@clear-capabilities/agentic-security-scanner 0.149.4 → 0.150.2

package/src/llm-validator/model-capabilities.js

@@ -0,0 +1,269 @@
+// Model family hints, RAM-aware memory profiles, and the memory-admission
+// check for the Ollama provider (agentic-security-ollama-offline-prd.md
+// §13, §14, §15, §22.3, §30).
+//
+// FAMILY HINTS ARE DEFAULTS, NEVER AUTHORITY (PRD §12/§13). A name like
+// `gemma4:e2b` tells us nothing Ollama itself won't confirm — it only lets the
+// harness suggest a sane default before any network call. If a model actually
+// installed under a family-hinted name lacks a capability the hint implied,
+// the runtime probe (model-probe.js, added when tool-calling/structured-output
+// probing lands) always wins. This module only classifies and estimates; it
+// never asserts a capability is present.
+//
+// MEMORY NUMBERS ARE ESTIMATES, NOT PROMISES (PRD §22.3, §14.1). Ollama
+// artifact sizes and this module's headroom reserves are best-effort figures
+// sourced from what Ollama currently publishes; they exist so the harness can
+// fail BEFORE an OS-level OOM, not so it can claim an exact answer. Every
+// admission decision leaves a stated safety margin rather than trying to pack
+// memory to the byte.
+
+import * as os from 'node:os';
+import { priorOOMFor } from './oom-feedback.js';
+
+// PRD §12/§13 FR-1203 — non-authoritative family hint from a model name.
+// Longest/most-specific pattern first so `qwen3.5:4b` doesn't fall through to
+// the bare `qwen` bucket.
+const FAMILY_PATTERNS = [
+  [/^qwen3\.5/i, 'qwen3.5'],
+  [/^qwen3-coder-next/i, 'qwen3-coder-next'],
+  [/^qwen3-coder/i, 'qwen3-coder'],
+  [/^qwen2\.5-coder/i, 'qwen2.5-coder'],
+  [/^qwen3/i, 'qwen3'],
+  [/^qwen/i, 'qwen'],
+  [/^gemma4/i, 'gemma4'],
+  [/^functiongemma/i, 'functiongemma'],
+  [/^gemma3/i, 'gemma3'],
+  [/^gemma/i, 'gemma'],
+];
+
+/** Non-authoritative family classification for defaults/messaging only. */
+export function classifyModelFamily(modelName) {
+  const name = String(modelName || '').trim();
+  for (const [re, family] of FAMILY_PATTERNS) if (re.test(name)) return family;
+  return 'unknown';
+}
+
+// PRD §13.1 — non-authoritative defaults per family, overridden by any real
+// runtime probe result (model-probe.js). `tools`/`structuredJson`/`thinking`
+// are 'unknown' where Ollama's own behavior varies by specific tag/quant
+// rather than by family alone.
+const FAMILY_CAPABILITY_HINTS = {
+  'qwen3.5': { chat: true, structuredJson: true, tools: true, thinking: 'unknown' },
+  qwen3: { chat: true, structuredJson: true, tools: true, thinking: 'unknown' },
+  'qwen3-coder': { chat: true, structuredJson: true, tools: true, thinking: false },
+  'qwen3-coder-next': { chat: true, structuredJson: true, tools: true, thinking: false },
+  'qwen2.5-coder': { chat: true, structuredJson: true, tools: 'unknown', thinking: false },
+  qwen: { chat: true, structuredJson: 'unknown', tools: 'unknown', thinking: 'unknown' },
+  gemma4: { chat: true, structuredJson: true, tools: true, thinking: 'unknown' },
+  functiongemma: { chat: true, structuredJson: 'unknown', tools: true, thinking: false },
+  gemma3: { chat: true, structuredJson: true, tools: false, thinking: false },
+  gemma: { chat: true, structuredJson: 'unknown', tools: 'unknown', thinking: 'unknown' },
+  unknown: { chat: true, structuredJson: 'unknown', tools: 'unknown', thinking: 'unknown' },
+};
+
+/**
+ * Build the PRD §13.1 ModelCapabilities object from a family hint alone
+ * (Layer B). Layer A (Ollama's own /api/show metadata) and Layer C (runtime
+ * probes) are applied by the caller and override these fields — this
+ * function only ever sets `source.familyHint: true`.
