Analyzing and optimizing OpenCode agent injected tokens

Key conclusion

OpenCode takes about 110,500 tokens as input on every request. Of that, conversation history is 91.5%, but the fixed injection (system prompt + rules + tool schema) is the base reused every time. Reducing the fixed injection raises the prompt cache hit rate and lowers total cost.

OpenCode has a hardcoded binary, so the system prompt/tools are hard to reduce directly. Instead, cleaning up rule files and MCP servers is the key saving point.

1. Overall injection map

[System prompt]   ~5,865 tok  (5.3%)  <- hardcoded in the binary
[Rule files]      ~1,612 tok  (1.5%)  <- .clinerules + .instructions.md
[Tool schema]     ~2,400 tok  (2.2%)  <- native + MCP
[Conversation history] ~101,200 tok (91.5%) <- accumulated in the session
-------------------------------------
Total            ~110,500 tok (100%)

1-1. System prompt (~5,865 tok / ~23,458 chars)

Hardcoded in the OpenCode binary (~/.opencode/bin/opencode, 184MB). Contents:

• Role definition: "You are opencode, an interactive CLI tool..."
• Tone/style: concise, markdown, no emojis, answers within 4 lines
• Tool policy: prefer the Task tool, encourage parallel calls
• Code style: search priority, file_path:line_number reference pattern
• Task management: TodoWrite usage rules

This script cannot be edited directly — it cannot be changed without rebuilding the binary.

1-2. Rule files (~1,612 tok)

1-3. Native tools (~12)

Tools built into OpenCode. Usable without MCP.

1-4. MCP tools

Tools provided by MCP servers defined in opencode.json.

Direct connection (opencode.json):

Via proxy (mcp-proxy.json, lazy mode):

Lazy mode behavior: only the bridge tools (tool_search, tool_describe, tool_call) are exposed up front. The actual server tools are loaded at the point of use.

1-5. tool-slim plugin

.config/opencode/plugins/tool-slim.js truncates tool descriptions:

• Tool description: limited to 80 chars
• Parameter description: 60 chars (top level), 40 chars (nested)

-> The clue that distinguishes tools disappears, which can cause the model to fail at tool selection.

2. Real measurements (production environment)

The figures below are real values extracted from the opencode DB for recent responses.

• With prompt caching, the system prompt + tools are reused via cache_read (~106,752 tok cache hit)
• Tokenizer: cl100k_base (OpenAI-compatible)
• Korean: ~1.1 chars/token, English: ~4.5 chars/token

3. How to optimize

3-1. Diet the rule files (applies immediately)

.clinerules and .instructions.md can be edited directly by the user.

Current composition:

• .clinerules (1,860 bytes): auto-run agent rules + web search + weather + chart
• .instructions.md (551 bytes): local AI environment info

Saving strategies:

Cautions:

• Deleting the MTP server address (127.0.0.1:18081) in .instructions.md breaks the server connection
• Deleting the chart generation rules makes the chart-render skill unusable
• A cache reset (session restart) is needed after rule changes

3-2. Clean up MCP servers (the biggest saving point)

Disabling unused servers out of the 7 in mcp-proxy.json reduces the tool schema.

Current MCP tool schema estimate:

Candidates for disabling:

How to disable: change the server block in mcp-proxy.json to "enabled": false.

Cautions:

• Disabling playwright makes browser automation impossible
• Disabling filesystem makes local file access via MCP impossible
• Since it is lazy mode, tools load only at use time, but the tool list is injected every time
• Always check the usage history of that server's tools before disabling

3-3. Adjust the tool-slim plugin

Increasing the truncation length in .config/opencode/plugins/tool-slim.js raises tool discrimination.

// current settings
const MAX_DESC_CHARS = 80;      // tool description
const MAX_PARAM_CHARS = 60;     // parameter description (top level)
const MAX_NESTED_PARAM = 40;    // nested parameter

// proposal: widen the tool description to 120 chars
const MAX_DESC_CHARS = 120;

Effect: keywords in the tool description survive, improving the model's tool-selection accuracy.

Cautions:

• Making the truncation length too long increases token consumption
• 120-150 chars is the appropriate range
• After changing, verify the truncation behavior through the debug log

3-4. Manage conversation history

Since history is 91.5%, history management has the biggest impact on total tokens.

compaction settings (opencode.json):

"compaction": {
  "auto": true
}

• auto: true: when the context window fills, automatically summarize old turns
• Summarization calls MiMo (an external model), so it costs money

History saving strategies:

Key: compaction does not "delete content" but "replaces old turns with a summary." The total content does not shrink, but the injected token count does.

3-5. Direct MCP vs proxy comparison

The current structure goes through the proxy server; switching to a direct connection can reduce overhead.

Recommendation: disabling unused proxy servers takes priority over switching to direct connections.

4. Cautions

4-1. Required checks before changing

1. Check the list of MCP tools currently in use: review each MCP server tool's usage history in an opencode session
2. Back up rule files: back up .clinerules and .instructions.md before changing
3. Test: verify normal operation with a simple task after the change

4-2. Required actions after changing

1. Restart the session: rule/MCP changes only take effect in a new session
2. Reset the cache: the prompt cache may remember the old settings
3. Monitor: watch for an increase in the tool-selection failure rate

4-3. How to recover

4-4. OpenCode vs Hermes comparison

5. Checklist

• [ ] Check and clean up unused rules in .clinerules
• [ ] Delete unnecessary environment info from .instructions.md
• [ ] Disable unused servers in mcp-proxy.json
• [ ] Review the tool-slim.js truncation length (80 to 120 chars)
• [ ] Verify normal operation after restarting the session
• [ ] Monitor the tool-selection failure rate

Measured in the operator's environment (opencode v1.x, cl100k_base tokenizer, Linux). Figures may differ by environment.