v2.0.0 — Modernization: loop-first runtime, Verification 2.0, producer skills, enforcing gates

.claude/agents/beamer-translator.md

@@ -2,7 +2,8 @@
 name: beamer-translator
 description: Specialist agent for translating Beamer LaTeX slides to Quarto RevealJS. Handles content translation, environment mapping, citation conversion, and formatting. Use as a subagent during the /translate-to-quarto workflow for the actual slide-by-slide translation work.
 tools: Read, Write, Edit, Grep, Glob, Bash
-model: inherit
+model: sonnet
+effort: medium
 ---
 
 You are a specialist in translating academic Beamer slides to Quarto RevealJS format.

.claude/agents/claim-verifier.md

@@ -2,7 +2,8 @@
 name: claim-verifier
 description: Fresh-context verifier for factual claims made by other agents or skills. Implements the Chain-of-Verification (CoVe) independence trick via context forking — the verifier never sees the original draft, only the extracted claims + the source material. Use when a skill has produced a draft that contains citations, numerical facts, named entities, or literature references that need hallucination-checking before returning to the user.
 tools: Read, Grep, Glob, WebFetch, WebSearch, Bash
-model: inherit
+model: opus
+effort: high
 ---
 
 <!-- Adapted from Dhuliawala et al. 2023, "Chain-of-Verification Reduces Hallucination in Large Language Models" (arxiv.org/abs/2309.11495). The core idea — answering verification questions in a context that does NOT contain the original draft — is architecturally enforced here by running the agent via Task with context: fork. -->
@@ -39,6 +40,11 @@ claims:
     text: "The method requires conditional parallel trends."
     source_hint: "same paper"
     verification_question: "What parallel trends assumption does the paper require — unconditional or conditional?"
+    author_alternative: ""        # OPTIONAL. If the author has pre-recorded a
+                                  # concrete named alternative that accounts for
+                                  # an expected numeric/directional gap, put it
+                                  # here; a contradiction then resolves to
+                                  # EXPLAINED instead of HIGH-WARN. Blank = none.
 ```
 
 ### Step 2: Answer each question independently
@@ -67,6 +73,7 @@ Each per-claim finding now carries a **severity tier** (v1.9.0):
 - **HIGH-WARN** — fabricated reference (cited paper does NOT exist at the named venue/year), direct contradiction between draft and source, or `not_found` retrieval that you interpret as a hallucinated citation. These block `/commit` via `/verify-claims`.
 - **MED-WARN** — transient infrastructure failure: DOI resolver timed out, partial PDF read, paywall the cache normally bypasses. The author should re-run the verification later or supply a local copy.
 - **LOW-WARN** — source genuinely inaccessible (paywalled with no cache hit, private dataset, pre-print server transient). Surface but do not gate-refuse — the claim may still be correct; the verifier just can't independently confirm.
+- **EXPLAINED** — a numeric or directional contradiction for which the request carries a *concrete named alternative* (`author_alternative`) that accounts for the gap — a different but defensible edition, specification, sample, or rounding convention. Surface with the evidence and the recorded alternative, but do **not** gate-refuse: the disagreement is documented, not a bug. EXPLAINED **never** applies to a fabricated/nonexistent citation (that stays HIGH-WARN), nor to a blank or vague alternative ("different version", "rounding" with no specifics).
 
 ```markdown
 ## Claim Verification Report
@@ -80,7 +87,7 @@ Each per-claim finding now carries a **severity tier** (v1.9.0):
 
 | ID | Claim (draft) | Independent answer | Evidence | Match? | Tier |
 |----|--------------|---------------------|----------|--------|------|
-| C1 | [quoted claim] | [what source says] | [quote + loc] | yes / partial / no / cannot-verify | — / LOW / MED / HIGH |
+| C1 | [quoted claim] | [what source says] | [quote + loc] | yes / partial / no / cannot-verify | — / LOW / MED / HIGH / EXPLAINED |
 
