prism-mcp-server 20.12.1 → 20.13.1

package/README.md

@@ -694,14 +694,20 @@ for manual configuration and host-specific paths.
 **Optional — local model fleet** for offline tool-routing. Pull whichever fits your hardware:
 
 ```bash
-ollama pull dcostenco/prism-coder:2b    # 2.3 GB · mobile / lightweight (99.1% on our routing suite)
-ollama pull dcostenco/prism-coder:4b    # 3.4 GB · verifier (100% on our routing suite)
-ollama pull dcostenco/prism-coder:9b    # 5.8 GB · default router (100% on our routing suite, Qwen3.5)
-ollama pull dcostenco/prism-coder:27b   # 16 GB  · complex tasks (100% on our routing suite)
+ollama pull dcostenco/prism-coder:2b    # 3.3 GB  · on-device / lowest RAM · sees images (100% on our routing suite)
+ollama pull dcostenco/prism-coder:4b    # 3.5 GB  · verifier · sees images (100%)
+ollama pull dcostenco/prism-coder:9b    # 6.7 GB  · default router · sees images (95.7%, reasons before answering)
+ollama pull dcostenco/prism-coder:27b   # 16.8 GB · complex code / quality · text only (100%)
 ```
 
 Prism detects both the namespaced (`dcostenco/prism-coder:9b`) and bare (`prism-coder:9b`) Ollama tags automatically.
 
+The 2b/4b/9b tiers carry a vision tower and accept screenshots through
+`prism_infer({ images: [...] })` — pass absolute paths or base64. Image
+requests are refused rather than answered blind when no tier (or the Layer 1
+safety classifier) can actually see the image, so a text-only model is never
+handed a prompt about a screenshot it never received. The 27b is text only.
+
 ---
 
 ## What it does
@@ -844,9 +850,9 @@ air-gap. **Enterprise** includes a HIPAA Business Associate Agreement.
 
 ## Models
 
-The `prism-coder` fleet uses Qwen3.5 for MCP tool-routing AND general inference. The 9B and 27B are fine-tuned with LoRA (r=128, all 64 layers including DeltaNet); the 2B and 4B use stock Qwen3.5-4B at different quantization levels. The 27B scored 100% on our internal 115-case tool-routing suite and 100% on an internal 15-problem coding eval, at $0 inference cost. These are self-run evaluations, not [BFCL](https://gorilla.cs.berkeley.edu/leaderboard.html) leaderboard submissions.
+The `prism-coder` fleet uses Qwen3.5 for MCP tool-routing AND general inference. The 9B and 27B are fine-tuned; the 2B and 4B use stock Qwen3.5-4B at different quantization levels. The 27B scored 100% on our internal 115-case tool-routing suite and 100% on an internal 15-problem coding eval, at $0 inference cost. These are self-run evaluations, not [BFCL](https://gorilla.cs.berkeley.edu/leaderboard.html) leaderboard submissions.
 
-`prism_infer` supports three modes: `route` (tool routing, fast, nothink), `chat` (conversation with thinking), and `code` (code generation with thinking). In chat/code modes, the model uses `<think>` blocks for chain-of-thought reasoning, which are stripped before the response is served. If the local model fails a quality gate (empty, think-only, or truncated), paid tiers automatically escalate to Gemini 3.6 Flash via the Synalux portal.
+`prism_infer` supports three modes: `route` (tool routing, fast), `chat` (conversation) and `code` (code generation). Reasoning is decided by the **tier**, not the mode: a tier carrying `MODEL_TIERS.prefersThinking` also carries a `minLocalTokens` floor so reasoning cannot crowd out the answer, and only those tiers use `<think>` blocks (stripped before the response is served). The 9B does; the 4B and 2B do not, because on those tiers reasoning drew down the same `num_predict` budget the answer needed and returned an empty response. An explicit `think: true` still overrides, for a caller who has sized `max_tokens` for it. If the local model fails a quality gate (empty, think-only, or truncated), paid tiers automatically escalate to Gemini 3.6 Flash via the Synalux portal.
 
 Every route-mode result is parsed locally and checked against `allowed_tools`
 before it reaches the host. Malformed or unadvertised calls become `NO_TOOL`.
@@ -856,16 +862,33 @@ draft that may need correction—to Synalux for authenticated deterministic
 correction. Advertised custom host tools remain local. Set
 `route_guard: "local"` for a fully on-device route path.
 
