Skip to content

CHANGELOG.md

Latest commit

 

History

History
906 lines (590 loc) · 113 KB

CHANGELOG.md

File metadata and controls

906 lines (590 loc) · 113 KB
Raw

Changelog

All notable changes to this template are documented here. We follow Keep a Changelog and use loose semantic versioning: major when fork upgrades require manual migration, minor for new skills or features that are additive, patch for fixes and docs.

If you have forked this template, see the Upgrading section at the bottom for how to pull updates without losing your customizations.

v2.1.0 — 2026-06-10

A currency + citability release, driven by a 48-agent web-verified audit ("is this actually up to date and the best for economists, today?"). The architecture audit came back clean — the fixes are facts, not structure.

Changed — model refresh (Fable 5)

  • model-versions.md SSoT — Claude Fable 5 (GA 2026-06-09) added as the top tier: claude-fable-5 (alias fable), opt-in, $10/$50 per MTok, 1M context, falls back to Opus 4.8 on flagged content, needs Claude Code ≥ 2.1.170. Opus 4.8 demoted from "newest" to "current Opus tier; API/account default". Marker, table, and verified-date updated per the file's own protocol.
  • model-routing.md — new section "Where Fable 5 fits — and where it does not": the fleet deliberately stays on Opus/Sonnet/Haiku (2× price on the judgment tier + launch-week forced-tool-protocol failures observed at 28/28 vs 0 on Opus 4.8); Fable 5 recommended only for interactive long-horizon sessions. Re-evaluate at point releases.
  • check-model-versions.sh hardened — version regex generalized (4.x → any major[.minor], so bare "Fable 5" is tracked), Fable added to the tier loop, and a new superlative-drift check (flags "newest/most capable" claims naming a non-top tier) — the exact class of claim the Fable launch falsified while the old gate stayed green. SSoT update protocol gains a manual superlative-grep step.
  • README model-lineup bullet rewritten (Fable 5 = most capable, Opus 4.8 = default + routed tier); CONTRIBUTING's Co-Authored-By example made version-free.

Fixed — Claude Code doc currency (3 real bugs)

  • scheduled-routines.md: the /schedule example used a fabricated flag syntax (--cron/--prompt) — replaced with the real natural-language form + /schedule update for precise cron, the 1-hour minimum interval, the daily run cap, and the committed-repos-only constraint. The MCP guardrail was inverted: cloud Routines include all connectors with write access by default (the risk is a fully-armed connector, not a missing one) — rewritten to least-privilege guidance.
  • templates/skill-template.md + guide: allowed-tools was taught as a restriction; per current docs it is a pre-approval list ("does not restrict which tools are available: every tool remains callable"). The actual restrictor is disallowed-tools — now documented, with a read-only-skill pattern (disallowed-tools: ["Edit","Write","Bash"] + AskUserQuestion for unattended loops) and new paths / when_to_use / arguments rows. Security note rewritten; guide gains an "allowed-tools does not sandbox" warning callout.

Added

  • /submission-disclosures — the submission-time disclosure block: AI-use disclosure matched to the target journal's verified-current policy (fetched at draft time, not remembered), CRediT roles, COI, and data-availability statements. Two independent audits flagged this as the most economist-salient 2026 norms gap (AEA-family journals now mandate AI-use statements). Distinct from /disclosure-check (statistical disclosure).
  • CITATION.cff — the repo is now citable (GitHub "Cite this repository"); Zenodo DOI is a follow-up (backlog).
  • .github/SECURITY.md + .github/CODE_OF_CONDUCT.md — community-health files a public template shipping hooks + autonomous loops should have.
  • GitHub discoverability — repo topics (was: zero) + homepage URL set.

Inventory at release: 52 skills, 18 agents, 32 rules, 7 hooks (was 51 / 18 / 32 / 7 at v2.0.0).

v2.0.0 — 2026-06-09

A paradigm-shift major release. The template moves from a prompt-craft contractor — craft a prompt, invoke a skill, read a report — to a verification-gated research lab: you state a goal, a fleet of specialist agents does the labor under gates that enforce themselves, and you act as the auditor of the disagreements they surface. Two ideas converge: "loops, not prompts" (Boris Cherny / the Claude Code team) and "ground truth is a process, not a dataset" (Amazon Science). The result modernizes the orchestration, not the substance — the passport, simulation contract, and journal-calibrated referees are untouched. Shipped against quality_reports/plans/2026-06-09_v2.0-modernization-master-plan.md.

⭐ Why upgrade — read this even if you've used the template before

If you forked v1.x, here is what is materially different now (one-stop summary):

Before (v1.x) Now (v2.0)
Quality gate ran only inside /commit — a direct git commit bypassed it A real git pre-commit hook (./scripts/install-hooks.sh) runs surface-sync + quality (≥80) on every commit
The review loop was "a pattern, not a runtime" — prose a human triggered A real runtime: fan-out → reduce → judge + hallucination gate → loop-until-dry (orchestrator-protocol.md)
A numeric mismatch was always a FAIL (so people disabled the gate) A defensible, named alternative is recorded as EXPLAINED and flows into your response-to-referees; real errors stay fail-closed
Hooks nagged ("you should update the log / compile") Hooks act: the Stop hook auto-writes the session log; git-guardrails blocks destructive git; claim-reconcile flags stale claims the moment analysis changes
17 of 18 agents ran on model: inherit (cost savings unrealized) Every agent pinned to model + effort (Opus referees / Sonnet reviewers / Haiku mechanical)
You audited reproducibility You produce the deposit: /replication-package builds the AEA DCAS / openICPSR package
Onboarding said "paste this starter prompt" Onboarding is goal-first; the /prompt and /prompt-only skills are retired (shaping is now ambient)

If you do only one thing after pulling v2.0: run ./scripts/install-hooks.sh to activate the pre-commit gate.

Inventory at release: 51 skills, 18 agents, 32 rules, 7 hooks (was 38 / 18 / 28 / 6 at v1.10.0). Net: +15 skills, −2 retired; +4 rules; +2 hooks, −1 retired. New: 9 references (3 added), 2 output-styles, CI.

Added — economist "producer" skills (the biggest gold-standard gap)

  • /replication-package — assembles a submission-ready AEA DCAS / openICPSR deposit: standard README, dataset manifest, computational-requirements capture, a Table/Figure → script:line map (from the passport), and a confidential-data deposit note. Gates on /audit-reproducibility (blocks on FAIL, allows EXPLAINED). Move from auditing reproducibility to producing the deliverable.
  • /capture-environment — snapshots the computational environment: renv.lock + sessionInfo (R), requirements.txt/uv.lock (Python), Stata version + ado list, seeds/RNG, and an optional pinning Dockerfile.
  • /did-event-study — a thin wrapper over canonical staggered-DiD packages (Callaway–Sant'Anna did, Sun–Abraham fixest::sunab, HonestDiD; Stata twins) that surfaces each package's native diagnostics and never reimplements an estimator. Warns on the never-treated vs not-yet-treated control choice (the contested decision the EXPLAINED disposition exists for).
  • /power-analysis — power / required-N / MDE for two-arm RCTs (clustering/ICC), multi-arm, and simulation-based designs; feeds /preregister.
  • /disclosure-check — statistical-disclosure-limitation pre-screen for restricted-data outputs (small cells, dominance, PII); gates on CRITICAL.
  • /grant-proposal + /data-management-plan — compose the confidential-data + environment-capture primitives into NSF/NIH/ERC-shaped drafts + requirements checklists.
  • /coauthor-brief — a collaborator/multi-machine handoff brief (git delta, per-artifact state, reproduce-locally + restricted-data access steps).
  • confidential-data.md rule — restricted/IRB-data protocol (never commit raw data; disclosure clearance; restricted-data-safe multi-author git topology). AEA Data Editor / DCAS policy added to journal-profiles.md.

Added — teaching-at-scale (the repo's origin, previously a blind spot)

  • /syllabus, /teach-from-paper, /respond-to-eval (the teaching analogue of /respond-to-referees), and /scaffold-exercises (graded problem sets with worked solutions + explainers, adapted from mattpocock/skills).

Added — autonomy, runtime, and authoring

  • Real orchestration runtime — .claude/references/orchestration-schemas.md (the FINDING/SCORECARD/RUN_CONFIG contracts + the post-judge hallucination gate) and agent-fleet.md (the 18-agent manifest with tiers). orchestrator-protocol.md rewritten from pattern to runtime.
  • Event-driven + scheduled autonomy — claim-reconcile.py (flags stale numeric claims when an analysis script/output changes), .claude/references/scheduled-routines.md + scripts/nightly-repro-check.sh, and /triage-inbox (schedulable, human-gated email/calendar triage + referee-obligations tracker).
  • /new-skill — scaffolds a convention-compliant skill that passes check-skill-integrity on the first try (adapted from mattpocock/skills' write-a-skill).
  • /diagnose — root-cause a wrong/failing empirical result with a disciplined reproduce → minimise → hypothesise → instrument → fix loop (the net-new idea from a fresh pass over mattpocock/skills, reshaped for research code where the bug is usually a silent wrong number, not a crash; competing-hypothesis fan-out via Task, --no-fix to localize without editing).
  • Output styles — .claude/output-styles/academic-writing.md and referee.md (the template shipped zero before).
  • CI — .github/workflows/gates.yml (surface-sync + integrity + model-version checks on every PR) and deploy.yml (re-render the guide → docs/ on push to main).
  • Enforcing pre-commit hook — .githooks/pre-commit + scripts/install-hooks.sh (opt-in via core.hooksPath).
  • git-guardrails.py (PreToolUse) — blocks git reset --hard, git clean -f, git push --force, and blanket git add -A; warns on hardcoded machine paths in code (CLAUDE_STRICT_PATHS=1 to hard-block).
  • Table-row sync gate — check-surface-sync.py now verifies that the enumerative skills/agents tables (not just prose counts) carry one row per item on disk — the drift that hid the v1.5.0 peer-review trio for three releases.

Changed

  • Verification 2.0 — "ground truth is a process." audit-reproducibility and claim-verifier gain an EXPLAINED disposition: an out-of-tolerance numeric/citation mismatch stays a fail-closed FAIL unless the author records a concrete named alternative (e.g. "never-treated vs not-yet-treated comparison group"), in which case it is surfaced, non-blocking, and carried into the response-to-referees. The hard floor holds (fabricated citations, unmatched claims, vague notes never downgrade). audit-reproducibility also no longer presumes the manuscript is the oracle — a mismatch means "one of {paper, code} must change."
  • Loop-until-dry. qa-quarto, review-paper --adversarial, deep-audit, and seven-pass-review converge when a round surfaces no new CRITICAL/MAJOR finding (2 dry rounds), with the old "max 5 rounds" demoted to a fallback cap; synthesizers run the post-judge hallucination gate (an editor can't desk-reject on a reason no referee raised).
  • Hooks act instead of nagging. log-reminder.py now auto-writes a structured session-log entry on each meaningful change-set (was a stderr reminder). pre-compact.py's DRAFT-plan block is ON by default. The statusline now shows context-% + dirty-file count + plan status.
  • Per-agent model routing realized. All 18 agents pinned to model: + effort: per model-routing.md (9 Opus / 8 Sonnet / 1 Haiku).
  • Goal-first onboarding. The README "How It Works" leads with goal-first, gate-enforced; the guide and landing page reframed accordingly. "Contractor mode" remains, augmented.

Removed

  • /prompt and /prompt-only skills — retired. Prompt-craft is 2023-era; a loop-first workflow shapes every ambiguous request automatically, so this is now the ambient prompt-shaping.md rule, not a command. (prompt-formatting-core.md re-pointed; /interview-me remains for multi-turn specification.)
  • verify-reminder.py hook — retired. Its per-edit "compile reminder" is superseded by the Stop-hook completion note in the auto-writing session log.

Upgrading (forkers pulling v2.0)

  1. ./scripts/install-hooks.sh — activate the pre-commit gate (the single highest-value step).
  2. If you customized /prompt or /prompt-only, move that logic into your own rule or /interview-me — the skills are gone.
  3. If you reference verify-reminder.py in a fork, remove it; the Stop hook (log-reminder.py) now writes logs.
  4. Pin model:/effort: in any custom agents (model-routing.md), and re-run ./scripts/check-surface-sync.sh after adding skills — it now checks table rows too.
  5. No data migrations; the passport schema is backward-compatible (EXPLAINED is additive).

Notes

  • Count surfaces (README, CLAUDE.md, guide source + rendered HTML, landing page, skill template) updated to 51 / 18 / 32 / 7; check-surface-sync.sh (now including the table-row gate), check-skill-integrity.py, and check-model-versions.sh all pass.
  • Explicit non-goals (documented, not omissions): no autonomous daemon, no plugin marketplace, no multi-estimator production fleet (one /did-event-study proof-of-concept), no challenger→auditor→ledger pipeline (the zero-cost EXPLAINED mechanism instead), MEMORY.md stays the committed memory backend.

v1.10.0 — 2026-05-31

A hub-expansion + currency-refresh minor release. The template gains a Monte Carlo simulation capability and an R package-development release gate, refreshes the model / effort / cost guidance for Opus 4.8, and reframes itself as a hub for an entire research program — not just slides and papers. Shipped against the plan at quality_reports/plans/2026-05-31_v1.10.0-simulation-and-hub.md (local-only per the quality_reports/plans/* ignore rule). No breaking changes.