+ */
+export function capabilitiesFromFamilyHint(modelName) {
+  const family = classifyModelFamily(modelName);
+  const hint = FAMILY_CAPABILITY_HINTS[family] || FAMILY_CAPABILITY_HINTS.unknown;
+  return {
+    chat: hint.chat,
+    structuredJson: hint.structuredJson,
+    tools: hint.tools,
+    thinking: hint.thinking,
+    vision: false,
+    contextTokens: undefined,
+    source: { metadata: false, familyHint: true, runtimeProbe: false },
+  };
+}
+
+// ── RAM-aware memory profiles (PRD §14.4, §15.2, §22.3, §30) ───────────────
+
+const MB = 1024 * 1024;
+const GB = 1024 * MB;
+
+// Best-effort artifact sizes as currently distributed by Ollama, used only to
+// pick a SENSIBLE STARTING recommendation — the real admission decision below
+// uses actually-free memory, not this table. Keep in sync with the PRD's own
+// cited figures; a stale entry only affects the suggested default, never the
+// admission math (which reads real os.freemem()).
+export const KNOWN_MODEL_SIZE_GB = Object.freeze({
+  'qwen3.5:2b': 1.7,
+  'qwen3.5:4b': 3.4,
+  'qwen3.5:9b': 6.6,
+  'gemma4:e2b': 7.2,
+  'gemma4:12b': 7.6,
+  'gemma4:latest': 9.6,
+});
+
+/** PRD §30 profile presets. `auto` picks between these by detected RAM. */
+export const MEMORY_PROFILES = Object.freeze({
+  '8gb': {
+    label: '8gb',
+    preferredModel: 'qwen3.5:4b',
+    fallbackModel: 'qwen3.5:2b',
+    initialContextTokens: 4096,
+    targetContextTokens: 8192,
+    maxConcurrency: 1,
+    minFreeRamMb: 1536,
+  },
+  '16gb-qwen': {
+    label: '16gb-qwen',
+    preferredModel: 'qwen3.5:9b',
+    fallbackModel: 'qwen3.5:4b',
+    initialContextTokens: 16384,
+    targetContextTokens: 32768,
+    maxConcurrency: 1,
+    minFreeRamMb: 2048,
+  },
+  '16gb-gemma': {
+    label: '16gb-gemma',
+    preferredModel: 'gemma4:e2b',
+    fallbackModel: 'qwen3.5:4b',
+    initialContextTokens: 8192,
+    targetContextTokens: 16384,
+    maxConcurrency: 1,
+    minFreeRamMb: 2048,
+  },
+});
+
+/**
+ * PRD §22.3 — detect total/available system RAM. Thin wrapper over `os` so
+ * tests can inject fake values without mocking the `os` module globally.
+ */
+export function detectSystemMemory({ totalBytes, freeBytes } = {}) {
+  return {
+    totalBytes: Number.isFinite(totalBytes) ? totalBytes : os.totalmem(),
+    freeBytes: Number.isFinite(freeBytes) ? freeBytes : os.freemem(),
+  };
+}
+
+/**
+ * Pick the RAM tier ('8gb' | '16gb') a machine falls into. Anything under
+ * ~9 GB total is treated as the 8 GB tier — real "8 GB" machines report
+ * slightly less than 8*1024^3 bytes to userspace (firmware/GPU reservations),
+ * so a hard `< 8*GB` cutoff would misclassify real 8 GB hardware as unknown.
+ */
+export function detectMemoryTier(totalBytes) {
+  if (!Number.isFinite(totalBytes) || totalBytes <= 0) return 'unknown';
+  if (totalBytes < 9 * GB) return '8gb';
+  return '16gb';
+}
+
+/**
+ * PRD §22.3 admission algorithm: does `contextTokens` at `model` fit in
+ * currently-free memory with the configured reserve intact?
