Golden path: the CI spine

Twin: the agent-learn init --preset ci scaffold (backing: [] — scaffold-backed by rule; proven offline by tests/test_init_golden_paths.py) · emits agent-learning.replay.v1 · offline, no credentials. A coding agent can complete this page from the frontmatter alone.

1. What you are testing

The ci preset scaffolds the smallest honest CI pipeline for an agent: a behavior run, a red-team gate, and a replay over both — in that order. The order is the lesson. Replay re-executes existing manifests and compares; it means nothing until baselines exist. So the spine is run first, red-team second, replay last, and the scaffold's README and init output list exactly that sequence. Your first replayable artifact lands at the second command — the path to a first artifact stays within three commands, offline.

agent-learn init . --preset ci writes both manifests at once: manifests/run.json (the deterministic scripted smoke run — a scripted agent, the local_text engine, one turn, an agent_report threshold of 0.7) and manifests/redteam.json (a tool-surface prompt-injection attack pack and campaign mapped to the OWASP LLM and agentic taxonomies, with a 0.9 evaluation threshold). These are the same manifests the single-purpose run and redteam presets produce; the ci preset exists to wire them into one directory that a pipeline can execute end to end and archive.

The final command, agent-learn replay manifests, re-runs every manifest in the directory and emits one combined verdict. In a pipeline you commit the manifests, run all three commands on every change, and archive artifacts/ — the replay artifact is the thing CI gates on.

2. Run it

CLI:

agent-learn init . --preset ci
agent-learn run manifests/run.json --output artifacts/run.json
agent-learn redteam manifests/redteam.json --output artifacts/redteam.json
agent-learn replay manifests --output artifacts/replay.json

SDK, same operations:

import asyncio
from fi.alk import redteam, simulate

run_result = asyncio.run(simulate.run_manifest_file("manifests/run.json"))
rt_result = asyncio.run(redteam.redteam_manifest_file("manifests/redteam.json"))
replay_result = simulate.replay_manifests(["manifests"])
assert replay_result["kind"] == "agent-learning.replay.v1"

3. What you built

Postcondition (machine-checkable — the same check the scaffold README carries):

python -c "import json; p=json.load(open('artifacts/replay.json')); assert p['kind']=='agent-learning.replay.v1', p['kind']; print('ok')"

Three artifacts now sit in artifacts/: run.json (agent-learning.run.v1, with status: "passed"), redteam.json (agent-learning.redteam.v1, findings and campaign coverage), and replay.json (agent-learning.replay.v1, the per-manifest replay results and the combined verdict). Each carries its own kind header, so a pipeline can verify every step with the same one-line check pattern.

4. When it fails

Symptom	First-mile class	Doctor check
vendored import failed from any command	infra	agent-learn doctor → summary.missing_engine_modules
key/credential errors on a platform-connected lane	config fault	agent-learn doctor → summary.api_key_configured
init would overwrite existing file(s); use --force	scaffold state	not a doctor fault — rerun with --force or use a clean directory

All four commands run with no API key set — required_env in the scaffolded manifests is CI metadata, and the per-preset golden-path test executes this sequence with the key explicitly unset.

5. Prove it / keep it

This spine becomes real the moment the scripted agents are replaced with your own and the manifests are committed. Then:

1. Promote passing runs and red-team findings into regressions/ so replay compares against frozen baselines instead of fresh manifests — ../simulate/regression-lifecycle.md is the full baseline → compare → replay → promote → shrink journey.
2. Grow the red-team half into real campaigns via ../redteam/red-team-anything.md.
3. When you also want eval suites and optimization in one gate, the all preset and ../prove/trinity-suite.md collapse the whole spine into a single agent-learn suite command.