CommunityWriting & Editinggithub.com

seldonframe/reelier

Agents make claims. Reelier writes receipts — record an agent's tool-call workflow once, replay it deterministically at 0 tokens, and diff runs to catch drift.

What is reelier?

reelier is a Claude Code agent skill that agents make claims. Reelier writes receipts — record an agent's tool-call workflow once, replay it deterministically at 0 tokens, and diff runs to catch drift.

Works with✓Claude Code~Codex CLI~Cursor
npx skills add seldonframe/reelier

Installed? Explore more Writing & Editing skills: steipete/notion, langchain-ai/langchain, bytedance/podcast-generation · View all 6 →

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

Reelier — replay instead of re-reasoning

A *.skill.md file is a recorded run that worked: tool calls with a per-step pass/fail result. Replaying it costs 0 LLM tokens, takes milliseconds, and produces a receipt. Re-reasoning it burns tokens and quietly drifts. Prefer the replay.

1. FIRST: check for an existing skill

Before manually re-executing a tool-call sequence, Glob the project for **/*.skill.md. If one matches the task, run it:

npx -y reelier run <file>.skill.md

Deterministic, 0 LLM tokens, read-only by default. Report the receipt summary to the user: steps passed/failed, replay time, tokens (0). If a step fails, the failing assertion is the why — report it plainly and fall back to doing the task manually. Never report a pass you did not see in the run record.

2. No skill yet? Compile the run you just did

Only after you completed the task successfully via tool calls:

npx -y reelier from-session <transcript.jsonl>
# don't know the transcript path? find it first:
npx -y reelier scan

Never hand-author steps you did not actually run — a skill is a recording, not a plan. If the compiler prints "Open questions", read them out verbatim: those are gaps it declined to guess about (literal dates, IDs that should be {{variables}}). If it reports nothing replayable, say so — do not manufacture a skill anyway.

3. Check for drift

npx -y reelier diff <skill-name>

Exit 0 = SAME. Exit 1 = DRIFTED, per step, with the failing assertion as the why. Use it to gate a scheduled replay or to vet a model upgrade against a frozen baseline.

4. Optional: durable receipts

One line: npx -y reelier push <file>.skill.md syncs the receipt to a shareable ledger permalink (free account at reelier.com) — opt-in, never required.

Guards — can't-fail rules

  • Read-only by default. Write steps (idempotent-write) are held back unless you pass --allow-writes; destructive steps never auto-run. Replaying never silently re-fires a write.
  • No tool, no receipt. If npx -y reelier is unavailable or errors, do the task normally and tell the user Reelier wasn't used — never fake a receipt or pretend a replay happened.
  • Honest scope. Reelier replays MCP/HTTP tool calls only — file edits, shell commands, and reads are not replayable. A session with nothing replayable yields an honest empty result, not a fabricated skill; that is expected, not an error.

Related Skills