+ *
+ * This is deliberately conservative and coarse (PRD "avoid pretending memory
+ * estimates are exact"): model residency is estimated from KNOWN_MODEL_SIZE_GB
+ * when available (falling back to a pessimistic 8 GB assumption for an
+ * unrecognized tag so an unknown model never LOOKS safer than a known large
+ * one), and KV-cache growth is approximated as a fixed per-1K-token cost
+ * rather than modeled per-architecture — real KV cache size depends on layer
+ * count/head count/quantization the harness cannot know without Ollama's own
+ * runtime numbers.
+ */
+const ESTIMATED_KV_CACHE_MB_PER_1K_TOKENS = 32; // conservative, model-independent approximation
+const RUNTIME_OVERHEAD_MB = 512; // Ollama server + OS scheduler slack, independent of model size
+
+export function evaluateMemoryAdmission({
+  modelName,
+  contextTokens,
+  freeBytes,
+  minFreeRamMb,
+  modelSizeGb,
+} = {}) {
+  const sizeGb = Number.isFinite(modelSizeGb) ? modelSizeGb : (KNOWN_MODEL_SIZE_GB[modelName] ?? 8);
+  const modelMb = sizeGb * 1024;
+  const kvCacheMb = (Number(contextTokens) || 0) / 1000 * ESTIMATED_KV_CACHE_MB_PER_1K_TOKENS;
+  const requiredMb = modelMb + kvCacheMb + RUNTIME_OVERHEAD_MB + (Number(minFreeRamMb) || 0);
+  const freeMb = (Number(freeBytes) || 0) / MB;
+  const admitted = freeMb >= requiredMb;
+  return {
+    admitted,
+    freeMb: Math.round(freeMb),
+    requiredMb: Math.round(requiredMb),
+    modelEstimateMb: Math.round(modelMb),
+    kvCacheEstimateMb: Math.round(kvCacheMb),
+    reserveMb: Number(minFreeRamMb) || 0,
+  };
+}
+
+/**
+ * Full recommendation flow (PRD §22 "Memory admission algorithm"):
+ * try the profile's preferred context, shrink it, then fall back to the
+ * profile's smaller model, before ever declaring the profile unusable.
+ * Never recommends cloud — the worst outcome this function can return is
+ * `{admitted:false}` with a human-readable explanation, which callers treat
+ * as "run deterministic-only" (PRD §23.4).
+ */
+export function recommendAdmission(opts = {}) {
+  const result = _recommendAdmissionCore(opts);
+  // Adversarial-review fix (2026-09): a memory-admission ESTIMATE that
+  // actually caused a real OOM (ollama-provider.js's callOllamaChat now
+  // records this via oom-feedback.js) used to have no way to affect a
+  // future admission decision for the SAME model on the SAME machine — the
+  // exact same "admitted: true" would repeat forever. This does not
+  // recalibrate the underlying estimate (that needs real hardware variety
+  // one machine's observed failures can't substitute for); it attaches an
+  // honest warning so the decision is no longer presented with unqualified
+  // confidence.
+  const chosenModel = result.model || opts.requestedModel;
+  const prior = chosenModel ? priorOOMFor(chosenModel) : null;
+  if (prior) {
+    return {
+      ...result,
+      priorOOMWarning: `'${chosenModel}' has previously failed with an out-of-memory error on this machine ` +
+        `(${prior.count} time${prior.count === 1 ? '' : 's'}, most recently ${new Date(prior.lastAt).toISOString()}). ` +
+        'The memory estimate below may be optimistic for your hardware.',
+    };
+  }
+  return result;
+}
+
+function _recommendAdmissionCore({ profile, freeBytes, requestedContextTokens, requestedModel } = {}) {
+  const p = MEMORY_PROFILES[profile];
+  if (!p) return { admitted: false, reason: `unknown memory profile '${profile}'` };
+
+  const model = requestedModel || p.preferredModel;
+  const attempts = [];
+
+  // 1. Requested (or target) context at the requested/preferred model.