Inventory at release: 38 skills, 18 agents, 28 rules, 6 hooks (was 36 / 16 / 26 / 6 at v1.9.0). Adds 2 skills (/simulation-study, /r-package-check), 2 agents (sim-reviewer, r-package-reviewer), and 2 rules (simulation-conventions, r-package-conventions).

Added

  • /simulation-study skill (.claude/skills/simulation-study/) — scaffolds and runs a reproducible Monte Carlo study: a parameterized DGP, an estimator grid, a seeded replication loop (L'Ecuyer-CMRG streams for parallel reps), and a summary of bias, RMSE, empirical SE, coverage, and size/power — each reported with its Monte Carlo standard error. Mirrors /data-analysis; saves per-replication raw results + a summary table to scripts/R/_outputs/.
  • sim-reviewer agent (.claude/agents/sim-reviewer.md) — a read-only reviewer for the simulation-specific layer that general R review misses: DGP↔estimand alignment, replication budget + Monte Carlo SE, coverage computed against the truth (not the point estimate), parallel-seed discipline, and claims↔tables parity. Builds on r-reviewer (Cat 9 error handling, Cat 11 numerical discipline) instead of duplicating it.
  • simulation-conventions rule (.claude/rules/simulation-conventions.md) — path-scoped Monte Carlo discipline (the DGP/estimand contract, L'Ecuyer seeding, MCSE reporting, coverage-vs-truth, raw-result storage). Sibling to r-code-conventions.md; auto-loads only on simulation files.
  • /r-package-check skill (.claude/skills/r-package-check/) — the R package release gate: regenerates docs (devtools::document()), runs the test suite, runs R CMD check --as-cran, triages every ERROR / WARNING / NOTE against CRAN policy, then runs r-package-reviewer. Produces a CRAN-submission checklist. Does not bump versions or submit.
  • r-package-reviewer agent (.claude/agents/r-package-reviewer.md) — reviews package source (DESCRIPTION / dependency hygiene, NAMESPACE & imports, roxygen completeness, testthat coverage, CRAN-policy red flags). Builds on r-reviewer (general numerical discipline) instead of duplicating it.
  • r-package-conventions rule (.claude/rules/r-package-conventions.md) — path-scoped package-source standards (no library() in R/, roxygen-generated NAMESPACE, Imports vs Suggests, testthat 3e, CRAN-policy red flags, semver + NEWS). r-code-conventions.md now carries an explicit "analysis scripts, not package source" scope banner pointing here.
  • "One repo, many project types" framing (README + landing page) — positions the template as a hub for courses, papers, replication, simulation studies, and the R package gate, with Stata/Python package checks and personal-productivity workflows on the roadmap (.claude/references/v2.0-backlog.md).

Changed — Opus 4.8 currency refresh

Brings the model / effort / cost guidance up to date — every claim re-verified against official Anthropic docs on 2026-05-31:

  • Model lineup → Opus 4.8. README, the guide (lineup callout + routing table + rendered HTML), model-routing.md, TROUBLESHOOTING.md (retirement migration target), statusline.sh, and the v2.0 backlog now lead with Opus 4.8 (claude-opus-4-8, GA 2026-05-28, API default, $5/$25, 1M context) as the newest model; Opus 4.7 is named as the prior generation. Historical CHANGELOG entries are left intact.
  • Effort guidance. Opus 4.8 defaults to high — its high does roughly what 4.7's xhigh did, for fewer tokens. The old "xhigh recommended for 4.7 / default to medium" framing is replaced with "high is the default; reserve xhigh for the hardest runs; ultracode (xhigh + dynamic workflows) for repo-scale autonomous tasks." model-routing.md gains a new effort-axis section (effort is the first cost lever, model tier the second).
  • Cost / cache. Prompt-cache TTL framed correctly: 5-min default on API keys; 1-hour automatic on Claude Pro/Max subscriptions. The 70/20/10 routing split is unchanged (Sonnet 4.6 / Haiku 4.5 remain the current tiers).
  • Auto mode. Now described as on Team / Enterprise / API and rolling out to Max, gated on Opus 4.6+ or Sonnet 4.6 (was "Max + Opus 4.7").
  • Version pin updated to v1.10.0 (2026-05-31).

Fixed — hook output channels, non-fatal scoring, + a model-version drift gate

  • context-monitor.py + verify-reminder.py were silent on the user side. Both fire on PostToolUse and print() plain ANSI-colored text — which reaches Claude's context as literal escape-code noise and never reaches the user at all. Both now emit the proper PostToolUse JSON contract: a clean systemMessage (shown to the user) and hookSpecificOutput.additionalContext (injected into Claude's context), ANSI-free. Verified against the hooks reference.
  • context-monitor.py calibration. The old "context %" was a bare tool-call counter (MAX_TOOL_CALLS = 150) that cried wolf far too early against a 1M-token window. It now estimates tokens from the transcript_path size against CLAUDE_CONTEXT_WINDOW_TOKENS (default 1,000,000), falls back to a configurable counter (CLAUDE_CONTEXT_MAX_TOOL_CALLS, default 400), and labels the figure a coarse proxy.
  • New model-version drift gate. .claude/references/model-versions.md is now the single source of truth for current model point versions, and scripts/check-model-versions.sh flags any superseded version presented as current in user-facing surfaces (historical CHANGELOG + "prior generation"/comparison lines are allowed via markers). Wired into scripts/check-surface-sync.sh as a third pre-commit gate, so the ~6-week model-staleness class that prompted this release can't silently recur.
  • post-compact-restore.py ANSI on SessionStart. The post-compaction restore message printed ANSI color codes, which on SessionStart land in Claude's context as literal escape-code noise. It now emits a clean, ANSI-free message via the SessionStart additionalContext contract.
  • quality_score.py timeouts were fatal. A quarto render (120s) or Rscript (10s) timeout returned the same as a real compilation failure → auto_fail, score 0, commit blocked. Timeouts (and a missing tool) are now a distinct could-not-verify state: the score reflects the static checks and the report notes which check was skipped. Both timeouts are env-configurable (QUALITY_QUARTO_TIMEOUT, QUALITY_RSCRIPT_TIMEOUT).

Notes

  • Count surfaces (README, CLAUDE.md, guide source + rendered HTML, landing page, skill template) updated to 38 / 18 / 28 / 6; check-surface-sync.sh passes.
  • The entire audit P0 (Opus 4.8 currency refresh, the two silent-hook fixes, the model-versions.md drift gate, the post-compact-restore.py SessionStart cleanup, and the quality_score.py could-not-verify fix) was folded into this release. The main remaining roadmap item is Stata/Python package dev — see .claude/references/v2.0-backlog.md.

v1.9.0 — 2026-05-20

A guide-refresh + ecosystem catch-up minor release shipped in five passes against the plan at quality_reports/plans/2026-05-20_v1.9.0-guide-refresh.md (local-only; not tracked in git per the standard quality_reports/plans/* ignore rule). No breaking changes.

Inventory at release: 36 skills, 16 agents, 26 rules, 6 hooks (was 30 / 14 / 24 / 6 at v1.8.0). The release adds 6 skills (/humanize, /prompt, /prompt-only, /compress-session, /promote-memory, /stata-replication), 2 agents (humanize-auditor, promote-memory-council), and 2 rules (model-routing.md, stata-code-conventions.md).

Pass 1 (PR #114, merged 2026-05-20) — guide refresh mechanical corrections

Eight mechanical corrections that bring the guide in line with Anthropic shipments through May 2026 (Weeks 17–20), reframe three lingering "automatic orchestrator" mentions, refresh the clo-author citation to its current MAS v2 architecture, and add a Models / API section to TROUBLESHOOTING covering the 2026-06-15 Sonnet 4 + Opus 4 retirement and the Agent SDK credit-pool split. No new skills, agents, rules, or hooks. On-disk inventory unchanged.

Fixed

  • guide/workflow-guide.qmd + TROUBLESHOOTING.md + CHANGELOG.md — /less-permission-prompts → /fewer-permission-prompts. Verified against Anthropic's Week 16 changelog and Boris Cherny's launch announcement: the skill shipped under /fewer-permission-prompts from day one. The previous template name was a typo, not a rename — historical-changelog wording rewritten for accuracy.
  • guide/workflow-guide.qmd:46–157 — three lingering framings of orchestration as automatic ("Claude automatically: → Runs X, → Runs Y...") rewritten to name the invoked skill explicitly ("Claude invokes /slide-excellence, which internally..."). Reinforces the orchestrator-protocol rule — there is no repo-wide daemon; orchestration lives inside the invoked skill.

Added — Anthropic shipments since 2026-04-27

  • Model lineup callout (Multi-Model Strategy section) — documents Opus 4.7 as the Max/Team Premium default (GA 2026-04-16, same pricing as 4.6: $5/$25 per MTok), Sonnet 4.6 as workhorse (1M context still available; the 4.5 1M beta retired 2026-04-30), Haiku 4.5 as the fast tier. Sources: Opus 4.7 announcement, Platform release notes.
  • xhigh effort level (Effort Levels table + How-to-set list) — introduced Week 16 (Apr 13–17, v2.1.105–113); recommended for Opus 4.7 coding work. Bare /effort now opens an interactive slider. Hooks see effort.level and $CLAUDE_EFFORT is set in Bash subprocess env (Week 19).
  • Auto mode flag retirement (Permission Modes table) — --enable-auto-mode no longer required for Max + Opus 4.7 users as of Week 16.
  • /goal <verifiable condition> (Session Management + Anthropic Utilities) — shipped Week 20, v2.1.139 (May 11). "Keep working until X holds" command, distinct from /loop (interval) and plan mode (strategy approval). Pairs with /commit quality gates. Source: docs.claude.com/en/goal.
  • claude agents dashboard (Cost-Conscious Parallelism + Anthropic Utilities) — also Week 20, v2.1.139. Single screen for all background sessions; aligned with the /review-paper --peer parallel-referee pattern.
  • /loop alias /proactive noted (Week 16).
  • worktree.baseRef setting (Pattern 12) — Week 19. New callout explains fresh (default; branch from remote default) vs head (branch from local HEAD). Surprises users with uncommitted in-flight edits — documented inline.

Added — TROUBLESHOOTING entries (Models and API section)

  • Sonnet 4 / original Opus 4 retire 2026-06-15 — migration checklist covering ANTHROPIC_MODEL env, .claude/settings.json model overrides, agent frontmatter pins, and CI claude -p --model calls. Recommended replacements: Sonnet 4.6 and Opus 4.7 respectively.
  • Agent SDK credit-pool split (2026-06-15) — heads-up that claude -p headless subprocess calls (used by /coarse-review and any other claude -p-based skill) draw from a separate monthly Agent SDK credit pool after the cutover. Anthropic posts cited: Apr 8 "Decoupling Brain from Hands", Platform release notes.

Changed — ecosystem references

  • clo-author (Ecosystem section) — new callout documenting v26.05 (2026-05-10): MAS v2 (second-generation multi-agent system), Skill-Centric Restructure (13 skills + 18 agents, up from 17), HTML Dashboard. The /checkpoint attribution to v4.2.0 remains accurate as historical record (that's where the pattern was adapted from); the callout makes clear that current forkers of clo-author get the MAS v2 / skill-centric layout.

Provenance

This pass is the first slice of a research-grounded refresh. Four parallel research agents (Anthropic ecosystem, community repos, cross-vendor coding agents, internal guide audit) plus two verification agents (Anthropic claim verification, ARS source-code audit) ran on 2026-05-20. Findings, ranked recommendations, and source URLs are in quality_reports/plans/2026-05-20_v1.9.0-guide-refresh.md (local-only; not tracked in git per the standard quality_reports/plans/* ignore rule). Passes 2–5 add capability (B Cost & Caching section, C poli-sci breadth woven into guide body, D /preregister + /checkpoint patterns, E /review-paper --variance N, F passport.yaml claims provenance, G architect/editor cost-routing, H /promote-memory five-critic council, I HIGH-WARN claim-faithfulness blocking, J /compress-session, K SDK credit awareness, L Anthropic engineering-post citations, M /prompt port from Blattman, N /humanize detect-and-flag skill, plus Stata-MCP integration). All deferred to follow-up commits; nothing in this commit adds new on-disk infrastructure.

Verification

  • ./scripts/check-surface-sync.sh — 26/26 assertions pass; counts unchanged (30 skills / 14 agents / 24 rules / 6 hooks)
  • ./scripts/check-palette-sync.sh — palette in sync
  • ./scripts/check-skill-integrity.py — all checks pass

Pass 2A — guide-only capability adds (2026-05-20)

Two guide-body additions: weaves the v1.8.0 political-science breadth into the user-facing guide body (previously documented only in the v1.8.0 CHANGELOG entry), and introduces a new "Cost-Conscious Composition" subsection that documents prompt-cache TTL behaviour, per-agent model routing (70/20/10 pattern), and /cost / /usage monitoring. No new skills, agents, rules, or hooks. On-disk inventory still unchanged.

