| name | cost-discipline |
| description | Use to enforce session cost limits, prevent verbose output, avoid retry loops, and batch operations efficiently. Triggers on: $10 limit, token cost, session budget, cost cap, retry loop, verbose output, redundant search, spend tracking. |
Cost Discipline
Role
Own session economics as measurable constraint enforcement, not aspirational frugality.
Working Mode
Map -> Separate evidence from hypothesis -> Smallest intervention -> Validate.
Focus Areas
- Session budget cap -- $10/session hard limit, STOP and confirm before exceeding
- Token efficiency -- prefer targeted reads over full file dumps
- Batch operations -- group related API calls, never loop one-at-a-time
- Retry discipline -- max 1 retry per approach, diagnose before retrying
- Search efficiency -- targeted grep/glob before broad Agent searches
- Output verbosity -- concise responses; skip preamble and trailing summaries
- Tool selection -- use dedicated tools (Grep, Glob, Read) not Bash grep/find
- Spend tracking -- report cumulative cost at natural session milestones
Quality Gates
- verify: Session cost stays under $10 (track via token counts or CC cost meter)
- confirm: Each retry is preceded by diagnosis, not blind repetition
- check: Batch ops used when 3+ similar items need processing
- ensure: No full file reads when only a section is needed (use offset/limit)
- call_out: Flag if approaching $8 threshold (80% of cap)
Output Format
Single status line at session milestones:
Cost check: $X.XX / $10.00 (XX%) โ [current_task] โ [items_remaining]
Constraints
- STOP and confirm with user if next action will push session over $10
- Never retry a failed approach more than once without diagnosing root cause
- Never search with Agent tool if Grep/Glob can find it in 1-2 queries
- Never read entire files when offset/limit suffices
- $0 cost operations (file reads, git log) are always preferred over API calls
Guard Rail
Do not initiate any action estimated to cost > $2 without confirming total session spend first.
