Skip to content

Changelog

All notable changes to Wren (create-wren) are documented here. Format follows Keep a Changelog; versions are integer-ish semver.

[0.8.0] - 2026-08-02

Adds headless scheduling + email and the weekly literature digest (roadmap milestone 0.8). /librarian-team (the heavier library-automation skill) is deferred to its own release, 0.8.1.

Added

  • Scheduled-jobs + email tier — scripts/setup/install_scheduled.py renders per-job runner scripts and a launchd agent (macOS) or cron line (Linux) from research-config.yml, and prints the exact activation command rather than registering the job itself. ./setup.sh --with-schedule now runs it. Cross-platform, opt-in, nothing scheduled until you enable and activate a job. Docs: Scheduled jobs.
  • /lit-digest [--since N] [--paper <id>] — a fast cross-paper sweep for new publications: one web search per paper, aggregated into a markdown + JSON digest under outputs/lit_digests/, with an optional Gmail draft (never sent). Headless-safe, so the weekly_digest scheduled job runs it unattended.

Changed

  • Gmail is wired as a Claude-account-connected integration; the digest job includes mcp__claude_ai_Gmail__create_draft in its --allowedTools only when integrations.gmail.enabled. Wren ships 18 skills.

[0.7.0] - 2026-08-02

Introduces the sub-agent tier, the two flagship agent-team review skills, and cost-aware model routing (roadmap milestone 0.7).

Added

  • Sub-agent tier (.claude/agents/) — focused personas Claude Code runs in isolated contexts via the Agent tool, with a README documenting the convention. Set model: per agent to pick a tier.
  • /qc-team [paper-id] [--skeptic-only] — three-agent adversarial QC: a Skeptic attacks the draft against its source data, a Responder defends, a Team Leader adjudicates into a Must-Fix / Should-Fix verdict. Agents qc-skeptic, qc-responder, qc-team-leader (all opus), run in isolation so the review stays honest.
  • /figure [paper-id] [--figure N] [--all] [--score-only] — a Designer → Critic loop that iterates a figure to publication quality. figure-designer (sonnet) edits and regenerates; scientific-graphics-reviewer (opus) scores on an 8-dimension rubric with hard gates and a ≥95% pass threshold.
  • Token management — .claude/lib/model-policy.mjs (critique/produce/triage → opus/sonnet/haiku), an always-on .claude/rules/token-budget.md rule, and scripts/utils/token_report.py (npm run tokens) reporting usage by model/agent from local session logs. No API, nothing leaves your machine.

Changed

  • The "Model Selection" note in CLAUDE_REFERENCE now points at the real model-policy + token-budget rule. Wren ships 17 skills; the sub-agent tier retires the "planned" caveat on /qc-team and /figure.

[0.6.0] - 2026-08-02

Three new skills, genericized from the private research system — quick rigor and reference wins (roadmap milestone 0.6).

Added

  • /methods-audit [paper-id] [--fix] — audits a paper's analysis scripts for exploration artifacts, stale intermediates, reproducibility gaps (missing seeds, absolute paths), inconsistent parameters, and statistical gaps, cross-checked against METHODS_LOG.md. A QC pass to run before the gold-standard freeze.
  • /ref-check [paper-id] [--format] [--gaps] [--zotero] — validates a draft's citations against your reference library (Zotero MCP or the .bib file), finds uncited claims, verifies sources support their claims, and checks formatting against the target journal.
  • /zotero-audit [--collection <name>] [--since <date>] [--fix] — scans a Zotero library for items missing metadata, tags, PDFs, or collections; classifies severity; and can auto-complete metadata from Crossref (per-item approval). Needs references.provider: zotero.

[0.5.0] - 2026-08-02

Brings the Wren logo to every surface it can reach.

Added

  • Colored wren at scaffold time. npm create wren now prints the pixel-wren in its banner — rendered in the logo's real colors via half-block characters, with a monochrome fallback under NO_COLOR and nothing when output is piped/CI. Generated into bin/wren-art.mjs from the same 16×16 grid as the logo, so it never drifts.
  • Docs site branding — header logo and browser-tab favicon (mkdocs-material theme.logo/theme.favicon), plus a logo hero on the docs home page.
  • Logo on the README / npm page.
  • GitHub social-preview card (assets/logo/wren-social.png, 1280×640) generated from the logo grid; upload once in repo Settings.

Changed

  • assets/logo/generate_logo.py is now the single source for the SVG, the PNG rasters, the terminal art, and the social card — all from one GRID/PALETTE.
  • Fresh-project nudge emoji 👋 → 🐦.

[0.4.0] - 2026-08-02

Adds a guided first-run experience so a new project starts with orientation and momentum instead of an empty paper portfolio.

