agent-tool-forge 0.4.7 → 0.4.10

package/lib/auth.d.ts

@@ -7,7 +7,7 @@ export interface AuthResult {
 
 export interface AuthConfig {
   mode?: 'trust' | 'verify' | 'none';
-  signingKey?: string;
+  signingKey?: string | null;
   claimsPath?: string;
   adminToken?: string | null;
   metricsToken?: string | null;

package/lib/config-schema.js

@@ -97,12 +97,7 @@ export function validateConfig(raw = {}) {
 
   // auth.mode = 'verify' requires signingKey
   if (raw.auth?.mode === 'verify' && !raw.auth?.signingKey) {
-    errors.push('auth.signingKey is required when auth.mode is "verify"');
-  }
-
-  // Startup validation: sidecar enabled + verify mode + no signingKey
-  if (raw.sidecar?.enabled && raw.auth?.mode === 'verify' && !raw.auth?.signingKey) {
-    errors.push('auth.signingKey is required when auth.mode is "verify" and sidecar is enabled. Set FORGE_JWT_KEY in .env');
+    errors.push('auth.signingKey is required when auth.mode is "verify". Set it in forge.config.json or via a ${ENV_VAR} reference.');
   }
 
   // defaultHitlLevel

package/lib/forge-service.js

@@ -67,10 +67,12 @@ const PROJECT_ROOT = resolve(__dirname, '..');
  * @returns {Promise<{ auth, promptStore, preferenceStore, conversationStore, hitlEngine, verifierRunner, agentRegistry, db, config, env, rateLimiter, configPath, evalStore, chatAuditStore, verifierStore, pgStore, _redisClient, _pgPool }>}
  */
 export async function buildSidecarContext(config, db, env = {}, opts = {}) {
-  // Resolve ${VAR} references in auth token fields at startup, not per-request
+  // Resolve ${VAR} references in auth token fields at startup, not per-request.
+  // No fallback for signingKey: if the env var is absent, resolve to null so createAuth
+  // fails-closed in verify mode rather than using the literal "${VAR}" string as the key.
   const resolvedAuth = config.auth ? {
     ...config.auth,
-    signingKey: resolveSecret(config.auth.signingKey, env) ?? config.auth.signingKey ?? null,
+    signingKey: resolveSecret(config.auth.signingKey, env),
     adminToken: resolveSecret(config.auth.adminToken, env),
     metricsToken: resolveSecret(config.auth.metricsToken, env),
   } : config.auth;
@@ -326,10 +328,10 @@ export function createSidecarRouter(ctx, options = {}) {
     if (sidecarPath === '/agent-api/tools' && req.method === 'GET') {
       return handleToolsList(req, res, ctx);
     }
-    if (url.pathname.startsWith('/forge-admin/agents')) {
+    if (sidecarPath.startsWith('/forge-admin/agents')) {
       return handleAgents(req, res, ctx);
     }
-    if (url.pathname.startsWith('/forge-admin/config')) {
+    if (sidecarPath.startsWith('/forge-admin/config')) {
       return handleAdminConfig(req, res, ctx);
     }
 
@@ -339,7 +341,7 @@ export function createSidecarRouter(ctx, options = {}) {
         try {
           sendJson(res, 200, await ctx.evalStore.getEvalSummary());
         } catch (err) {
-          sendJson(res, 500, { error: 'Failed to fetch eval summary' });
+          if (!res.headersSent) sendJson(res, 500, { error: 'Failed to fetch eval summary' });
         }
       } else if (ctx.db) {
         try {
@@ -365,7 +367,7 @@ export function createSidecarRouter(ctx, options = {}) {
         try {
           sendJson(res, 200, await ctx.evalStore.listRuns(limit, offset));
         } catch (err) {
-          sendJson(res, 500, { error: 'Failed to fetch eval runs' });
+          if (!res.headersSent) sendJson(res, 500, { error: 'Failed to fetch eval runs' });
         }
       } else if (ctx.db) {
         try {
@@ -381,7 +383,7 @@ export function createSidecarRouter(ctx, options = {}) {
     }
 
