| name | context-budget |
| description | Bound repository context reads and semantic-search output without weakening source, test, or evidence verification. |
Context Budget
Goal
Limit reads to the minimum high-signal surface required to implement the task correctly, then widen only when evidence is missing.
Workflow
- Read
AGENTS.md.
- Identify exactly which module and command path is affected.
- Open one source file, one test file, and one skill that governs the area.
- Implement with tight scope.
- If uncertainty remains, open one more implementation file and one more test file only.
Hard limits
- Avoid bulk reading unrelated directories and generated artifacts.
- Avoid opening entire modules for tangential concerns.
- Prefer
rg for symbol search over opening large files.
- Keep semantic routing at the five-result default; ten results is the hard
maximum. Refine the query or use path/language filters before widening.
- Treat measured UTF-8 output bytes as a context-volume proxy only. Do not
infer or publish token savings without a tokenizer-specific measurement.
Exit condition
Before coding the next step, verify that remaining assumptions are either covered by
existing tests or addressed by newly added tests.