-| Model | Ollama tag | Size | Routing accuracy¹ | Role | Automatic routing tier |
-|---|---|---|---|---|---|
-| Qwen3.5-4B Q3_K_M | `prism-coder:2b` | 2.3 GB | 99.1% × 3 seeds | iPhone / mobile first gate | Free |
-| Qwen3.5-4B Q4_K_M | `prism-coder:4b` | 3.4 GB | 100% × 3 seeds | Verifier | Free |
-| Qwen3.5-9B (LoRA) | `prism-coder:9b` | 5.8 GB | 100% × 3 seeds | Default router | Standard+ |
-| Qwen3.5-27B (LoRA) | `prism-coder:27b` | 16 GB | 100% × 3 seeds | Quality tier (DeltaNet, 28.5 tok/s) | Advanced+ |
-
-¹ Self-run on a narrow 115-case MCP tool-selection suite, 3 seeds. It says these
-models pick the right tool on our own eval, nothing more — not a general capability
-measure, and not an independent benchmark result. Full methodology caveats below.
+| Model | Ollama tag | Size | Vision | Routing accuracy¹ | Role | Automatic routing tier |
+|---|---|---|---|---|---|---|
+| Qwen3.5-4B Q4_K_S | `prism-coder:2b` | 3.3 GB | ✅ | 100% | On-device / lowest RAM (4.5 GiB free) | Free |
+| Qwen3.5-4B Q4_K_M | `prism-coder:4b` | 3.5 GB | ✅ | 100% | Verifier (5.2 GiB free) | Free |
+| Qwen3.5-9B (LoRA) | `prism-coder:9b` | 6.7 GB | ✅ | 95.7%² | Default router / workhorse (9 GiB free) | Standard+ |
+| Qwen3.5-27B (LoRA) | `prism-coder:27b` | 16.8 GB | — | 100% | Complex code / quality (21 GiB free) | Advanced+ |
+
+¹ Self-run on a narrow 115-case MCP tool-selection suite, `temperature: 0`,
+measured through the call path `prism_infer` actually uses (`/api/chat`, each
+model's own template). It says these models pick the right tool on our own eval,
+nothing more — not a general capability measure, and not an independent
+benchmark result. Earlier revisions of this table quoted 99.1–100% from a
+harness that hand-rolled a ChatML prompt with `raw: true`, bypassing the
+template; those numbers described a path no caller exercises. Full methodology
+caveats below.
+
+² The 9B is the one tier that reasons before answering, and it is measured with
+reasoning enabled: 95.7% with thinking, 83.5% without. `prism_infer` sets this
+per-tier (`MODEL_TIERS.prefersThinking`), so callers get the 95.7% path by
+default. Reasoning costs roughly 600 tokens, which is why the 9B also carries a
+2,048-token local floor.
+
+**Vision.** The 2B/4B/9B tags ship a separate `projector` layer (0.68–0.92 GB)
+and read images; the 27B is text-only. `prism_infer` probes for that layer and
+skips a tier with no vision rather than sending it an image — asked directly, a
+text-only model will still answer confidently about pixels it never received.
+Exercised against the real models in `tests/integration/visionScreenshot.test.ts`.
 
 These tiers control automatic `prism_infer` selection, not Ollama itself. Any
 user can run any downloaded on-device model directly through Ollama on every
@@ -920,10 +943,17 @@ reliability, not general model capability.
 
 | Model | Routing accuracy | Notes |
 |---|---|---|
-| prism-coder:2b (Q3_K_M) | 99.1% × 3 seeds | 1 failure: regex→knowledge_search |
-| prism-coder:4b / 9b / 27b | 100% × 3 seeds | Perfect on all 115 cases |
+| prism-coder:2b (Q4_K_S) | 100% | The 2B was requantised when vision shipped; the old 99.1% was Q3_K_M |
+| prism-coder:4b | 100% | |
+| prism-coder:9b | 95.7% with reasoning | 83.5% without — the only tier where this differs |
+| prism-coder:27b | 100% | |
 | Claude (frontier, same eval) | ~98% | Stronger everywhere outside this narrow task |
 
+Measured through `/api/chat` with each model's own template — the path
+`prism_infer` uses. `temperature: 0`, so the three seeds only reshuffle case
+order and cannot disagree; earlier revisions cited that agreement as
+confirmation, which it never was.
+
 **Memory uplift (LoCoMo-Plus, self-published).** A separate long-context dialogue benchmark ([dcostenco/Locomo-Plus](https://github.com/dcostenco/Locomo-Plus)) measures how much structured memory helps a base model retain multi-day context. Results show large gains when a model is paired with Prism memory versus running raw. Note this benchmark is authored, run, and LLM-judged by this project — treat it as a reproducible demonstration, not an independent third-party result, and run it yourself with the commands in that repo.
 
