Token Warning Interpretation
When to Use
Use when the harness emits a warning like "Read results using 881.4k tokens (88%) → save ~264.4k" and you're unsure whether to change your reading strategy.
Decision
| The warning means | The warning does NOT mean |
|---|---|
| Don't re-read the same file repeatedly | Always prefer grep |
| Don't read a 2000-line file to check one constant | Never do full reads |
| Don't read generated code, lockfiles, or vendor code that grep could answer | Documentation/audit tasks are wasteful |
| Don't read a file whole when you only need one section AND know the structure | Full reads are always avoidable |
Pattern
The warning flags wasteful patterns:
- Re-reading the same file multiple times in a session → trust prior reads + targeted re-reads
- Reading a huge file to check one constant → use
offset/limit - Reading generated/vendor/lockfile code → use
Grepfor the one symbol you need - Reading whole when you already know the structure and need one section → use
offset/limit
It does NOT flag: a thorough read of 15 small source files to document a module. That is the correct cost for Type B, not waste.
Common Mistakes
- Wrong: Switching a Type B task to grep-first because the harness warned about token usage → Right: Verify you weren't doing one of the 4 wasteful patterns; if not, the warning doesn't apply to your strategy
- Wrong: Treating "88% context used" as a hard stop mid-audit → Right: If remaining work would push over budget, dispatch remaining exploration to a subagent
- Wrong: Using the warning to justify skipping parent class reads → Right: Missing parent class content causes audit/doc failures that cost more than tokens
See Also
- Task Type A — Targeted Edits — where token warnings usually apply
- Type C — Large Exploration — when real budget constraint needs delegation
- Reference: https://code.claude.com/docs/en/best-practices