Tool descriptions consume excessive tokens in system prompt

[MartinWie]

Problem

The tool description files (bash.txt, todowrite.txt, task.txt) are included in the system prompt for every message. Currently they are verbose and consume ~3,000-4,000 tokens per message unnecessarily.

Proposal

Condense the tool descriptions while preserving:

• All safety rules (git safety, permission handling)
• Core usage instructions
• Essential examples

Remove:

• Redundant explanations
• Lengthy examples that don't add value
• Verbose formatting

Expected Impact

~69% reduction in tool description token usage, reducing per-message costs without changing tool behavior.

[4rzumanyan]

I am also experiencing this same issue. In fact, I have observed an even higher rate of token consumption, approximately 8,000 to 10,000 tokens. To troubleshoot, I have removed all files from the OpenCode configurations and disabled all MCPs, skills, plugins, and agents, leaving only the standard, default configuration. Even in this minimal setup, a simple "Hello World" input consumes around 10,000 tokens.

[DrDexter6000]

Ran into this same problem and did some exact measurements. Downloaded all 16 tool .txt files from packages/opencode/src/tool/ and counted characters:

bash.txt is the worst offender at 9,288 chars (~2,654 tokens). What's wild is that 63% of it (5,874 chars / ~1,678 tokens) is git commit and PR creation instructions that have nothing to do with actually executing bash commands.

The key insight I found: tool descriptions are NOT cacheable, but system prompt text IS (Anthropic prompt caching, OpenAI cached prefix). Those git/PR instructions get paid for on every single API turn, even when the user never touches git.

Full breakdown:

• bash.txt: 9,288 chars — 63% git/PR instructions
• task.txt: 3,799 chars
• multiedit.txt: 2,406 chars
• edit.txt: 1,369 chars
• apply_patch.txt: 1,092 chars
• read.txt: 1,158 chars
• lsp.txt: 1,040 chars
• websearch.txt: 976 chars
• batch.txt: 998 chars
• codesearch.txt: 802 chars
• webfetch.txt: 750 chars
• question.txt: 657 chars
• grep.txt: 689 chars
• write.txt: 623 chars
• glob.txt: 545 chars
• ls.txt: 314 chars

Total: ~26,500 chars (~7,573 tokens) for built-in tool descriptions alone. In my setup (5 MCP servers + plugin), a fresh session hits ~40K tokens before the first user message.

I opened #21345 with a specific proposal to move the git/PR sections out of bash.txt and into the system prompt where they'd be cacheable. That alone would save ~1,678 tokens per request.

[cranknet]

single prompt like: Say Hi
costs about 30000 tokens.

[MartinWie]

Opened #24202 addressing this.

TL;DR: condenses all 17 tool description files (packages/opencode/src/tool/*.txt) by ~66% — 33,163 → 11,302 chars, estimated ~6,250 tokens/message saved.

Preserves all git safety rules, placeholders (${os}, ${shell}, ${chaining}, ${maxLines}, ${maxBytes}, {{year}}), the apply_patch format block, 9 LSP ops, 4 todo states, and commit/PR workflow steps.

Largest cuts: bash.txt 70%, todowrite.txt 81%, task.txt 72%. Behavioral anchor examples retained where model behavior depends on concrete patterns (workdir good/bad pairs, Edit line-prefix example, todowrite use/skip examples).

Complementary to #21345 (@DrDexter6000's proposal to move git/PR instructions out of bash.txt into the cacheable system prompt) — the two changes stack.

[MartinWie]

For anyone wanting to see where their context is going, PR #24210 adds a /context command that surfaces:

• authoritative usage from the last assistant turn (input / output / cache)
• per-section breakdown: system prompt segments, skills, rules / memory, agent, built-in tools, MCP tools (grouped per server)
• messages currently in context vs full pre-compaction transcript (with rolled-up count + lifetime cost)

Useful when you're trying to figure out what's eating the window before deciding to compact, prune skills, or trim MCP servers.