 **Code generation evaluation.** In a small July 2026 deterministic execution
@@ -1053,9 +1083,14 @@ prism_infer({
 
 | Mode | Think | Model | Use case |
 |------|-------|-------|----------|
-| `route` | Off (fast) | 9B default | MCP tool routing |
-| `chat` | On | 27B preferred | Conversation, reasoning |
-| `code` | On | 27B preferred | Code generation, debugging |
+| `route` | Off (fast) — except a tier that reasons better, e.g. 9B | 9B default | MCP tool routing |
+| `chat` | Per tier: on for 9B, off for 4B/2B | 27B preferred | Conversation, reasoning |
+| `code` | Per tier: on for 9B, off for 4B/2B | 27B preferred | Code generation, debugging |
+
+Think is a **tier** property, not a mode property. Tiers with
+`prefersThinking` also declare a `minLocalTokens` floor that reserves budget for
+the answer; tiers without it spend the whole `num_predict` allowance inside
+`<think>` and return nothing. Pass `think` explicitly to override either way.
 
 Full TypeScript signatures live in [`src/tools/`](src/tools/); architecture in [`docs/ARCHITECTURE.md`](docs/ARCHITECTURE.md).
 

package/dist/storage/synalux.js

@@ -29,6 +29,8 @@
  *     - getHistory        → POST /api/v1/prism/memory  action=memory_history
  *     - patchLedger       → POST /api/v1/prism/memory  action=save_embedding
  *     - getEntriesMissingEmbeddings → POST /api/v1/prism/memory  action=list_missing_embeddings
+ *     - listProjects      → POST /api/v1/prism/memory  action=list_projects
+ *     - exportLedger      → POST /api/v1/prism/memory  action=export_memory (paginated)
  *
  *   Methods still falling through to SupabaseStorage (Phase 3 Tier B+):
  *   save_experience direct entrypoint, compactLedger, image ops,
@@ -449,6 +451,58 @@ export class SynaluxStorage extends SupabaseStorage {
         const entries = Array.isArray(result.entries) ? result.entries : [];
         return entries;
     }
+    // ─── Project inventory + export ──────────────────────────────
+    // Both portal actions shipped in Phase 3 but the client was never
+    // wired: listProjects and the export path fell through to
+    // SupabaseStorage and threw "Supabase not configured" on every
+    // paid-tier install (2026-08-18 audit).
+    async listProjects() {
+        const result = await this.portalPost("/api/v1/prism/memory", {
+            action: "list_projects",
+        });
+        // Strict on drift: a 200 without a projects ARRAY is a contract change,
+        // not "no projects" — the portal returns projects:[] for genuinely none.
+        // Coercing drift to [] would make callers report empty inventories while
+        // claiming success. (R1 adversarial review 2026-08-18.)
+        if (!Array.isArray(result.projects)) {
+            throw new Error("[SynaluxStorage] list_projects: portal response missing projects[] — contract drift");
+        }
+        return result.projects
+            .map((p) => (typeof p === "string" ? p : p?.project))
+            .filter((name) => typeof name === "string" && name.length > 0);
+    }
+    async exportLedger(project) {
+        // action=export_memory is paginated (EXPORT_PAGE_SIZE=1000). Follow
+        // next_offset until has_more is false, capped at 10 pages to match the
+        // local path's 10k-row OOM guard.
+        //
+        // Strict on drift (R1 adversarial review 2026-08-18): this feeds a
+        // BACKUP. A 200 missing ledger[] or page{} must throw, not degrade —
+        // coercing either would write an empty or silently-truncated export
+        // file that reports ✅ success, which is worse than any error.
+        const rows = [];
+        let offset = 0;
+        for (let page = 0; page < 10; page++) {
+            const result = await this.portalPost("/api/v1/prism/memory", {
+                action: "export_memory",
+                project,
+                offset,
+                limit: 1000,
+            });
+            if (!Array.isArray(result.ledger)) {
+                throw new Error("[SynaluxStorage] export_memory: portal response missing ledger[] — contract drift");
+            }
+            const pageInfo = result.page;
+            if (typeof pageInfo?.has_more !== "boolean") {
+                throw new Error("[SynaluxStorage] export_memory: portal response missing page.has_more — refusing a possibly-truncated export");
+            }
+            rows.push(...result.ledger);
+            if (!pageInfo.has_more || typeof pageInfo.next_offset !== "number")
+                break;
+            offset = pageInfo.next_offset;
+        }
+        return rows;
+    }
     // ─── Time Travel ─────────────────────────────────────────────
     // Phase 3 Tier B: route memory_history through portal instead of
     // falling through to SupabaseStorage (which requires a direct