Added — political-science breadth surfaced in guide body

  • 5-Lens Framework table — added Political Science as a third worked example column alongside Economics and Physics. Lens entries reflect poli-sci norms (ignorability, conjoint AMCEs under Hainmueller–Hopkins–Yamamoto, cjoint/survey::svyglm defaults, manipulation-check fidelity).
  • Domain-reviewer ship status callout — documents that the template ships two concrete customizations of .claude/agents/domain-reviewer.md (econ + poli-sci), both viable starting points for forkers' own fields.
  • /review-paper --peer [journal] callout (Research Skills section) — new explainer covering the editor + 2-referee + editorial-decision pipeline, the journal-profiles starter set (AER, QJE, JPE, ECMA, JoE, APSR, AJPS, JOP), and the full 6-paper-type methods-referee taxonomy with sanity checks per type (reduced-form, structural, theory+empirics, descriptive, formal-theory, survey-experiment). The v1.8.0 formal-theory + survey-experiment additions are now discoverable from the main guide, not just from the CHANGELOG.
  • "What ships preloaded vs. what you customize" callout (§7 Customizing for Your Domain) — explains what already ships for econ + poli-sci (journal profiles, paper types, discipline cards) and what forkers add for psych / sociology / public-health (~3 profiles + 2–3 paper types + 1 card per discipline). Points at .claude/references/v1.9-backlog.md as the candidate-next-breadth roadmap.

Added — Cost-Conscious Composition subsection (§ between Multi-Model Strategy and Permissions)

  • Prompt-cache TTL guidance — documents the 60 min → 5 min default-TTL change and the opt-in ENABLE_PROMPT_CACHING_1H environment variable on API/Bedrock/Vertex/Foundry plans. Cites Apr 2026 community cost-analysis showing 30–60% effective input-cost increase for long pipelines without the 1-hour opt-in. Documents cache_miss_reason (public beta 2026-05-13) as the diagnostic.
  • 70/20/10 model-routing pattern — explicit table of Haiku (70%, mechanical work) / Sonnet (20%, review and critique) / Opus (10%, high-judgment work). Maps each tier to specific template agents (TikZ extraction → Haiku; r-reviewer → Sonnet; editor/methods-referee → Opus). Cites Anthropic's Apr 8 2026 "Decoupling brain from hands" post for primary-source endorsement.
  • Effort-budgeting guidance — reach for xhigh (Opus 4.7) only when verified that the cheaper tier failed.
  • /cost and /usage monitoring callout — when to run each and how to read cache hit-rate as a TTL-misconfiguration signal.
  • Agent SDK credit-pool 2026-06-15 callout — links to the TROUBLESHOOTING Models / API section landed in Pass 1.

Verification — Pass 2A

  • ./scripts/check-surface-sync.sh — 26/26 assertions pass; counts still unchanged (30/14/24/6)
  • ./scripts/check-skill-integrity.py — all checks pass
  • quarto render guide/workflow-guide.qmd — clean render
  • python3 scripts/quality_score.py guide/workflow-guide.qmd — 100/100 [EXCELLENCE]

Pass 2B — /preregister + /checkpoint promoted to Pattern entry (2026-05-20)

Promotes the v1.8.0 /preregister and /checkpoint skills from appendix entries to a first-class Pattern 16: Preregistration and Submission Discipline in the Workflow Patterns section. ~85 lines of new prose. Still no new skills, agents, rules, or hooks — Pattern 16 documents what already ships.

Added — Pattern 16: Preregistration and Submission Discipline

  • When-preregistration-matters checklist — launch of experiment, observational analysis before outcome inspection, R&R PAP request, funding pre-submission.
  • Registry-comparison table — OSF / AsPredicted / AEA RCT side-by-side (field, length, editability). Public-health / clinical-trial registries explicitly out of scope (on v1.9-backlog).
  • End-to-end workflow diagram — /interview-me spec → /preregister --style → user uploads → /checkpoint preregistration-submitted → data collection begins.
  • Per-registry mandatory-field matrix — 10 rows × 3 styles, showing MUST / SHOULD / MAY for each field. Captures registry-specific differences (e.g., AEA mandates intervention description and data-sharing plan; AsPredicted allows looser covariate specification).
  • /checkpoint pairing protocol — explicit pre-submit → submitted [registration-id] → data-arrived snapshot sequence so the boundary between confirmatory and exploratory is structurally visible.
  • Post-flight /verify-claims integration — every cited reference in the preregistration runs through Chain-of-Verification before submission. Hallucinated citations are the most common preregistration failure mode.
  • Anti-pattern protections enumerated — HARKing, p-hacking, forking paths, selective reporting — and which fields of the preregistered document defend against each.
  • Field-specific guidance — pointers to Christensen & Miguel (2018) for development economics, Monogan (2015) for political science, OSF Preregistration Guide for psychology.

Verification — Pass 2B

  • ./scripts/check-surface-sync.sh — 26/26 assertions pass; counts unchanged (30/14/24/6)
  • ./scripts/check-skill-integrity.py — all checks pass
  • quarto render guide/workflow-guide.qmd — clean render
  • python3 scripts/quality_score.py guide/workflow-guide.qmd — 100/100 [EXCELLENCE]

Pass 2C — /review-paper --variance N reviewer-disposition variance mode (2026-05-20)

Adds a fourth --peer mode to /review-paper: --variance (with integer N, default 3) runs N referees with independently sampled dispositions from the 6-way taxonomy, then the editor synthesizes the results into a decision distribution (not a point estimate). Motivated by AgentReview (ACL 2024, arXiv:2406.12708), which found ~37% of paper decisions vary purely from reviewer-disposition sampling.

Changed — /review-paper SKILL.md

  • New --variance flag in argument-hint and the sub-flags list under "Peer-review mode."
  • New "Variance mode (--peer --variance N)" subsection: motivation (AgentReview), procedure (sample with replacement + stratification override that always includes a SKEPTIC if none drawn for N ≥ 3), output files (referee_1.md … referee_N.md, decision_distribution.md, editor_synthesis.md), cost discipline (referee-tier cost × N relative to default --peer; hard cap N=5; route referees to Sonnet for variance runs), and mutual-exclusivity rules (cannot combine with --stress or --r2/--r3).
  • When-to-reach-for-it guidance: pre-submission "how confidently will this survive," cross-journal target comparison, post-rejection sanity check.

Changed — editor.md agent

  • Phase 1b (Referee selection) now branches: default 2-referee deliberate-diversity sampling (existing) vs. variance-mode N-referee independent sampling with replacement (new). Stratification rule: if N ≥ 3 and the realised draw includes no SKEPTIC, replace one randomly chosen referee with a SKEPTIC. Realised-disposition table and stratification-override metadata are recorded for decision_distribution.md.
  • New "Variance synthesis mode" section documents the two-file output (decision_distribution.md + editor_synthesis.md). The distribution file contains a realised-disposition table, a verdict-distribution table (counts and shares), a concern-frequency table (K-of-N), and an interpretation guide (tight modal majority + robust concerns; wide spread + high-K skeptic objection; bimodal "love-it-or-hate-it"). The editorial letter explicitly references the variance instead of collapsing it.

Verification — Pass 2C

  • ./scripts/check-surface-sync.sh — 26/26 assertions pass; counts unchanged (30/14/24/6)
  • ./scripts/check-skill-integrity.py — all checks pass (forward + reverse flag-parity OK after one cycle of integrity feedback — original `--variance N` single code-span fixed to --variance standalone code-span with N documented in prose)
  • No on-disk inventory change

Pass 2D — /humanize detect-and-flag skill (2026-05-20)

Ships a new skill and a new agent to audit academic prose for AI-voice tells. Detect-only by design — there is no --rewrite mode. The author edits manually after reading the report; auto-rewriting prose to strip AI tells degrades quality (cross-vendor research finding) and introduces new AI tells.

Added — new skill

  • .claude/skills/humanize/ — /humanize [file] [--severity low|med|high] runs a read-only audit on .tex, .qmd, or .md prose against 10 detection categories: (1) boilerplate transitions, (2) AI-cliché lexicon, (3) em-dash/punctuation overuse, (4) symmetric paragraph shapes, (5) tricolon abuse, (6) hedging stacking, (7) "Not only X, but also Y" frames, (8) formulaic openers, (9) hyphenation excess, (10) sycophancy/self-important framing. Output: quality_reports/humanize_<filename>_report.md (gitignored) with per-finding line numbers, severity (LOW / MED / HIGH), and suggested rewrites. Carries disable-model-invocation: true. Documents the no---rewrite anti-pattern explicitly in the SKILL body.

Added — new agent

  • .claude/agents/humanize-auditor.md — read-only auditor that runs the 10-category protocol and returns a structured report (per-category counts, per-finding table, concentration analysis identifying the top-3 most-affected paragraphs, recommendation thresholds for "rewrite" vs "strip" vs "cosmetic"). Calibration heuristics built in (~1000 words: 0 HIGH clean; 3 HIGH manageable; 8+ HIGH recommend rewrite). Reads style-profile.md if present to respect documented author preferences. Will not auto-flag single-mention idioms or discipline-legitimate constructions.

Changed — count-bearing surfaces

  • Inventory: 31 skills, 15 agents, 24 rules, 6 hooks (was 30 / 14 / 24 / 6). All count-bearing surfaces updated: README.md, CLAUDE.md, guide/workflow-guide.qmd (×3 places: capability table, You-Don't-Need-All-Of-This callout, Customizing-Skills section), docs/index.html (landing page), templates/skill-template.md. Verified via ./scripts/check-surface-sync.sh.
  • CLAUDE.md Skills Quick Reference — added /humanize row.
  • guide/workflow-guide.qmd Appendix — added Humanize Auditor row to All Agents; added /humanize row to All Skills.

Verification — Pass 2D

  • ./scripts/check-surface-sync.sh — 26/26 assertions pass; counts now 31 / 15 / 24 / 6
  • ./scripts/check-skill-integrity.py — all checks pass on the new SKILL.md
  • quarto render guide/workflow-guide.qmd — clean render

Pass 3A — claims provenance + HIGH-WARN gate + /prompt port (2026-05-20)

Strategic additions to the paper-pipeline lens: (a) machine-readable claims provenance via a new passport.yaml contract, (b) HIGH/MED/LOW-WARN severity tiers on /verify-claims with HIGH-WARN gate-refusing /commit, and (c) two new skills (/prompt, /prompt-only) ported from Chris Blattman's claudeblattman v2.1 with the Blattman-specific tool-routing and council elements stripped.

Added — new template

  • templates/passport-template.yaml — starter passport file for per-paper claims provenance. Forkers copy once per paper to quality_reports/passports/<paper-slug>.yaml. Schema records claim text, manuscript location, source script + line, output file + field, tolerance overrides, last-verified timestamp, and PASS / FAIL / STALE / UNVERIFIED status per claim.

Added — new skills (2)

  • .claude/skills/prompt/ — /prompt [text] [depth:light|standard|deep] reformats an informal or dictated request into a structured six-section prompt (Role / Task / Context / Constraints / Output format / Bookend) then executes it immediately. Depth heuristic: Light (< 40 words, no jargon) emits Role + Task + Output + Bookend; Standard (40–200 words, 1+ domain term) adds Context + Constraints + Assumptions block; Deep (> 200 words, specific paper/dataset/submission target) adds Investigation pre-step. Always shows the formatted prompt to the user before executing.
  • .claude/skills/prompt-only/ — /prompt-only [text] [depth] [--save path] is the same skill without execution. Emits the formatted prompt as a reusable artifact. Optional --save writes to disk. Use for prompts you'll run later (different conversation, different model, recurring task).

Both ported with attribution from chrisblattman/claudeblattman v2.1. Deliberately stripped: Blattman's tool-routing table (ChatGPT / Perplexity / Gemini dispatch) and his council token (his /council skill is not in this template). Documented in .claude/references/prompt-formatting-core.md "What we don't ship" section.

Added — new reference

  • .claude/references/prompt-formatting-core.md — shared reference used by /prompt and /prompt-only. Defines the six-section skeleton (Role / Task / Context / Constraints / Output format / Bookend), the depth-calibration heuristic, a worked example, and the explicit boundary with /interview-me (single-shot input shaping vs. multi-turn project specification).

Changed — claims provenance integration

  • .claude/rules/replication-protocol.md — new "Claims Provenance: passport.yaml" section. Documents the YAML schema (paper metadata + per-claim entries with id, claim, location, source_file, source_line, output_file, output_field, tolerance, last_verified_on, status, notes), the status semantics (PASS / FAIL / STALE / UNVERIFIED), integration with /audit-reproducibility (passport-mode), /commit (advisory by default, gate-refuse on --strict-passport), and /review-paper (summary section). Attributes the pattern to Imbad0202/academic-research-skills "Material Passport" while scoping our schema to numeric-claim provenance only (theirs threads ~13 contracts through ~6 agents; ours stays narrow).
  • .claude/skills/audit-reproducibility/SKILL.md — new "Passport-mode (v1.9.0)" section. When a passport file exists, the skill reads + updates + rewrites it in place rather than emitting a one-shot report. Updates status per claim (PASS / FAIL / STALE), records discrepancies in notes, refreshes timestamps, refuses to auto-populate (passport scope is author-curated to avoid bad inferences).