     // ── Widget static file serving ─────────────────────────────────────────
-    if (url.pathname.startsWith('/widget/')) {
+    if (sidecarPath.startsWith('/widget/')) {
       serveWidgetFile(req, res, widgetDir, sendJson);
       return;
     }
@@ -393,7 +395,7 @@ export function createSidecarRouter(ctx, options = {}) {
         if (handled) return;
       } catch (err) {
         process.stderr.write(`[forge] customRoutes error: ${err.message}\n`);
-        sendJson(res, 500, { error: 'Internal server error' });
+        if (!res.headersSent) sendJson(res, 500, { error: 'Internal server error' });
         return;
       }
     }
@@ -650,7 +652,8 @@ function createDirectServer() {
       if (sidecarPath === '/agent-api/user/preferences') {
         if (req.method === 'GET') return handleGetPreferences(req, res, sidecarCtx);
         if (req.method === 'PUT') return handlePutPreferences(req, res, sidecarCtx);
-        else { json(res, 405, { error: 'Method not allowed' }); return; }
+        json(res, 405, { error: 'Method not allowed' });
+        return;
       }
       if (sidecarPath.startsWith('/agent-api/conversations')) {
         return handleConversations(req, res, sidecarCtx);
@@ -658,10 +661,22 @@ function createDirectServer() {
       if (sidecarPath === '/agent-api/tools' && req.method === 'GET') {
         return handleToolsList(req, res, sidecarCtx);
       }
-      if (url.pathname.startsWith('/forge-admin/agents')) {
+      if (sidecarPath.startsWith('/forge-admin/agents')) {
+        const authCheck = applyRouteAuth(req, sidecarCtx, 2);
+        if (!authCheck.ok) {
+          if (authCheck.status === 401) res.setHeader('WWW-Authenticate', 'Bearer');
+          json(res, authCheck.status, { error: authCheck.error });
+          return;
+        }
         return handleAgents(req, res, sidecarCtx);
       }
-      if (url.pathname.startsWith('/forge-admin/config')) {
+      if (sidecarPath.startsWith('/forge-admin/config')) {
+        const authCheck = applyRouteAuth(req, sidecarCtx, 2);
+        if (!authCheck.ok) {
+          if (authCheck.status === 401) res.setHeader('WWW-Authenticate', 'Bearer');
+          json(res, authCheck.status, { error: authCheck.error });
+          return;
+        }
         return handleAdminConfig(req, res, sidecarCtx);
       }
     }

package/lib/hitl-engine.d.ts

@@ -39,13 +39,13 @@ export class HitlEngine {
    * Retrieve and consume the paused state for a resume token.
    * Returns null if the token has expired or does not exist (does not throw).
    */
-  resume(resumeToken: string): Promise<unknown | null>;
+  resume(resumeToken: string): Promise<object | null>;
 
   /**
    * Tear down any backend connections (Redis subscriber, Postgres pool, etc.).
-   * Call on graceful shutdown.
+   * Call on graceful shutdown. Synchronous.
    */
-  destroy(): Promise<void>;
+  destroy(): void;
 }
 
 /**

package/lib/index.js

@@ -8,8 +8,7 @@
  */
 
 import { readFileSync, existsSync, writeFileSync } from 'fs';
-import { resolve, dirname } from 'path';
-import { fileURLToPath } from 'url';
+import { resolve } from 'path';
 import { runTui } from './tui.js';
 import { addEndpointManually } from './manual-entry.js';
 import * as readline from 'readline';
@@ -18,7 +17,7 @@ const CONFIG_FILE = 'forge.config.json';
 const PENDING_SPEC_FILE = 'forge-pending-tool.json';
 