package/dist/tools/compactionHandler.js

@@ -9,7 +9,7 @@
 import { getStorage } from "../storage/index.js";
 import { PRISM_USER_ID } from "../config.js";
 import { getLLMProvider } from "../utils/llm/factory.js";
-import { callLocalLlm } from "../utils/localLlm.js";
+import { inferText } from "./prismInferHandler.js";
 import { PRISM_LOCAL_LLM_ENABLED, PRISM_STRICT_LOCAL_MODE } from "../config.js";
 import { debugLog } from "../utils/logger.js";
 // ─── Constants ────────────────────────────────────────────────
@@ -111,7 +111,10 @@ async function summarizeEntries(entries) {
     // ── Path 1: Local LLM (prism-coder:9b) ───────────────────────────
     if (PRISM_LOCAL_LLM_ENABLED) {
         debugLog(`[compact_ledger] Attempting local LLM summarization (${entries.length} entries)`);
-        const localResponse = await callLocalLlm(prompt);
+        // Ledger content goes through the full ladder so Layer 1 screening, the
+        // quality gate and the entitlement ceiling all apply — this summarises
+        // session records that may carry sensitive material.
+        const localResponse = await inferText(prompt, { mode: "chat" });
         if (localResponse) {
             debugLog(`[compact_ledger] Local LLM summarization succeeded`);
             return parseCompactionResponse(localResponse, "local-llm");