Changed — HIGH-WARN claim-faithfulness tier

  • .claude/skills/verify-claims/SKILL.md Phase 4 — three severity tiers introduced (HIGH-WARN: fabricated reference / direct contradiction / not-found retrieval interpreted as hallucination; MED-WARN: transient retrieval failure; LOW-WARN: source genuinely inaccessible). Tier aggregation table maps to PASS / PARTIAL / FAIL outcomes. HIGH-WARN gate-refuses /commit for files the skill was just run against unless --no-fail-closed or verifyClaims.allowHighWarn: true in settings.
  • .claude/agents/claim-verifier.md — output schema extended to include a per-claim Tier column (HIGH / MED / LOW / —). New "Tier-assignment rules" section codifies the assignment logic: fabricated citation → HIGH; numerical contradiction → HIGH; directional contradiction → HIGH; transient retrieval failure → MED; genuine inaccessibility → LOW; reasonable paraphrase → no tier. "Be conservative on HIGH-WARN" — false positives erode the gate's authority.

Changed — count-bearing surfaces

  • Inventory: 33 skills, 15 agents, 24 rules, 6 hooks (was 31 / 15 / 24 / 6 after Pass 2D). Both new skills (/prompt, /prompt-only) propagated through README.md, CLAUDE.md, guide capability table, You-Don't-Need-All-Of-This callout, Customizing-Skills section, docs/index.html, templates/skill-template.md, guide appendix tables.
  • CLAUDE.md Skills Quick Reference — added /prompt and /prompt-only rows.
  • Guide Appendix — added /prompt and /prompt-only rows to All Skills table.

Verification — Pass 3A

  • ./scripts/check-surface-sync.sh — 26/26 assertions pass; counts now 33 / 15 / 24 / 6
  • ./scripts/check-skill-integrity.py — all checks pass on both new SKILL.md files
  • quarto render guide/workflow-guide.qmd — clean render
  • python3 scripts/quality_score.py guide/workflow-guide.qmd — 100/100 [EXCELLENCE]

Pass 3B — per-agent model routing + /compress-session + /promote-memory (2026-05-20)

Three additions to the cost-discipline and memory-management lenses:

  • G — new .claude/rules/model-routing.md codifying the 70/20/10 architect/editor split (Haiku for mechanical work; Sonnet for review/critique; Opus for high-judgment work). Cites the Anthropic Apr 8 "Decoupling brain from hands" post for primary-source endorsement.
  • J — new /compress-session skill that distils a long-session into a structured note before auto-compaction (motivated by Drew Breunig's "How Long Contexts Fail" four-mode taxonomy). Distinct from /checkpoint (explicit stop-point) and from auto-compaction (lossy truncation).
  • H — new /promote-memory skill that runs a five-critic council (generality / staleness / redundancy / evidence / format) on candidate [LEARN] entries in .claude/state/personal-memory.md to decide what graduates to MEMORY.md. Pattern adapted from Chris Blattman's claudeblattman v2.1 five-critic council with attribution.

Added — new rule

  • .claude/rules/model-routing.md — path-scoped on .claude/agents/**/*.md and .claude/skills/**/SKILL.md. Documents the 70/20/10 pattern with per-tier recipe (mechanical → Haiku 4.5; review/critique → Sonnet 4.6; high-judgment → Opus 4.7). Tags agents per tier (e.g., editor / methods-referee / claim-verifier / quarto-critic → Opus; r-reviewer / slide-auditor / proofreader / humanize-auditor → Sonnet; quarto-fixer / mechanical TikZ extraction → Haiku). Documents two anti-patterns: pushing Opus down a tier (defeats hallucination defence) and the self-as-architect-and-editor pairing (same-model self-pairings produce correlated errors).

Added — new skills (2)

  • .claude/skills/compress-session/ — /compress-session [slug] distils the current session into a structured note (Active state, Decisions made, Files touched, Open questions, Next actions, Discarded as noise, Proposed [LEARN] entries) and writes to quality_reports/session_logs/YYYY-MM-DD_compression_<slug>.md. Defends against Drew Breunig's four failure modes: poisoning, distraction, confusion, clash. Explicit "Discarded as noise" section is the novel contribution — failed hypotheses don't carry forward as ghost context. Optional PreCompact-hook integration that surfaces a reminder (does NOT auto-invoke — preserves the user's review step).
  • .claude/skills/promote-memory/ — /promote-memory [filter] runs a five-critic council in parallel (forked contexts via Task) on candidate [LEARN] entries. Each critic votes YES/NO on one dimension; majority (3+ of 5) promotes. Critics run on Haiku by default (per model-routing.md); the user is the final gate even on 5-of-5 unanimous votes. Saves an audit file at quality_reports/memory_promotion_<date>.md for forensics. Pattern adapted from Chris Blattman's claudeblattman v2.1 with attribution.

Added — new agent

  • .claude/agents/promote-memory-council.md — implements the five-critic protocol. One agent file, five role specs (Generality / Staleness / Redundancy / Evidence / Format), dispatched in parallel via Task with context: fork. Each critic returns a strict **Vote:** YES | NO + one-sentence rationale; the calling skill aggregates. Architectural isolation: each critic sees only its dimension's relevant context (no cross-pollination, no groupthink).

Changed — count-bearing surfaces

  • Inventory: 35 skills, 16 agents, 25 rules, 6 hooks (was 33 / 15 / 24 / 6 after Pass 3A). All 6 count-bearing surfaces updated: README.md, CLAUDE.md, guide capability table, You-Don't-Need-All-Of-This callout, Customizing-Skills section, docs/index.html, templates/skill-template.md, guide appendix tables.
  • CLAUDE.md Skills Quick Reference — added /compress-session and /promote-memory rows.
  • Guide Appendix — added Promote-Memory Council row to All Agents, /compress-session + /promote-memory rows to All Skills, Model Routing row to path-scoped All Rules.

Verification — Pass 3B

  • ./scripts/check-surface-sync.sh — 26/26 assertions pass; counts now 35 / 16 / 25 / 6
  • ./scripts/check-skill-integrity.py — all checks pass on both new SKILL.md files
  • quarto render guide/workflow-guide.qmd — clean render
  • python3 scripts/quality_score.py guide/workflow-guide.qmd — 100/100 [EXCELLENCE]
  • ./scripts/check-skill-integrity.py — all checks pass on both new SKILL.md files
  • quarto render guide/workflow-guide.qmd — clean render
  • python3 scripts/quality_score.py guide/workflow-guide.qmd — 100/100 [EXCELLENCE]

Pass 4 — Stata expansion: stata-mcp + /stata-replication + audit-reproducibility extension (2026-05-20)

User-driven expansion of the template to support Stata-first projects (correction to the original v1.9.0 plan: AEA does not mandate Stata; the expansion targets users whose pipelines are Stata-first for reasons of audience reach or original-replication-package fidelity). Three sub-items:

Added — new skill

  • .claude/skills/stata-replication/ — /stata-replication [paper-or-data] scaffolds a numbered Stata pipeline (00_install.do through 99_run_all.do) in scripts/stata/ and executes via the stata-mcp MCP server. Mirrors /data-analysis for R-first projects. --from-r flag ports an existing R pipeline. --no-execute produces scaffolding only. Phase 0 pre-flight halts if stata-mcp is not registered (with install instructions); Phase 4 (optional) runs R cross-check on --from-r translations to surface clustering-df / default-option drift.

Added — new rule

  • .claude/rules/stata-code-conventions.md — path-scoped on **/*.do and scripts/stata/**. Codifies: standard header (version 18, clear all, set seed, set sortseed, cap log close, log capture); numbered pipeline (00–99); outputs convention (scripts/stata/_outputs/ with sessionInfo.txt); esttab for publication-ready tables with \input{} mechanical-update pattern; significance-stars convention (* 0.10 ** 0.05 *** 0.01); clustering / SE discipline (reghdfe, cluster bootstrap for < 50 groups); balance + attrition via iebaltab; graph export to vector + raster; AEA Data Editor compliance checklist. Common Stata-to-R-or-AEA traps table.

Changed — /audit-reproducibility source-language coverage

  • .claude/skills/audit-reproducibility/SKILL.md — new "Source-language coverage" section documents the three supported ecosystems (R / Stata / Python) with default outputs directories and read-output method per language. Stata-specific notes added: .dta outputs read via haven::read_dta() (R) or pyreadstat.read_dta() (Python); esttab .tex table-cell values as strongest provenance signal when manuscript uses \input{}; clustering-df discrepancy diagnosis (reghdfe vs reg, cluster()).

Added — TROUBLESHOOTING entry

  • TROUBLESHOOTING.md — new "/stata-replication halts at 'stata-mcp not registered'" section. Documents the one-command install (claude mcp add stata-mcp --scope user -- uvx stata-mcp), the uv prerequisite, and the verify step (claude mcp list).

Added — Ecosystem entry

  • guide/workflow-guide.qmd Ecosystem section — new "stata-mcp: MCP server for Stata execution" subsection (before ClaudeCodeTools). Documents SepineTam's project, install command, why it matters for v1.9.0's Stata expansion, and the parallel-skill relationship between /data-analysis (R) and /stata-replication (Stata).

Changed — count-bearing surfaces

  • Inventory: 36 skills, 16 agents, 26 rules, 6 hooks (was 35 / 16 / 25 / 6 after Pass 3B). All 6 count-bearing surfaces updated. Note: rules count rises only by 1 (the Stata convention rule); skills count rises only by 1 (/stata-replication itself); agents unchanged.
  • CLAUDE.md Skills Quick Reference — added /stata-replication row.
  • Guide Appendix — added /stata-replication row to All Skills, Stata Code Conventions row to path-scoped All Rules.

Verification — Pass 4

  • ./scripts/check-surface-sync.sh — 26/26 assertions pass; counts now 36 / 16 / 26 / 6
  • ./scripts/check-skill-integrity.py — all checks pass on the new SKILL.md
  • quarto render guide/workflow-guide.qmd — clean render
  • python3 scripts/quality_score.py guide/workflow-guide.qmd — 100/100 [EXCELLENCE]

Pass 5 — verification follow-ups: SDK credit awareness + Anthropic engineering-post citations (2026-05-20)

Closes the v1.9.0 cycle with two small but load-bearing additions:

K — Agent SDK credit-pool split (2026-06-15)

Already partially shipped in Pass 1 (TROUBLESHOOTING "Models and API" section) and Pass 2A (Cost-Conscious Composition callout in the guide). Pass 5 confirms the coverage is complete — the /coarse-review skill is a Claude Code plugin (not in this template's .claude/skills/), so no SKILL.md callout was needed; the TROUBLESHOOTING entry and the guide callout already document the cutover, the affected pattern (claude -p headless subprocesses), and the diagnostic ("check the Agent SDK credit balance separately if /coarse-review fails with credit-exhaustion errors even though your interactive session works"). K is closed.

L — Anthropic engineering-post citations

  • Apr 8, 2026 — "Scaling Managed Agents: Decoupling the brain from the hands." Already cited as primary-source endorsement of the architect/editor split in .claude/rules/model-routing.md (Pass 3B) and in the guide's Cost-Conscious Composition section (Pass 2A). No additional citation needed.
  • Apr 23, 2026 — "An update on recent Claude Code quality reports." New callout added to the guide between the "Mandatory Verification" / "Don't Skip Verification" section and "Creating Your Own Domain Reviewer." Frames the post as a reminder that model quality can regress and that the template's verification patterns (/verify-claims with forked verifier, /audit-reproducibility with passport.yaml, cross-artifact review, HIGH-WARN gate-refuse, /review-paper --variance N) all assume drift rather than treating any specific checkpoint as a stable baseline.

Verification — Pass 5 (and final v1.9.0)

  • ./scripts/check-surface-sync.sh — 26/26 assertions pass; counts at release: 36 skills / 16 agents / 26 rules / 6 hooks
  • ./scripts/check-skill-integrity.py — all checks pass
  • quarto render guide/workflow-guide.qmd — clean render
  • python3 scripts/quality_score.py guide/workflow-guide.qmd — 100/100 [EXCELLENCE]

Release summary

Pass PR Net additions to disk What it shipped
1 #114 0 Mechanical guide refresh, Anthropic Apr–May 2026 catch-up, clo-author v26.05 citation, Models / API TROUBLESHOOTING section
2A #115 0 Poli-sci breadth woven into guide body, Cost-Conscious Composition subsection
2B #116 0 Pattern 16: Preregistration and Submission Discipline
2C #117 0 /review-paper --variance N reviewer-disposition variance mode
2D #118 +1 skill, +1 agent /humanize detect-and-flag for AI-voice tells
3A #119 +2 skills, +1 template passport.yaml claims provenance, HIGH-WARN claim-faithfulness gate, /prompt + /prompt-only port from Blattman
3B #120 +2 skills, +1 agent, +1 rule Model-routing rule (70/20/10), /compress-session, /promote-memory five-critic council
4 #121 +1 skill, +1 rule Stata expansion: /stata-replication, stata-code-conventions.md, audit-reproducibility source-language coverage, stata-mcp ecosystem entry
5 (this PR) 0 Apr 23 Anthropic post citation in verification section; closes SDK-credit awareness loop

Total v1.9.0 additions: 6 skills, 2 agents, 2 rules, 1 reference, 1 template. No breaking changes. All count-bearing surfaces verified in sync. Provenance: every addition traceable to the research-grounded plan at quality_reports/plans/2026-05-20_v1.9.0-guide-refresh.md (local-only).

v1.8.0 — 2026-04-27

A disciplinary breadth + audit-hardening + Apr 2026 incorporation minor release. The cycle landed in two passes: (1) infrastructure-only audit-hardening (mechanical parity checks via check-skill-integrity.py, living pet-peeves catalogue, PreCompact blocking, Routines awareness) and (2) capability work (two new skills /checkpoint and /preregister, political-science breadth via three journal profiles + two paper types + a discipline-cards reference, and Apr 2026 documentation: auto mode promotion, protected-paths gate explainer, session-management commands, Computer Use sidebar, Monitor tool integration, disable-model-invocation discipline). No breaking changes; counts updated across all monitored surfaces.