+  const primaryContext = Number.isFinite(requestedContextTokens) ? requestedContextTokens : p.targetContextTokens;
+  let check = evaluateMemoryAdmission({ modelName: model, contextTokens: primaryContext, freeBytes, minFreeRamMb: p.minFreeRamMb });
+  attempts.push({ model, contextTokens: primaryContext, ...check });
+  if (check.admitted) return { admitted: true, model, contextTokens: primaryContext, attempts };
+
+  // 2. Reduce context to the profile's conservative initial value first —
+  //    PRD FR-2104: "shrink context before declaring an otherwise compatible
+  //    model unusable."
+  if (primaryContext !== p.initialContextTokens) {
+    check = evaluateMemoryAdmission({ modelName: model, contextTokens: p.initialContextTokens, freeBytes, minFreeRamMb: p.minFreeRamMb });
+    attempts.push({ model, contextTokens: p.initialContextTokens, ...check });
+    if (check.admitted) return { admitted: true, model, contextTokens: p.initialContextTokens, attempts, reducedContext: true };
+  }
+
+  // 3. Fall back to the profile's smaller model at its initial context.
+  if (p.fallbackModel && p.fallbackModel !== model) {
+    check = evaluateMemoryAdmission({ modelName: p.fallbackModel, contextTokens: p.initialContextTokens, freeBytes, minFreeRamMb: p.minFreeRamMb });
+    attempts.push({ model: p.fallbackModel, contextTokens: p.initialContextTokens, ...check });
+    if (check.admitted) {
+      return {
+        admitted: true, model: p.fallbackModel, contextTokens: p.initialContextTokens, attempts,
+        reducedContext: true, fellBackToSmallerModel: true,
+      };
+    }
+  }
+
+  // 4. Nothing fits — deterministic-only, never cloud.
+  return {
+    admitted: false,
+    attempts,
+    reason: `No local model/context combination fit in available memory with the configured reserve. ` +
+      `Recommend deterministic-only scanning, or free memory before retrying.`,
+  };
+}

package/src/llm-validator/model-probe.js

@@ -0,0 +1,232 @@
+// PRD §13.2 — the three-layer model capability detection strategy.
+//
+// LAYER A (metadata) is the cheapest and most authoritative: Ollama's own
+// `/api/show` response, when it reports a `capabilities` array, is not a
+// guess. LAYER B (model-capabilities.js's family hint) is a non-authoritative
+// default used only where Layer A is silent. LAYER C (this module's
+// `probeStructuredOutput`/`probeToolCalling`) is the most expensive — it
+// consumes real inference time — so it is OPT-IN (the caller decides when
+// "necessary" per the PRD's own wording), never run implicitly on every
+// `models doctor`/`models inspect` invocation.
+//
+// PRECEDENCE: Layer C overrides Layer A overrides Layer B, field by field. A
+// field only ever gets overridden by a MORE authoritative layer that actually
+// has an opinion — a probe that couldn't run (offline/timeout) leaves the
+// field exactly as the layer below it set it, it never downgrades to
+// 'unknown'.
+//
+// CACHE KEY = Ollama version + model digest + model name (PRD §13.2 exactly).
+// Digest is load-bearing: `ollama pull` replacing a tag's underlying weights
+// must invalidate the cache even though the name/tag string is unchanged.
+// Same disk-cache directory convention as sca/sigstore-verify.js and
+// engine.js's OSV cache (`~/.claude/agentic-security/<name>/`).
+//
+// TTL + force-reprobe (adversarial-review fix, 2026-09). The key-based
+// invalidation above is real but not complete: this module's own comment
+// used to claim the entry is safe "forever" because the key changes when
+// the model does — but `/api/show` doesn't expose a digest on every Ollama
+// version (falls back to model NAME alone then, a few lines below), so a
+// same-tag re-pull, or simply an unlucky single-trial probe the first time
+// (see probeStructuredOutput/probeToolCalling's own single-call design),
+// had no way to ever self-correct short of a user manually deleting a file
+// under `~/.claude/agentic-security/`. Two independent fixes, since either
+// alone leaves a real gap: a default TTL as a safety net for the case
+// nobody notices, and an explicit `force` option (`models test --force`)
+// for the case someone DOES suspect a stale answer and wants it right now.
