v8.19.3 (2026-07-01, PATCH — arm-classifier: keyword-inferred read-only runs no longer arm [recurring phantom-arm deadlock fix]): a natural-language review/audit/analysis request ('faça revisão adversarial', 'faça uma auditoria', 'análise') was classified FULL/Audit and armed the pipeline-arm-pending marker — but read-only work never opens an exec window, so it hit INLINE_WORK_BLOCKED, aborted, and left an orphan marker that bricked the next task with PIPELINE_NOT_ARMED (observed live in interview-coach 2026-06-30 and 07-01). FIX: lib/pipeline-arm.cjs::writeArmPending now also skips arming for the read-only TYPES (Audit, UX Simulation) when they were INFERRED FROM KEYWORDS (source kw:), while an EXPLICIT --audit/--ux flag stays a deliberate governed run and still arms (preserving the STEP_1_7 direct-entry evidence, v8.16.0). Safe direction per the documented arm trade-off: a missed arm is recoverable (re-invoke); a phantom arm deadlocks. TDD tests/regression/v8.19.3/F1-readonly-types-no-arm (RED 4-fail → GREEN); 2 deliberate v8.7.0 pins updated to the --audit-flag path. Iron Law: only lib/pipeline-arm.cjs + tests/ touched; Gates/Registry UNTOUCHED; no .codex mirror. Suite 252/5 (5 pre-existing Langfuse/env/F4). Lockstep across all version surfaces. v8.19.2 (2026-07-01, PATCH — edit-guard-hook robustness [Cursor-crash hardening, Claude-Code-scoped]): the PreToolUse edit-guard no longer crashes into the opaque external catch ('edit-guard-hook internal error — failing closed') when .pipeline/sessions is present-but-unreadable (a file where a dir is expected → ENOTDIR/EACCES) — getActiveLock + getActiveExecWindow now guard readdirSync and return null instead of throwing, and the CLI external catch surfaces the REAL sanitized error (single-line, printable-ASCII, capped 200) instead of the opaque message, so a genuine failure stays fail-closed but debuggable. Posture UNCHANGED (still deny when the guard cannot evaluate); NO foreign tool-name vocabulary added — cross-runti
Output.ai workflow development toolkit for Claude Code. Adds 5 specialist agents (planner, builder, debugger, prompt writer, quality reviewer), 40+ slash-command skills covering scaffolding, debugging, evaluation, and credential management, plus a SessionStart hook that auto-loads Output SDK conventions so Claude understands the framework before the first prompt.
v2.43.0 (2026-07-28) - REMOVED: the H-Sutra header Stop enforcement layer. hooks/h-sutra-enforce.sh is deleted and its Stop registration is gone. Nothing blocks or forces a redo for a missing or malformed header on any profile. The header itself survives as a convention: /core:start still documents the format, per-turn-discipline-prompt.sh still asks for it every turn, and the 9-cell classification log rail (holding/state/interaction/log.jsonl or \u002esutra/h-sutra.jsonl) is untouched -- it was always written by per-turn-discipline-prompt.sh, not by the enforcer. What dies with the hook: the .enforcement/h-sutra-audit.jsonl enforcement telemetry and the SUTRA_HSUTRA_ENFORCE_DISABLED / ~/.h-sutra-enforce-disabled kill-switches (nothing left to disable). Stop floors that REMAIN hard: per-turn-hard-gate.sh (Input Routing, Depth, BLUEPRINT) and flow-stop-check.sh (Flow, profile-gated). sutra-defaults.json now records per_turn_blocks.human_sutra_header.enforcement = convention_only, and the /core:start CLAUDE.md template no longer claims the header is hard-enforced. Founder direction 2026-07-28. + v2.42.0 (2026-07-28) - Two Stop-layer fixes reach the fleet. (1) blueprint-check.sh v3 is TEXT-FIRST (#81): it validates the BLUEPRINT block in the turn's response text -- the block the user actually sees -- instead of the .claude/blueprint-registered marker. Emitting a correct block now unblocks Edit/Write; the marker becomes a per-turn cache the hook writes for itself. PreToolUse enforcement narrows to FOUNDATIONAL paths (now configurable via per_turn_blocks.blueprint.foundational_paths in sutra-defaults.json, overridable per repo with blueprint_foundational_paths[] in .claude/sutra-project.json); ordinary files are floored at Stop by per-turn-hard-gate.sh on any turn that mutated a governed file. Closes the divergence behind #68, the 2026-07-08 Testlify incident and its 2026-07-27 repeat. (2) h-sutra-enforce.sh v8/v9 + flow-stop-check.sh honor .profile (#72): DIRECTION-VERB
Highly opinionated Product → Blueprint → Plan → Execute workflow for solo developers or small teams — enforced workspace structure with a selectable stack-template menu checked by /vwf:doctor, a flow-first blueprint (entities as data contracts, OpenAPI + JSON-Schema artifacts) with whole-product coherence review, chained dependency plans, an execute that asks only when the plan is silent, with review/security/acceptance/UX subagents behind one final merge gate that stamps blueprint implementation state, post-deploy verify with production release freezing + API backward-compat enforcement, and feedback intake.
A coordinated AI scrum team. The /build command orchestrates specialist agents (specification writer, scrum master, UI/UX designer, implementation engineer, infrastructure engineer, test runner, and code reviewer) directly from the main conversation, plus reusable skills for development workflows.
Enough process for the work at hand. No more. Compass assesses each piece of work by risk, familiarity and size, then picks the approach to match: what to write, what to check, and when to ask a human. A typo fix stays light. A payments rewrite gets specs, tests and review. Five default guardrails apply to all work that ships code, such as a passing test for each change.
The complete Vishwakarma skill catalog: 21 frontend engineering skills that teach an agent to build interfaces that look designed, not generated — with evidence-carrying rules and self-review verification built in.
Complete AEP-X skill set for aepx-core: one skill per service, the connector bus and each connector category, plus SDK, CLI, consoles, blockchain, governance, RFCs/conformance, database, CI/Jenkins, docs and architecture.
Full-SDLC automation for Rust microservices built from the Rust service template: one-command service creation, bounded-context scaffolding, GitHub issue creation, BMAD-method planning, bmalph/Ralph implementation with Claude CLI, multi-lens code review with the BMAD FR/NFR gate, black-box REST and GraphQL QA, CI auto-fix, and AI review comment resolution, looping until the PR is finished.
Eighteen workflow skills for coding agents (develop, debugging, code-review, verification, intake, research, and more). Hooks route each prompt to one skill, announce the workflow risk, and keep a review reminder open until review evidence arrives, while a git guard blocks destructive git commands. A Sonnet worker runs the workflow skills; read-only reviewer, auditor, and researcher subagents report back.
This plugin provides a specialized suite of skills for data engineers and database practitioners working on Google Cloud. It acts as an expert assistant, allowing you to use natural language prompts in your preferred coding agent to architect complex data pipelines, transform data with dbt, write Spark and BigQuery SQL notebooks, and orchestrate end-to-end workflows across GCP's data ecosystem.
Structured data from Google results, AI Mode, Maps, Amazon, Walmart, Zillow, Indeed, Yelp, YouTube, TikTok and flights, or any URL as markdown. Connects to HasData for you. Commands for local lead lists, price checks and company enrichment.
Full-SDLC automation for PHP backend engineering: GitHub issue creation, BMAD-method planning, bmalph/Ralph autonomous implementation with Claude CLI, multi-skill code review with BMAD FR/NFR gate, QA verification, CI auto-fix, and AI review comment resolution — looping until the PR is fully finished.