Diff two agent runs with reelier diff (exit 1 on drift)

00Before you start

• Reelier installed — npm i -g reelier (or swap npx -y reelier@latest for reelier in every command below).
• A compiled skill with at least TWO recorded runs. reelier diff reads .reelier/runs/<skill>.jsonl and needs two run records to compare — replay the skill twice (or let a schedule tick twice). If you don't have a skill yet, compile one from a real session first.
• Nothing else. The diff itself calls no model and makes no network calls — it only reads the receipts already on disk.

01Get two runs into the record

reelier run re-executes a compiled skill's recorded tool calls and checks each assertion. At the default --max-level 0 no model is constructed, so the replay costs 0 tokens and reproduces only the deterministic core of the workflow. Every run appends one receipt line to .reelier/runs/<skill>.jsonl.

Run the skill twice — the first receipt becomes the baseline, the second is the candidate the diff will compare against it. (Replay is read-only by default; a recorded write step never re-fires unless you pass --allow-writes, and a destructive one still needs --yes.) The full record-and-replay loop is a guide of its own, linked below.

reelier run registry-check.skill.md

✓ Step 1 — Call http.get [passed] 164ms

PASSED: 1/1 steps ok, 0 failed, 164ms total

02Diff the last two runs

reelier diff <skill> aligns the last two runs step by step and compares each step's outcome and escalation level. When every step matches, it prints one = line per step and a SAME summary, and exits 0.

It never compares the data a step returned — only the shape of what happened. The compiled skill asserts a shape (status == 200), not a pinned value, so two honest replays seconds apart return different response bodies and still diff SAME. That is the point: same steps, fresh data is not drift.

reelier diff registry-check

  = Step 1 "Call http.get": passed, unchanged.

✓ SAME — 1 step(s), every outcome and escalation level identical to the baseline.

03Read the per-step verdict

Each step gets one marked line. = means the step is unchanged. ✗ is hard drift — the outcome flipped (passed → failed), or a step was added or removed, i.e. the run did something structurally different. ~ is soft drift — same outcome, but the step had to escalate to a different level to reach it, so the world moved underneath and healing covered it (soft drift only appears once you opt into BYOK escalation above Level 0; the deterministic core never escalates).

The summary line tallies them: DRIFTED splits the count into outcome/structure changes and steps that needed a different escalation level, so a strict CI gate can fail on any drift while a lenient one allows soft drift only. Timing is never counted as drift.

04Fail CI on the exit code

The exit code is the whole integration surface: 0 on SAME, 1 on DRIFTED (and 1 when the two records aren't even the same skill). Add reelier diff <skill> as a CI step after the replay and a non-zero exit fails the job — no extra assertions to wire up.

To see a real DRIFTED verdict below, we forced one: we tightened the recorded contract from status == 200 to status == 404, so the next replay against the live registry genuinely flipped passed → failed. In production the identical verdict fires on its own when an upstream API actually changes. Wiring this into a GitHub Actions job is covered on the platform page and the CI-gate guide, both linked below.

reelier diff registry-check

  ✗ Step 1 "Call http.get": passed → failed — status == 404 failed: got 200.

⚠ DRIFTED — 1 outcome/structure change(s): #1 outcome-changed.

05The receipt

$ reelier diff registry-check
  = Step 1 "Call http.get": passed, unchanged.

✓ SAME — 1 step(s), every outcome and escalation level identical to the baseline.

$ reelier diff registry-check
  ✗ Step 1 "Call http.get": passed → failed — status == 404 failed: got 200.

⚠ DRIFTED — 1 outcome/structure change(s): #1 outcome-changed.

Both verdicts of the same skill, verbatim. The first diff ran over two clean replays of registry-check against the live npm registry and exited 0; then the recorded assertion was tightened from status == 200 to status == 404 to force a genuine passed → failed flip, so the second diff printed DRIFTED and exited 1. Produced on 2026-07-22 with reelier v0.14.0 in a scratch directory; the skill was compiled by reelier from-session from a one-call Claude Code-style transcript (a single http.get, shaped per reelier/test/session.test.ts). The diff itself constructed no model and made no network call — it only read the two receipts in .reelier/runs/registry-check.jsonl.