manychat-ai-agent 0.23.1 → 0.23.3

package/CHANGELOG.md

@@ -1,5 +1,39 @@
 # Changelog
 
+## [0.23.3](https://github.com/pedronastasi/manychat-ai-agent/compare/v0.23.2...v0.23.3) (2026-10-14)
+
+
+### Bug Fixes
+
+* **deps:** update patch dependencies ([8790819](https://github.com/pedronastasi/manychat-ai-agent/commit/87908192e30abafdf11f02a49692048f1177a10b))
+* **deps:** update patch dependencies ([e19e412](https://github.com/pedronastasi/manychat-ai-agent/commit/e19e41258dc91d7187a1cf476f992de10311f7f0))
+* price openai:gpt-6-luna at its list price for the budget cap ([76d5f9a](https://github.com/pedronastasi/manychat-ai-agent/commit/76d5f9aca411f528198edfd4cccc1934c6df2a1b))
+* price the common chat models of each hosted provider ([f6fd2c7](https://github.com/pedronastasi/manychat-ai-agent/commit/f6fd2c7d740d0f90d354449d2b935394f677641c))
+* price the common chat models so the budget cap tracks real spend ([0c3a3ea](https://github.com/pedronastasi/manychat-ai-agent/commit/0c3a3ead45102b8e5f6879ced90fd4d3b76f717e))
+
+
+### Specs and Docs
+
+* specify tenant overrides for system prompt sections ([4cc2f6e](https://github.com/pedronastasi/manychat-ai-agent/commit/4cc2f6eb26d714cba26afa68374ba50c9156ba09))
+* specify tenant overrides for system prompt sections ([6ca8001](https://github.com/pedronastasi/manychat-ai-agent/commit/6ca800150bfbad24c7af76788bbf738b543728f8))
+
+
+### Dependencies
+
+* **deps:** lock file maintenance ([1190025](https://github.com/pedronastasi/manychat-ai-agent/commit/11900256fc78d9244eabc521f5433cc35a981da0))
+* **deps:** lock file maintenance ([cc15348](https://github.com/pedronastasi/manychat-ai-agent/commit/cc15348e34a23fddf8832dddc71229664b32a2e1))
+* **deps:** update dependency mermaid to v12.2.0 ([c01cf91](https://github.com/pedronastasi/manychat-ai-agent/commit/c01cf9140f919f36d3df125c7a0a27fbb0001b50))
+* **deps:** update dependency mermaid to v12.2.0 ([5b3d363](https://github.com/pedronastasi/manychat-ai-agent/commit/5b3d363a8a804309caa284dcfd1f10b6c27b6a77))
+* **deps:** update pnpm to v12.11.2 ([8f565b5](https://github.com/pedronastasi/manychat-ai-agent/commit/8f565b59fbf949ae0de3f2ed81c659d607ea967a))
+* **deps:** update pnpm to v12.11.2 ([73dee04](https://github.com/pedronastasi/manychat-ai-agent/commit/73dee04d192aed76279fe17e2f31775bcab8096f))
+
+## [0.23.2](https://github.com/pedronastasi/manychat-ai-agent/compare/v0.23.1...v0.23.2) (2026-10-13)
+
+
+### Bug Fixes
+
+* **agent:** honor tenant conversation policies and Luna reasoning ([6d16fa0](https://github.com/pedronastasi/manychat-ai-agent/commit/6d16fa078c908ec3c3a96be4ad1019d7cef42589))
+
 ## [0.23.1](https://github.com/pedronastasi/manychat-ai-agent/compare/v0.23.0...v0.23.1) (2026-10-13)
 
 