+
+import * as fs from 'node:fs';
+import * as path from 'node:path';
+import * as os from 'node:os';
+import * as crypto from 'node:crypto';
+import { callOllamaStructured, callOllamaChat, showOllamaModel, getOllamaVersion } from './ollama-provider.js';
+import { capabilitiesFromFamilyHint } from './model-capabilities.js';
+
+const CACHE_DIR = path.join(os.homedir(), '.claude', 'agentic-security', 'ollama-capability-cache');
+
+// Default safety-net TTL: 30 days. Not the primary invalidation mechanism
+// (the key is) — a backstop for the cases the key can't see: a same-tag
+// re-pull on an Ollama version that doesn't expose a digest, or a single
+// unlucky probe trial that happened to pass/fail against the model's true
+// behavior. Overridable for anyone who wants a tighter or looser bound.
+export const DEFAULT_CACHE_TTL_MS = 30 * 24 * 60 * 60 * 1000;
+
+function _ensureCacheDir() { try { fs.mkdirSync(CACHE_DIR, { recursive: true }); } catch {} }
+function _cacheKey(ollamaVersion, modelDigest, modelName) {
+  return crypto.createHash('sha256').update(`${ollamaVersion}::${modelDigest}::${modelName}`).digest('hex');
+}
+function _cachePath(key) { return path.join(CACHE_DIR, key + '.json'); }
+
+/**
+ * @returns {object|null} the cached probe RESULT (not the envelope), or
+ *   `null` on a miss, a parse failure, OR an entry older than `ttlMs`.
+ */
+function _readProbeCache(key, ttlMs) {
+  let envelope;
+  try { envelope = JSON.parse(fs.readFileSync(_cachePath(key), 'utf8')); } catch { return null; }
+  // Backward-compatible with a pre-TTL cache file that was just the bare
+  // result object (no `probedAt`) — treat an entry with no timestamp as
+  // fresh rather than discarding every cache written before this fix.
+  if (envelope && typeof envelope === 'object' && 'probedAt' in envelope && 'result' in envelope) {
+    if (Number.isFinite(ttlMs) && ttlMs > 0 && Date.now() - envelope.probedAt > ttlMs) return null;
+    return envelope.result;
+  }
+  return envelope;
+}
+function _writeProbeCache(key, value) {
+  _ensureCacheDir();
+  try { fs.writeFileSync(_cachePath(key), JSON.stringify({ probedAt: Date.now(), result: value })); } catch {}
+}
+
+/**
+ * PRD §13.2 Layer A — parse `/api/show`'s response into the subset of
+ * ModelCapabilities it can actually speak to. A field this layer has no
+ * opinion on is omitted (not set to `false`) so the caller's merge never
+ * mistakes silence for a negative.
+ */
+export function capabilitiesFromShowMetadata(show) {
+  const out = { source: { metadata: true } };
+  if (Array.isArray(show?.capabilities) && show.capabilities.length > 0) {
+    const caps = show.capabilities;
+    out.chat = caps.includes('completion') || caps.includes('chat');
+    out.tools = caps.includes('tools');
+    out.vision = caps.includes('vision');
+    out.thinking = caps.includes('thinking');
+  }
+  const modelInfo = show?.modelInfo;
+  if (modelInfo && typeof modelInfo === 'object') {
+    const ctxKey = Object.keys(modelInfo).find((k) => k.endsWith('.context_length'));
+    if (ctxKey && Number.isFinite(modelInfo[ctxKey])) out.contextTokens = modelInfo[ctxKey];
+  }
+  return out;
+}
+
+/**
+ * PRD §13.2 Layer C — structured-output probe. A tiny schema, a request for
+ * `{"ok": true}`, verified end to end through the SAME
+ * callOllamaStructured() bounded-retry path every real structured call uses
+ * (not a bespoke lighter-weight check that could disagree with production
+ * behavior).
