StackMap
Subscribe
Explore / prompt-cache-skills
OnlyTerp

prompt-cache-skills

13 drop-in skills that fix prompt-caching bugs in OSS agent harnesses (Cline, Roo, Continue, OpenCode, Aider) — point your coding agent at the repo, it patches and verifies on the wire.

107 7 Python NOASSERTIONupdated 6 days ago
Curator's take

If you run Cline, Roo Code, Continue, Aider or OpenCode against your own API keys, this likely pays for itself in an afternoon: audited findings with file:line citations, each fix a 5-15 line diff, and a wire-level verify step (check_cache.py fires the request cold+warm and diffs the cache_* token fields — trust but verify is built in). Skip it on Claude Code or Codex CLI, both audited as already-correct reference implementations, and on managed backends (Devin, Windsurf, Antigravity) where nothing is patchable. Audits are dated 2026-05-27; harnesses move fast, so re-run the verify step after upgrading yours.

Mapped by ShipWithAI editors · links verified
README.md

prompt-cache-skills — Drop-in prompt-caching fixes for every LLM agent harness

Get Started 13 Skills 13 Completed Audits 10x Savings

License Audit Date PRs Welcome


Most popular OSS agent harnesses (Cline, Roo Code, Continue, OpenCode, Aider) leave 30-90% off your API bill on the table because their prompt-caching code is subtly wrong, off-by-default, or just missing for some providers.

This repo is a set of drop-in skills that any AI coding agent (Claude Code, Codex, Cline, Cursor, Devin, Gemini CLI, OpenCode…) can read and apply on its own.

You don't read the diffs. You point your agent at this repo and say:

"Apply every skill in this repo that matches the harnesses I use."

The agent reads each SKILL.md, checks if it applies to your setup, lands the diff, and verifies the fix on the wire. You go from broken or partial caching to 80-99% cache hit rates without doing the research yourself.


What you actually save

One row per completed audit, so the coverage matches the scorecard:

Harness Finding Cost impact today Fix / status
Claude Desktop Code Default Desktop Code launches embedded Claude Code; clean Mac logs show non-zero cache read/create counters by default Already gets Anthropic cache benefits; no prompt-caching fix needed No skill; working baseline
Codex CLI Correct OpenAI cache design: stable thread_id cache key Already gets OpenAI cache benefits No skill; reference implementation
Aider --cache-prompts off by default; 5min TTL/keepalive overhead Many users get 0% cache reads unless they opt in; shorter cache window Skills: default-on caching + 1h TTL
OpenCode Strong Anthropic path, but proxy/Bedrock edge cases exist Some OpenAI-compatible→Anthropic/Bedrock routes miss cache Skills: proxy detection + Bedrock doc-block fix
Roo Code Anthropic volatile-message bug; Bedrock custom ARN gap Wastes breakpoints; custom ARNs can drop to 0% cache reads Skills: volatile-msg fix + Bedrock custom ARN fix
Cline Anthropic volatile-message bug; OpenAI lacks prompt_cache_key Wastes Anthropic breakpoint; OpenAI native can get 0% cache reads Skills: volatile-msg fix + OpenAI cache key + timestamp pin
Continue Cache opt-in default; Gemini explicit caching missing; volatile-message bug Many users get 0% cache reads; Gemini relies on implicit luck Skills: default-on + volatile-msg + Gemini explicit cache
Hermes / Nous Multi-provider cache plumbing works; xAI wire showed cached tokens No verified savings bug in this audit No skill; working audit
Codex Desktop ChatGPT Codex backend cache-scope headers observed/inferred No verified savings bug in this audit No skill; inferred working
Devin CLI Raw CLI model path is opaque Codeium/Devin protobuf Cache behavior not inspectable from public CLI capture No skill; unver

Continue your stack

What teams reach for next — and why each earns a place beside prompt-cache-skills. Ranked by curator confidence.