Added

  • Guided onboarding — /setup wren. A conversational first-run flow: a short tour of the system (four-layer memory, the 14-phase pipeline and gold-standard gate, the skill map, and the audit/QC/peer-review functions), an interview about the researcher's interests, data requirements, and paper ideas, tool + plugin setup (database backend, reference manager, integrations) matched to those needs, and their first paper — or a planned paper series. Honors the one-at-a-time, no-unapproved-tasks, and rigor rules; flags /qc-team and /figure as planned-not-shipped.
  • docs/welcome.md — a single-source quick guide the tour narrates from, added to the docs nav.
  • One-time fresh-project nudge (session-start-welcome.py SessionStart hook): points a new project at /setup wren, then self-disables via a .wren/onboarded marker once onboarding completes. Silent outside a Wren project.

Changed

  • /setup now has two modes. /setup wren runs the guided onboarding above; bare /setup keeps the plain reconfigure-the-renderer behavior.
  • Post-scaffold handoff leads with /setup wren. The npm create wren CLI message and the verify_install success line now point at the guided tour (with /paper as the jump-straight-in alternative); QUICKSTART, Getting Started, and the README first step updated to match.

[0.3.0] - 2026-08-02

Changed

  • Renamed the project to Wren. The npm package is now create-wren; install with npm create wren@latest my-research. The previous package name create-research-assistant is superseded. Documentation moved to https://rchoularton.github.io/wren/.

[0.2.0] — 2026-08-01

Fixes a scaffold-time install failure reported on a real machine, hardens the setup pipeline, and ships four writing skills.

Fixed

  • Install no longer fails at verify. scrub_check (a maintainer publish gate) was running as a hard check during a user's install; with no git repo in the fresh project it scanned the gitignored research-config.yml and flagged the user's own home path. Removed it from verify_install; the scaffolder now runs git init; and scrub_check's no-git fallback excludes rendered/user files.
  • PyYAML bootstrap in setup.sh no longer aborts on PEP-668 "externally-managed" Python — it tries --user / --break-system-packages and prints clear guidance instead of hard-failing.
  • Config writing is YAML-safe. Interactive answers containing quotes, backslashes, or newlines no longer produce invalid research-config.yml.
  • paper:build / paper:extract now print a clear "run npm run paper:setup" message when python-docx / markdown aren't installed, instead of a traceback; QUICKSTART documents the one-time paper:setup step.
  • Memory slug now matches Claude Code's real project-dir encoding (all non-alphanumerics → -), so cross-session memory loads for project paths that contain spaces or other punctuation. Single shared sanitizer across the scripts.

Added

  • Four writing skills, genericised from the source system: /cover-letter, /revision-response, /peer-review, /literature-scan. These make the drafting → submission phases of the pipeline real.

Changed

  • Docs are honest about what ships: /qc-team and /figure are labelled planned (not live pipeline steps); removed the dead qc_suite / figure_suite / librarian config flags that nothing consumed.
  • Made the external-sync wording in paper:status provider-neutral (the sync folder is optional; it previously named one cloud provider).

[0.1.0] — 2026-08-01

Initial public release. A generic, agentic research assistant for Claude Code, extracted and genericised from a disaster-risk-finance research system with all domain content, identity, hosting choices, and secrets stripped.

Added

  • npm create research-assistant scaffolder — copies the kit into a new project directory, prompts for the handful of fields that make it yours, writes research-config.yml, then renders templates + installs hooks via the Python renderer.
  • Config-driven templating — research-config.yml drives CLAUDE.md / CLAUDE_REFERENCE.md / .mcp.json rendering; conditional MCP blocks per the chosen data and reference backends.
  • Paper workflow — Markdown⇄Word build/extract engine, per-paper STATUS.md, and the paper:* npm scripts.
  • Corpus-memory tier (tiers.research_memory) — a two-layer episodic + semantic memory over your literature and projects, with:
  • memory:bootstrap — idempotent scaffolder/repair for the memory tree.
  • memory:audit — offline health/completeness check (JSON log + severity findings).
  • Behavior rules, hooks, and skills — the rigor/tooling/one-at-a-time rule set, session hooks, and the core skills (paper, draft, gold-standard, research-memory, research, retro).
  • Pluggable backends — database (files / SQLite / DuckDB / local Directus / managed) and reference manager (BibTeX / Zotero / Mendeley) selected at setup.
  • Safety — a secret-scrub gate (setup:scrub-check) and protected-write-path guards.

Requirements

  • Node.js ≥ 18 (scaffolder is zero-dependency).
  • Python ≥ 3.9 on PATH (template renderer; auto-installs PyYAML on first run).

Not yet shipped (see ROADMAP.md)

  • Automated corpus-memory ingestion (the Zotero-coupled nightly engine).
  • Scheduled headless jobs, the QC/figure/peer-review agent suites, and a local database tier.

This page reuses the root CHANGELOG.md, which remains the canonical, editable copy.