 ### HIGH-WARN (gate-refuse `/commit`)
 
@@ -102,12 +109,12 @@ Each per-claim finding now carries a **severity tier** (v1.9.0):
   - "I retrieved the source and it contradicts the claim" → HIGH-WARN
   - "I cannot retrieve the source because of a transient failure" → MED-WARN
   - "I cannot retrieve the source because it's genuinely inaccessible" → LOW-WARN
-- A cited paper that does NOT exist (no DOI, no arXiv id, no venue search hit) is **always HIGH-WARN** — this is the canonical hallucination signature. Do not soften to MED-WARN on the grounds that "maybe my search missed it."
-- A numerical contradiction (draft says X, source says Y, X ≠ Y within rounding) is **always HIGH-WARN**.
-- A directional contradiction (draft says "positive effect", source says "negative effect") is **always HIGH-WARN**.
+- A cited paper that does NOT exist (no DOI, no arXiv id, no venue search hit) is **always HIGH-WARN** — this is the canonical hallucination signature. Do not soften to MED-WARN or EXPLAINED on the grounds that "maybe my search missed it." This is the hard floor: a fabricated citation is never downgradable.
+- A numerical contradiction (draft says X, source says Y, X ≠ Y within rounding) is **HIGH-WARN by default** — *unless* the claim carries a concrete `author_alternative` that names a defensible reason for the gap (different edition/table, specification, sample, rounding convention), in which case record it as **EXPLAINED** (surfaced, non-gating). A blank or vague `author_alternative` does not soften — stay HIGH-WARN.
+- A directional contradiction (draft says "positive effect", source says "negative effect") is **HIGH-WARN by default** — same EXPLAINED escape: only a concrete named alternative for that claim downgrades it; otherwise HIGH-WARN.
 - A paraphrase mismatch where the draft's gloss is a reasonable summary of the source — not HIGH-WARN. Flag as `partial` with no tier (or LOW if you want to surface the gloss difference).
 
-Be conservative on HIGH-WARN. It blocks `/commit`. False positives erode the gate's authority; false negatives let known-bad claims ship.
+Be conservative on HIGH-WARN. It blocks `/commit`. False positives erode the gate's authority; false negatives let known-bad claims ship. The EXPLAINED escape exists only to stop a *defensible, author-documented* disagreement from gate-blocking — it is never a way to wave through a fabricated citation or an undocumented contradiction.
 
 ## What you DO NOT do
 

.claude/agents/domain-referee.md

@@ -2,7 +2,8 @@
 name: domain-referee
 description: Substantive referee for a manuscript. Reviews contribution, literature positioning, substantive argument, external validity, and journal fit. Calibrated to a target journal and primed with a disposition + pet peeves by the editor agent. Used by `/review-paper --peer`.
 tools: Read, Grep, Glob
-model: inherit
+model: opus
+effort: high
 ---
 
 <!-- Adapted from Hugo Sant'Anna's clo-author (github.com/hugosantanna/clo-author),

.claude/agents/domain-reviewer.md

@@ -2,7 +2,8 @@
 name: domain-reviewer
 description: Substantive domain review for lecture slides. Template agent — customize the 5 review lenses for your field. Checks derivation correctness, assumption sufficiency, citation fidelity, code-theory alignment, and logical consistency. Use after content is drafted or before teaching.
 tools: Read, Grep, Glob
-model: inherit
+model: opus
+effort: high
 ---
 
 <!-- AUTO-DETECT-TEMPLATE-MARKER — do not remove unless you have customized

.claude/agents/editor.md

@@ -2,7 +2,8 @@
 name: editor
 description: Journal editor who desk-reviews manuscripts, selects two referees with deliberately different dispositions, calibrates to a target journal from `.claude/references/journal-profiles.md`, and synthesizes an editorial decision (FATAL / ADDRESSABLE / TASTE). Used by `/review-paper --peer [journal]`.
 tools: Read, Grep, Glob, WebSearch, WebFetch
-model: inherit
+model: opus
+effort: high
 ---
 
 <!-- Adapted from Hugo Sant'Anna's clo-author (github.com/hugosantanna/clo-author),

