artifact-audit 2026-07-25: monthly drift pass (6 fixes, 12 flags)

[espi]

Artifact drift

First scheduled run of the monthly artifact-audit skill. Four parallel audit agents (skills / templates / guardrails / runbooks+docs) checked every practice artifact against the knowledge base (last verified 2026-07-20). No prior artifact-audit PR was open; no update-knowledge PR was in flight. Per the skill: fixes below are backed by High-confidence KB entries or verified internal cross-references; everything touching a hard stop, protected file, or Medium/unverifiable evidence is flagged, not rewritten.

Fixed

File	What was stale	Evidence
.claude/skills/new-loop/SKILL.md	"Use /loop only for recurring/scheduled work" steered scheduled/unattended jobs to /loop, which is session-scoped (no laptop-closed runs; recurring tasks expire after 7 days). Now states the limits and points laptop-closed scheduling at a Routine (/schedule, alias /routines).	Primer §4 /loop + Routines entries — High (code.claude.com/docs/en/routines read directly)
.claude/skills/new-loop/SKILL.md	"/goal with a Haiku validator" → "separate validator model (defaults to Haiku)" — Haiku is the default, not fixed.	Primer §4 — High
guardrails/README.md	Tool-agnostic enforcement section predated Claude Code's native runaway-loop caps (v2.1.212, Jul 17: WebSearch cap 200/session, subagent-spawn cap 200/session, MCP >2 min auto-background). Added as complements to — not replacements for — the three hard stops (none is a dollar ceiling; defaults are far above a sane per-loop cap).	Primer §4 / sources.md changelog entry — High
guardrails/README.md	Receipts list missing the newest named-company hard cap: Tesla $200/week effective Jul 6, 2026 (third after Uber, Microsoft).	sources.md — High (multi-outlet)
guardrails/checklist.md	Permissions checklist item didn't warn that auto mode is on by default on Bedrock/Vertex/Foundry since v2.1.207 (disableAutoMode to opt out) — exactly the unintended-autonomy case the item exists to catch.	Primer §4 — High (changelog)
templates/loop/loop.md	Starter invocation /loop 15m /review-pr 1234 calls a skill that doesn't exist → /review (the bundled PR-review skill).	Internal cross-reference check
runbooks/first-loop.md	"Durable asset" pointer cited primer §5; the skills-as-durable-asset material is §4.	Internal cross-reference check
runbooks/ui-audit-overnight.md	Garbled sentence ("Rename FINDINGS.md is already in place.").	Internal cross-reference check

Flagged for you (not changed — human decision needed)

Hard-stop / protected areas:

1. templates/ralph/run.sh — the budget ceiling is a disarmed placeholder by default (highest-severity finding). cost_so_far() returns ${LOOP_COST_USD:-0}, which nothing in the loop updates — so hard stop knowledge: 2026-06-22 update pass (7-day delta from Jun 15) #3 evaluates 0 >= 10 forever unless the user wires a meter. Two decisions: (a) the comment's example claude cost command is not documented anywhere in the KB (documented spend surfaces are /usage, the Rate Limits API, the Analytics Admin API); (b) consider making the script refuse to start when the meter is unwired, per the repo's own "never run an uncapped loop." Related non-hard-stop nit: eval "$SUCCESS_CHECK" evals an env-overridable string.
2. .claude/skills/artifact-audit/SKILL.md (this skill's own guardrails) — "the native CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION cap bounds the ~4-agent fan-out" overstates the default: it's 200 natively (v2.1.212), no practical bound unless set ≈12 in the Routine env as runbooks/staying-current.md instructs. Suggested wording: "…the CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION cap (set ≈12 in the Routine env — native default is 200) bounds the ~4-agent fan-out."
3. guardrails/README.md hard-stops section never states that max_turns / max_budget_usd (Agent SDK) and --max-iterations (ralph-wiggum) all default to unlimited — KB-High and load-bearing, but inside the protected three-hard-stops text. (Also max_budgets_usd typo → max_budget_usd.)
4. templates/ui-audit/PROMPT.md references a RUN.md that doesn't exist in the template dir ("stop at the caps set in RUN.md") — the caps actually live in runbooks/ui-audit-overnight.md's invocation. Flag-only because it defines where the loop's caps live.
5. runbooks/ui-audit-overnight.md — the recommended interactive lane's budget stop is a morning /usage check, i.e. an alert, not enforcement (the runbook is honest about it and offers run.sh as the hard path).

Possibly-dead or KB-unverifiable claims (re-verify against live docs; candidates for the KB re-verify backlog):

6. runbooks/babysit-prs.md — the laptop-closed path rests on a "cloud auto-fix-PR feature", a "web CI bar", and /autofix-pr, none of which the KB tracks. The KB-verified mechanism is a Routine with a GitHub-event (pull_request) trigger via /schedule. If /autofix-pr isn't real, this runbook's core path is a dead end — likely needs a rewrite around Routines. (High severity if confirmed.)
7. Stop commands — Ctrl+C for /goal and /cancel-ralph appear in guardrails/checklist.md, runbooks/first-loop.md, and runbooks/ui-audit-overnight.md, but neither is in the KB (only Esc for /loop is verified).
8. runbooks/staying-current.md — four unbacked claims: /schedule disabled in web sessions; a /schedule update subcommand; a "per-account daily routine-run cap" (load-bearing — both maintenance skills cite that "Budget & caps" section as their budget hard stop); Routine network tiers (Trusted/Full/Custom).
9. templates/goal/recipes.md — operational details not in the KB (alias list stop/off/reset/cancel, ~4,000-char condition limit, trust-dialog/disableAllHooks prerequisites, headless claude -p "/goal …"). Not contradicted, just unverified.

Behavior-shift notes (v2.1.198, subagents background-by-default + auto-commit/push/draft-PR on worktree finish):

10. templates/loop/loop.md and runbooks/babysit-prs.md both say "use a worktree agent to fix them" with no note that worktree agents now auto-commit/push/open draft PRs without pausing — decide whether to annotate or constrain.

Descriptive-only observations:

11. guardrails/README.md intro says Anthropic's billing layer "does not auto-disable" — still true per the KB, but now understates the Jul 2 Claude Enterprise spend controls (alerts/entitlements, Medium-High, not confirmed hard-stop). Human to decide whether/how to acknowledge without softening the rule.
12. CLAUDE.md (protected) — its Routines note omits that Routines run with no permission prompts (KB-High, safety-relevant).

Clean (no drift found)

loop-guardrails and update-knowledge skills, templates/ralph/PROMPT.md, templates/goal/recipes.md (load-bearing claims), the ui-audit PREFLIGHT/PAGES/FINDINGS files, runbooks/overnight-build.md, README.md, budget.env, and both self-edit-guard files (all sentinel regions, paths, and the 10-line cap verified on disk).

No self-edit: commits in this PR; nothing here touches knowledge/. Human reviews and merges — never merged to main by the routine.

🤖 Generated with Claude Code

https://claude.ai/code/session_01KzGwSojW7vAPQiM6pDggD3

---

Generated by Claude Code