manta-token-economy

manta-token-economy

Every turn re-reads the cached prefix. Cache reads count fully against the session quota. Task cost ≈ N_turns × (initial_prefix + cumulative additions).

Discipline:

• Batch independent tool calls in one turn. Issue multiple Read/Grep/Bash in parallel within a single message, not back-to-back turns. One turn ≫ three turns.

• Compound shell commands. cmd1 && cmd2 && cmd3 (or ; for non-dependent) for multi-step inspection — never sequential single-command bash turns.

• Don't re-Read after Edit. Edit validates old_string; trust the prior Read. Re-reading the same file adds thousands of tokens to every subsequent turn's cache_read.

• Use offset/limit on large Reads. A 36KB file is ~9k tokens × every later turn. Read only the relevant section.

• No --output json | head / huge JSON dumps from bash. Use --output table for inspection; project specific fields with --jq/python3 -c when JSON is needed.

• Compound git ops. git status && git diff --stat && git log -3 in one bash call, not three turns.

• Prefer Grep over Bash for content search. Don't run grep or rg via Bash when the Grep tool exists — it's faster and returns structured results without parsing overhead.

• Batch file reads when inspecting a directory. If you need to understand the structure of src/main/, src/renderer/, etc., read the directory listing once and use it to plan targeted Reads — don't read files one at a time in separate turns.

• Avoid redundant verification turns. If you just ran npm run typecheck && npm test and both passed, don't run them again to "confirm" before posting results. Trust the output you just captured.