Added — new skills

  • .claude/skills/checkpoint/ — /checkpoint produces a structured state snapshot (active plan, recent decisions, file pointers with line numbers, open questions, next 1–3 actions) into quality_reports/checkpoints/YYYY-MM-DD_<slug>.md. Companion to (NOT replacement for) the narrative session-log workflow under quality_reports/session_logs/. Carries disable-model-invocation: true (writes to a persistent state file — must be user-intent). Pattern adapted from Hugo Sant'Anna's clo-author v4.2.0 with permission; reimplemented in original prose against this template's narrative-session-log + plan-on-disk + auto-memory architecture. Attribution header on the SKILL.md.
  • .claude/skills/preregister/ + templates/preregistration-template.md — /preregister drafts a registry-ready preregistration document in OSF, AsPredicted, or AEA RCT Registry style. Extracts hypotheses, design, sampling plan, exclusions, and analysis plan from a research spec (/interview-me output) or free-form description. MUST/SHOULD/MAY clarity annotation per section. Pre-flight cross-checks: directional hypothesis, named estimator, ex-ante exclusion rules, sample-size stopping rule. Post-flight CoVe verification of any cited literature (re-uses /verify-claims). Output: quality_reports/preregistrations/YYYY-MM-DD_<slug>.md (gitignored). Refuses retrospective preregistration (description containing realised results). Greenfield — not ported.

Added — political-science breadth

  • .claude/references/journal-profiles.md +3 profiles: APSR (highest theoretical bar; THEORY-disposition pool weight 0.30; methods-referee tilts toward formal-theory comparative-static sharpness), AJPS (methods-emphasis; CREDIBILITY 0.30, replication policy enforced), JOP (clarity-of-contribution bar; SKEPTIC 0.25). Each profile is ~40–45 lines following the existing schema.
  • .claude/agents/methods-referee.md +2 paper types: formal-theory (pure theory; weights tilt toward model originality and comparative-static sharpness; sanity checks include equilibrium existence, assumption tractability, robustness to assumption relaxation) and survey-experiment (vignette/conjoint/list/factorial; weights tilt toward design, sampling, and attrition+manipulation checks; sanity checks include balance, manipulation-check pass rate, attrition asymmetry, sampling-frame validity). The 4 prior paper types (reduced-form / structural / theory+empirics / descriptive) are unchanged — additions are purely additive.
  • .claude/agents/domain-reviewer.md template-marker comment now ships two customization examples (econ + poli-sci) to illustrate that the 5-lens structure is field-agnostic. Lens content under each example reflects field-specific norms (poli-sci: ignorability, conjoint AMCE algebra, Hainmueller-Hopkins-Yamamoto for cross-reference, cjoint/survey::svyglm package defaults).
  • .claude/references/discipline-cards.md — new reference. Two cards: econ and poli-sci. Each card: paper-type frequency table, dominant journals (cross-referenced to journal-profiles.md), preregistration norms (cross-referenced to /preregister --style), method conventions (significance-stars conventions, SE conventions, dominant code language). Read by /research-ideation, /interview-me, /preregister, and the editor agent when paper-type or discipline is given without a target journal. Forkers extend for psych / sociology / public-health / etc.

Changed — research-side skill paper-type awareness

  • .claude/skills/research-ideation/SKILL.md — new Step 3 tags each generated RQ with a likely paper type from the 6-type taxonomy. Output format adds a **Paper type:** field per RQ. discipline-cards.md informs the default distribution.
  • .claude/skills/interview-me/SKILL.md — Phase 1 ("Big Picture") now optionally asks the researcher what kind of paper they envision (same 6-type taxonomy). Saved to spec frontmatter as paper_type: so downstream skills (/preregister, /data-analysis, /review-paper --peer) can read it.

Added — Apr 2026 feature documentation

  • TROUBLESHOOTING.md +2 sections under Permissions:
    • Bypass mode still prompts on protected paths — explains the protected-path list (.git, .vscode, .idea, .husky, .claude with carve-outs for commands/agents/skills/worktrees) and that auto mode is the only mode that routes protected paths through the classifier instead of prompting. Documents the auto-mode requirements (Max/Team/Enterprise/API + Sonnet 4.6 / Opus 4.6 / Opus 4.7 + Anthropic API). Includes two workarounds for users without auto-mode access: edit through Bash (python3 heredoc) or move edits out of .claude/.
    • .vscode/settings.json key typo — claudeCode.allowDangerouslySkipPermissions is silently ignored; the canonical key is allowDangerouslySkipPermissions (no claudeCode. prefix). The typo leaves the protected-paths gate active even with broad CLI bypass. Reload window after fixing.
  • .vscode/settings.json — fixed the typo: allowDangerouslySkipPermissions: true (was incorrectly prefixed). Takes effect on next VSCode window reload.

Changed — disable-model-invocation audit

  • .claude/skills/create-lecture/, .claude/skills/new-diagram/, .claude/skills/learn/ — added disable-model-invocation: true to frontmatter. Rationale: each writes a load-bearing persistent file (a new lecture .tex, a new TikZ source, a new SKILL.md respectively) that should only be created on explicit user intent. /deep-audit also gains the flag in this release (its body writes audit reports + applies fixes). The new /checkpoint and /preregister skills carry the flag too.

Added — Apr 2026 doc additions (guide + onboarding)

  • guide/workflow-guide.qmd — new ### Session Management subsection under "Settings — Permissions and Hooks": /btw for side questions outside conversation history, /rewind and Esc+Esc for checkpoint navigation, /clear and /compact <instruction> for context resets, Ctrl+G for in-editor plan editing, claude --continue / --resume / /rename for cross-session continuity, plus /checkpoint <slug> (this template). Three composition patterns documented.
  • guide/workflow-guide.qmd — added auto mode row to the permission-modes table; documented bypass mode's protected-path gate inline (.git, .vscode, .idea, .husky, .claude minus commands/agents/skills/worktrees carve-outs).
  • guide/workflow-guide.qmd — Computer Use callout in the Adversarial Pattern section (Apr 2026 Week 14, research preview, optional). Frames Computer Use as the visual loop extension when text-level /qa-quarto and /visual-audit aren't enough.
  • guide/workflow-guide.qmd — Monitor-tool subsection in "Cost-Conscious Parallelism" (Apr 2026 Week 15). Replaces polling-loop anti-pattern for long-running R fits, replication batch reruns, etc.
  • guide/workflow-guide.qmd — new "Anthropic-Shipped Apr 2026 Utilities" section in the ecosystem area: /team-onboarding, /autofix-pr, /powerup, Ultraplan, /fewer-permission-prompts. Framed as off-ramps when this template's scope doesn't fit.
  • README.md — Quick Start callout pointing forkers heavily diverging from academic content at Anthropic's /init to re-derive CLAUDE.md.
  • templates/skill-template.md — new "When to set disable-model-invocation: true" subsection codifying the rule (write-load-bearing-persistent-file → set the flag) and the new "CLAUDE.md @import syntax" subsection documenting the Anthropic Apr 2026 import feature, with explicit guidance that this template's CLAUDE.md deliberately does NOT use it (under-150-line CLAUDE.md is better monolithic).
  • .claude/skills/data-analysis/SKILL.md, .claude/skills/audit-reproducibility/SKILL.md — new "Long-running fits / batch reruns: use the Monitor tool" subsections. Document the background-launch + Monitor pattern for jobs that take more than a couple of minutes.

Surface-sync

  • Skills 28 → 30 (/checkpoint, /preregister). Agents unchanged at 14. Rules unchanged at 24. Hooks unchanged at 6. Counts updated across all 6 monitored surfaces (README, CLAUDE.md, guide .qmd + rendered .html, docs/index.html, docs/workflow-guide.html, templates/skill-template.md). scripts/check-surface-sync.sh clean; scripts/check-skill-integrity.py clean. Guide re-rendered with Quarto 1.8.x; docs/workflow-guide.html synced.

Attribution

/checkpoint shape: adapted from Hugo Sant'Anna's clo-author v4.2.0 with permission. Attribution header on .claude/skills/checkpoint/SKILL.md.

Pre-v1.8.0 infrastructure (folded into this release)

Two themes that landed in the working tree before the v1.8.0 capability work: audit-hardening (mechanical parity checks + living pet-peeves catalogue that close classes of bug the agent-based /deep-audit was missing) and selective incorporation of Claude Code Apr 2026 features (Routines for AFK scheduling, PreCompact blocking, /fewer-permission-prompts as a sibling to our /permission-check).

Added — mechanical integrity checks

  • scripts/check-skill-integrity.py — deterministic parity checks that run in under a second. Four checks: (1) frontmatter allowed-tools ↔ body tool-invocation parity — catches the v1.7.0 PR #92 class of bug where 4 skills promised Task in their body but had no Task permission; (2) argument-hint ↔ body flag parity — documented flags advertised in the one-line hint; (3) internal markdown anchor resolution — no broken [text](path#anchor) links (the #category-11-numerical-discipline miss on PR #87); (4) rule paths: ↔ skill implementation parity — rule claims skill follows protocol, so skill body must reference the protocol's keywords (the /interview-me miss on PR #92).
  • scripts/check-surface-sync.sh now runs both check-surface-sync.py (count assertions) and check-skill-integrity.py (parity). Either gate failing blocks /commit Step 0b.
  • .claude/skills/slide-excellence/SKILL.md argument-hint updated to advertise --fast, --skip-substance, --acknowledge-template-domain-reviewer — the first real P2 the new check caught on baseline.

Added — living pet-peeves catalogue

  • .claude/references/audit-pet-peeves.md — 12-entry seed document cataloguing classes of bug review bots (Copilot / Codex) have caught on recent PRs that /deep-audit missed. Each entry: example + how to catch + why deep-audit missed it + when to apply. Grows with each PR — the meta-fix for the "audit agents keep missing this one thing" pattern.
  • .claude/skills/deep-audit/SKILL.md now has a Phase 0 that runs the mechanical checks first (cheapest, most precise), and every agent in Phase 1 is instructed to read the pet-peeves file before reporting clean.

Changed — documentation

  • MEMORY.md [LEARN:audit] entry on the mechanical-vs-agent tradeoff (catch classes of bug where precision matters with deterministic checks; use agents for judgment calls).
  • TROUBLESHOOTING.md section for check-skill-integrity failures — what each P0/P1/P2 means, how to resolve or tune.

Fixed

  • Self-referential false positive in the pet-peeves doc: example [text](path#anchor) markdown inside prose was matched by the anchor-resolution check. Script now strips inline code spans and fenced code blocks before scanning, so illustrative examples don't trigger the check.

Added — Apr 2026 Claude Code incorporation

  • PreCompact hook can now block compaction (.claude/hooks/pre-compact.py). Opt-in via env var CLAUDE_PRECOMPACT_BLOCK_ON_DRAFT=1: blocks once per DRAFT plan so the user can approve before losing mid-plan context. Uses the modern Claude Code block protocol (exit 0 + JSON {"decision":"block","reason":"..."} on stdout). Fires at most once per plan path — no lock-out loops. Default off; existing users get no change.
  • MEMORY.md [LEARN:scheduling] + [LEARN:hooks] capturing two lessons from Apr 2026: (a) CronCreate is session-only in practice — use Claude Code Routines (launched Apr 14) for any autonomous work that must survive session termination; (b) PreCompact hooks can now block, which is the right primitive for "don't lose this context."
  • .claude/references/audit-pet-peeves.md entry 17 — don't use CronCreate for long-delay autonomous work; Routines is the right primitive.
  • TROUBLESHOOTING.md scheduling section — explains CronCreate vs Routines tradeoff (short-delay in-session vs AFK work on web infra) and documents the PreCompact blocking guard. Plus a pointer to the built-in /fewer-permission-prompts skill as a sibling to our /permission-check (diagnose with /permission-check, remediate with /fewer-permission-prompts).

No stale model references audited — all 14 agents already use model: inherit, so they auto-adapt to Opus 4.7 / Sonnet 4.6 / Haiku 4.5 without changes.

v1.7.0 — 2026-04-16

A discipline-patterns minor release. Additive infrastructure for anti-drift (summary-parity) and anti-hallucination (Post-Flight Verification / Chain-of-Verification). Full details below; no breaking changes for forks.

