OnlyTerp

Aider uses 5min TTL by default and works around long pauses with keepalive pings. Wire up the 1h TTL beta instead.

Aider's --cache-prompts is off by default. Flip it on for supported models.

Stop Cline from burning a cache breakpoint on the volatile current user message every turn.

Cline OpenAI native provider sends no prompt_cache_key. Add a stable per-task key so cached_tokens stops being zero.

Cline's system prompt includes a timestamp that may be recomputed per request, invalidating the system-prompt cache.

Continue's prompt caching is opt-in via config and off by default. Flip the default to systemAndTools.

Continue inherits the Cline-family volatile-message cache bug. Same fix shape.

Continue's Gemini provider doesn't use the cachedContents API at all. Add explicit caching for sessions over the minimum token threshold.

Delegate a PR review to Claude Code with a scoped read-only GitHub PAT

Build human-readable release notes from a range of commits or merged PRs

Weekly LLM cost breakdown by provider / gateway / skill, posted to private DM

Sweep inbox (email + Slack + Telegram DMs) and produce a prioritized action list with suggested replies

Produce a weekly "What's new in Hermes" digest by summarizing merged PRs + active issues from NousResearch/hermes-agent

Run `hermes backup`, encrypt, upload to remote storage, prune old backups

Classify inbound Telegram DMs, autoreply low-stakes, escalate high-stakes to you

Audit dependencies across configured repos for security advisories, open triage issues

Use ast-grep for structural (AST-aware) code search and refactoring instead of plain ripgrep when the intent depends on code structure — not just text matching. Triggers on refactors, codemods, "find all X that don't Y", or any search where naive grep would miss edge cases.

Guides proactive /compact usage with preservation instructions to prevent context rot in long Cascade sessions. Triggers when context reaches ~50% of window, before large tool outputs, or when the user mentions "context", "compact", "drift", or "forgetting".

Use when any non-trivial task (>20 LOC or >1 file) would otherwise be done without a persistent plan. Produces and maintains a markdown plan file that survives session compaction and restarts. Based on the Manus-style planning pattern popularized by OthmanAdi/planning-with-files (18.4k⭐).

Turn a feature branch into a clean, reviewable PR. Squash WIP commits, polish messages, verify green checks, draft PR description. Invoke right before opening a PR.

Scan a diff, branch, or working tree for leaked secrets, API keys, tokens, private keys, and PII before commit. Blocks the operation on CRITICAL findings, warns on MEDIUM. Pairs with the pre_commit hook.

Fill test coverage gaps to hit the project's target (AGENTS.md says ≥80% on src/). Reads existing tests to match style, prioritizes public-API and error paths, runs the suite after every batch.

Read the project wiki (vault/) before starting any non-trivial task. Surfaces relevant decisions, ownership, incidents, and glossary entries for the task at hand.

Maintain the project wiki in vault/. Use at the end of any session that produced a decision, fact, or state change. Extracts entities, updates topic pages, adds cross-links, never deletes. Pairs with the post_cascade_response_with_transcript hook.

GPU-native real-time screen capture via DXGI Desktop Duplication. Gets frames from the GPU compositor, uses DXGI dirty-rect metadata as change hints, refines coalesced rectangles into readable crop PNGs, and emits WSL-usable JSON paths. Use to watch the screen in real time, detect what changed where, capture HDR desktops, or grab frames of GPU-composited apps. Complements the screenshot skill (which provides input control).