.claude/agents/humanize-auditor.md

@@ -2,7 +2,8 @@
 name: humanize-auditor
 description: Read-only auditor for AI-voice tells in academic prose. Reviews `.tex`, `.qmd`, `.md` files for the 10 detection categories defined in `/humanize` (boilerplate transitions, AI-cliché lexicon, em-dash overuse, symmetric paragraph shapes, tricolon abuse, hedging stacking, "not only X but also Y" frames, formulaic openers, hyphenation excess, sycophancy/self-important framing). Produces a structured report without editing. Use when invoked by `/humanize`.
 tools: Read, Grep, Glob
-model: inherit
+model: sonnet
+effort: high
 ---
 
 You are a read-only auditor for AI-voice tells in academic prose. Your job is to detect statistically conspicuous LLM patterns in the user's manuscript and report them — **never edit**.

.claude/agents/methods-referee.md

@@ -2,7 +2,8 @@
 name: methods-referee
 description: Methodology referee for a manuscript. Paper-type-aware (reduced-form / structural / theory+empirics / descriptive / formal-theory / survey-experiment), each with its own dimension weights and mandatory sanity checks. Calibrated to a target journal and primed with a disposition + pet peeves. Used by `/review-paper --peer`.
 tools: Read, Grep, Glob
-model: inherit
+model: opus
+effort: high
 ---
 
 <!-- Adapted from Hugo Sant'Anna's clo-author (github.com/hugosantanna/clo-author),

.claude/agents/pedagogy-reviewer.md

@@ -2,7 +2,8 @@
 name: pedagogy-reviewer
 description: Holistic pedagogical review for academic slides. Checks narrative arc, prerequisite assumptions, worked examples, notation clarity, and deck-level pacing. Use after content is drafted.
 tools: Read, Grep, Glob
-model: inherit
+model: sonnet
+effort: high
 ---
 
 You are an expert pedagogy reviewer for academic lecture slides. Your audience is advanced students learning specialized material for the first time.

.claude/agents/proofreader.md

@@ -2,7 +2,8 @@
 name: proofreader
 description: Expert proofreading agent for academic lecture slides. Reviews for grammar, typos, overflow, and consistency. Use proactively after creating or modifying lecture content.
 tools: Read, Grep, Glob
-model: inherit
+model: sonnet
+effort: high
 ---
 
 You are an expert proofreading agent for academic lecture slides.

.claude/agents/quarto-critic.md

@@ -2,7 +2,8 @@
 name: quarto-critic
 description: Adversarial QA agent that compares Quarto HTML against Beamer PDF benchmark. Produces harsh, actionable criticism. Does NOT edit files — read-only analysis only.
 tools: Read, Grep, Glob
-model: inherit
+model: opus
+effort: high
 ---
 
 You are a **harsh, uncompromising quality auditor** for academic presentation slides.

.claude/agents/quarto-fixer.md

@@ -2,7 +2,8 @@
 name: quarto-fixer
 description: Implements fixes from the quarto-critic agent. Applies changes to QMD files, re-renders slides, and verifies fixes. Does NOT make independent decisions — follows critic instructions exactly.
 tools: Read, Edit, Write, Bash, Grep, Glob
-model: inherit
+model: sonnet
+effort: medium
 ---
 
 You are a **precise implementer** for Quarto slide fixes.

.claude/agents/r-package-reviewer.md

@@ -2,7 +2,8 @@
 name: r-package-reviewer
 description: R package source reviewer. Checks the things that decide whether a package passes R CMD check --as-cran and survives CRAN review — DESCRIPTION/dependency hygiene, NAMESPACE and imports, roxygen documentation completeness, testthat coverage, and CRAN-policy red flags. Use after writing or modifying package source (R/, tests/, DESCRIPTION, NAMESPACE), or as the review pass inside /r-package-check.
 tools: Read, Grep, Glob, Bash
-model: inherit
+model: sonnet
+effort: high
 ---
 
 You are a **CRAN-savvy R package maintainer** who has shepherded packages through `R CMD check --as-cran` and CRAN's human review many times. You know exactly which patterns trigger a NOTE, which trigger a WARNING, and which get a package archived.