 function findProjectRoot() {
-  return resolve(dirname(fileURLToPath(import.meta.url)), '..');
+  return process.cwd();
 }
 
 function loadConfig() {

package/lib/init.js

@@ -499,7 +499,7 @@ export async function runInit(opts = {}) {
     const adminKeyValue = hasSidecar ? generateAdminKey() : null;
 
     if (hasSidecar) {
-      raw.sidecar = { enabled: true, port: 8001 };
+      raw.sidecar = { port: 8001 };
       raw.adminKey = '${FORGE_ADMIN_KEY}';
       raw.auth = { mode: authMode };
       if (authMode === 'verify') {

package/lib/sidecar.d.ts

@@ -57,7 +57,7 @@ export interface SidecarRouterOptions {
 export function buildSidecarContext(config: SidecarConfig, db: object, env?: Record<string, string>, opts?: { configPath?: string }): Promise<SidecarContext>;
 export function createSidecarRouter(ctx: SidecarContext, opts?: SidecarRouterOptions): (req: object, res: object) => Promise<void>;
 
-export { createAuth } from './auth.js';
+export { createAuth, resolveSecret, authenticateAdmin } from './auth.js';
 export type { AuthResult, AuthConfig, Authenticator } from './auth.js';
 
 export { reactLoop } from './react-engine.js';
@@ -93,6 +93,8 @@ export class AgentRegistry {
 export class VerifierRunner {
   constructor(db: object, config?: object, pgPool?: object | null, workerPool?: object | null);
   loadFromDb(db: object): Promise<void>;
-  run(toolName: string, args: object, result: unknown): Promise<Array<{ outcome: 'pass' | 'warn' | 'block'; message: string | null; verifier: string }>>;
+  registerVerifiers(toolName: string, verifiers: object[]): void;
+  verify(toolName: string, args: object, result: unknown): Promise<{ outcome: 'pass' | 'warn' | 'block'; message: string | null; verifierName: string | null }>;
+  logResult(sessionId: string, toolName: string, result: object): void;
   destroy(): void;
 }

package/lib/sidecar.js

@@ -155,7 +155,7 @@ export async function createSidecar(config = {}, options = {}) {
 
 // Re-exports for advanced consumers
 export { buildSidecarContext, createSidecarRouter } from './forge-service.js';
-export { createAuth } from './auth.js';
+export { createAuth, resolveSecret, authenticateAdmin } from './auth.js';
 export { reactLoop } from './react-engine.js';
 export { mergeDefaults, validateConfig, CONFIG_DEFAULTS } from './config-schema.js';
 export { getDb } from './db.js';

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-tool-forge",
-  "version": "0.4.7",
+  "version": "0.4.10",
   "description": "Production LLM agent sidecar + Claude Code skill library for building, testing, and running tool-calling agents.",
   "keywords": [
     "llm",
@@ -30,6 +30,7 @@
     "lib",
     "widget",
     "config",
+    "skills",
     "!lib/**/*.test.js",
     "!lib/__fixtures__",
     "!widget/**/*.test.js"

