From 50ad3f4c802fbf30c73b81db5307bbcac892c1e4 Mon Sep 17 00:00:00 2001
From: JoshRob297 <99092299+JoshRob297@users.noreply.github.com>
Date: Mon, 19 Oct 2026 22:14:08 -0600
Subject: [PATCH] fix: raise the declared output limit from 16k to 32k (v1.0.4)

OpenCode forwards limit.output as max_completion_tokens, so the hardcoded 16,384
capped every Command Code response regardless of what the model supports. The
DeepSeek V4.1 Flash 1017 build declares 393,216 output tokens.

- Default is now 32,768, the highest value safe across the whole catalog: the
  minimum declared output in models.dev's commandcode provider is exactly 32,768.
- COMMANDCODE_OUTPUT_TOKENS overrides it per process (verified: 131072 reaches the
  wire as max_completion_tokens).
- The 28 static model entries in opencode.json carried output: 16384, which won over
  the plugin; the limits are now owned by the plugin alone.
- Verified end to end with a local echo provider: default 32768, override 131072.
---
 CHANGELOG.md | 13 +++++++++++++
 README.es.md |  1 +
 README.md    |  1 +
 package.json |  2 +-
 src/index.ts | 20 +++++++++++++++++++-
 5 files changed, 35 insertions(+), 2 deletions(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index f8c223e..c062904 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -5,6 +5,19 @@ Format: [Keep a Changelog](https://keepachangelog.com/en/1.1.0/) - SemVer.
 
 ## [Unreleased]
 
+## [1.0.4] - 2026-10-20
+
+### Fixed
+
+- The declared output limit was hardcoded to 16,384 tokens for every model. OpenCode
+  forwards `limit.output` as `max_completion_tokens`, so every response was capped at
+  16k even though the catalog supports far more (the DeepSeek V4.1 Flash 1017 build
+  declares 393,216). The default is now 32,768, the highest value that is safe across the
+  whole catalog (the minimum declared output in models.dev's commandcode provider is
+  exactly 32,768), and `COMMANDCODE_OUTPUT_TOKENS` overrides it per process.
+- The static model entries in `opencode.json` carried `output: 16384`, which took
+  precedence over the plugin. Those limits are now owned by the plugin alone.
+
 ## [1.0.3] - 2026-10-19
 
 ### Added
diff --git a/README.es.md b/README.es.md
index 526f060..b2fbce1 100644
--- a/README.es.md
+++ b/README.es.md
@@ -210,6 +210,7 @@ Cada ajuste es una variable de entorno con default seguro. Referencia completa:
 | `CC_NO_CONNECTION_CLOSE` | `0` | Poner en `1` para dejar de inyectar `Connection: close` |
 | `COMMANDCODE_STATE_DIR` | `$XDG_STATE_HOME/opencode` | Donde viven los logs y la cache de modelos |
 | `COMMANDCODE_LIVE_TESTS` | `0` | Poner en `1` para correr la suite de tests con red |
+| `COMMANDCODE_OUTPUT_TOKENS` | `32768` | Limite de tokens de salida declarado por modelo (OpenCode lo envia como `max_completion_tokens`). Subilo si un modelo soporta mas |
 
 Kill switch en runtime para la higiene de conexion: crea un archivo centinela llamado
 `.disable-connection-close` dentro del directorio del plugin.
diff --git a/README.md b/README.md
index 047dbc4..32d462c 100644
--- a/README.md
+++ b/README.md
@@ -204,6 +204,7 @@ Every setting is an environment variable with a safe default. Full reference:
 | `CC_NO_CONNECTION_CLOSE` | `0` | Set to `1` to stop injecting `Connection: close` |
 | `COMMANDCODE_STATE_DIR` | `$XDG_STATE_HOME/opencode` | Where logs and the model cache live |
 | `COMMANDCODE_LIVE_TESTS` | `0` | Set to `1` to run the network test suite |
+| `COMMANDCODE_OUTPUT_TOKENS` | `32768` | Output token limit declared per model (OpenCode sends it as `max_completion_tokens`). Raise it for models that support more |
 
 Runtime kill switch for connection hygiene: create a sentinel file named
 `.disable-connection-close` inside the plugin directory.
diff --git a/package.json b/package.json
index 14d1b27..c138fef 100644
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "opencode-commandcode",
-  "version": "1.0.3",
+  "version": "1.0.4",
   "description": "Production-grade OpenCode plugin for Command Code: dynamic model discovery, stability hooks and quota telemetry",
   "main": "index.ts",
   "type": "module",
diff --git a/src/index.ts b/src/index.ts
index 38f5650..1c7c61f 100644
--- a/src/index.ts
+++ b/src/index.ts
@@ -18,6 +18,24 @@ import { describeRetryEvent } from "./telemetry.ts";
 import { buildReleaseDate, freeCostEntry } from "./catalog.ts";
 
 export const PROVIDER_ID = "commandcode";
+
+/**
+ * Max completion tokens declared for every Command Code model.
+ *
+ * OpenCode forwards `limit.output` as `max_completion_tokens`, so this value caps
+ * every response. The catalog API only exposes the context window, but every model
+ * in models.dev's commandcode provider declares an output limit of at least 32,768
+ * (the minimum is exactly that, for Qwen3.8-27B and Ling 3.1 Flash), so 32,768 is
+ * the highest value that is safe across the whole catalog.
+ *
+ * Override with COMMANDCODE_OUTPUT_TOKENS when a specific model needs more (the
+ * DeepSeek V4.1 Flash 1017 build supports 393,216).
+ */
+function outputLimit(): number {
+  const raw = Number(process.env.COMMANDCODE_OUTPUT_TOKENS);
+  if (Number.isFinite(raw) && raw >= 1024) return Math.floor(raw);
+  return 32768;
+}
 const WATCHDOG_TIMEOUT_MS = 60_000;
 
 function logPath(name: string): string {
@@ -82,7 +100,7 @@ function applyModelCatalog(editor: any, models: any[]): { registered: number; sk
         draft.enabled = true;
         draft.limit = {
           context: model.context_length || 200000,
-          output: 16384,
+          output: outputLimit(),
         };
         draft.capabilities = {
           tools: true,