Added — Post-Flight Verification (Chain-of-Verification)

Mirror of v1.6.0's Pre-Flight Reports, at the output side. Where Pre-Flight proves inputs were read before work, Post-Flight proves factual claims hold after drafting — before the skill returns to the user. Adapted from Dhuliawala et al. 2023, "Chain-of-Verification Reduces Hallucination in Large Language Models" (arXiv:2309.11495). The core CoVe idea — answer verification questions in a context that does NOT contain the original draft — is architecturally enforced here via context: fork on the verifier agent.

  • .claude/rules/post-flight-verification.md — new path-scoped rule. Defines the 4-step CoVe protocol (draft → extract claims → generate verification questions → answer independently in fresh context → reconcile). Scoped to skills that generate factual claims. Fail-closed: if the verifier errors or times out, the draft is surfaced as provisional rather than shipped silently. Opt-out via --no-verify.
  • .claude/agents/claim-verifier.md — new forked agent. Never sees the draft (enforced by context: fork at the Task boundary). Receives only: claims, verification questions, source-material pointers. Uses Read, Grep, Glob, WebFetch, WebSearch, Bash. Returns a structured verification report (PASS / PARTIAL / FAIL per claim, with evidence quotes + source locations).
  • .claude/skills/verify-claims/ — new user-facing skill. Runs Post-Flight on any draft the user hands to it (a .md, .qmd, .tex, .txt file). Accepts --source <path-or-url> to point at source material. Callable directly by users or spawned by other skills via Task.
  • Four skill integrations — /lit-review, /research-ideation, /respond-to-referees, /review-paper (novelty probe in --peer mode). Each now runs Post-Flight internally before returning. Skip conditions documented per skill (e.g., --no-verify, user pre-verifies).

Rationale: /lit-review citations from WebSearch were hallucination-prone (already flagged in SKILL.md); /research-ideation negative-literature claims (e.g., "no prior work studies X") and dataset-structure claims are classic hallucination vectors; /respond-to-referees "we added X on page Y" claims can be wrong or out-of-date after revision; /review-paper --peer editor novelty probe depends on WebSearch.

Added — anti–whack-a-mole rule (from earlier in v1.7.0 cycle)

  • .claude/rules/summary-parity.md — path-scoped rule preventing surgical word-level fixes on summary paragraphs from introducing new drift elsewhere in the same paragraph. Triggers on CHANGELOG ledes, README taglines, PR ## Summary blocks, skill/rule/agent frontmatter description fields, guide section abstracts, MEMORY.md [LEARN] headlines. Core heuristic: two review-bot flags on the same paragraph = rewrite structurally (abstract up), don't patch. Motivated by 3 consecutive Copilot findings on the v1.6.1 CHANGELOG opening (PRs #88–#90), each surgical fix introducing a new drift elsewhere.

Added — audit learnings

  • MEMORY.md new [LEARN:audit] entries capturing the whack-a-mole anti-pattern (summary-parity) and the CoVe vs critic-fixer vs cross-artifact distinction (three complementary verification mechanisms at three different architectural levels). Future sessions inherit both lessons on fresh context.

Added — guide coverage

  • /verify-claims and the new claim-verifier agent added to the guide's "All Skills" / "All Agents" tables. post-flight-verification.md and summary-parity.md added to "All Rules." Pre-existing gap in the rules table closed as well: content-invariants.md and cross-artifact-review.md were on disk but missing from the table; now listed. Path-scoped rule callout updated (16 → 20 path-scoped rules total).

Changed — counts

  • Skills 27 → 28 (/verify-claims). Agents 13 → 14 (claim-verifier). Rules 22 → 24 (summary-parity.md + post-flight-verification.md). Hooks unchanged at 6. Surface-sync gate propagates the new counts across all 6 monitored surfaces (README, CLAUDE.md, guide .qmd + .html, docs/index.html, docs/workflow-guide.html, templates/skill-template.md, /commit SKILL internal example).

Changed — drift-proofing the v1.6.1 CHANGELOG lede (applied the new summary-parity rule)

  • Rewrote the v1.6.1 opening paragraph abstraction-first ("No breaking changes. No new directories were added to .claude/; existing infrastructure was revised …"). First test of the new summary-parity rule on an existing entry. GitHub release notes for v1.6.1 synced to match.

Governance note

v1.7.0 establishes the discipline-pattern trilogy:

  1. v1.6.0 Pre-Flight (input discipline) — inputs were read before work.
  2. v1.6.1 summary-parity (framing discipline) — summaries don't drift from bodies.
  3. v1.7.0 Post-Flight (output discipline) — factual claims hold before work ships.

All three share a shape: fail-closed, structured output block, honest fallbacks. All three address classes of bugs a human reviewer would catch but the agent loop should catch earlier.

v1.6.1 — 2026-04-16

A framing honesty + hook friction patch release. No breaking changes. No new directories were added to .claude/; existing infrastructure was revised to address two classes of issue surfaced by a multi-round audit:

  1. Claim-vs-reality drift: v1.6.0 docs and rules described the "orchestrator" as if it were a repo-wide daemon that activates automatically after plan approval. In reality, the 6-step loop (IMPLEMENT → VERIFY → REVIEW → FIX → RE-VERIFY → SCORE) is a pattern implemented by specific skills (/commit, /qa-quarto, /review-paper --adversarial, /slide-excellence, /create-lecture, /data-analysis, /review-paper --peer). Plan approval does NOT trigger an auto-loop. Similarly, quality thresholds are advisory inside /commit, not enforced by a repo-wide git pre-commit hook.

  2. Hook blocking fatigue: the Stop hook log-reminder.py used {"decision": "block"} to force session-log creation. Effective for discipline, disruptive for autonomous flows. Now exits 0 with stderr-only advisories.

Changed — honest framing

  • .claude/rules/orchestrator-protocol.md rewritten: opens with "This rule describes the contract that skills implement. The 6-step loop is a pattern, not a runtime." Adds a skill-by-skill implementation table (which skill covers which steps; notes where auto-fixing happens and where it doesn't). "Just Do It" mode clarified to explicitly NOT authorize commits on its own — /commit invocation is still required.
  • .claude/rules/quality-gates.md renamed to "Quality Review & Scoring Rubrics" in practice (header kept for URL stability). Opens with an advisory-framing callout: enforcement is by the /commit skill only (halt + ask to override); a direct git commit bypasses the review.
  • .claude/rules/cross-artifact-review.md clarifies that detection is pattern-based (\input{scripts/...} / %% source: / filename matches). If the manuscript has none of those signals, nothing auto-invokes — and --no-cross-artifact is a no-op. Removed a reference to an unimplemented --with-scripts forcing flag.
  • .claude/rules/beamer-quarto-sync.md adds a precedence-with-SSOT section for the case where the Quarto file has manual post-translation edits: Beamer remains authoritative; presentation-only divergence (HTML-specific callouts) is allowed; structural drift is a bug.
  • guide/workflow-guide.qmd 4 sections rewritten: "The Orchestrator" → "A Pattern, Not a Daemon"; "Quality Scoring" now advisory; the "Skills vs Orchestrator" callout acknowledges both paths invoke the pattern inside a skill. Removed 5 occurrences of "automatic orchestrator" overselling across the document.
  • README.md contractor-mode framing updated: "runs the orchestrator pattern internally" instead of "runs autonomously." Quality Review section adds an explicit framing-honesty note: "advisory at the harness level — if you bypass the skill, you bypass the review."
  • docs/index.html landing-page bullets reworded: "Contractor mode via skills" (not "Contractor mode orchestrator"), quality-scoring bullet describes halt-and-override inside /commit.
  • CLAUDE.md Quality Thresholds table title now reads "(advisory)" with a one-line footnote clarifying /commit-only enforcement.

Changed — hook friction relief

  • .claude/hooks/log-reminder.py no longer blocks. Both block-return branches (no-log-exists and 15-response-counter) converted to stderr-only advisories. THRESHOLD raised from 15 → 50. Docstring updated to match the new semantics. The old blocking behavior was effective but disrupted /loop and batched-fix flows — stderr reminders preserve the nudge without halting execution.
  • .claude/hooks/verify-reminder.py throttle bumped from 60s → 300s (5 minutes). Same reminder, less noise during iterative .tex / .qmd / .R edits.

Added — orphan wiring + skill disambiguation

  • templates/decision-record.md (v1.6.0 addition) now wired into /interview-me: when the researcher explicitly chooses among alternatives during an interview (e.g. DiD vs IV vs RDD, admin vs survey data), the skill produces an ADR alongside the research spec. Skipped when there's a single uncontested path.
  • Decision trees added to the top of /review-paper, /seven-pass-review, and /slide-excellence SKILL.md files. Users can now pick the right skill at a glance: review-paper = most drafts; seven-pass = submission-ready / R&R; slide-excellence = slides; plus pointers to single-lens skills.
  • .claude/agents/domain-reviewer.md and domain-referee.md both prefixed with a scope-disambiguation block. domain-reviewer is the general (not disposition-primed) substance reviewer used by /slide-excellence and /seven-pass-review. domain-referee is the disposition-primed variant used by /review-paper --peer. Same domain expertise, different calibration.
  • .claude/rules/r-code-conventions.md adds Section 8: Numerical Discipline with the project epsilon (eps <- 1e-12) for CDF clamping plus 7 headline rules, cross-referenced to r-reviewer Category 11. The checklist gains a "numerical discipline" row.

Added — documentation

  • TROUBLESHOOTING.md +5 sections for v1.5/v1.6 features:
    • Permissions / bypass / statusline (6-layer stack diagnosis, why /permission-check gates host-global reads, statusline parse-failure recovery)
    • Peer-review pipeline (missing journal profile, cloned referee reports, R&R continuation chain breaks)
    • Surface-sync gate (count drift resolution, adding a new skill breaks the gate by design)
    • Pre-Flight Reports (fail-closed semantics, first-lecture fallback in /create-lecture)
    • Decision records (where to save, gitignore behavior)
  • README.md Quick Start adds two callouts: Python/R/markdown-only users can skip XeLaTeX/Quarto; MEMORY.md vs personal-memory.md distinction introduced early (was previously session-2 discovery only).
  • MEMORY.md +6 new [LEARN] entries: framing (orchestrator is pattern, quality gates advisory, cross-artifact pattern-based), dogfooding (empty quality_reports/ dirs is a red flag — Stop hook caught it mid-session), audit (claim-vs-reality is the highest-ROI lens for a governance-heavy template repo). Template inventory refreshed from 6 → 10 files + tikz-snippets/.
  • guide/workflow-guide.qmd "All Skills" table adds /seven-pass-review and /permission-check (were missing).

Fixed — review-driven polish (from Copilot + Codex on PR #87)

  • Broken anchor: r-code-conventions.md linked to #category-11-numerical-discipline, but the actual heading in r-reviewer.md is ### 11. NUMERICAL DISCIPLINE. Dropped the anchor, references by name.
  • Unimplemented flag: beamer-quarto-sync.md advised running /translate-to-quarto --diff [file], but the skill has no --diff option. Replaced with "regenerate into a scratch path, diff manually."
  • Contradictory scope: domain-reviewer.md claimed "slides only" while also stating "used by /seven-pass-review" (a manuscript skill). Reframed as general reviewer for both artifacts; domain-referee.md is the disposition-primed manuscript variant.
  • Stale docstring: log-reminder.py docstring described blocking behavior after the code had been converted to advisory. Updated.
  • Stale docstring: verify-reminder.py said "within 60s" after throttle was bumped to 5 min. Updated.
  • Daemon phrasing in TROUBLESHOOTING: sessionInfo.txt fix referenced "the orchestrator" as if it were a daemon. Points at 00_run_all.R via /data-analysis or the user's pipeline runner instead.
  • /context-status regression averted: intermediate commit had unwired context-monitor.py from PostToolUse, but /context-status reads the cache that hook writes. Codex flagged the dependency; the hook was re-wired before merge.
  • CHANGELOG upgrade example: pin example updated from v1.3.0 (2026-04-13) to v1.6.0 (2026-04-15) to reflect current state.
  • Guide frontmatter date: stale 2026-03-20 → 2026-04-15 (v1.6.0 release date).
  • v1.6.0 Pre-Flight claim: CHANGELOG line about "fail-closed if inputs can't be read" now acknowledges /create-lecture's first-lecture fallback (documented exception, not a contradiction).

Governance note

v1.6.1 establishes the "claim-vs-reality" audit lens as a first-class review category going forward. When a template repo oversells itself, the gap is the first thing forkers notice when reality bites. The [LEARN:audit] entry in MEMORY.md captures this: for any governance-heavy feature, the audit question "is the claim mechanically enforced, or is it prose?" catches more real bugs than skill/doc consistency checks.

The Stop hook's conversion from blocking to advisory is a philosophical tradeoff: discipline vs. autonomy. v1.6.0 chose discipline (block until the user complies); v1.6.1 chooses autonomy (nudge but don't halt) because the blocking version disrupted /loop, batched audits, and autonomous-ship flows. The nudge survives in stderr, the discipline now depends on the user's own habit. If a future release finds this has quietly caused session-log neglect, the blocking form can be restored behind an opt-in setting.

Verification

Pre-merge deep audit launched 4 parallel agents (guide content, hook code, skills/rules consistency, cross-doc). 1 genuine finding (stale docstring), 1 false alarm (log-reminder fail-open is correct). Fixed on branch before merge.

