nodal-agents 0.4.2 → 0.4.4

package/README.md

@@ -16,12 +16,6 @@ connectors, and delegate to each other to get the job done.
 to install, runs on any machine with Node 22+ (Mac, PC, Linux), and works
 with **any LLM** — frontier or local, paid or free.
 
-> **Run the models others choke on.** DeepSeek V4 works out of the box —
-> including the cheap, fast OpenRouter routes — because the runtime quietly
-> handles its quirks (non-standard tool-call payloads) instead of failing.
-> Mix the right model per agent: a frontier model where judgment matters, a
-> near-free one where it doesn't.
-
 ---
 
 ## Why Nodal-Agents
@@ -29,7 +23,7 @@ with **any LLM** — frontier or local, paid or free.
 | | |
 | --- | --- |
 | 🏠  **Local-first** | Single binary, embedded Postgres, zero cloud dependency. Your conversations, memory, and credentials stay on your machine. |
-| 🔌  **Any model, per agent** | Anthropic, OpenAI, Google, Groq, Mistral, OpenRouter, or any local model (LM Studio, Ollama). Setup is just an API key per provider — **each agent picks its own model**, so you can run Claude for the orchestrator and **DeepSeek V4** for the workers. The quirks of OSS frontier models (DeepSeek's non-spec tool args, Kimi/Qwen/GLM XML tool formats) are handled automatically. |
+| 🔌  **Any model, per agent** | Anthropic, OpenAI, Google, Groq, Mistral, OpenRouter, or any local model (LM Studio, Ollama). Setup is just an API key per provider — **each agent picks its own model**, so you can run Claude for the orchestrator and **DeepSeek V4** for the workers. The quirks of OSS frontier models (DeepSeek's non-spec tool args, Kimi/Qwen/GLM XML tool formats, and round-tripping a reasoning model's chain-of-thought across tool calls so MiniMax M3 / DeepSeek / Gemini 3 don't degrade mid-task) are handled automatically. |
 | 🛟  **Provider failover** | Give an agent a backup key chain — if its provider 5xx's, times out, or hits quota mid-job, the runner fails over to the next one and keeps going (and fails loud only when the whole chain is down). |
 | 🧠  **Memory that compounds** | Persistent facts (entity-scoped, auto-injected into every job) and chat-thread continuity (your agent remembers what it said 30 seconds ago — and what it said yesterday). |
 | 🤝  **Orchestrators that finish** | Router and planner orchestrators delegate to specialist sub-agents, then resume on completion to wrap up the answer. |
@@ -90,7 +84,7 @@ When a newer version is available, `nodal-agents up` also prints a one-line
 notice:
 
 ```
-ℹ v0.4.2 available — run `nodal-agents update`
+ℹ v0.4.4 available — run `nodal-agents update`
 ```
 
 ### Build from source
@@ -218,7 +212,7 @@ pnpm deps:check   # runs locally and in CI before every release
 
 ## Status
 
-**Current release:** `0.4.2` on npm `latest`. Used daily by the
+**Current release:** `0.4.4` on npm `latest`. Used daily by the
 maintainer, stable enough for personal production. Pre-1.0 — breaking
 changes are still possible between minors.
 
@@ -227,14 +221,20 @@ changes are still possible between minors.
 - Multi-LLM, **per-agent model selection** — provider setup is just an API key
   (one per provider); each agent chooses its own model on top. Frontier and
   local providers, including **DeepSeek V4** (non-spec tool-call args normalized
-  automatically), plus native tool-call parsing for Kimi K2 / Qwen3-Coder / GLM
+  automatically) and **reasoning models like MiniMax M3** (their hidden
+  chain-of-thought is round-tripped across tool calls so they keep reasoning on
+  multi-turn tasks), plus native tool-call parsing for Kimi K2 / Qwen3-Coder / GLM
 - **Provider failover** — an opt-in ordered key chain per agent; on a 5xx /
   timeout / quota mid-job the runner fails over to the next key, and fails loud
   (`all_providers_failed`) only when the whole chain is exhausted
 - **Reliability guards (generic, model-agnostic)** — per-job token budget +
-  no-progress detector (kill runaway loops), and a no-false-success guard that
+  no-progress detector (kill runaway loops), a no-false-success guard that
   refuses to complete a job as "success" when a tool action failed and was never
-  resolved (fail loud, never fake it)
+  resolved (fail loud, never fake it), and context compaction that evicts stale
+  tool output before the model's context window overflows
+- **Diagnosable failures** — every failed job persists its full transcript and
+  the real upstream provider error (not an opaque "provider returned error"), so
+  you can see exactly what the agent did and why it stopped
 - Persistent memory (sanitation, dedup, importance ranking, auto-injection,
   feedback loop)
 - Session-thread continuity on chat channels (Telegram today)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "nodal-agents",
-  "version": "0.4.2",
+  "version": "0.4.4",
   "description": "Local-first AI agent platform with a web dashboard — install in one command.",
   "license": "MIT",
   "repository": {
@@ -46,6 +46,7 @@
     "@ai-sdk/openai": "^3.0.63",
     "@ai-sdk/openai-compatible": "^2.0.47",
     "@ai-sdk/provider": "^3.0.10",
+    "@openrouter/ai-sdk-provider": "^2.9.0",
     "ollama-ai-provider-v2": "^3.5.0",
     "googleapis": "^171.4.0",
     "@notionhq/client": "^2.3.0",