+ */
+const PROBE_SCHEMA = { type: 'object', required: ['ok'], properties: { ok: { type: 'boolean' } } };
+
+export async function probeStructuredOutput({ host, model, timeouts, keepAlive } = {}) {
+  const r = await callOllamaStructured({
+    host, model,
+    messages: [{ role: 'user', content: 'Reply with ONLY a JSON object: {"ok": true}' }],
+    schema: PROBE_SCHEMA,
+    validateFn: (obj) => (obj && obj.ok === true ? { ok: true, value: obj } : { ok: false }),
+    keepAlive, timeouts,
+  });
+  if (r.ok) return { supported: true };
+  // A transport-level failure (server unreachable, timed out) tells us
+  // nothing about the MODEL's capability — leave it 'unknown' rather than
+  // reporting a false negative for an offline/slow server.
+  if (['ollama-unreachable', 'ollama-not-running', 'ollama-timeout', 'ollama-model-not-installed'].includes(r.code)) {
+    return { supported: 'unknown', reason: r.reason || r.code };
+  }
+  return { supported: false, reason: r.reason || r.code };
+}
+
+/**
+ * PRD §13.2 Layer C — tool-calling probe. One harmless `echo_capability_probe`
+ * function; success is Ollama returning a structured `tool_calls` entry
+ * naming it, not a check on what the model chose to reply with in prose.
+ */
+const PROBE_TOOL = {
+  type: 'function',
+  function: {
+    name: 'echo_capability_probe',
+    description: 'Echo back the given value. Used only to test whether this model supports tool calling.',
+    parameters: { type: 'object', required: ['value'], properties: { value: { type: 'string' } } },
+  },
+};
+
+export async function probeToolCalling({ host, model, timeouts, keepAlive } = {}) {
+  const r = await callOllamaChat({
+    host, model,
+    messages: [{ role: 'user', content: 'Call the echo_capability_probe function with value set to "probe-ok". Reply with nothing else.' }],
+    tools: [PROBE_TOOL],
+    keepAlive, timeouts,
+  });
+  if (!r.ok) {
+    if (['ollama-unreachable', 'ollama-not-running', 'ollama-timeout', 'ollama-model-not-installed'].includes(r.code)) {
+      return { supported: 'unknown', reason: r.reason || r.code };
+    }
+    return { supported: false, reason: r.reason || r.code };
+  }
+  const calls = r.result.toolCalls || [];
+  const called = calls.some((c) => c?.function?.name === 'echo_capability_probe');
+  return called ? { supported: true } : { supported: false, reason: 'model did not emit a tool_calls entry for the probe function' };
+}
+
+function _mergeLayer(base, overlay, sourceFlag) {
+  const merged = { ...base };
+  let touched = false;
+  for (const field of ['chat', 'structuredJson', 'tools', 'thinking', 'vision', 'contextTokens']) {
+    if (overlay[field] !== undefined) { merged[field] = overlay[field]; touched = true; }
+  }
+  if (touched) merged.source = { ...merged.source, [sourceFlag]: true };
+  return merged;
+}
+
+/**
+ * Orchestrates all three layers (PRD §13.2) with caching (PRD: "so startup
+ * does not repeatedly consume inference time"). `probe: true` opts into
+ * Layer C — omitted or false, this returns Layer A+B only, which is what
+ * every non-probing caller (models list/inspect/doctor's default path)
+ * should use, since Layer C spends real inference time on the user's
+ * machine.
+ *
+ * `force: true` (adversarial-review fix, 2026-09 — `models test --force`)
+ * skips reading the cache — always runs a fresh probe and overwrites
+ * whatever was there. `ttlMs` (default 30 days, `DEFAULT_CACHE_TTL_MS`)
+ * bounds how long a cached entry is trusted without either; pass `0`/
+ * `Infinity` to disable the TTL safety net entirely and rely on the key
+ * alone, matching this module's original design intent.
