Add context control for subagents

.opencode/agent/code-reviewer.md

@@ -0,0 +1,35 @@
+---
+description: Reviews code changes for quality and best practices
+mode: subagent
+model: ollama/qwen2.5:1.5b
+temperature: 0.2
+context:
+  mode: filtered
+  includeToolResults: ["read", "edit", "write"]
+  includeFileChanges: true
+  maxMessages: 10
+  maxTokens: 8000
+tools:
+  write: false
+  edit: false
+  bash: false
+---
+
+Review the provided code for:
+
+1. **Code Quality**
+   - Readability and maintainability
+   - Proper naming conventions
+   - Code organization and structure
+
+2. **Best Practices**
+   - Design patterns and architecture
+   - Error handling
+   - Performance considerations
+
+3. **Potential Issues**
+   - Bug risks
+   - Security concerns
+   - Edge cases
+
+Provide specific, actionable feedback with line references when possible.

.opencode/agent/doc-reader.md

@@ -0,0 +1,25 @@
+---
+description: Reads documentation and summarizes key insights
+mode: subagent
+model: ollama/llama3.2:1b
+temperature: 0.1
+context:
+  mode: filtered
+  includeMessageTypes: ["user"]
+  includeToolResults: ["read"]
+  maxTokens: 2000
+tools:
+  write: false
+  edit: false
+  bash: false
+---
+
+Extract key points, APIs, and caveats from documentation.
+Provide a short summary and a Q&A with likely follow-ups.
+
+Focus on:
+
+- Main concepts and their purpose
+- API endpoints and their parameters
+- Common pitfalls and best practices
+- Integration examples

.opencode/agent/log-analyzer.md

@@ -0,0 +1,34 @@
+---
+description: Analyzes log files for errors and patterns
+mode: subagent
+model: ollama/llama3.2:1b
+temperature: 0.0
+context:
+  mode: summary
+  includeFileChanges: false
+  maxTokens: 1000
+tools:
+  write: false
+  edit: false
+  bash: false
+  webfetch: false
+---
+
+Analyze the provided log file and identify:
+
+1. **Errors and Warnings**
+   - Critical errors with timestamps
+   - Warning patterns
+   - Frequency of issues
+
+2. **Performance Metrics**
+   - Response times
+   - Resource usage patterns
+   - Bottlenecks
+
+3. **Anomalies**
+   - Unusual patterns
+   - Spike in errors
+   - Unexpected behavior
+
+Provide a concise summary with actionable insights.

.opencode/agent/quick-search.md

@@ -0,0 +1,23 @@
+---
+description: Fast search for specific code patterns or functions
+mode: subagent
+model: ollama/qwen2.5:1.5b
+temperature: 0.0
+context:
+  mode: none
+tools:
+  write: false
+  edit: false
+  bash: false
+  webfetch: false
+---
+
+Search the codebase for the requested pattern, function, or code snippet.
+
+Return:
+
+- File paths and line numbers
+- Brief context around each match
+- Total number of matches found
+
+Keep responses concise and focused on the search results.

.opencode/agent/test-context-filtered.md

@@ -0,0 +1,16 @@
+---
+description: Test context control with filtered mode
+mode: subagent
+model: anthropic/claude-sonnet-4
+context:
+  mode: filtered
+  includeToolResults: ["read"]
+  maxTokens: 2000
+tools:
+  write: false
+  edit: false
+  bash: false
+---
+
+Summarize what you see very concisely.
+List only the key points.

.opencode/agent/test-context-none.md

@@ -0,0 +1,13 @@
+---
+description: Test with no context (baseline)
+mode: subagent
+model: anthropic/claude-sonnet-4
+context:
+  mode: none
+tools:
+  write: false
+  edit: false
+  bash: false
+---
+
+Answer the question concisely.

.opencode/agent/test-context-summary.md

@@ -0,0 +1,15 @@
+---
+description: Test context control with summary mode
+mode: subagent
+model: anthropic/claude-sonnet-4
+context:
+  mode: summary
+  includeFileChanges: true
+  maxTokens: 1000
+tools:
+  write: false
+  edit: false
+  bash: false
+---
+
+Based on the session summary, answer the question.