package/dist/agent/prompt.js

@@ -70,19 +70,28 @@ options = {}) {
         persona.trim(),
         '',
         PROMPT_MARKERS[0],
-        '1. Answer only with information from the CATALOG. If it is not there, escalate.',
+        "The persona above defines this tenant's business policy and conversation flow.",
+        'Its explicit response, escalation and closing-question policies take precedence over',
+        'the conversational defaults below. This does not override SECURITY, factual grounding,',
+        'the output schema or tool execution constraints.',
+        '1. Ground factual answers in the CATALOG. If it cannot answer, follow an explicit',
+        '   response or escalation policy in the persona; otherwise escalate.',
         '2. Never invent prices, dates, schedules, discounts or policies.',
         '3. Discounts, "is that the best price?" or haggling: escalate with "price_negotiation".',
         '   Instalments or a payment plan: present the PAYMENT OPTIONS that cover it. If none',
         '   does, escalate with "price_negotiation".',
         '4. Complaints, disputes or refund requests: escalate with "complaint".',
         '5. A request to speak to a person: escalate with "explicit_request".',
-        '6. Anything the catalog cannot answer: escalate with "out_of_scope".',
+        '6. If neither the CATALOG nor a response policy in the persona covers the question,',
+        '   escalate with "out_of_scope".',
         '7. When unsure: escalate with "low_confidence". Escalating is correct; guessing is not.',
         '8. Read short replies in context. When the contact answers your previous question — picks an option you offered, says yes/no, names a preference — that is a valid conversational answer: continue the conversation with high confidence. Never escalate a direct answer to your own question.',
-        '9. If asked whether you are a bot, say yes plainly and offer to pass them to someone.',
+        '9. If asked whether you are a bot, acknowledge the assistance honestly. Follow the',
+        "   persona's wording and next step; absent that guidance, offer to pass them to someone.",
         '10. Never restate a reply you already sent. If the information was already given, acknowledge briefly instead of repeating it.',
-        '11. When the contact closes — thanks, bye, ok, perfect — reply with a brief, natural acknowledgement and an offer to help further. Never answer a closer by replaying your previous reply.',
+        "11. For thanks, bye, ok, perfect, follow the persona's policy for pending versus",
+        '    completed conversations. Without that policy, acknowledge briefly and offer help.',
+        '    Never answer a closer by replaying your previous reply.',
         '12. A contact who says they have paid, or sends a receipt: escalate with "payment_reported".',
         '    You cannot see payments, so never confirm one.',
         '13. Never invent urgency or scarcity ("only two places left", "the price goes up on Friday")',
@@ -116,6 +125,7 @@ options = {}) {
         'end of `messages`. Send null only when a question does not belong — a handoff,',
         'a delicate or health matter, a contact who already has the payment link,',
         'someone who has declined twice, or a turn whose opening flow asks it.',
+        'Follow any additional closing-question exceptions explicitly defined in the persona.',
         'Null is a decision, not a way to skip the field.',
         ...(withTools ? ACTIONS_SECTION : []),
         ...(options.pluginTools ? PLUGIN_TOOLS_LINES : []),
@@ -239,6 +249,8 @@ function intentSection(tools) {
                 'reply does not greet the contact, introduce you or repeat what the flow says. If the',
                 'flow asks a question, do not ask it again, in any words: unless their message already',
                 'answers it, send closing_question null.',
+                'Apply this opening exception only after an actual openingQueued result or OPENING',
+                'note in this turn. A prospect write without that notice does not announce an opening.',
             ]
             : []),
     ];