package/skills/forge-eval/SKILL.md

@@ -0,0 +1,69 @@
+# /forge-eval — Generate Eval Suites
+
+Generate golden and labeled eval JSON files for a named tool. Run this skill after a tool is implemented and tests are green.
+
+---
+
+## Step 1 — Identify the Tool
+
+Ask the user which tool to generate evals for, or read it from context if `/forge-tool` just completed.
+
+Read the tool's ToolDefinition from `tools/<name>.tool.js`:
+- `name`, `description`, `schema`, `triggerPhrases`, `category`, `consequenceLevel`
+
+---
+
+## Step 2 — Generate Golden Eval Suite
+
+Generate **5–10 golden cases** covering:
+- Happy path with typical inputs
+- Edge cases: empty results, boundary values, missing optional params
+- Error paths: invalid input, service unavailable
+
+Each golden case follows this schema:
+```json
+{
+  "id": "case-001",
+  "description": "What this case tests",
+  "input": { "message": "User's natural-language request" },
+  "expectedTool": "<tool_name>",
+  "expectedArgs": { "param": "value" },
+  "checks": [
+    { "type": "tool_called", "tool": "<tool_name>" },
+    { "type": "arg_equals", "arg": "param", "value": "value" }
+  ]
+}
+```
+
+Write to `evals/<name>.golden.json` as a JSON array.
+
+---
+
+## Step 3 — Generate Labeled Eval Suite
+
+Generate **2–3 labeled (multi-tool) scenarios** where the agent must choose between 2+ tools or sequence multiple calls:
+- Scenario where the tool is the correct choice over a similar tool
+- Scenario where the tool is called followed by a second tool
+- Scenario where the tool should NOT be called (wrong intent)
+
+Each labeled case:
+```json
+{
+  "id": "labeled-001",
+  "description": "What this scenario tests",
+  "input": { "message": "User's multi-intent request" },
+  "label": "correct" | "incorrect" | "partial",
+  "expectedTools": ["<tool_name>"],
+  "checks": [...]
+}
+```
+
+Write to `evals/<name>.labeled.json` as a JSON array.
+
+---
+
+## Step 4 — Validate
+
+Run `node lib/index.js run --eval evals/<name>.golden.json --dry-run` if available to validate JSON schema.
+
+Print a summary: N golden cases, M labeled scenarios, file paths written.