Surface-sync gate: 27 skills / 13 agents / 22 rules / 6 hooks matched across 6 surfaces. quality_score.py on guide/workflow-guide.qmd: 100/100.

v1.6.0 — 2026-04-15

A discipline-layer release: the template's infrastructure now actively catches the class of bugs that produced reviewer-driven fix PRs in v1.5.x. Also adds two observability/diagnosis tools (statusline, /permission-check), doubles the referee pet-peeve pools, and ports three quality patterns from clo-author (Pre-Flight Reports, Content Invariants, Numerical Discipline). No breaking changes.

Added — observability + diagnostics

  • .claude/scripts/statusline.sh — always-visible mode badge ([BYPASS] / [PLAN] / [AUTO-EDIT] / [PROMPT]) + model + git branch. Renders on every turn. Wired via .claude/settings.json statusLine. Parses session JSON in a single python3 invocation (per-turn perf).
  • .claude/skills/permission-check/ — new /permission-check skill. Read-only diagnostic: reads repo-local settings layers auto, requires explicit user confirmation before reading host-global (~/.claude/settings.json, VSCode user settings). Redacts unrelated keys. Surfaces drift across the 6-tier permission stack (VSCode user / workspace / CLI user / project / project-local / in-session runtime).
  • Six-Layer Permission Stack + Plan→Bypass Handoff in the guide. Troubleshooting checklist for "prompts fire despite bypass." Explicit callout-warning that plan approval is NOT an enforcement boundary — exiting plan mode returns to defaultMode with full bypass authority.

Added — surface-sync gate

  • scripts/check-surface-sync.py + scripts/check-surface-sync.sh — cross-document count consistency gate. Counts .claude/{skills,agents,rules,hooks} on disk, scans 6 surfaces (README, CLAUDE.md, guide .qmd + .html, docs/index.html, skill-template) for count assertions using compound regex patterns (avoids false positives on unrelated phrases like "3 parallel agents" or attribution lines). Fails closed on drift — no "commit anyway" override. Wired into /commit as Step 0b.
  • Addresses the systemic drift pattern that produced PRs #70, #76, #78 in v1.5.x (adding a skill updated .claude/ but left stale counts in prose).

Added — referee quality polish

  • Expanded editor pet-peeve pools: 25 → 29 critical peeves (added: notation drift, seed-dependent results, covariate balance absent, overlap/common-support missing); 20 → 25 constructive peeves (added: "what this paper does not show" paragraph, raw-data figures, alternative specs, notation tables, careful attribution). Now exceeds clo-author v3.1 baseline (27/24).
  • quality_reports/decisions/ + templates/decision-record.md — ADR-style decision records. Template with Status / Problem / Options considered / Decision + rationale / Consequences / Rejected alternatives. Gitignored like plans/specs.

Added — discipline patterns (ported from clo-author, adapted)

  • Pre-Flight Reports in /data-analysis, /create-lecture, /review-paper --peer. Each skill now requires a structured output block proving inputs were read before doing work (dataset fields, project conventions, notation registry checks, journal profile, cross-artifact status). Fail-closed if inputs can't be read, with a documented first-lecture fallback in /create-lecture (proposes a minimal knowledge base when the template is still empty, to avoid deadlocking fresh forks).
  • .claude/rules/content-invariants.md — new rule, path-scoped to Slides/**/*.tex, Quarto/**/*.qmd, Quarto/**/*.scss, Preambles/header.tex, scripts/R/**/*.R. Defines INV-1 through INV-12: palette sync, Beamer↔Quarto notation parity, Quarto CSS override contract, TikZ-as-SVG, single bibliography, no \pause, max 2 boxes per slide, motivation-before-formalism, set.seed once, relative paths only, transparent-bg figures, project theme on all plots. Critics can now cite invariants by number.
  • r-reviewer agent — category 11 "Numerical Discipline": no float ==, CDF clamping to open interval with named epsilon (not [0,1] — exact 0/1 to qnorm yields ±Inf), integer literals for counts (1L), pre-allocated vectors, deterministic bootstrap seeding, explicit na.rm, no T/F shorthands.

Changed — skill trigger descriptions

  • 17 .claude/skills/*/SKILL.md frontmatter rewrites for reliable auto-invocation. Verb+object + "Use when: …" trigger phrases + disambiguation from sibling skills (e.g., /interview-me explicitly says NOT for lit review, pairs with /research-ideation). Follows the deep-audit gold-standard pattern. Cold-prompt auto-invocation is now reliable for /commit, /deploy, /proofread, /data-analysis, and siblings.
  • commit skill triggers tightened after Codex flagged vague end-of-task phrases as risky: now only explicit commit intent ("commit", "ship it", "push this", "open a PR", "merge to main", "let's commit this"). Removed "wrap up these changes" and end-of-task-signal trigger.
  • Guide: "Writing Effective Trigger Descriptions" expanded with the 3-part pattern (verb+object → "Use when:" phrases → disambiguation), A/B rewrite example, and a pre-ship checklist.

Changed — documentation

  • Counts: skills 26 → 27 (added /permission-check), rules 21 → 22 (added content-invariants). Agents and hooks unchanged. Synced across all 6 surfaces via the new sync gate.
  • Guide re-rendered with the new sections (statusline, permission stack, Plan→Bypass handoff, expanded trigger-description guidance). docs/workflow-guide.html synced.

Fixed — systemic quality during development

This release absorbed an unusual volume of reviewer-driven fixes from Codex and Copilot. Representative samples:

  • Count drift: a single replace_all on "26 skills" missed "26 skills, and 21 rules" (extra "and"), missed "26 slash commands", missed "template's 26". The deep-audit skill now documents the phrasing-variant trap; the surface-sync gate prevents the class of bug.
  • Stop-hook block protocol: some audit guidance implied non-zero exit codes are required to block. Actually, modern Claude Code accepts BOTH exit 2 + stderr reason (legacy) AND exit 0 + JSON {"decision":"block","reason":"..."} on stdout (modern — what log-reminder.py uses). Deep-audit skill now documents both protocols explicitly so future audit agents don't re-flag log-reminder.py.
  • Statusline parse-failure fallback: parse error emitted cwd="." which wasn't empty, bypassing the pwd fallback. Fixed to emit empty third line and tightened the bash guard to treat "." as invalid.
  • notify.sh robustness: best-effort notification now fails open on missing jq AND on malformed JSON input (defaults before jq attempt, silent stderr on parse).
  • Plan→Bypass framing: initial guide text said "combines safety and prompt-free execution." Codex correctly flagged this as overselling — plan approval doesn't bind later execution to the approved plan. Reworded as "review-before-execute convenience" with a callout warning.
  • /permission-check privacy boundary: first draft read ~/.claude/ and VSCode user settings unconditionally on ambiguous prompts like "why am I getting prompts?". In a shared/corporate environment this could leak host-global config. Restructured into Phase A (repo-local, auto) + Phase B (host-global, explicit user confirmation + key redaction).
  • CDF clamping math bug in the new Numerical Discipline checklist: initial draft said pmin(1, pmax(0, p)) but exact 0/1 to qnorm yields ±Inf. Fixed to open interval with named eps.
  • Content-invariants path globs: initial frontmatter used bare directories; other rule files use quoted glob patterns. Aligned.
  • Seed format conflict in /data-analysis: Phase 1 required YYYYMMDD (per r-code-conventions.md) but the template example still showed set.seed(42). Made self-consistent.

Attribution

The discipline patterns in this release (Pre-Flight Reports, Content Invariants, Numerical Discipline rules) are ported from Hugo Sant'Anna's clo-author with adaptation to our lecture-shaped surface. Hugo's v4.1.x has moved to a 10-verb skill consolidation + full paper-type branching we deliberately didn't port (doesn't fit our primary artifact). Invariants have lecture-specific codes (INV-1..INV-12) rather than clo-author's paper-centric ones.

Governance note

v1.6.0 establishes the "discipline layer" as a first-class template concern. The surface-sync gate makes count drift a pre-commit error, not a reviewer catch. The statusline + /permission-check turn permission debugging from detective work into a 2-second glance. The Pre-Flight Report pattern makes "agent hallucinated your variable names" a category of failure that can't happen silently.

v1.5.0 — 2026-04-14

Added — Simulated peer review

A new --peer [journal] mode for /review-paper that runs a full editorial pipeline: editor desk review → referee selection (2 different dispositions from a 6-way taxonomy) → 2 blind referees in parallel → editorial synthesis (FATAL / ADDRESSABLE / TASTE). Calibrated per journal.

  • .claude/agents/editor.md — journal editor persona. Desk-reviews (with optional WebSearch novelty probe, ON by default — opt out with --no-novelty-check), selects two referees with deliberately different dispositions from the 6-way taxonomy (STRUCTURAL / CREDIBILITY / MEASUREMENT / POLICY / THEORY / SKEPTIC), assigns each referee 1 critical + 1 constructive pet peeve, then synthesizes their reports into an editorial decision with classification and "MUST / SHOULD / MAY push back" response-planning block.
  • .claude/agents/domain-referee.md — substance referee. 5 weighted dimensions (contribution 30 / lit positioning 25 / substance 20 / external validity 15 / journal fit 10). Disposition-primed. Requires "What would change my mind: [specific ask]" on every MAJOR concern — a discipline that separates adversarial review from productive review.
  • .claude/agents/methods-referee.md — methodology referee with paper-type branching: reduced-form / structural / theory+empirics / descriptive. Each type has its own dimension weights and mandatory pre-scoring sanity checks (sign / magnitude / parameter plausibility / construct validity / etc.). Same "What would change my mind" requirement.
  • .claude/references/journal-profiles.md — NEW directory. Ships with 5 econ journals (AER / QJE / JPE / ECMA / ReStud), each with Focus / Bar / Domain-referee adjustments / Methods-referee adjustments / Typical concerns / Referee-pool weights / optional Table-format override. Plus a "Field adaptation" section with detailed instructions for non-econ users.
  • templates/journal-profile-template.md — skeleton for adding your own journal/field. Includes a disposition reference and field-specific paper-type guidance (e.g., biology: observational/experimental/computational/review; political science: case-study/comparative/formal-model/survey).

Changed

  • /review-paper — now supports three modes: default (single-pass), --adversarial (critic-fixer loop from v1.4.0), and the new --peer <journal> pipeline. Sub-flags: --peer --r2 / --peer --r3 (R&R continuation — reloads prior reports, classifies concerns Resolved / Partial / Not addressed, max 3 rounds), --peer --stress (hostile editor — forces SKEPTIC dispositions, doubles critical peeves), --no-novelty-check (skip WebSearch probes). Output directory: quality_reports/peer_review_[paper]/ with desk_review.md, referee_domain.md, referee_methods.md, editorial_decision.md.
  • .claude/rules/cross-artifact-review.md — adds --peer ordering clause: in --peer mode, cross-artifact review runs before the editor's desk review so reproducibility FAIL can be cited as desk-reject grounds. In default and --adversarial modes, cross-artifact still runs at Step 6b (after the paper review).
  • Counts: 10 → 13 agents (editor, domain-referee, methods-referee). Skills (26) and rules (21) unchanged. Synced across README, docs/index.html, guide, templates, CLAUDE.md.

Attribution

The simulated-peer-review pipeline is adapted from Hugo Sant'Anna's clo-author with his explicit permission. Hugo's work contributed:

  • The pipeline shape (editor desk → 2 referees → editorial synthesis).
  • The 6-way disposition taxonomy (STRUCTURAL / CREDIBILITY / MEASUREMENT / POLICY / THEORY / SKEPTIC).
  • The journal-calibration schema (Focus / Bar / Typical concerns / Referee-pool weights / Table-format overrides).
  • The paper-type branching in the methods referee (reduced-form / structural / theory+empirics / descriptive) with per-type dimension weights and sanity checks.
  • The "What would change my mind" requirement on every major concern.
  • The R&R continuation pattern (reload prior round, classify Resolved / Partial / Not addressed).

We reimplemented all files rather than copying verbatim (clo-author has no LICENSE file at time of adaptation; our version is original text under MIT). All new files carry an attribution header pointing back to clo-author.

Thanks to Hugo for both the inspiration and the permission. The fork is doing great work on the authoring side of the pipeline — complementary to our template's authoring + review orientation.

Governance note

With v1.5.0, the template's review story is now the strongest single feature: single-pass review (fast feedback), adversarial loop (iterative revision), seven-pass parallel review (broad coverage), simulated peer review (journal-calibrated editorial pipeline), R&R continuation (R&R rehearsal), stress mode (hostile editor). Four complementary review modes, one skill.

v1.4.0 — 2026-04-14

Added — review-skills hardening

  • .claude/skills/audit-reproducibility/ — enforces replication-protocol.md by cross-checking numeric claims in a manuscript (ATT = -1.632 (0.584), N = 2,847, p-values, percentages) against the actual R / Stata / Python outputs. Tolerance-graded PASS/FAIL per claim. Usable as a /commit gate (exit 1 on FAIL). Addresses the "I updated the analysis but forgot to update Table 2" bug.
  • .claude/skills/seven-pass-review/ — mechanizes Pattern 15. Seven forked subagents, one per lens (abstract, intro, methods, results, robustness, prose, citations), run in parallel, then a synthesizer produces a prioritized revision checklist with cross-lens contradictions surfaced.
  • .claude/rules/cross-artifact-review.md — paper ↔ code dependency-graph protocol. When /review-paper runs, auto-invokes /review-r on referenced scripts and /audit-reproducibility on the pair. Surfaces critical cross-artifact findings (code bug invalidates paper claim) at the top of the review report. Opt-out: --no-cross-artifact.