.claude/agents/r-reviewer.md

@@ -2,7 +2,8 @@
 name: r-reviewer
 description: R code reviewer for academic scripts. Checks code quality, reproducibility, figure generation patterns, and theme compliance. Use after writing or modifying R scripts.
 tools: Read, Grep, Glob
-model: inherit
+model: sonnet
+effort: high
 ---
 
 You are a **Senior Principal Data Engineer** (Big Tech caliber) who also holds a **PhD** with deep expertise in quantitative methods. You review R scripts for academic research and course materials.

.claude/agents/sim-reviewer.md

@@ -2,7 +2,8 @@
 name: sim-reviewer
 description: Monte Carlo simulation reviewer. Checks the parts of a simulation study that general R review misses — DGP/estimand alignment, replication budget and Monte Carlo standard error, coverage computed against the truth, parallel-seed discipline, and whether headline simulation claims match the generated tables. Use after writing or modifying a Monte Carlo simulation script, or as the review pass inside /simulation-study.
 tools: Read, Grep, Glob
-model: inherit
+model: opus
+effort: high
 ---
 
 You are a **methodologist who referees simulation evidence for top journals**. You have caught simulation bugs that flipped a paper's headline conclusion, and you know that a beautiful results table built on a mis-seeded loop or a coverage-against-the-estimate bug is worse than no table at all.

.claude/agents/slide-auditor.md

@@ -2,7 +2,8 @@
 name: slide-auditor
 description: Visual layout auditor for RevealJS and Beamer slides. Checks for overflow, font consistency, box fatigue, and spacing issues. Use proactively after creating or modifying slides.
 tools: Read, Grep, Glob
-model: inherit
+model: sonnet
+effort: high
 ---
 
 You are an expert slide layout auditor for academic presentations.

.claude/agents/tikz-reviewer.md

@@ -2,7 +2,8 @@
 name: tikz-reviewer
 description: Harsh devil's advocate reviewer for TikZ diagrams. Checks every label position, overlap, visual consistency, and aesthetic appeal. Use after creating or modifying any TikZ code. The calling agent must iterate with this reviewer until all issues are resolved.
 tools: Read, Grep, Glob
-model: inherit
+model: opus
+effort: high
 ---
 
 You are a **merciless visual critic** for TikZ diagrams in academic slides. Your job is to find EVERY visual flaw, no matter how small. You have extremely high standards — a diagram is not done until it is perfect.

.claude/agents/verifier.md

@@ -2,7 +2,8 @@
 name: verifier
 description: End-to-end verification agent. Checks that slides compile, render, deploy, and display correctly. Use proactively before committing or creating PRs.
 tools: Read, Grep, Glob, Bash
-model: inherit
+model: opus
+effort: high
 ---
 
 You are a verification agent for academic course materials.

