guest path **`/mnt/models/`**) and materialize into a host cache for llama.cpp load:
+
+- Host cache: **`$BARE_OS_HOST_DATA/qvac/models`** (default **`~/.bare-os/qvac/models`**)
+- Runtime config: **`$BARE_OS_HOST_DATA/qvac/qvac.config.json`** (`QVAC_CONFIG_PATH`, absolute `cacheDirectory`)
+
+First load downloads via the QVAC registry into the **host cache**, then mirrors into HDMS. If HDMS already has the GGUF, Bare OS materializes it to the host cache before load. If auto-create fails, run **`hdms create models`** after unlock.
+
 - `backend` — `qvac` (default) or `rest`
 - `qvac_profile` — `lite` / `recommended` / `strong` / `tool-tiny`
 - `qvac_model` — registry id (e.g. `QWEN3_1_7B_INST_Q4`)
-- `qvac_ctx_size` — optional context window override (profile defaults: lite/tool-tiny 4096, recommended/strong 8192; max 32768)
+- `qvac_ctx_size` — optional context window override. Profile defaults match model cards: Qwen3 profiles **32768**, tool-tiny (Llama 3.2 1B) **131072** (max override 131072). Tool-enabled profiles floor at 8192 (tool schemas alone are ~4.5k tokens). Cap further with host env `BARE_OS_QVAC_MAX_CTX` if VRAM is tight.
 - `qvac_device` — `gpu` (default) or `cpu` (bypass Vulkan; slower)
 - `qvac_main_gpu` — `auto` (default: pick highest-VRAM Vulkan GPU, else CPU), `dedicated`, `integrated`, or device index (`0`, `1`, …)
 - `qvac_gpu_layers` — optional layer offload count (`0` = CPU weights)