@vizuh/sabi 0.2.0 → 0.2.1

package/README.md

@@ -1,9 +1,24 @@
-# @vizuh/sabi: Command Code adapter
-
-This npm artifact is one Sabi adapter, not the whole Sabi product. Sabi also supports a local
-OpenAI-compatible proxy for OpenCode, Hermes, Prime Agent, Kilo and other clients, plus an
-experimental controller surface for Claude Code, Codex and Orca. See the
-[adapter directory](../../../docs/adapters/README.md) for the product map and evidence boundaries.
+# @vizuh/sabi: Sabi inference adapter (Command Code mod)
+
+Sabi is harness-independent adaptive inference scheduling for coding-agent trajectories:
+per-round routing of model, effort and provider from trajectory evidence, while each host
+harness keeps its own loop. Sabi promotes no single harness; every host below is a
+peer surface with its own install path.
+
+This npm artifact is the Command Code mod — one Sabi surface, not the whole product.
+See the [adapter directory](../../../docs/adapters/README.md) for the product map and
+evidence boundaries.
+
+| Harness | Sabi surface | Install |
+| --- | --- | --- |
+| Claude Code | Controller hooks (task/session, fail-open) | `npm install --global @vizuh/sabi-controller` then `sabi setup` |
+| Codex | Controller hooks (task/session, fail-open) | `npm install --global @vizuh/sabi-controller` then `sabi setup` |
+| Command Code | This package: in-process mod, per-round model + reasoning effort | `cmd mods add -g npm:@vizuh/sabi` |
+| Hermes | Local proxy + native `llm_request` middleware | Checkout-based setup in [the install guide](../../../docs/install.md) |
+| Oh My Pi | Local proxy via OpenAI-compatible extension provider | Checkout: `omp --extension packages/adapters/oh-my-pi/src/sabi-extension.mjs --model sabi/sabi-code` |
+| OpenCode | Local proxy, plus optional controller hook | Checkout-based proxy (`npm start` + `npm run connect:opencode`); hooks via `@vizuh/sabi-controller` |
+| Orca | Controller plugin + inventory/dispatch bridge | `npm install --global @vizuh/sabi-controller` then `sabi setup` |
+| DeepSeek Harness, Kilo, Cline, Prime Agent | Proxy or bundle paths | See the [adapter directory](../../../docs/adapters/README.md) |
 
 Adaptive inference scheduling for [Command Code](https://commandcode.ai): a mod that plans each
 continuing round, including model and reasoning effort, from the trajectory's own state (tool calls and
@@ -17,7 +32,7 @@ The two paths are independent:
 
 - The Command Code mod needs no Sabi provider key or proxy. It routes the subscription already
   available to Command Code.
-- The local proxy works with OpenCode, Hermes, Kilo, and other OpenAI-compatible clients. It uses
+- The local proxy works with Hermes, Oh My Pi, OpenCode, and other OpenAI-compatible clients (including Cline, Kilo and Prime Agent paths). It uses
   OpenRouter, Ollama, or another configured upstream. In the shipped default the OpenRouter
   upstream is **free-models-only** (`paidModelsAllowed: false`): a priced model id is refused
   before the request leaves the process, so the proxy cannot spend on its own.

package/mod/sabi.mjs

@@ -1,4 +1,4 @@
-// @vizuh/sabi 0.2.0 — generated by pack.mjs from packages/adapters/command-code/mod/sabi.ts.
+// @vizuh/sabi 0.2.1 — generated by pack.mjs from packages/adapters/command-code/mod/sabi.ts.
 // Source and docs: https://github.com/vizuh/sabi
 