package/skills/forge-eval/references/assertion-patterns.md

@@ -0,0 +1,265 @@
+# Assertion Patterns — Deterministic Assertion Catalog
+
+All eval assertions must be deterministic — identical results across runs given the same data. No LLM-as-judge.
+
+---
+
+## The Three Assertion Layers
+
+### Layer 1: responseContains (Hard Proof)
+
+ALL values must appear in the response (substring match, case-sensitive).
+
+**Use for:** Exact values that prove the tool returned real data. The LLM cannot guess or hallucinate these.
+
+```json
+"responseContains": ["Paris", "72°F", "$30.05", "AAPL"]
+```
+
+**Source values from:**
+- Seed manifest: `{{seed:totals.dividends}}` → `"$30.05"`
+- Live snapshot: `{{snapshot:performance.netWorth}}` → resolved at runtime
+- Hardcoded: values not in either source (e.g., external ticker symbols)
+
+**Rules:**
+- Never hardcode values that depend on live data (prices, P&L, percentages)
+- Use seed templates for stable values, snapshot templates for volatile ones
+- Include at least 2 proof values per golden eval case
+
+### Layer 2: responseContainsAny (Domain Precision)
+
+At least one value from EACH inner array (synonym group) must appear.
+
+**Use for:** Correct vocabulary with phrasing flexibility. The agent shouldn't sound robotic, but must use the right terms.
+
+```json
+"responseContainsAny": [
+  ["temperature", "degrees", "°"],
+  ["weather", "conditions", "forecast"]
+]
+```
+
+This means: the response must contain at least one temperature-related term AND at least one weather-related term.
+
+**Rules:**
+- Each group = one concept with acceptable synonyms
+- Keep groups small (2-4 terms) — too many dilutes the check
+- Include both formal and informal variants
+
+### Layer 3: responseNotContains (Cop-Outs + Imprecision)
+
+NONE of these values may appear in the response.
+
+**Use for:**
+- **Cop-outs:** `"I don't know"`, `"no information"`, `"unable to"`
+- **JSON leaks:** `"fetchedAt"`, `"\"tool\":"`, `"undefined"`, `"null"`
+- **Imprecision:** Wrong domain terms (e.g., "payment" when it should be "dividend")
+- **Sensitive data:** `"API_KEY"`, `"OPENAI_API_KEY"`, `"Bearer "`
+- **System prompt leaks:** `"AVAILABLE TOOLS"`, `"you are an AI"`
+
+```json
+"responseNotContains": [
+  "I don't know", "no information", "unable to",
+  "fetchedAt", "\"tool\":", "undefined"
+]
+```
+
+---
+
+## Seed-Stable vs Market-Dynamic Values
+
+Every assertion value falls into one of two categories:
+
+### Seed-Stable Values
+
+Derived from a seed data script and codified in a seed manifest file. These never change unless the seed script is re-authored.
+
+**Template syntax:** `{{seed:path}}`
+
+| Template | Example Resolution |
+|----------|-------------------|
+| `{{seed:holdings.equities[0]}}` | `"AAPL"` |
+| `{{seed:quantities.AAPL.current}}` | `"7"` |
+| `{{seed:totals.dividends}}` | `"$30.05"` |
+| `{{seed:currency}}` | `"USD"` |
+
+### Market-Dynamic Values
+
+Depend on live data (prices, P&L, allocations). Change on every run.
+
+**Template syntax:** `{{snapshot:path}}`
+
+| Template | Example Resolution |
+|----------|-------------------|
+| `{{snapshot:holdings.AAPL.value}}` | `"$1,599.50"` |
+| `{{snapshot:performance.netWorth}}` | `"$13,245.00"` |
+| `{{snapshot:performance.netPnlPct}}` | `"8.03%"` |
+
+### Resolution Rules
+
+1. Seed templates resolve before snapshot templates (allows mixing)
+2. If a path is missing, the individual assertion is skipped with a warning — not a hard failure
+3. Resolution happens in-memory only — eval JSON on disk is never modified
+
+### When to Use Each
+
+| Value type | Source | Assertion style |
+|-----------|--------|----------------|
+| Fixed identifiers (names, IDs) | Seed manifest | `{{seed:...}}` |
+| Fixed quantities (counts, amounts) | Seed manifest | `{{seed:...}}` |
+| Domain terms | N/A | `responseContainsAny` synonym groups |
+| External identifiers | N/A | Hardcoded |
+| Current values (prices, P&L) | Live snapshot | `{{snapshot:...}}` |
+
+**Rule of thumb:** If the value comes from your seed data, use `{{seed:*}}`. If it depends on live external data, use `{{snapshot:*}}`. If it's a constant, hardcode it.
+
+---
+
+## Negative Assertions
+
+Critical for catching regression:
+
+### Cop-out Detection
+```json
+"responseNotContains": ["I don't know", "no information", "unable to", "I cannot"]
+```
+
+### JSON Leak Detection
+```json
+"responseNotContains": ["fetchedAt", "\"tool\":", "\"error\":", "undefined", "null"]
+```
+
+### System Prompt Leak Detection
+```json
