Replay an agent skill on every PR with the Reelier GitHub Action

00Before you start

• A recorded skill committed to the repo you want to guard — a plain .skill.md file. Record it first on your coding agent (freeze a Claude Code session, or wrap an MCP server with reelier mcp --wrap on Cursor, Codex or Windsurf); the linked guides cover recording.
• A read-only skill makes the best gate. Reelier replays MCP and tool-call steps plus the http.get / http.post builtins — not Bash, Read, Edit, Write or Grep — so pick a run whose value was in the API calls, like a data pull or a status check.
• A GitHub repository with Actions enabled. No account, API key or model is needed at --max-level 0 — the replay costs 0 tokens, and a cloud-key is optional.

01Prove the skill passes locally first

The Action runs one command for you — reelier run <skill> --max-level 0 — so run it yourself before wiring CI. It re-executes the recorded tool calls against the live systems and checks each recorded assertion. The repo's bundled npm-info.skill.md (a single read-only http.get to the npm registry) reproduces the receipt below exactly.

The skill asserts minimally — reelier's minimal-assertion rule: a step asserts a shape, not a brittle value, and only when it recorded a clean result. Here the http.get step asserts status == 200 and nothing else. At --max-level 0 — the default — no LLM client is ever constructed, so the replay costs 0 tokens and exits 0 when every assertion holds. That exit code is what the check gates on.

reelier run npm-info.skill.md --max-level 0

✓ Step 1 — Fetch registry metadata and extract version/license/tarball [passed] 461ms

PASSED: 1/1 steps ok, 0 failed, 461ms total

02Add the seldonframe/reelier@v1 workflow

Create .github/workflows/reelier.yml. The Action is a composite step published at seldonframe/reelier, tag @v1; after an actions/checkout@v4 it installs reelier@^0.12, runs the same reelier run <skill> --max-level 0 you just ran, writes the receipt as the job summary, and fails the check on the run's own exit code — you never wire the pass/fail yourself.

Trigger it on: pull_request so a drifted step blocks the merge, and scope it with paths: ["my-job.skill.md"] so it only fires when the skill changes; add a schedule: cron for a standing dated proof on a public repo. In the with: block, set skill: to your committed file and keep max-level: "0" — the default that needs no credentials. The full copy-paste YAML, the inputs table, and the passed / run-record-path outputs live on the platform page; this guide links it rather than restating it.

03Open a pull request and read the receipt

Push the workflow and the skill on a branch and open a PR. The Action replays the skill and posts a receipt as the run's job summary: the skill name, a PASSED / FAILED verdict, and a per-step table of passed / unchecked / skipped / failed with the duration and token spend (0 at level 0).

Read the unchecked column honestly — a step that ran without error but asserted nothing is reported as unchecked, never folded into passed. A green check means every recorded assertion actually held, not merely that the CLI ran. The exact markdown the Action writes is the closing receipt below.

04Watch it turn red when the tool calls drift

The gate earns its keep on the day the recorded run stops matching. If an upstream contract changes — the endpoint returns a 500, a field moves — the recorded assertion (status == 200) no longer holds, the step fails, reelier run exits non-zero, and the Action's final step re-raises that exit code. The check turns red and the PR cannot merge on a drifted tool call — regression testing for agent tool calls, with no model in the loop.

When the change was intentional, update the baseline as you would a snapshot: re-record and recompile the .skill.md, or edit the failing assertion directly — it is markdown in your repo. Replay stays read-only by default, so a PR or scheduled gate never re-fires a recorded write; a write step only re-executes under --allow-writes (idempotent writes only — a destructive step still needs --yes), which the Action does not pass.

05The receipt

## Reelier replay: `npm-info.skill.md`

**Skill:** `npm-info`  
**Result:** ✅ PASSED

| steps | passed | unchecked | skipped | failed | duration |
| --- | --- | --- | --- | --- | --- |
| 1 | 1 | 0 | 0 | 0 | 461ms |

> `unchecked` = ran without error but asserted nothing — the honest-success rule (SPEC.md §4.3) means it is never counted as a pass.

_Ledger: not pushed (no `cloud-key` input set)._

The receipt the Action posts as its GitHub job summary — the markdown you see on the PR's checks tab. Produced on 2026-07-22 with reelier v0.14.0 by running the Action's own .github/scripts/gha-summary.mjs against the .reelier/runs/npm-info.jsonl record that reelier run npm-info.skill.md --max-level 0 wrote (the 461ms matches Step 1 — one record, two views). The collapsible per-step detail block is trimmed here for space; nothing else is edited. 0 tokens; cloud-key was unset, so the replay was local-only and still wrote this summary.