token-budget-coach

Token Budget Coach

Use this skill when there's a token-usage signal worth acting on. Sources of signal:

• /sustain:status output (session token tally + cache hit rate).
• The Stop hook's per-phase tokens: line.
• Subjective drift: responses getting vaguer, references to earlier turns failing.

Reading the numbers

The session JSONL records four token streams:

Field	What it is	What to optimize
input_tokens	Fresh prompt tokens (not cached).	Fewer is better. Driven by file reads and tool-result echoes.
cache_creation_input_tokens	New cache writes.	Spikes early; should plateau.
cache_read_input_tokens	Cached prompt reuse.	Higher is better — the prompt cache is paying off.
output_tokens	Generated tokens.	Bound by R4 (concise responses); large output usually means a long search/summary should have been a subagent.

Cache hit rate = cache_read / (cache_read + cache_creation + input). Healthy long sessions: > 70%.

Decision tree

Is cache hit rate < 50% AND session is long?
├─ Probably context churn (lots of new file reads / re-searches).
│  → Fix R5.1: re-read what's already in context instead of re-searching.
│  → Apply R1.1/R1.3: offset+limit reads, not full files.

Is output_tokens of any single turn > ~5K?
├─ You echoed a file or wrote a long explanation.
│  → R4.1/R4.3 — be concise. If it was a search summary, delegate to a subagent next time.

Is input_tokens climbing turn over turn (no plateau)?
├─ Context is filling with raw tool output.
│  → Time to /compact at the next phase boundary.
│  → Or dispatch the remaining work to a subagent (R3.1) so the bulk lives in *its* context.

Is the user's task long-running and we've just finished a phase?
├─ → Apply phase-self-check (the 6 questions).
│  → Especially Q5 (compact?) and Q6 (anything worth saving as memory?).

Concrete moves, in order of cost

1. Free — switch to subagent_type: "Explore" for the next codebase question. Bulk discovery happens in their context, not yours.
2. Free — use model: "haiku" for a simple lookup (R3.3).
3. Cheap — /compact at the next natural phase boundary (don't compact mid-task — Iron-1 friction).
4. Cheap — persist current learnings via memory-write-router then /clear and resume in a fresh session with the search-bridge.
5. Last resort — stop and write a plan, then start a new session against the plan.

Anti-patterns

• Calling /compact reflexively every N turns. Compaction has a cost (it summarises, dropping detail). Use it when the data shows pressure, not on a clock.
• Switching to haiku mid-debugging. R3.3 is for lookups, not reasoning.
• Persisting "what I just did" as memory. That's git-history's job. Persist only durable rules / project facts / external pointers (see memory-write-router).