+ *
+ * @returns {{ok:true, capabilities:object, cached:boolean} | {ok:false, code, reason}}
+ */
+export async function getModelCapabilities({ host, model, env = process.env, probe = false, force = false, ttlMs = DEFAULT_CACHE_TTL_MS, timeouts, keepAlive } = {}) {
+  let capabilities = capabilitiesFromFamilyHint(model);
+
+  const show = await showOllamaModel({ host, model, timeouts });
+  if (show.ok) {
+    capabilities = _mergeLayer(capabilities, capabilitiesFromShowMetadata(show), 'metadata');
+  }
+
+  if (!probe) {
+    return { ok: true, capabilities, cached: false };
+  }
+
+  const versionResult = await getOllamaVersion({ host, timeouts });
+  const ollamaVersion = versionResult.ok ? versionResult.version : 'unknown-version';
+  // The digest is whatever Layer A's /api/show reported under `details`
+  // (Ollama does not expose it on /api/show consistently across versions —
+  // fall back to the model name alone, which still invalidates on a tag
+  // change, just not on a same-tag re-pull).
+  const modelDigest = show.ok && show.details?.digest ? show.details.digest : 'unknown-digest';
+  const cacheKey = _cacheKey(ollamaVersion, modelDigest, model);
+
+  const cached = force ? null : _readProbeCache(cacheKey, ttlMs);
+  if (cached) {
+    return { ok: true, capabilities: _mergeLayer(capabilities, cached, 'runtimeProbe'), cached: true };
+  }
+
+  const [structured, tools] = await Promise.all([
+    probeStructuredOutput({ host, model, timeouts, keepAlive }),
+    probeToolCalling({ host, model, timeouts, keepAlive }),
+  ]);
+
+  const probeResult = {};
+  if (structured.supported !== 'unknown') probeResult.structuredJson = structured.supported;
+  if (tools.supported !== 'unknown') probeResult.tools = tools.supported;
+
+  // Only cache a probe that actually resolved something — an all-'unknown'
+  // result (server unreachable mid-probe) would otherwise poison the cache
+  // with a permanent non-answer.
+  if (Object.keys(probeResult).length > 0) _writeProbeCache(cacheKey, probeResult);
+
+  return { ok: true, capabilities: _mergeLayer(capabilities, probeResult, 'runtimeProbe'), cached: false };
+}
+
+export const _internals = { CACHE_DIR, _cacheKey, _cachePath };

package/src/llm-validator/model-status.js

@@ -64,3 +64,30 @@ export function summarizeModelStatus(findings) {
   }
   return { counts, notApplicable, total: (findings || []).length };
 }
+
+/**
+ * agentic-security-ollama-offline-prd.md §26 — turn a summarizeModelStatus()
+ * result into the "AI stages" call/success/refused/failed breakdown a report
+ * can render. Reuses the SAME five-state taxonomy every other status
+ * consumer already trusts, rather than adding new instrumentation: COMPLETED
+ * + MALFORMED + UNAVAILABLE are all outcomes of an ATTEMPTED call (the model
+ * was actually asked); POLICY_BLOCKED never reached the network at all, so
+ * it is reported as "refused" rather than folded into "calls". DISABLED
+ * (nothing configured) contributes to neither — a caller checks that
+ * separately (status.counts[MODEL_STATUS.DISABLED] === status.total means
+ * "don't render this stage at all").
+ *
+ * @returns {{calls:number, success:number, refused:number, failed:number}}
+ */
+export function stageSummaryFromModelStatus(status) {
+  const c = status?.counts || {};
+  const completed = c[MODEL_STATUS.COMPLETED] || 0;
+  const malformed = c[MODEL_STATUS.MALFORMED] || 0;
+  const unavailable = c[MODEL_STATUS.UNAVAILABLE] || 0;
+  return {
+    calls: completed + malformed + unavailable,
+    success: completed,
+    refused: c[MODEL_STATUS.POLICY_BLOCKED] || 0,
+    failed: malformed + unavailable,
+  };
+}