.opencode/agent/test-debug.md

@@ -0,0 +1,19 @@
+---
+description: Debug agent to test context control
+mode: subagent
+model: anthropic/claude-sonnet-4
+context:
+  mode: filtered
+  includeToolResults: ["read", "list", "glob"]
+  includeMessageTypes: ["user"]
+  maxMessages: 5
+  maxTokens: 3000
+tools:
+  write: false
+  edit: false
+  bash: false
+---
+
+You are a debug agent testing context control.
+Just respond with "I received context" if you got any context, or "No context" if not.
+Then list what you can see.

AGENTS.md

@@ -15,6 +15,44 @@
 
 - To test opencode in the `packages/opencode` directory you can run `bun dev`
 
+## Context Control for Subagents
+
+Subagents can be configured with context filters to control what parent session context they receive. This enables better performance for lightweight local models.
+
+### Configuration
+
+Add a `context` field to your agent configuration:
+
+```yaml
+context:
+  mode: filtered
+  maxTokens: 2000
+  includeToolResults: ["read", "edit"]
+  includeMessageTypes: ["user"]
+```
+
+### Modes
+
+- `none` (default) - No parent context
+- `summary` - Compact summary with message counts and tool usage
+- `filtered` - Selective inclusion by message type and tool results
+- `full` - All context with token/message limits
+
+### Example
+
+```yaml
+---
+mode: subagent
+model: ollama/llama3.2:1b
+context:
+  mode: filtered
+  includeToolResults: ["read"]
+  maxTokens: 2000
+---
+```
+
+See `.opencode/agent/` for example configurations.
+
 ## Tool Calling
 
 - ALWAYS USE PARALLEL TOOLS WHEN APPLICABLE. Here is an example illustrating how to execute 3 parallel file reads in this chat environment:

packages/opencode/src/agent/agent.ts

@@ -33,6 +33,7 @@ export namespace Agent {
       prompt: z.string().optional(),
       tools: z.record(z.string(), z.boolean()),
       options: z.record(z.string(), z.any()),
+      context: Config.ContextFilter.optional(),
     })
     .meta({
       ref: "Agent",
@@ -148,7 +149,20 @@ export namespace Agent {
           tools: {},
           builtIn: false,
         }
-      const { name, model, prompt, tools, description, temperature, top_p, mode, permission, color, ...extra } = value
+      const {
+        name,
+        model,
+        prompt,
+        tools,
+        description,
+        temperature,
+        top_p,
+        mode,
+        permission,
+        color,
+        context,
+        ...extra
+      } = value
       item.options = {
         ...item.options,
         ...extra,
@@ -169,6 +183,7 @@ export namespace Agent {
       if (top_p != undefined) item.topP = top_p
       if (mode) item.mode = mode
       if (color) item.color = color
+      if (context) item.context = context
       // just here for consistency & to prevent it from being added as an option
       if (name) item.name = name
 

packages/opencode/src/config/config.ts

@@ -349,6 +349,43 @@ export namespace Config {
   })
   export type Command = z.infer<typeof Command>
 
+  export const ContextFilter = z
+    .object({
+      mode: z
+        .enum(["none", "summary", "filtered", "full"])
+        .optional()
+        .default("none")
+        .describe(
+          "Context mode: 'none' (no parent context), 'summary' (compact summary), 'filtered' (selective context), 'full' (all parent context)",
+        ),
+      maxTokens: z
+        .number()
+        .int()
+        .positive()
+        .optional()
+        .describe("Maximum tokens to include from parent session context"),
+      maxMessages: z
+        .number()
+        .int()
+        .positive()
+        .optional()
+        .describe("Maximum number of messages to include from parent session"),
+      includeToolResults: z
+        .array(z.string())
+        .optional()
+        .describe("Which tool results to include (e.g., ['read', 'edit', 'bash'])"),
+      includeMessageTypes: z
+        .array(z.enum(["user", "assistant"]))
+        .optional()
+        .describe("Which message types to include from parent session"),
+      includeFileChanges: z.boolean().optional().describe("Include file changes made in parent session"),
+    })
+    .optional()
+    .meta({
+      ref: "ContextFilterConfig",
+    })
+  export type ContextFilter = z.infer<typeof ContextFilter>
+
   export const Agent = z
     .object({
       model: z.string().optional(),
@@ -364,6 +401,7 @@ export namespace Config {
         .regex(/^#[0-9a-fA-F]{6}$/, "Invalid hex color format")
         .optional()
         .describe("Hex color code for the agent (e.g., #FF5733)"),
+      context: ContextFilter.describe("Control what context from the parent session is passed to this subagent"),
       permission: z
         .object({
           edit: Permission.optional(),