Changed

  • /review-paper — new --adversarial mode runs the critic-fixer loop from /qa-quarto (up to 5 rounds). Single-pass default unchanged. Now also auto-invokes cross-artifact review (Step 6b).
  • /slide-excellence — conditional dispatch: only spawns subagents that can produce useful output (tikz-reviewer only on TikZ-bearing files, content-parity only when both .tex and .qmd counterparts exist, R reviewer only when R chunks are present). Pre-flight check refuses to run domain-reviewer if the agent is still the shipped template. New flags: --skip-substance, --acknowledge-template-domain-reviewer, --fast. Cuts token cost ~50% on typical .qmd-only files.
  • /validate-bib — new --semantic mode: citation-drift detection (duplicate .bib entries for the same paper), crossref DOI verification with caching and rate limiting, within-file citation-style consistency, optional --cite-claim abstract surfacing. Structural mode unchanged. --skip-doi for offline.
  • .claude/agents/domain-reviewer.md — adds AUTO-DETECT-TEMPLATE-MARKER so /slide-excellence can detect un-customized templates and warn before running generic substance review.
  • Counts: 24 → 26 skills, 20 → 21 rules. Synced across README, docs/index.html, guide, templates, and docs/workflow-guide.html.

Fixed — Codex round-2 regressions

  • Trust-boundary porousness (PR U1): The permissions.deny on .claude/settings.json and .claude/hooks/** was bypassable via allowlisted shell tools (Bash(python3 *), Bash(cp *), etc.). Narrowed the broad shell allows and added a PreToolUse guard. (Later removed in the bypass-mode directive — see "Removed" below.)
  • TikZ prevention bypasses (PR U2): check_p3 missed scale={0.8} and scale=\myscale; check_p4 incorrectly flagged align=left (treated left as a direction). Rewrote with a brace/bracket-balanced tokenizer that matches standalone option keys. Parse errors now exit 2 (visible) instead of silent 0.
  • R stale-state leak (PR U3): scripts/R/03_analyze.R used exists("df") without inherits = FALSE, allowing a stale globalenv to satisfy the guard. Added inherits = FALSE to match the contract already applied to 02_clean.R and 05_figures.R.
  • R template silent-skip (PR U4): 05_figures.R silently switched to base-R PDF if ggplot2 was missing and silently skipped SVG if svglite was missing. Made ggplot2 a hard dependency (fail loudly); kept svglite optional but emits an explicit "SKIPPED" warning. Rewrote scripts/R/README.md to document hard vs. optional deps.

Removed

  • .claude/hooks/protect-sensitive-paths.sh — added in PR U1, removed the same release under explicit user directive: "bypass mode means bypass." Bypass permissions (defaultMode: bypassPermissions) is the user's chosen high-autonomy workflow; re-adding approval gates during autonomous runs is a regression. Persisted to memory as feedback_bypass_permissions.md so this decision isn't re-litigated.

Governance note

When adversarial reviewers (Codex, Copilot) flag the absence of approval gates as a risk under bypass mode, the template now treats that as a documented tradeoff, not a bug. Hardening under bypass mode is limited to non-blocking signals (PostToolUse reminders, notifications, logging). See feedback_bypass_permissions.md in the project's memory directory.

v1.3.0 — 2026-04-13

Added — TikZ story overhaul

Ported the best parts of Scott Cunningham's MixtapeTools TikZ infrastructure and wired them into our pipeline end-to-end.

  • .claude/rules/tikz-prevention.md — 6 authoring rules (P1–P6) that stop collisions at write-time: explicit node dimensions, coordinate-map comments, prohibition on scale=, directional keyword on every edge label, use the snippet gallery, one tikzpicture per idea.
  • .claude/rules/tikz-measurement.md — six-pass protocol with concrete formulas: Bézier max_depth = (chord/2)·tan(bend/2), character-width table by font size, label-gap calculation, 0.4 cm shape-boundary rule, matplotlib arc3 Bézier helpers, full margin matrix.
  • templates/tikz-snippets/ — 8 production-ready standalone .tex diagrams (DAG basic, DAG mediation, two-period DiD, event study, timeline, regression scatter, 3-step flowchart, supply-demand). Every snippet compiles on the first try and passes the prevention grep checks.
  • Preambles/header.tex — production-ready Beamer preamble (previously empty): 11-color palette matching the SCSS, shared TikZ styles (dag-node, decision-node, observed-edge, counterfactual-edge, confound-edge, observed-dot, counterfactual-dot), Beamer theme assignments, convenience macros (\muted, \key, \good, \bad, \transitionslide).
  • Preambles/README.md — usage + palette contract + inventory.
  • scripts/check-palette-sync.sh — greps Preambles/header.tex and Quarto/theme-template.scss, enforces that the five core palette names exist on both sides. Wired into validate-setup.sh.
  • .claude/skills/new-diagram/ — scaffold a new TikZ diagram from the gallery with prevention checks pre-applied; compiles standalone, invokes tikz-reviewer with measurement citations, loops until APPROVED (max 5 rounds).

Changed

  • /extract-tikz — mandatory Step 1 prevention pre-check (greps for bare edge labels and scale=) before the expensive compile + SVG cycle.
  • tikz-reviewer agent now requires citing the specific pass and formula from tikz-measurement.md for every CRITICAL/MAJOR finding. Vague reports are rejected.
  • settings.json allowlist expanded substantially (+23 Bash tools, +36 Edit/Write path rules): read-only tools (grep/cat/head/tail/awk/find/tree/basename/dirname/file), file ops (cp/mv/touch/mktemp), pipeline tools (pandoc/docx2txt/pdftotext), missing git/gh subcommands (tag/rm/mv/remote, issue/release), and Edit/Write pre-approvals for every directory we normally edit (.claude/, templates/, guide/, docs/, scripts/, Preambles/, Slides/, Quarto/, Figures/, quality_reports/, explorations/, master_supporting_docs/, .github/**, plus CLAUDE.md, README.md, CHANGELOG.md, MEMORY.md, .gitignore).
  • Counts: 23 → 24 skills, 18 → 20 rules, 7 → 6 hooks. Synced across README, docs/index.html, guide body, guide appendix, and CLAUDE.md.
  • Guide Step 3 "Adapt Your Theme" rewritten to document the two-surface palette contract and the sync script.

Removed

  • .claude/hooks/protect-files.sh and its PreToolUse registration. The hook used to block Edit/Write on Bibliography_base.bib and settings.json unless bypass was signalled. With the explicit Edit(...) / Write(...) allow-rules added to settings.json (above), Claude Code's permission system handles this cleanly and the extra hook was redundant friction. Removing it also cuts a failure mode (earlier sessions had to work around the hook with python3 -c writes).

Attribution

TikZ prevention + measurement rules adapted from tikz_rules.md in scunning1975/MixtapeTools. The source repo has no LICENSE file; its README says "Use freely. Attribution appreciated but not required." Both ported rule files cite Scott at the top.

Copilot review follow-up (same day)

This release shipped as a sequence of small PRs (#53–#56) plus an end-of-day Copilot-review cleanup PR that fixed 15 issues Copilot caught across those four PRs, removed protect-files.sh per user preference, and unified cross-skill grep patterns. Summary of the biggest catches:

  • LaTeX load order in Preambles/header.tex (xcolor before hyperref; \makeatletter around \@ifclassloaded; inputenc gated by \ifPDFTeX).
  • protect-files.sh env-var bypass moved above $(cat) so it exits immediately.
  • /extract-tikz and /new-diagram grep patterns unified character-for-character.
  • P1 scoped to boxed nodes; P3 reconciled with the scale=1.1 convention.
  • check-palette-sync.sh uses absolute paths ($(dirname "$0")/..) and a stable exit-code contract.
  • Removed Bash(npm *) from the allowlist (too broad — npm run executes arbitrary scripts).

v1.2.0 — 2026-04-13

Added

  • /respond-to-referees skill — parses a referee report, classifies each concern (addressed / partially / deferred / disagreement), points to specific revisions, and drafts a complete response document using templates/response-to-referees.md. Use during the R&R stage of paper revision.
  • HelloWorld sample — Slides/HelloWorld.tex and Quarto/HelloWorld.qmd — minimal decks that compile/render on a fresh fork before any project customization.
  • scripts/validate-setup.sh — colored dependency checker for Claude Code, XeLaTeX, Quarto, R, Python, git, gh, and hook permissions.
  • GitHub templates — .github/CONTRIBUTING.md, issue templates, and PR template.
  • CHANGELOG.md — this file.

Changed

  • /commit skill — now runs scripts/quality_score.py and the verifier agent on changed files as a pre-commit gate. Halts on score < 80 unless the user explicitly overrides.
  • /extract-tikz skill — now invokes the tikz-reviewer agent after SVG generation and loops on revisions to the Beamer source until APPROVED.
  • /slide-excellence skill — domain-reviewer agent is now MANDATORY for .tex files (was optional).
  • CLAUDE.md — example rows in the Beamer environments and Quarto CSS classes tables are now visible (not hidden in HTML comments). Added links to MEMORY.md and quality_reports/ so Claude knows where cross-session context lives.
  • README.md — added "Verify Your Setup" step in the Quick Start; replaced "Work in progress" disclaimer; added badges and CHANGELOG link.
  • docs/index.html — added SEO metadata (description, keywords, Open Graph, JSON-LD SoftwareApplication schema).
  • Skill count: 22 → 23 across all surfaces.

Fixed

  • scripts/quality_score.py — Quarto compilation check no longer doubles the path when cwd is set to the file's parent (was producing spurious 0/100 scores).
  • scripts/validate-setup.sh — git config check now guards behind command -v git to avoid misleading warnings when git is missing.
  • Slides/HelloWorld.tex — added a citation and bibliography slide so /compile-latex's 3-pass + bibtex pipeline completes cleanly on the onboarding sample.

v1.1.0 — 2026-03-20

Added

  • /deep-audit skill — repository-wide consistency audit (4 parallel specialist agents).
  • 2026 Claude Code feature support — effort levels (/effort low|medium|high|max), 5 permission modes, 4 hook handler types, 11 new hook events documented.
  • Skill frontmatter reference — effort, context: fork, agent, hooks, dynamic !\command`` injection.
  • Pattern 15: Sequential Adversarial Audits — seven-audit protocol for paper review (inspired by ClaudeCodeTools).
  • Ecosystem section — autoresearch (Karpathy), ClaudeCodeTools, clo-author, claudeblattman, MixtapeTools.
  • Prerequisites section — install command (curl -fsSL https://claude.ai/install.sh | bash), Node.js, Claude account, cost notes.
  • plansDirectory setting — explicit quality_reports/plans/ location.
  • Automatic "Last Modified" date — date-modified: last-modified in guide YAML.

Changed

  • Major guide refresh: 2400+ lines, all 25 factual claims verified against official docs across two deep-audit rounds.
  • Template cleanup for fork-friendliness — removed project-specific session logs, emptied Bibliography_base.bib, renamed Emory SCSS to generic theme-template.scss.

Fixed

  • All 5 Python hooks: from __future__ import annotations, fail-open try/except, ~/.claude/sessions/ storage, hash length consistency.
  • pre-compact.py exit code (2 → 0) and stdout → stderr (PreCompact ignores stdout).
  • post-compact-restore.py reads source field (was reading type, never ran).

v1.0.0 — 2026-02-28

Initial Release

  • 10 specialized agents: proofreader, slide-auditor, pedagogy-reviewer, r-reviewer, tikz-reviewer, beamer-translator, quarto-critic, quarto-fixer, verifier, domain-reviewer.
  • 22 skills covering LaTeX, Quarto, R, reproducibility, research, and meta-workflows.
  • 18 rules (4 always-on, 14 path-scoped) for quality gates, verification, and domain standards.
  • Hooks for notifications, context monitoring, session logging, and compaction state.
  • Orchestrator protocol (contractor mode) with adversarial critic-fixer loop (max 5 rounds).
  • Plan-first workflow with on-disk plan persistence across context compaction.
  • Three-tier memory system: CLAUDE.md (project), MEMORY.md (auto-memory), session logs.
  • GitHub Pages deployment via scripts/sync_to_docs.sh.

Upgrading Your Fork

If you forked this repo and want to pull our updates:

git remote add upstream https://github.com/pedrohcgs/claude-code-my-workflow.git
git fetch upstream
git merge upstream/main           # or: git rebase upstream/main

Files you almost certainly customized — CLAUDE.md, Bibliography_base.bib, Quarto/theme-template.scss, your lecture files in Slides/ and Quarto/, .claude/agents/domain-reviewer.md — may produce merge conflicts. Resolve in favor of your customizations; pull only the infrastructure improvements.

To pin to a specific version: git checkout v1.10.0 (latest as of 2026-05-31).