Arthur

Harness-aligned Claude Code marketplace: behavioral steering, context management, memory architecture, evaluation, orchestration, and multi-agent decomposition.

Behavioral steering via modular hooks: 13 principle-based rules, destructive command blocking, scope compliance guard, self-review enforcement, output style safety

Context management: progressive pressure tracking, drift detection, edit safety, large-file warnings, plan mode hooks, MCP monitoring, prompt budget awareness, failure-escalation with safe-mode flag.

Long-running session continuity: init.sh bootstrap, append-only claude-progress.txt log, features.json testable-requirements list, /session-resume briefing.

Three-tier memory architecture: pointer index, on-demand topics, searchable transcripts. Version-aware updates snapshot prior topic contents to the lineage ledger. Includes /lineage-audit for ledger invariant inspection.

Evaluator-optimizer pattern: adversarial review, verification gates (TaskCompleted), static analysis, postmortems, reference-fidelity check, self-evolution proposal assessment.

Hook-driven agentic workflow orchestrator. Automatic routing (shell or LLM-hybrid), sprint-contract enforcement, TDD loop, session bootstrap, handoff nudges. Self-evolution SEPL loop (propose/assess/commit/rollback) over versioned resources. Composes agents/evaluator/context-engine/memory rather than duplicating them.

Multi-agent decomposition: planner/generator/reviewer pipeline triad with tool-isolated capability boundaries, plus a read-only researcher for pre-planning investigation. Includes /worktree-team for physical worktree isolation, directory-ownership enforcement (opt-in), and output-schema checks on generator exit.

Power-user reference: extended thinking, unix pipes, parallel execution patterns

Execution trace collection, analysis, and harness evolution. Stores structured JSONL traces for pattern detection and failure-driven improvement. Feeds /evolve as a proposal source, plus per-session clarification-waste analysis.

Always-on compressed communication. Drops articles, filler, pleasantries from responses. Reduces output tokens ~65% while keeping technical accuracy. Loaded at session start, survives compaction.

On-demand rewrite that strips AI-writing tells from a piece of text so it reads natural and human. Detects 30+ patterns from Wikipedia's "Signs of AI writing" guide and can match your own voice from a sample. Explicit-invocation, no hooks.

Prevents token waste: duplicate read detection, large output warnings, session token audit

Codebase health scanning: documentation drift, registration gaps, convention violations, stale memory, HARNESS_SPEC invariant compliance. /validate-marketplace performs pre-commit mechanical correctness checks (distinct from /entropy-scan). /startup-profile reports SessionStart hook latency. /policies-list prints the cross-plugin enforcement registry (FS01–FS42).

Policy Gateway (TRAE §5.2.4) between planner and execution: PII/secrets scan, prompt-injection detection, sensitive-ops audit. Same permissionDecision:deny contract as behavioral-core block-destructive. Rules live in rules.d/ so SEPL can evolve them.

Mechanical enforcement of read-before-edit: blocks Edit/Write on files not Read in the current session. Based on issue #42796 data showing 33.7% of edits made to unread files.

Auto-install rtk-ai/rtk (Rust token killer) on first session and register its global Bash rewrite hook. Second SessionStart hook surfaces a remediation warning when the binary is missing or unhealthy. Opt-out: FORGE_RTK_DISABLED=1.

Auto-install colbymchenry/codegraph. Claude Code queries a tree-sitter structural graph of the current repo through MCP instead of re-reading files on every task. Second SessionStart hook surfaces a remediation warning when the MCP server is missing or unhealthy. Claude Code only. Opt-out: FORGE_CODE_GRAPH_DISABLED=1.

Curated color themes for Claude Code: Catppuccin Mocha, Tokyo Night, Nord. Pick via /theme; Ctrl+E to fork into ~/.claude/themes/ for editing.

Parallel work across sibling repos with result aggregation

Self-evolution boundary: change-manifest writer, evolution-history ledger, session-digest, auto-tune-skill outer loop, controllability invariant (POLICY.md).

Native-git stacked-PR management: push guard (blocks wrong-branch and bare --force pushes), stack-graph state, restack/reparent/submit workflows built on git rebase --update-refs and --onto. No third-party stacking binary.