v0.2.0

[0.2.0] — 2026-06-11

Features

• Reserve M/B/X static keys + dup-key validation + truncate_at opt-in

---

Generated by git-cliff

v0.1.6

[0.1.6] — 2026-06-10

Bug Fixes

• M1-M5 + R1/S1 menu-rendering bugs
• Reword pyproject comment to drop a fullwidth homograph that tripped a skillaudit FP

---

Generated by git-cliff

v0.1.5

[0.1.5] — 2026-05-17

Bug Fixes

• Correct \.git→\.git doubled-backslash in _current_repo_slug regex

Documentation

• Add version badge for publish.py auto-update

Features

• Install canonical CPV pipeline + fix all validate findings

Miscellaneous Tasks

• Sync uv.lock to v0.1.4

Styling

• Ruff format publish.py + suppress upstream-template lint noise

---

Generated by git-cliff

v0.1.4

Fixed

• Critical install bug — .claude-plugin/plugin.json declared
  "hooks": "./hooks/hooks.json", which pointed the manifest override
  at the auto-discovered default file. Claude Code rejects this with
  Duplicate hooks file detected AND silently disables the plugin's
  hooks (and any sibling MCP servers). v0.1.3 and earlier would not
  load on any Claude Code instance. Fix: remove the hooks field
  entirely — the canonical hooks/hooks.json is auto-discovered.

  This is empirical plugin-loading bug #2 catalogued by CPV
  (claude-plugins-validation) under RC-PLUGIN-LOAD-002. The very
  bug-class that motivated CPV's loader-audit rules. Ironic.

Changed

• README rewritten to lead with the value proposition: "The
  AskQuestion prompt is not good for a big number of choices. Just
  run this command and you get a nice menu with as many entries as
  you need." Followed by a single copy-paste snippet. ~240 lines
  trimmed to ~80.

v0.1.3

Removed

• skills/render-menu/SKILL.md — the skill path is gone. Benchmarks
  in v0.1.2 (under CLAUDE_CODE_EFFORT_LEVEL=max opus) showed the skill
  fork took 15-25 seconds end-to-end vs 25 ms for direct Bash to
  menu_write.py, dominated by extended-thinking model inference no
  skill-frontmatter knob could escape. effort: low, removing model:,
  removing allowed-tools:, removing context: fork — none helped on
  a max-effort session. No user would wait that long for a menu.

Changed

• README rewritten to lead with Option 1 (direct Bash invocation).
  Plugin authors call python3 $CLAUDE_PLUGIN_ROOT/scripts/menu_write.py spec.json
  from any Bash block in their orchestrator. ~25 ms. Reliable. No fork.
• commands/menu-test.md updated to invoke menu_write.py directly
  instead of the skill.
• commands/menu-render.md updated to invoke menu_write.py directly
  instead of the skill.
• Plugin now ships only hooks + scripts + commands + examples.

Why

The skill existed for the use case "consumer wants to delegate menu
rendering to a fork so the parent doesn't pay any tokens". In practice:
(a) the fork's wall time was unacceptable, (b) the parent was going to
spend Bash-tool tokens anyway, and (c) direct Bash is a 25 ms call —
faster than any fork could be. The skill was solving the wrong problem.

Migration for consumers

If you were using Skill({skill: "claude-menu-system:render-menu", args: "spec.json"}):

# Replace with:
python3 "$CLAUDE_PLUGIN_ROOT/scripts/menu_write.py" "spec.json"

If your plugin is in another marketplace, resolve the cache path:

CMS_ROOT=$(ls -d ~/.claude/plugins/cache/emasoft-plugins/claude-menu-system/*/ | sort -V | tail -1)
python3 "$CMS_ROOT/scripts/menu_write.py" spec.json

v0.1.2

Fixed

• Critical UX bug — emitted systemMessage now prepends a \n so
  the menu box starts on its own line beneath Claude Code's Stop says:
  banner. v0.1.1 emitted the box's first row inline with the banner,
  shifting every subsequent row right and breaking visual alignment.
  Regression-pinned by test_systemMessage_always_starts_with_newline.
• render-menu skill body — rewritten to NOT rely on the
  $ARGUMENTS placeholder (slash-command syntax, not skill syntax).
  The haiku fork now receives explicit instructions to take the user
  message as the path and invoke the Bash tool with it. v0.1.1's
  body left the fork describing what it would do rather than actually
  executing the script.

v0.1.1

Added

• /menu-test slash command — self-test that renders a demo menu so
  users can verify the plugin is wired correctly after install.
• /menu-render <spec-path> slash command — ad-hoc rendering from any
  JSON spec file on disk; useful for plugin authors testing their specs.
• scripts/menu_test_specs.py — helper script that produces canonical
  demo specs for all 8 modes. Used by /menu-test and re-used to seed
  the examples/ gallery.
• examples/ directory with 8 canonical JSON specs (one per render
  mode) — plugin authors can copy-paste these as starting points.

Notes

• All changes additive — no behavior change to v0.1.0 hooks / skill /
  renderer. Consumer plugins built against v0.1.0 work unchanged on v0.1.1.

v0.1.0

Initial release. First-class, reusable menu rendering for Claude Code
plugins via the systemMessage-from-Stop-hook pattern empirically
proven against Emasoft/token-reporter-plugin.

Added

• render-menu skill (context: fork, model: haiku,
  agent: general-purpose) — other plugins invoke via
  Skill({skill: "claude-menu-system:render-menu", args: "<spec>"}).
• Stop / StopFailure hooks that emit the queued menu as
  systemMessage at main-session turn end.
• SubagentStop hook that no-ops (systemMessage from subagent-stop
  routes to AI context, not the user terminal — documented).
• 8 render modes: menu, summary, breakdown, status_table,
  panel, multi_box, progress, confirm.
• 5 box styles: heavy (default for menus), rounded (default for
  panels), light, double, ascii. Auto-downgrade to ascii when
  the terminal can't render unicode box-drawing.
• Semantic ANSI palette (border / header / value / success / warning /
  error / info / muted + severity tiers). Auto-strip when NO_COLOR=1,
  CLAUDE_MENU_COLOR=0, or TERM=dumb.
• Display-width-aware padding via unicodedata.east_asian_width —
  correct alignment for emoji + CJK + box-drawing glyphs.
• ANSI-safe text wrapping that preserves color state across continuation
  lines (ported from token-reporter _wrap_ansi).
• Auto-terminal-width detection (COLUMNS env → shutil.get_terminal_size
  → 80 default, capped at CLAUDE_MENU_MAX_WIDTH default 120).
• 10K hook-output cap with tiered truncation (SMALL untouched / BIG
  row-dropped / OVERFLOW title-stubbed) and line-safe truncation
  (cut only at \n boundaries, never mid-ANSI-escape).
• Session-isolated queue dir
  (${TMPDIR}/claude-menu-system/<session_id>/).
• flock-protected concurrent emit (multiple Stop hooks don't race).
• TTL-based stale cleanup (default 60s, configurable via
  CLAUDE_MENU_TTL_SEC).
• 188 tests (≥95% line coverage on every module), all running in
  under 1 second via pytest-xdist.

Notes

• Runtime dependency: stdlib only (no third-party deps).
• Plugin-scoped hooks — picked up by /reload-plugins, no full session
  restart needed.
• Consumer plugins must hard-require claude-menu-system to be enabled.
  Pre-flight check pattern in the README.