agentxp: Skill for Claude Code

Skill: /readouts

Purpose

Every readout produced by the presentation spine appends one entry to experiments/<id>/readouts/catalog.jsonl — an append-only hash-chained ledger. This skill walks the catalog (per-experiment or cross-experiment), validates the chain, and regenerates the static HTML index navigator.

The catalog is the only surviving hash chain in v2; tampering breaks it independently of git.

When to invoke

Direct:

• /readouts <exp_id> — list this experiment's renders
• /readouts --index — regenerate the cross-experiment HTML navigator
• /readouts <exp_id> --validate — walk the chain, exit non-zero on break

Plain-English routing:

Phrase	What to do
"Show me the renders for exp_001"	/readouts exp_001
"Rebuild the audit index"	/readouts --index
"Is the catalog intact"	/readouts <exp_id> --validate
"What readouts have I made across all experiments"	/readouts --index

Procedure

Catalog mode

from pathlib import Path
from agentxp.workflows.readouts import list_catalog

entries = list_catalog(Path.cwd() / "experiments" / args.exp_id)
for i, entry in enumerate(entries):
    payload = entry.payload
    print(f"{i:>3}  {entry.timestamp}  {payload.event:<26}  ...")

Index mode

from agentxp.workflows.readouts import build_index

rows = build_index(Path.cwd())
# Each row: experiment_id, n_renders, worst_status, latest_render_type, latest_render_at

Render an HTML table from the rows and write to readouts/index.html. The worst-case status per experiment uses the cascade: UNVERIFIABLE > DRAFT_UNVERIFIED > VERIFIED.

Validate mode

from agentxp.render.catalog import validate_catalog, CatalogChainBreak

try:
    validate_catalog(Path.cwd() / "experiments" / args.exp_id / "readouts" / "catalog.jsonl")
    print("catalog chain ok")
except CatalogChainBreak as exc:
    print(f"catalog chain break: {exc}", file=sys.stderr)
    raise SystemExit(1)