.claude/hooks/claim-reconcile.py

@@ -0,0 +1,128 @@
+#!/usr/bin/env python3
+"""
+Numeric-Claim Reconciliation Hook (PostToolUse)
+
+Event-driven half of the cross-artifact dependency graph: the moment an
+analysis script or an `_outputs/` artifact changes, the manuscript's
+numeric claims that depend on it may be STALE. Instead of waiting for the
+nightly reproducibility Routine, this hook surfaces the staleness
+immediately so the author re-runs /audit-reproducibility before relying
+on the affected tables.
+
+Fires on Write/Edit to:
+  - scripts/**/*.{R,do,py,jl}        (analysis code)
+  - scripts/**/_outputs/**           (regenerated outputs)
+when a passport (quality_reports/passports/*.yaml) exists. It counts the
+passport claims whose `source_file`/`output_file` mentions the changed
+file and emits a one-line systemMessage + additionalContext. Throttled to
+once per changed file per session (so a burst of edits is one nudge).
+
+PostToolUse output: exit 0 + JSON {"systemMessage", "hookSpecificOutput":
+{"additionalContext"}}. Fail-open: any error → exit 0, silent.
+
+For external regenerations (a user running Rscript outside Claude), the
+broader `FileChanged` event can drive the same logic — see
+.claude/references/scheduled-routines.md.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+import sys
+import time
+import hashlib
+from pathlib import Path
+
+WATCH = re.compile(r"(^|/)scripts/.*\.(R|r|do|py|jl)$|(^|/)scripts/.*/_outputs/")
+THROTTLE_S = 300
+
+
+def state_dir() -> Path:
+    pd = os.environ.get("CLAUDE_PROJECT_DIR", "")
+    h = hashlib.md5(pd.encode()).hexdigest()[:8] if pd else "default"
+    d = Path.home() / ".claude" / "sessions" / h
+    d.mkdir(parents=True, exist_ok=True)
+    return d
+
+
+def main() -> int:
+    try:
+        data = json.load(sys.stdin)
+    except (json.JSONDecodeError, EOFError):
+        return 0
+
+    ti = data.get("tool_input", {}) or {}
+    fp = ti.get("file_path", "") or ""
+    if not fp or not WATCH.search(fp):
+        return 0
+
+    project_dir = os.environ.get("CLAUDE_PROJECT_DIR", "") or data.get("cwd", "")
+    if not project_dir:
+        return 0
+    passports = sorted((Path(project_dir) / "quality_reports" / "passports").glob("*.yaml"))
+    if not passports:
+        return 0  # no claims tracked → nothing to reconcile
+
+    # Match + throttle on the project-relative PATH, not the bare basename —
+    # otherwise scripts/R/results.rds and scripts/stata/results.rds throttle
+    # each other, and clean.R spuriously matches data_clean.R.
+    try:
+        changed = str(Path(fp).resolve().relative_to(Path(project_dir).resolve()))
+    except Exception:
+        changed = Path(fp).name
+
+    # Throttle: one nudge per changed file per THROTTLE_S.
+    st_path = state_dir() / "claim-reconcile-state.json"
+    try:
+        st = json.loads(st_path.read_text())
+    except Exception:
+        st = {}
+    now = time.time()
+    if now - st.get(changed, 0) < THROTTLE_S:
+        return 0
+    st[changed] = now
+    try:
+        st_path.write_text(json.dumps(st))
+    except Exception:
+        pass
+
+    # Count passport claims that reference this file (best-effort text match).
+    affected = []
+    for p in passports:
+        try:
+            text = p.read_text(encoding="utf-8", errors="replace")
+        except Exception:
+            continue
+        hits = sum(1 for ln in text.splitlines()
+                   if ("source_file" in ln or "output_file" in ln) and changed in ln)
+        if hits:
+            affected.append((p.name, hits))
+    if not affected:
+        return 0
+
+    total = sum(h for _, h in affected)
+    where = ", ".join(f"{name} ({h})" for name, h in affected)
+    msg = (f"⟳ {changed} changed — {total} passport claim(s) may be STALE [{where}]. "
+           f"Run /audit-reproducibility before relying on the affected tables.")
+    json.dump({
+        "systemMessage": msg,
+        "hookSpecificOutput": {
+            "hookEventName": "PostToolUse",
+            "additionalContext": (
+                f"A tracked analysis input ({changed}) was just modified. "
+                f"{total} numeric claim(s) recorded in {where} depend on it and are now "
+                f"potentially stale. Before presenting or committing those numbers, run "
+                f"/audit-reproducibility to re-verify them against the regenerated outputs."
+            ),
+        },
+    }, sys.stdout)
+    return 0
+
+
+if __name__ == "__main__":
+    try:
+        sys.exit(main())
+    except Exception:
+        sys.exit(0)  # fail open

.claude/hooks/context-monitor.py

@@ -177,6 +177,12 @@ def run_context_monitor() -> int:
     # Estimate current context usage (coarse proxy)
     percentage = estimate_context_percentage(hook_input)
 
+    # Persist the latest estimate so the status line can surface it (best-effort).
+    try:
+        (get_session_dir() / "context-pct.txt").write_text(f"{percentage:.0f}")
+    except Exception:
+        pass
+
     # Check throttling
     if is_throttled(percentage):
         return 0