audit

Skill: /audit

Purpose

The audit surface is two things: experiments/<id>/log.md (append-only human-readable log) and git log (every commit_artifact runs a git commit). This skill walks both, optionally diffs two experiments, and optionally renders HTML.

There is no separate event log in v2 — git is the chain.

When to invoke

Direct:

• /audit <exp_id> — text timeline
• /audit <exp_id> --diff <other_exp_id> — pairwise diff
• /audit <exp_id> --html — self-contained HTML

Plain-English routing:

Phrase	What to do
"Why did exp_007 halt?"	/audit exp_007 then look for halt-related entries
"Compare exp_001 to exp_007"	/audit exp_001 --diff exp_007
"Send the audit to my director"	/audit <id> --html and surface the file path
"What dispatches landed for exp_004"	/audit exp_004 then filter for dispatch entries

Procedure

Text mode (default)

from pathlib import Path
from agentxp.workflows.audit import walk_log

for entry in walk_log(Path.cwd() / "experiments" / args.exp_id):
    print(f"{entry.timestamp}  {entry.message}")

Optionally interleave git log --oneline -- experiments/<id>/ so each commit SHA appears next to its log entry. The two are kept in sync by commit_artifact.

Diff mode

from agentxp.workflows.audit import diff_logs

for d in diff_logs(exp_a, exp_b):
    if d.kind == "changed":
        print(f"line {d.line_no}: {d.a.message}  ->  {d.b.message}")
    elif d.kind == "only_in_a":
        print(f"only in {exp_a.name}: {d.a.message}")
    elif d.kind == "only_in_b":
        print(f"only in {exp_b.name}: {d.b.message}")

HTML mode

Render a self-contained HTML page that includes the log timeline + the renders catalog summary + the integrity-lock receipt from brief.sealed.yaml. Use agentxp.render.report for the HTML adapter; the data comes from walk_log and list_catalog.

Tools you call

• walk_log / diff_logs from agentxp.workflows.audit
• list_catalog from agentxp.workflows.readouts (for the renders catalog summary in HTML mode)
• git log --oneline -- experiments/<id>/ via Bash for commit SHAs