feat: add configurable context_window action input

action.yml

@@ -12,6 +12,9 @@ inputs:
   model:
     description: "Ollama model to use. Default: qwen3.6:35b"
     required: false
+  context_window:
+    description: "Ollama context window size in tokens. Default: 262144"
+    required: false
   cwd:
     description: "Working directory for the agent (defaults to GITHUB_WORKSPACE)"
     required: false

agents/ollama.ts

@@ -119,6 +119,7 @@ async function runOllamaLoop(ctx: AgentRunContext): Promise<AgentResult> {
   }
 
   const model = ctx.model ?? process.env.OLLAMA_MODEL ?? DEFAULT_MODEL;
+  const numCtx = ctx.payload.contextWindow ?? 262144;
 
   log.info(`» connecting to Ollama at ${ollamaHost}, model ${model}`);
 
@@ -170,7 +171,7 @@ async function runOllamaLoop(ctx: AgentRunContext): Promise<AgentResult> {
           tools,
           keep_alive: -1,
           think: false,
-          options: { num_ctx: 262144, temperature: 0.1 },
+          options: { num_ctx: numCtx, temperature: 0.1 },
         }),
         {
           delaysMs: [3_000, 8_000],
@@ -196,11 +197,11 @@ async function runOllamaLoop(ctx: AgentRunContext): Promise<AgentResult> {
 
     if (promptTokens !== undefined) {
       const total = promptTokens + (evalTokens ?? 0);
-      const pct = Math.round((total / 262144) * 100);
+      const pct = Math.round((total / numCtx) * 100);
       log.info(
-        `  context: ${promptTokens} prompt + ${evalTokens ?? 0} eval = ${total} tokens (${pct}% of limit)`,
+        `  context: ${promptTokens} prompt + ${evalTokens ?? 0} eval = ${total} tokens (${pct}% of ${numCtx} limit)`,
       );
-      if (promptTokens > 200_000) {
+      if (promptTokens > numCtx * 0.77) {
         messages = pruneToolMessages(messages);
       }
     }

utils/payload.ts

@@ -11,6 +11,7 @@ const PushPermissionInput = type.enumerated("disabled", "restricted", "enabled")
 export const Inputs = type({
   prompt: "string",
   "model?": type.string.or("undefined"),
+  "context_window?": type.number.or("undefined"),
   "timeout?": type.string.or("undefined"),
   "push?": PushPermissionInput.or("undefined"),
   "shell?": ShellPermissionInput.or("undefined"),
@@ -58,8 +59,10 @@ export function resolvePromptInput(): ResolvedPromptInput {
 }
 
 function resolveNonPromptInputs() {
+  const contextWindowRaw = core.getInput("context_window");
   return Inputs.omit("prompt").assert({
     model: core.getInput("model") || undefined,
+    context_window: contextWindowRaw ? parseInt(contextWindowRaw, 10) : undefined,
     timeout: core.getInput("timeout") || undefined,
     cwd: core.getInput("cwd") || undefined,
     push: (core.getInput("push") as "disabled" | "restricted" | "enabled") || undefined,
@@ -93,6 +96,12 @@ export function resolvePayload(
   const model =
     str(jsonPayload, "model") ?? inputs.model ?? repoSettings.model ?? process.env.OLLAMA_MODEL ?? undefined;
 
+  const contextWindowStr = str(jsonPayload, "context_window");
+  const contextWindow =
+    (contextWindowStr ? parseInt(contextWindowStr, 10) : undefined) ??
+    inputs.context_window ??
+    undefined;
+
   const isNonCollaborator = !isCollaborator(event);
   const repoShell = repoSettings.shell ?? "restricted";
   const inputShell = inputs.shell;
@@ -110,6 +119,7 @@ export function resolvePayload(
   return {
     "~shockbot": true as const,
     model,
+    contextWindow,
     prompt,
     triggerer:
       str(jsonPayload, "triggerer") ?? process.env.GITHUB_ACTOR ?? undefined,