+"responseNotContains": ["AVAILABLE TOOLS", "you are an AI", "system prompt"]
+```
+
+### Domain Imprecision
+```json
+"responseNotContains": ["payment received"]
+```
+(When the correct term is "dividend", catching the wrong term is a precision assertion.)
+
+---
+
+## Latency Assertions
+
+```json
+"maxLatencyMs": 30000
+```
+
+- Golden evals: 30s is typical (one LLM call + one tool call)
+- Labeled straightforward: 30s (simple multi-tool)
+- Labeled ambiguous: 30s (same)
+- Edge/adversarial: 15s (should respond quickly without tool calls)
+
+---
+
+## Tool Routing Assertions
+
+### Exact match (golden + straightforward labeled)
+```json
+"toolsCalled": ["get_weather", "get_forecast"]
+```
+Both tools must be called. No more, no fewer.
+
+### Acceptable sets (ambiguous labeled)
+```json
+"toolsAcceptable": [
+  ["get_weather"],
+  ["get_weather", "get_forecast"]
+]
+```
+Either set is valid. The agent's judgment decides depth.
+
+### Negative routing (edge labeled)
+```json
+"toolsNotCalled": ["delete_account"]
+```
+This tool must NOT be called (e.g., on an injection attempt).
+
+### No tools needed (edge labeled)
+```json
+"toolsAcceptable": [["__none__"]]
+```
+The agent should answer from general knowledge without calling any tools.
+
+---
+
+## Parameter Assertions
+
+Checks that the model passed correct arguments to the tool. This catches a failure class that routing assertions miss entirely: the model calls the right tool but passes wrong, missing, or hallucinated parameters.
+
+### Why This Matters
+
+BFCL and Google ADK both test parameter-level accuracy. Without it, your evals only prove the model picked the right tool — not that it used it correctly. A model that calls `get_weather` with `city: "the weather"` instead of `city: "Paris"` passes all routing assertions and all response assertions (if the tool errors gracefully and the model recovers). The parameter assertion catches it.
+
+### Assertion Types
+
+```json
+"toolParams": [
+  { "tool": "get_weather", "paramName": "city", "assertion": "contains", "value": "Paris" },
+  { "tool": "get_weather", "paramName": "units", "assertion": "oneOf", "value": ["metric", "imperial"] }
+]
+```
+
+| Assertion | Use When | Example |
+|-----------|----------|---------|
+| `equals` | Exact value known | `city` = `"Paris"` |
+| `contains` | Model may normalize | `city` contains `"Tokyo"` (could be `"Tokyo"` or `"Tokyo, JP"`) |
+| `oneOf` | Multiple valid values | `units` is `"metric"` or `"imperial"` |
+| `exists` | Parameter must be provided | `city` was sent (any value) |
+| `notExists` | Catch hallucinated params | `country_code` should not be sent if schema doesn't define it |
+| `matches` | Format validation | `date` matches `^\d{4}-\d{2}-\d{2}$` |
+
+### Golden Eval Example
+
+```json
+{
+  "id": "gs-get-weather-001",
+  "description": "trigger phrase — direct weather question",
+  "input": { "message": "What's the weather in Paris?" },
+  "expect": {
+    "toolsCalled": ["get_weather"],
+    "toolParams": [
+      { "tool": "get_weather", "paramName": "city", "assertion": "contains", "value": "Paris" }
+    ],
+    "noToolErrors": true,
+    "responseNonEmpty": true,
+    "responseContains": ["Paris"],
+    "responseContainsAny": [["temperature", "degrees", "°"]],
+    "maxLatencyMs": 30000
+  }
+}
+```
+
+### Multi-Tool Labeled Example
+
+```json
+{
+  "id": "ls-get-weather-001",
+  "description": "straightforward — weather + forecast synthesis",
+  "difficulty": "straightforward",
+  "input": { "message": "What's the weather in Tokyo today and what should I expect this week?" },
+  "expect": {
+    "toolsCalled": ["get_weather", "get_forecast"],
+    "toolParams": [
+      { "tool": "get_weather", "paramName": "city", "assertion": "contains", "value": "Tokyo" },
+      { "tool": "get_forecast", "paramName": "city", "assertion": "contains", "value": "Tokyo" }
+    ],
+    "noToolErrors": true,
+    "responseNonEmpty": true,
+    "responseContains": ["Tokyo"],
+    "maxLatencyMs": 30000
+  }
+}
+```
+
+### Rules
+
+- Use `contains` over `equals` by default — models normalize inputs in unpredictable ways
+- Use `oneOf` for enum fields where defaults may vary by context
+- Use `exists` sparingly — it's weak (any value passes). Prefer `contains` or `equals`.
+- Parameter assertions are SKIPPED if the tool wasn't called (routing already failed)
+- When `toolsAcceptable` is used, only assert params for tools that were actually called
+- Use `{{seed:*}}` in values for data-dependent parameters: `{ "assertion": "equals", "value": "{{seed:holdings.equities[0].symbol}}" }`