[FEATURE]: Code Mode needs a token budget and a per-tool breakdown #58708

I measured on my own machine before asking for anything. OpenCode v2.1.4, Linux, one MCP server with 207 tools.

rea is one MCP server I have configured (a reverse-engineering toolkit). Its tools/list response has 207 tools. Dumped as JSON and counted with tiktoken o200k_base, those definitions are 3,423,785 characters and 1,033,360 tokens. One tool, address_name, is 2,721 tokens by itself.

Traditional tool calling injects every schema into every request, so this single server would spend 1.03M tokens before I type anything. That fits no 200k window and eats a 1M window whole. With search(), the tool, the system prompt and the two signatures the task actually needed came to about 6.6k tokens.

Then I ran the same task twice in two real sessions on this machine. Fetch 8 URLs, report which ones fail or mention "token". Session A used one tool call per URL and returned each body whole. Session B ran one program that returned a filtered array. Same model, same day, billing read from session_v2.

27.4x fewer tokens, 13.6x cheaper. The dollar gap is smaller than the token gap because the repeated prefix hits prompt cache.

A also got 3 of the 8 answers wrong. Its bodies were truncated at 50k characters, which cut off the word "token" on three pages. B searched the full text inside the sandbox and got all 8 right. Truncating observations changed three answers and the price.

A synthetic run of the same task counts 114,320 tokens for one call per URL against 2,436 for one program, 46.9x apart, and 24.5x apart once the prefix is priced at a 10% cache rate. That cost grows with N² because every observation is resent in every later request: at the 2,752-token average, 20 URLs cost 578k tokens and 50 cost 3.5M.

None of this is news to Cloudflare or Anthropic. Kenton Varda and Sunil Pai published the same approach in September 2025 (https://blog.cloudflare.com/code-mode/), and Anthropic shipped programmatic tool calling in November 2025 with a reported 37% average drop on research tasks. OpenCode already ships the mechanism; a default policy and a per-tool breakdown are missing.

I read #8625 (MCP search, shipped via #12520) and #53121 (per-step cost, open) before writing this. Both sit below what I am asking for.

What I would like to see, and the questions behind each item

1. A default for large MCP servers. Above roughly 20 tools, should search + execute become the default exposure instead of every schema? Right now a user has to know this, and the fallback degrades any model with a smaller window. Would mcp.servers.<name>.mode: "search" | "direct" fit the config, or is there a reason to keep schemas inline?

2. Per-tool and per-schema tokens. #53121 asks for per-step and per-turn cost. I want the split inside one step: system prompt, tool schemas, each tool observation, reasoning, output. Without it I cannot tell whether a session burned 40k on a page fetch or on 180 unused schemas. Is that already stored per message, and would a TUI view or an opencode api field be acceptable?

3. Output caps that spill instead of truncate. shell and webfetch already keep the full output in managed storage. read pages at 50 KiB and MCP results come back whole. Could a configurable cap (say 4k tokens per tool result, full output in a file, path returned) become the default for every tool, MCP included?

4. Does execute reach built-in tools? #57720 and #57627 read like this is still settling. If execute could call read, grep, edit and shell as functions, a six-step file task becomes one inference pass. Is that planned, or does Code Mode stay MCP-only on purpose?

5. Compaction accounting. One Code Mode step can hide 50 underlying calls. Does compaction weigh it as one message when it decides what to keep? If yes, a summary can throw away everything the program found. If no, large code steps can exceed compaction.keep.tokens on their own.

6. Sandbox and audit. My execute has fetch but no require('fs') (#56811). What are the actual limits on network, filesystem, memory and time, and can a user inspect or replay the code that ran? This decides how much I trust it as the default path.

What I would like from the community

If you run an MCP server with a lot of tools, please post three numbers: how many tools it exposes, what its tools/list payload costs in tokens, and your session cost before and after switching to search + execute. MCP servers with 50+ tools are the interesting ones.

I can turn the benchmark into opencode mcp cost <name> if you want it in the repo, or ship it as a plugin if core is not the place for it.

Benchmark, raw numbers and both measurement scripts: https://gist.github.com/RicSchonfelder/268a51cb86395db36eaac0aa018686cd