package/dist/tools/ingestHandler.js

@@ -54,50 +54,60 @@ function chunkSource(content, chunkSize, source) {
         totalChars: content.length,
     };
 }
-// ─── Q&A Generator (Claude Haiku) ───────────────────────────────
+// ─── Q&A Generator (via prism_infer) ────────────────────────────
+/**
+ * Generate training Q&A pairs for a chunk.
+ *
+ * This used to POST straight to api.anthropic.com with a key read from
+ * ANTHROPIC_API_KEY **or ~/.anthropic_key**. That is an ambient-credential
+ * path: any process with the variable set bills the owner, which is how a
+ * local test run once burned ~$500 against a key exported in a shell profile.
+ * It also sat outside every gate prism applies to model calls.
+ *
+ * Ingest is training-data generation rather than user-facing inference, which
+ * is why it was left alone at first — but "not inference" does not make a raw
+ * billing surface a good idea, and prism_infer already offers the same
+ * capability local-first with an entitlement-gated cloud fallback. So it goes
+ * through the ladder like everything else, and prism no longer reads a
+ * provider key from the environment at all.
+ *
+ * PHI redaction still runs BEFORE the model sees anything: a chunk may carry
+ * client names in file paths, inline identifiers, or clinical notes, and the
+ * local tier is not a licence to skip that.
+ */
 async function generateQAPairs(chunk, source) {
-    const apiKey = process.env.ANTHROPIC_API_KEY ||
-        (existsSync(`${process.env.HOME}/.anthropic_key`)
-            ? readFileSync(`${process.env.HOME}/.anthropic_key`, "utf-8").trim()
-            : null);
-    if (!apiKey) {
-        debugLog("[ingest] No ANTHROPIC_API_KEY — skipping Q&A generation, storing raw chunks");
-        return [{ prompt: `What does this ${source} code do?`, response: chunk.slice(0, 500) }];
-    }
-    // PHI redaction BEFORE sending to cloud LLM — the chunk may contain
-    // client names in file paths, inline identifiers, or clinical notes.
+    const fallback = [{ prompt: `What does this ${source} code do?`, response: chunk.slice(0, 500) }];
     const { scanAndRedactPHI } = await import("../utils/phiGuard.js");
     const redactedChunk = scanAndRedactPHI(chunk).redacted;
+    const { inferText } = await import("./prismInferHandler.js");
+    const text = await inferText(`Source: ${source}\n\`\`\`\n${redactedChunk.slice(0, 5000)}\n\`\`\``, {
+        system: 'Generate 3 Q&A training pairs as JSON array: [{"prompt":"...","response":"..."}]. Focus on what the code does, how it works, and key patterns.',
+        mode: "chat",
+        maxTokens: 2048,
+    });
+    if (!text) {
+        debugLog("[ingest] no model output — storing raw chunk");
+        return fallback;
+    }
     try {
-        const res = await fetch("https://api.anthropic.com/v1/messages", {
-            method: "POST",
-            headers: {
-                "Content-Type": "application/json",
-                "x-api-key": apiKey,
-                "anthropic-version": "2023-06-01",
-            },
-            body: JSON.stringify({
-                model: "claude-haiku-4-5-20251001",
-                max_tokens: 2048,
-                system: 'Generate 3 Q&A training pairs as JSON array: [{"prompt":"...","response":"..."}]. Focus on what the code does, how it works, and key patterns.',
-                messages: [{ role: "user", content: `Source: ${source}\n\`\`\`\n${redactedChunk.slice(0, 5000)}\n\`\`\`` }],
-            }),
-        });
-        if (!res.ok) {
-            debugLog(`[ingest] Claude API error: ${res.status}`);
-            return [];
-        }
-        const data = await res.json();
-        const text = data.content?.[0]?.text || "";
         const match = text.match(/\[.*\]/s);
-        if (match) {
-            return JSON.parse(match[0]);
+        if (!match) {
+            debugLog("[ingest] model output had no JSON array — storing raw chunk");
+            return fallback;
         }
+        const parsed = JSON.parse(match[0]);
+        if (!Array.isArray(parsed))
+            return fallback;
+        // Only keep well-formed pairs; a malformed element must not reach storage
+        // as an undefined prompt/response.
+        const pairs = parsed.filter((p) => !!p && typeof p.prompt === "string"
+            && typeof p.response === "string");
+        return pairs.length > 0 ? pairs : fallback;
     }
     catch (err) {
-        debugLog(`[ingest] Q&A generation error: ${err}`);
+        debugLog(`[ingest] Q&A parse error: ${err}`);
+        return fallback;
     }
-    return [];
 }
 // ─── Main Ingest Pipeline ───────────────────────────────────────
 export async function ingestKnowledge(args) {

package/dist/tools/ledgerHandlers.js

@@ -2719,12 +2719,24 @@ export async function sessionExportMemoryHandler(args) {
             debugLog(`[session_export_memory] Exporting project "${project}" as ${format}`);
             // Fetch handoff (live context)
             const ctx = await storage.loadContext(project, "deep", PRISM_USER_ID);
-            // Fetch full ledger (all non-deleted entries, capped at 10k as OOM guard)
-            const ledger = await storage.getLedgerEntries({
-                project: `eq.${project}`,
-                order: "created_at.asc",
-                limit: "10000",
-            });
+            // Fetch full ledger (all non-deleted entries, capped at 10k as OOM guard).
+            // Portal-backed installs export via action=export_memory (exportLedger);
+            // local/direct backends keep the getLedgerEntries assembly. Without this
+            // branch, paid thin-client installs threw "Supabase not configured" here.
+            const ledger = (typeof storage.exportLedger === "function"
+                ? await storage.exportLedger(project)
+                : await storage.getLedgerEntries({
+                    project: `eq.${project}`,
+                    // R1 adversarial review 2026-08-18: without this filter the
+                    // direct/local paths exported TOMBSTONED rows — content the user
+                    // had asked session_forget_memory to erase shipped in every
+                    // backup (GDPR Art. 17 leak; the portal export always excluded
+                    // them, so the two paths also disagreed). Both PostgREST and the
+                    // sqlite filter parser support is.null.
+                    deleted_at: "is.null",
+                    order: "created_at.asc",
+                    limit: "10000",
+                }));
             // Strip raw embedding vectors from the export (large binary / not human-useful)
             // embedding: raw float32 JSON array (~12KB/entry)
             // embedding_compressed: TurboQuant binary blob (~400B/entry, base64 in JSON)