package/dist/agent/registry.js

@@ -15,14 +15,33 @@ const ollama = createOpenAI({
 const registry = createProviderRegistry({ anthropic, openai, google, ollama });
 /**
  * Prices drift, and a stale table would silently mis-cap spend. Treated as a
- * best-effort estimate: `PRICING_OVERRIDES` lets an operator correct a value
- * without a release, and an unknown model falls back to a deliberately
- * pessimistic estimate so the cap errs toward stopping early.
+ * best-effort estimate: an unknown model falls back to a deliberately
+ * pessimistic estimate so the cap errs toward stopping early. That fallback
+ * prices a cheap model at many times its list price, so a model in regular use
+ * belongs in this table.
  */
 const PRICING = {
-    'anthropic:claude-haiku-4-5': { inputPerMTok: 1, outputPerMTok: 5, cacheReadPerMTok: 0.1 },
+    // Standard list prices as of October 2026. Where a model has a long-prompt
+    // tier, this is the short-prompt price: a prompt past that threshold (100k
+    // tokens on Haiku 5.5, 272k on GPT-6) is under-counted.
+    'anthropic:claude-haiku-5-5': { inputPerMTok: 0.1, outputPerMTok: 0.5, cacheReadPerMTok: 0.01 },
+    'anthropic:claude-sonnet-5-5': { inputPerMTok: 2, outputPerMTok: 10, cacheReadPerMTok: 0.1 },
+    'anthropic:claude-opus-5-5': { inputPerMTok: 4, outputPerMTok: 20, cacheReadPerMTok: 0.2 },
     'anthropic:claude-sonnet-5': { inputPerMTok: 2, outputPerMTok: 10, cacheReadPerMTok: 0.2 },
     'anthropic:claude-opus-5': { inputPerMTok: 5, outputPerMTok: 25, cacheReadPerMTok: 0.5 },
+    'anthropic:claude-sonnet-4-6': { inputPerMTok: 3, outputPerMTok: 15, cacheReadPerMTok: 0.3 },
+    'anthropic:claude-sonnet-4-5': { inputPerMTok: 3, outputPerMTok: 15, cacheReadPerMTok: 0.3 },
+    'anthropic:claude-haiku-4-5': { inputPerMTok: 1, outputPerMTok: 5, cacheReadPerMTok: 0.1 },
+    'openai:gpt-6-sol': { inputPerMTok: 2, outputPerMTok: 10, cacheReadPerMTok: 0.2 },
+    'openai:gpt-6-luna': { inputPerMTok: 0.1, outputPerMTok: 0.5, cacheReadPerMTok: 0.01 },
+    'openai:gpt-5': { inputPerMTok: 1.25, outputPerMTok: 10, cacheReadPerMTok: 0.125 },
+    'openai:gpt-5-mini': { inputPerMTok: 0.25, outputPerMTok: 2, cacheReadPerMTok: 0.025 },
+    'openai:gpt-5-nano': { inputPerMTok: 0.05, outputPerMTok: 0.4, cacheReadPerMTok: 0.005 },
+    // Introductory pricing halves this until 2027; the table holds the standard
+    // price, so the cap errs toward stopping early meanwhile. Cache reads at 10%
+    // of input, Google's ratio on its other Flash models.
+    'google:gemini-3.8-flash': { inputPerMTok: 1.5, outputPerMTok: 7.5, cacheReadPerMTok: 0.15 },
+    'google:gemini-3.5-flash-lite': { inputPerMTok: 0.3, outputPerMTok: 2.5, cacheReadPerMTok: 0.03 },
 };
 const FALLBACK_PRICING = {
     inputPerMTok: 5,
@@ -35,8 +54,8 @@ export function pricingFor(modelId) {
         return ZERO_PRICING;
     return PRICING[modelId] ?? FALLBACK_PRICING;
 }
-/** OpenAI's reasoning families: the o-series and gpt-5. */
-const REASONING_MODEL = /^openai:(o\d|gpt-5)/;
+/** OpenAI's reasoning families: the o-series and GPT-5/GPT-6. */
+const REASONING_MODEL = /^openai:(o\d|gpt-[56])/;
 /**
  * Whether a model accepts `temperature` at all.
  *

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "manychat-ai-agent",
-  "version": "0.23.1",
+  "version": "0.23.3",
   "description": "A provider-agnostic conversational AI agent gateway that bridges no-code chat platforms to any LLM.",
   "license": "MIT",
   "repository": {
@@ -36,7 +36,7 @@
   "engines": {
     "node": ">=22"
   },
-  "packageManager": "pnpm@12.10.1",
+  "packageManager": "pnpm@12.11.2",
   "scripts": {
     "dev": "node --experimental-strip-types --disable-warning=ExperimentalWarning --env-file=.env --watch src/main.ts",
     "bootstrap": "node scripts/setup.mjs",
@@ -71,7 +71,7 @@
     "@ai-sdk/anthropic": "^4.0.53",
     "@ai-sdk/google": "^4.0.69",
     "@ai-sdk/openai": "^4.0.66",
-    "@ai-sdk/provider": "4.0.26",
+    "@ai-sdk/provider": "4.0.27",
     "@fastify/helmet": "^13.1.1",
     "@fastify/rate-limit": "^11.2.0",
     "@fastify/swagger": "^9.8.1",
@@ -97,7 +97,7 @@
     "@electric-sql/pglite": "^0.5.8",
     "@eslint/js": "^10.0.1",
     "@types/node": "^24.0.0",
-    "@vitest/coverage-v8": "5.0.3",
+    "@vitest/coverage-v8": "5.0.4",
     "drizzle-kit": "^0.31.10",
     "eslint": "^10.10.0",
     "mermaid": "^12.0.0",