v0.4.2 — Rename advisor-skill → pipeline-suggest

UX disambiguation release. The lightweight auto-trigger skill is now /skill-advisor:pipeline-suggest.

Why

Users could not distinguish the skill from the /skill-advisor:advisor command — both shared the "advisor" prefix and the previous name was generic.

Path	Behavior
/skill-advisor:pipeline-suggest (this skill)	Quick prose dry-run, single confirmation
/skill-advisor:advisor (the command)	Full router→gate flow with per-step AskUserQuestion picker (recommendation + alternatives at every phase)

Changes

• skills/advisor-skill/ → skills/pipeline-suggest/ (directory rename)
• SKILL.md frontmatter name: advisor-skill → pipeline-suggest
• SKILL.md description rewritten with comparison table pointing at /skill-advisor:advisor
• CLAUDE.md plugin-structure note updated
• New regression test file: tests/skill-rename-pipeline-suggest.test.js (6 cases)

Tests

654 tests pass (was 648). New test acts as a regression fence: future PRs can't reintroduce the old name without explicit allowlist update.

Migration

Replace any hardcoded /skill-advisor:advisor-skill references in your scripts/prompts with /skill-advisor:pipeline-suggest. The /skill-advisor:advisor command is unchanged.

Upgrade

claude plugin marketplace update FX-studio-AI
claude plugin update skill-advisor@FX-studio-AI

🤖 Generated with Claude Code

v0.3.1 — Gate Delegation Hardening

Runtime-Critical Fixes

This release ships two critical fixes that together complete the approval-gate delegation refactor.

1. Gate Delegation Hardening (commit 94294e5)

Eleven adversarial findings addressed across three batches, each followed by an independent review and fix-pass:

• Batch 1 (CRITICAL + HIGH) — prompt-injection defense contract (Section 6.1 + Rule 12), SSOT routing table for decision → action, single enforcement owner for clarification+planning phases, CLAUDE.md developer docs updated.
• Batch 2 (MEDIUM hardening) — Step 7 pre-check expanded from 4 to 8 items; re-spawn retry policy with gate_respawn_count and hint field; cooldown script error observability + shell-injection fix (node -e path via env var); Other free-text retry cap with accent-folded negation-aware semantic mapping.
• Batch 3 (LOW polish) — Step 5 score explainer aligned with router's actual output; gate_token naming clarity notes; 4 tech-debt items tracked in TODOS.md.

36 findings total across 3 adversarial rounds. 431/431 tests pass.

2. Subagent tools frontmatter fix (commit b24f027)

Runtime testing revealed that the gate subagent had no access to AskUserQuestion — a deferred tool not in the default subagent tool set — causing it to silently fall back to the ASCII-box prompt that the delegation refactor was supposed to replace.

Fix: one line in agents/advisor-gate.md frontmatter:

```yaml
tools: AskUserQuestion, Agent, Skill, Read
```

This bug was not caught by three independent static reviews — only by actual end-to-end execution. Lesson: runtime smoke tests matter for LLM-driven control flow.

What users see

• /advisor gate approval menus now render as native arrow-key selectable menus (via AskUserQuestion) instead of ASCII-box prompts.
• Better error handling: failed subagent responses trigger bounded re-spawns instead of silent loops.
• Downstream agent spawns (router re-spawn for Alterar, planning skill for spec generation) now enforce the same escaping contract as the top-level spawn.

Upgrade

```bash
/plugin marketplace update FX-studio-AI
/plugin update skill-advisor
```

Then restart Claude Code to pick up the new subagent frontmatter.

v0.3.0 — AskUserQuestion for all user interactions

Interaction pattern overhaul: every user prompt in this plugin now goes through the AskUserQuestion tool (selectable list with arrow-key navigation) instead of prose prompts. Claude Code automatically appends an "Other" option for free-text fallback.

Changes

• agents/advisor-gate.md — Moment 1, Moment 2, 3-alternatives submenu, and legacy-mode confirmation all use AskUserQuestion. New Interaction Policy section + Rule 11 forbidding manual "Other" options.
• commands/advisor-feedback.md — the four feedback questions batched into a single AskUserQuestion call with options pulled from the last loadout.

Backward compatibility

• Gate JSON output contract unchanged — validateGateOutput test still passes
• 431/431 tests passing, 0 regressions
• Hook behavior unchanged (only UX pattern of prompts evolved)

Update path

Marketplace users: just update via your plugin manager. The marketplace manifest (FX-Studio-AI) now reports v0.3.0 and Claude Code will pull HEAD of main which matches this tag.

See v0.2.1 release for the hook stdin JSON fix that must be in every install.

v0.2.1 — Hook stdin fix + metadata alignment

Critical fix: the UserPromptSubmit hook was reading process.env.CLAUDE_USER_PROMPT (which Claude Code never sets), causing it to silently no-op on every prompt. Hook now reads the JSON payload from stdin using fs.readFileSync(0) (synchronous, honors the "no async" constraint).

Fixes

• SA-003 (critical): hook now receives real prompts via stdin JSON. Env var preserved as fallback for tests/manual invocation.
• SA-004: skills/advisor-skill/SKILL.md frontmatter name aligned (advisor-skill, matches directory + catalog).
• SA-005: package.json and plugin.json versions synced to 0.2.1.
• SA-009: guard against missing/empty candidate.sequence in replay hint.

Tests

• 424 → 431 passing (7 new tests covering stdin JSON contract)
• No regressions

Why you want this

Before v0.2.1 the hook produced zero output for real user prompts in any Claude Code install. This release is the difference between a silent no-op and a working nudge.