rlaope/oh-my-hermes / .claude/agent-skills/omh-context-budget-review
omh-context-budget-review
Verified[omh] Context window or token budget at risk: plan compact context…
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 8.8 KB |
| references/cache-placement.md | markdown | 1.8 KB |
references/cache-placement.md
View on GitHub# Cache Placement Discipline
Every major serving stack caches prompt prefixes by exact bytes: Anthropic prefix caching (explicit breakpoints, discounted reads), OpenAI automatic prefix caching, Gemini implicit caching, DeepSeek context caching. A single changed byte at position N re-bills everything at and after N. OMH never calls a provider; this card disciplines the text OMH generates and the guidance it prepares for the host.
## Placement rules
1. **Stable prefix ordering.** Assemble every instruction surface in a fixed section order, most-stable content first. Regeneration must be byte-stable: same inputs, same bytes.
2. **Volatile bytes never above the fold.** Dates, token counts, git state, status lines, and per-session values never belong in files loaded at session start; they ride the first user turn or the message tail.
3. **Changes travel as appended messages.** Mid-run skill, state, or instruction changes are appended conversation messages, never edits to the system prompt or to a session-start file — a mid-run system-prompt rewrite rebuilds the whole cache (the failure mode behind NousResearch/hermes-agent#13631 and #4319).
4. **Tool surface stays stable mid-session.** Choose the tool set at session start; avoid mid-session connect/disconnect of tool servers; prefer deferred tool loading where the host supports it; serialize tool payloads deterministically (sorted keys).
5. **Fan-outs share a byte-identical preamble.** Sibling prompts lead with the same shared bytes, unit-specific content appended after; stagger dispatch so the first request writes the cache the siblings read.
## Evidence boundary
Cache hit and creation counters are provider or host telemetry. Never claim a hit rate, a saving, or "cache-safe" as observed fact without the host's usage counters; prepared placement is prepared_not_observed.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| receiving-code-reviewobra/superpowersUse when receiving code review feedback, before implementing suggestions, especially if feedback seems unclear or technically questionable - requires technical rigor and verification, not performative agreement or blind implementation | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| requesting-code-reviewobra/superpowersUse when completing tasks, implementing major features, or before merging to verify work meets requirements | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| using-git-worktreesobra/superpowersUse when starting feature work that needs isolation from current workspace or before executing implementation plans - ensures an isolated workspace exists via native tools or git worktree fallback | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| verification-before-completionobra/superpowersUse when about to claim work is complete, fixed, or passing, before committing or creating PRs - requires running verification commands and confirming output before making any success claims; evidence before assertions always | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |