Releases · hebstr/claude-code-plugins

audit: the calibration-memory redirect stub was matched only as Canonical index:, so a memory store reached through a Canonical location: stub never resolved and sweep/walkthrough reviews silently ran uncalibrated. Both labels now match (case-insensitive), and the extracted path is normalized (surrounding backticks and whitespace stripped, trailing punctuation removed, leading ~ expanded).

audit: calibration loading globs scoped feedback_review_severity*.md variants (e.g. _personal) instead of a single fixed filename, reading every match.
audit: walkthrough-only mode now loads prior calibration once before the finding loop, deriving the project root by upward walk; previously only orchestrator mode loaded calibration.

Initial release of the hebstr marketplace: two plugins (audit, workflow) covering 10 skills.

walkthrough: interactive, point-by-point walkthrough of any review report. Orchestrator mode (target + reviewer) and walkthrough-only mode (existing report). Re-evaluates findings, proposes fixes, checks impacted files for regressions.
sweep: full-coverage project review with parallel specialist agents (architecture, quality, tests, docs), consolidated into a single deduplicated report sorted by severity.
blindspot: circularity-aware orchestrator. Detects when a reviewer shares its target's codebase, prompts, or model family, then injects cross-model judging via OpenRouter and convergence analysis. Reviewer is discovered at runtime via the sibling walkthrough scan-reviewers.py script and confirmed with the user. Explicit-invocation only.
skill-adversary: adversarial critic for Claude Code skills. Reports trigger edge cases, instruction ambiguities, contradictions, cross-file coherence issues, and gaps.
mcp-adversary: adversarial critic for MCP servers. Reports inter-tool discrimination issues, schema anti-patterns, semantic drift, error handling inconsistencies, and undocumented workflow dependencies.

sync: scan files for staleness relative to recent changes and propose targeted updates. Always runs a cross-repo semantic consistency scan with parallel agents. Explicit-invocation only.
write: strip AI writing patterns and rewrite prose to sound human. Routes to a French or English reference based on the text. Includes a bilingual review mode (FR↔EN parity, typography, faux amis). Explicit-invocation only.
continue: flush durable facts to memory, update .claude/PLAN.md, and print a minimal continuation prompt. No handoff document is written; PLAN.md and memory are the authoritative stores. Explicit-invocation only.
reco: deep-mode recommendation backed by external sources. Spawns two parallel agents (official documentation via WebFetch, community practice via WebSearch), then synthesizes a structured recommendation with verified citations. URLs are verified before citing; source disagreement is surfaced, not papered over. Explicit-invocation only.
doc-structure: audit project documentation layout (CLAUDE.md vs README.md), propose verbatim migrations of misplaced prose, and update the CLAUDE.md index. 5-phase workflow (Discovery, Audit, Migration, Index updates, Verification). Pre-bundle reference check flags broken intra-file anchors, relative paths, and image references. Explicit-invocation only.

skill-adversary and mcp-adversary: sub-agents use path-based file access instead of embedded prompts, removing prompt-injection surface.

Provide feedback