artifact-audit: Sep 2026 pass — billing-claim drift, native dollar ceiling, ralph caveats (3 fixed, 8 flagged)

.claude/skills/new-loop/SKILL.md

@@ -27,7 +27,7 @@ that returns clear pass/fail (e.g. `pytest test/users`, `npm test`,
 | Pattern | Use when | Mechanism |
 |---|---|---|
 | **`/goal`** (default) | Iterate-to-completion on a well-scoped task with a provable success condition. | `/goal` with a separate validator model (defaults to Haiku). Requires Claude Code v2.1.139+. |
-| **ralph** (`/ralph-loop` or bash) | Long autonomous build where each iteration should reset context to anchor files. | Official `ralph-wiggum` plugin, or `templates/ralph/run.sh`. |
+| **ralph** (`/ralph-loop` or bash) | Long autonomous build where each iteration should reset context to anchor files. | Official `ralph-wiggum` plugin (`--max-iterations` defaults to **unlimited** — always set it), or `templates/ralph/run.sh`. |
 | **`/loop`** | Recurring maintenance on an interval (babysit PRs, poll a deploy). | Bundled `/loop` skill (v2.1.72+); cron-backed, session-scoped; recurring tasks expire after 7 days. |
 
 Default to `/goal` for "do this task once, correctly." Use ralph for
@@ -55,8 +55,11 @@ Create `loops/<slug>/` (slug derived from the task) containing:
 2. **No-progress detection** — instruct the loop to bail if N consecutive
    iterations produce no git diff / no test-state change (default N=3), and
    document blockers instead of looping forever.
-3. **Budget ceiling** — set `MAX_BUDGET_USD` (ralph/SDK) or remind the user of
-   the session budget. State the expected cost order-of-magnitude.
+3. **Budget ceiling** — set `MAX_BUDGET_USD` (ralph `run.sh` / Agent SDK
+   `max_budget_usd`), or pass `--max-budget-usd` on the Claude Code invocation
+   (since v2.1.217 it also halts background subagents at the cap). A reminder
+   about the session budget is an alert, not a ceiling. State the expected
+   cost order-of-magnitude.
 
 Pull the exact defaults from `guardrails/budget.env`. Reference
 `guardrails/checklist.md` and confirm every box is satisfied before finishing.

README.md

@@ -57,8 +57,11 @@ bakes in guardrails.
 ## The one rule
 
 **Never run an uncapped loop.** Every loop gets all three hard stops — a max
-iteration count, no-progress detection, and a token/dollar budget ceiling. The
-billing layer will *not* save you (soft alerts only). See
+iteration count, no-progress detection, and a token/dollar budget ceiling.
+Anthropic now ships some real spend enforcement (Enterprise Spend Limits API,
+gateway spend limits, Managed Agents session budgets), but the gateway layer
+fails open by default and none of it reaches a plain Claude Code loop
+automatically — the ceiling still has to live in your harness. See
 [`guardrails/`](guardrails/).
 
 ## Staying current

templates/ralph/run.sh

@@ -12,13 +12,19 @@
 set -euo pipefail
 
 # ---- CONFIG (edit me) ------------------------------------------------------
+# NOTE: models that take more turns per unit of work (e.g. Fable 5.1-class
+# long-horizon models) trip iteration/stall caps tuned on older models earlier.
+# An early stop on such a model is a re-tuning signal, not proof of failure
+# (see knowledge/00-primer.md §4, Fable 5.1 breaking changes).
 PROMPT_FILE="PROMPT.md"
 MAX_ITERATIONS=20                 # hard stop #1: iteration cap
 MAX_BUDGET_USD=10                 # hard stop #3: dollar ceiling (enforced below)
 STALL_LIMIT=3                     # hard stop #2: bail after N no-progress passes
 COMPLETION_MARKER="<promise>COMPLETE</promise>"
 SUCCESS_CHECK="${SUCCESS_CHECK:-false}"   # e.g. "pytest -q" / "npm test"
 # Agent command. --dangerously-skip-permissions only for sandboxed/headless use.
+# Narrower alternative (v2.1.259+): --permission-prompts none denies anything
+# that would prompt instead of skipping permissions wholesale.
 AGENT_CMD=(claude -p --dangerously-skip-permissions)
 # ---------------------------------------------------------------------------