 // packages/core/src/telemetry.ts
@@ -323,7 +323,11 @@ var TRANSPORT_LIMIT_PATTERNS = [
   { re: /too many requests/i, label: "rate-limited" },
   { re: /\b(?:session|usage|weekly|monthly|daily|hourly|subscription|plan)\s+limit\b/i, label: "quota-exceeded" },
   { re: /quota[- ]?exceeded/i, label: "quota-exceeded" },
-  { re: /insufficient_quota|insufficient quota/i, label: "quota-exceeded" }
+  { re: /insufficient_quota|insufficient quota/i, label: "quota-exceeded" },
+  // A dropped connection is the transport failing. Without this, the wording a client prints after
+  // a transport deadline reads as a task failure and the next round escalates straight back to the
+  // tier that just timed out — ten consecutive 120s rounds in one real session.
+  { re: /socket (?:connection )?(?:was )?closed|socket hang up|ECONNRESET|other side closed|connection reset/i, label: "connection-closed" }
 ];
 var TRANSPORT_PATTERNS = [
   { re: /\b429\b/, label: "rate-limited" },
@@ -863,6 +867,12 @@ function validateConfig(value, source = "<inline>") {
     if (upstream.paidModelsAllowed !== void 0 && typeof upstream.paidModelsAllowed !== "boolean") {
       throw new Error(`Sabi config ${source}: upstream '${name}'.paidModelsAllowed must be a boolean`);
     }
+    if (upstream.auth !== void 0 && upstream.auth !== "passthrough") {
+      throw new Error(`Sabi config ${source}: upstream '${name}'.auth must be 'passthrough'`);
+    }
+    if (upstream.auth === "passthrough" && upstream.apiKey !== void 0 && upstream.apiKey !== false) {
+      throw new Error(`Sabi config ${source}: upstream '${name}' is auth:passthrough and must not declare apiKey`);
+    }
   }
   if (!Object.keys(models).length) throw new Error(`Sabi config ${source}: no models declared`);
   for (const [name, model] of Object.entries(models)) {
@@ -904,6 +914,21 @@ function validateConfig(value, source = "<inline>") {
       throw new Error(`Sabi config ${source}: transportFallback.enabled must be a boolean`);
     }
   }
+  const passthrough = config.passthrough;
+  if (passthrough !== void 0) {
+    if (!isObject(passthrough)) throw new Error(`Sabi config ${source}: passthrough must be an object`);
+    for (const field of Object.keys(passthrough)) {
+      if (field !== "alias") throw new Error(`Sabi config ${source}: passthrough.${field} is not a supported field`);
+    }
+    if (passthrough.alias !== void 0) {
+      if (typeof passthrough.alias !== "string" || !passthrough.alias.trim()) {
+        throw new Error(`Sabi config ${source}: passthrough.alias must be a nonempty string`);
+      }
+      if (aliases[passthrough.alias] !== "auto") {
+        throw new Error(`Sabi config ${source}: passthrough.alias '${passthrough.alias}' must name an alias targeting 'auto'`);
+      }
+    }
+  }
   const judge = config.judge;
   if (judge !== void 0) {
     if (typeof judge !== "object" || judge === null || typeof judge.enabled !== "boolean") {
@@ -1210,6 +1235,9 @@ function sanitizeDecisionRecord(record, config) {
     state,
     cache: record.cache,
     judge: record.judge,
+    // Effort labels are short provider-declared strings, never transcript text; still bounded.
+    ...typeof record.effort === "string" && record.effort.trim() ? { effort: record.effort.trim().slice(0, 64) } : {},
+    ...record.effortSource === "client" || record.effortSource === "scheduled" || record.effortSource === "unspecified" ? { effortSource: record.effortSource } : {},
     usage: record.usage,
     cost: record.cost,
     latencyMs: record.latencyMs,

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@vizuh/sabi",
-  "version": "0.2.0",
-  "description": "Adaptive inference scheduling for Command Code: one bundled mod that routes each continuing round by model, effort and trajectory state.",
+  "version": "0.2.1",
+  "description": "Sabi inference adapter for Command Code, part of harness-independent adaptive routing for Claude Code, Codex, Command Code, Hermes, Oh My Pi, OpenCode and Orca.",
   "type": "module",
   "license": "MIT",
   "repository": {
@@ -11,13 +11,20 @@
   },
   "homepage": "https://github.com/vizuh/sabi#readme",
   "keywords": [
-    "commandcode",
+    "sabi",
+    "ai-agents",
+    "claude-code",
+    "codex",
     "command-code",
+    "commandcode",
+    "hermes",
+    "inference-scheduling",
+    "llm-routing",
     "mod",
-    "llm",
-    "routing",
-    "ai-agents",
-    "inference-scheduling"
+    "oh-my-pi",
+    "opencode",
+    "openai-compatible",
+    "orca"
   ],
   "publishConfig": {
     "access": "public"

package/sabi.config.json

@@ -1,5 +1,5 @@
 {
-  "provenance": "Operator rule (2026-09-22): OpenRouter may serve FREE models only. The openrouter upstream declares paidModelsAllowed:false, so any model whose price is unknown or non-zero is refused at dispatch (packages/core/src/compatibility.ts) — a config mistake cannot spend money. Every openrouter model id below carries the :free variant and cost 0. Ids, context windows, output ceilings, modalities and zero pricing verified live from https://openrouter.ai/api/v1/models and a real completion per id on 2026-09-22. Jev (TypeSafe judge) is a separate upstream and unaffected by this rule.",
+  "provenance": "Operator rule (2026-09-22): OpenRouter may serve FREE models only. The openrouter upstream declares paidModelsAllowed:false, so any model whose price is unknown or non-zero is refused at dispatch (packages/core/src/compatibility.ts) — a config mistake cannot spend money. Every openrouter model id below carries the :free variant and cost 0. Ids, context windows, output ceilings, modalities and zero pricing verified live from https://openrouter.ai/api/v1/models and a real completion per id on 2026-09-22. Jev (TypeSafe judge) is a separate upstream and unaffected by this rule. Operator rule (2026-09-23): Claude Code and Codex are used through their own subscriptions — Sabi installs hooks into them and never writes a provider base URL, an API key or a model override into either harness, so neither can become per-token spend through Sabi. Guarded by packages/controller/test/hooks.test.ts.",
   "server": { "host": "127.0.0.1", "port": 8787 },
   "upstreams": {
     "openrouter": {
@@ -47,6 +47,7 @@
       "upstream": "ollama",
       "model": "qwen2.5-coder:7b",
       "contextWindow": 32768,
+      "maxOutputTokens": 32768,
       "capabilities": { "inputModalities": ["text"] },
       "cost": { "input": 0, "output": 0 }
     }