v1.6.1 — maintenance fixes and a model-era refresh

Maintenance: three routine fixes, a model-era refresh, and verdict-first
completion reports.

Changed

• rules/api-cost-optimization.md covers the current model tiers (Haiku 4.5,
  Sonnet 5, Opus 5.5, Fable 5.1); every claim was checked against the source
  (docs/model-era-review-2026-09-27.md).
• Completion reports start with Done: yes or Done: no — missing: …,
  scoped per visible surface (rules/slop-prevention.md).
• A correction about how you work with the agent (not about one codebase) is
  stored once, user-global, in ~/.claude/rules/preferences.local.md
  (rules/domain-docs-convention.md, template
  templates/preferences.local.md.template).
• scripts/ledger-append-proposed.sh recomputes the ledger's header counters
  on every write and offers --recompute-only / --check.

Fixed

• The subagent watchdog no longer reads a bare number such as a character
  count of 529 as an overload error; it needs error context around 429/529.
• Daily docs: the desktop scan counts unparseable lines instead of aborting,
  and file symlinks inside the scanned tree are tolerated without double
  counting (docs/adr/0004-desktop-audit-lone-surrogate-tolerance.md).
• Signal rotation writes an explicit empty shard for days without signals,
  merges late entries instead of overwriting them, and exits loudly when
  continuity cannot be proven; daily metrics book such days as verified empty.

v1.6.0 — English throughout, self-improvement front and center

English throughout, and self-improvement front and center.

Changed

• The framework, its docs and ADRs (renamed: docs/adr/0002-rcode-plan-follows-practice.md,
  docs/adr/0003-vault-and-gate.md) and the Cockpit are English throughout.
  The Hausbau output style's instruction text is English and still answers
  in the asker's language. German input keywords (skill triggers, report-only
  classifier phrases) remain accepted aliases alongside their English forms.

Added

• Self-improvement is presented as a headline feature on the website home
  page and in the README.
• CLAUDE_DAILY_DOCS_LANG (en default, or de) in ~/.claude/env.local.sh
  sets the language of the daily-docs logbook: its headings and the
  "not capturable" sentences logbook-count.sh emits. JSON keys are unchanged.

Fixed

• scripts/tests/routine-run-regression.sh ran its real-claude parser probe
  in the project folder, which started a short-lived Claude session there. It
  now runs in a throwaway folder.
• Cockpit (companion repo claude-cockpit): the pane follows the session in
  its folder that was most recently active, not the one that started last, so
  a short-lived background claude -p call can no longer take it over and
  leave every card waiting.

v1.5.0 — plain developer language by default, a Cockpit that always shows the wave

Plain developer language by default, and a Cockpit that always shows the
current wave of work.

Changed

• Default output style: none. A fresh install answers in ordinary
  software-developer language. The Hausbau output style (every change
  explained through a house-building metaphor, for readers who are not deeply
  technical) still ships in output-styles/hausbau.md and is opt-in:
  /output-style Hausbau. Until 1.4.1 it was switched on for everyone.
• /team-lead and the control agent record every wave as a task list
  (TaskCreate, set to in_progress at dispatch and completed in the wave
  review), so the wave is visible next to the session while it runs.

Added

• Publish step removes the maintainer's personal outputStyle from the
  published settings.json; regression suite
  hooks/tests/publish-strip-regression.sh.
• Cockpit (companion repo claude-cockpit): card 4 derives a "Welle" (wave)
  from the subagent start/stop hook events, one row per subagent, in every
  session that dispatches subagents — independent of whether the model keeps
  a task list, which appears above it when present.

v1.4.1 — the Cockpit, public and installable

The Cockpit is now public and installable, and the website and README show it
as it runs today. No rule, hook, skill, agent or script behavior changed for a
public install.

Added

• Cockpit setup, step by step: the Cockpit page ("Set it up"), the install
  page ("Optional: the Cockpit") and the README carry the one-line installer
  of the companion repo claude-cockpit. It lists every change before making
  it (Ghostty, tmux, Node and jq via Homebrew; the Cockpit under
  ~/.claude/cockpit; the cockpit command; ten entries in
  ~/.claude/settings.json; Ghostty keybindings and the rcode color theme),
  asks once, backs up every file it edits, and supports --dry-run.

Changed

• Cockpit product images: the Subagents card names model and effort for
  every subagent and adds a status line (its task while running, "idle" once
  done); the Workflows card lists every task with its own bar.
• Website Cockpit page describes both cards.

Fixed

• The publish step also removes the private Cockpit subagentStatusLine
  setting, so a public install never points at a script it does not have.

v1.4.0 — R.Code for Claude Code

A name that follows Anthropic's rules, a website, and a front page that shows
what the framework does. No rule, hook, skill, agent, script or setting changed.

Added

• Website at https://rcode-for-claude-code.vercel.app/: home, install,
  how it works, Cockpit and vault pages. Static files in site/, served by
  Vercel; deployed by the maintainer with each release (docs/PUBLISHING.md).
• Two animated terminal demos in docs/assets/demo/: the Autonomy Arbiter
  stopping gh pr merge until you answer, and /team-lead running three
  agents while the Cockpit shows their progress.
• Brand assets in docs/assets/brand/: wordmark, banners for light and dark
  themes, social preview image (1280×640), favicon, and the architecture
  diagram (architecture-{dark,light}.svg) in the brand's own style.
• Cockpit product images in docs/assets/cockpit/: Claude Code and the
  Cockpit side by side in Ghostty, and the Cockpit column up close. Demo
  data only.
• The block composition from the banner (colored, modular blocks) is the
  brand's primary motif and runs through every page of the website.

Changed

• The project is named R.Code for Claude Code. Anthropic does not allow
  "Claude" or "Claude Code" as part of a third-party product name. The
  repository name claude-rcode stays.
• README rebuilt as a front page: banner, tagline "Build smarter. Build
  better.", four benefits, demos, one-line install, generated component counts.
• Trademark notice on the README and every site page.

v1.3.0 — Pseudonymized by design

Pseudonymized by design: real names, machine paths and personal identifiers
never enter the framework. They live only in a local vault on the
developer's machine; the framework carries tokens. Decision record:
docs/adr/0003-tresor-und-tor.md.

Added

• Local vault ~/.claude/vault/ (git-ignored, secret 0600) as the single
  source of real values, with scripts/vault/vault.sh
  (init | add | check | tokenize | resolve | token | status | doctor | prune-public).
  Tokens are keyed HMAC-SHA256 prefixes (proj-…, acct-…, id-…) or fixed
  role tokens (<dir>, <email>, <private-repo-url>, <BAUHOF>).
• One matcher for every consumer: scripts/vault/lib.sh.
• Write gate hooks/vault-write-gate.sh: refuses real values in Write, Edit,
  MultiEdit and Serena writes to tracked files of a framework repo.
• Commit gate scripts/git-hooks/pre-commit, also run as pre-merge-commit;
  install with bash scripts/install-git-hooks.sh --only pre-commit. Blocks
  staged real values and force-added ignored files.
• Vault gate on the improvement-ledger write path (scripts/ledger-append-vault-gate.sh).
• Contribution path for third parties: templates/imp-submission.template.md
  and scripts/imp-submit.sh (validates a submission against the vault, no
  network access). See CONTRIBUTING.md.
• Templates: templates/env.local.sh.template (machine-specific runtime
  values), templates/reminders.local.md.template,
  templates/automode-environment.template.json.
• scripts/vault/public-names.txt: names that are public by decision and
  never treated as vault terms.

Changed

• Machine paths and personal names in prose, rules, skills, hooks, scripts and
  scheduled tasks replaced by role tokens or environment variables; runtime
  values come from ~/.claude/env.local.sh, git remote, or the system.
• autoMode is an installation-only settings key; it is no longer shipped in
  settings.json and survives claude-deploy.
• Routine timers use generic launchd labels (com.claude-code.routine-<task>);
  scripts/install-routine-timers.sh migrates the previous per-user labels.
• scripts/scrub-check.sh uses the vault matcher for every PII finding and
  redacts values in CI (GITHUB_ACTIONS, CI, or --redact).
• In commit mode (--staged) the rebrand check only counts added lines, so
  touching a file with accepted historical mentions no longer blocks a commit.
• Publish transforms 30-placeholder-scan and 50-pseudonymize are no-ops on
  the now-clean tree.
• This changelog replaces the previous stub that pointed to GitHub Releases.

Removed

• Publish transforms 60-declaw-scrub-check-pii and
  65-generalize-reminders-section.
• Hard-coded detection literals in scripts/scrub-check.sh.

Fixed

• publish-manifest.txt excluded every directory named vault/, which also
  dropped the vault tool scripts/vault/ from the published tree; the pattern
  is now anchored to the root (/vault/).
• Renamed files were skipped by the staged scan (--diff-filter=AMRC).
• The HMAC secret is never passed on a command line.

Security

• CI output of scrub-check.sh never prints a matched value.

Upgrade notes

1. bash scripts/vault/vault.sh init, then add your own private terms with
   vault.sh add.
2. Copy templates/env.local.sh.template to ~/.claude/env.local.sh and fill
   in your values.
3. bash scripts/install-git-hooks.sh --only pre-commit in your working copy.
4. bash ~/.claude/scripts/install-routine-timers.sh if you use the routines.

v1.1.0

v1.1.0 — R.Code "Plan folgt Praxis", routine liveness, hardened publishing

First release since v1.0.1 (2026-08-06). 162 files changed.

R.Code workflow

• /team-lead is now the single documented entrance for R.Code project work; /plan-team … /launch-team remain as thin aliases that force a Stage. Stage playbooks live in rcode/stages/, read by reference instead of being copied into five commands.
• Two tracker modes: github (issues) or plan (P-NNN checkbox units in BRAINSTORM.md) — GitHub is no longer a hard prerequisite.
• Binding glossary (Phase / Stage / Step / Work unit / Tracker / Check trio) and a two-tier ADR model; CONTEXT.md template added. Rationale: docs/adr/0002-rcode-plan-folgt-praxis.md.
• /rcode-upgrade now commits the applied rail update after one explicit y/n; /decompose is idempotent and gated before creating GitHub objects.

Hooks and routines

• Session-start liveness check for scheduled routines: two consecutive failed runs surface as a blocker line at session start instead of in a quarterly audit.
• Background-agent watchdog: every SubagentStop is logged; troops without an end signal are reported at the next prompt (three-state: end_turn, abnormal, unknown).
• Return channel for automatic security reviews: findings are persisted and announced at session start instead of vanishing with the parent session.
• general-purpose dispatch gate asks for a rationale that names and rejects a specialist; a series of such dispatches is escalated.
• New visual-qa-agent (read-only browser inspection); research-agent restored to the live roster.
• Routine runner fix: the SKILL.md prompt is passed via stdin (the --- frontmatter line was parsed as an option), with a regression suite.

Publishing pipeline

• The scrub gate runs against the staging tree with the private, unmodified scanner and requires a maintainer-local term list; missing list = no publish.
• Maintainer-specific transforms are applied but no longer shipped; private hook registrations (dashboard, external binaries) are stripped from the public settings.json.
• The public publish.sh always ships disabled.

Known limitations

• The shipped settings.json still carries the maintainer's model/effort defaults; adjust after install.
• Some public docs reference maintainer-only transform scripts that are intentionally not shipped.