forge

Plugin forge from marketplace kraken in KrakenNet/kraken-plugins. Commands read from the repo's manifest — 1 month ago.

Skills (2)

• graph

  Spec graph layer for forge. SQLite-backed index over Phase 1 artifacts (prd.md, shared.md, interviews, research, reviews) + Phase 2 task array (prd.json) + outcome edges from Ralph Loop. Agents query graph instead of re-reading whole files for 5-10x token reduction.

• pipeline

  Foundation skill for forge. Defines 17-stage pipeline, artifact schemas, agent roster, gate contracts, and when each piece fires. Read by every forge command and agent at start.

Slash commands (8)

• /forge:graph

  Forge spec graph operations — rebuild index, query, get task context, list open questions, export to graphify.

• /forge:init-ci

  Install Phase 3 GitHub Actions workflow — ensemble reviewers, E2E gauntlet, PR formatter, auto-fixer.

• /forge:lessons

  View, prune, or query forge lessons (.forge/lessons.md). Reflexion-style log of generalizable insights from past task runs.

• /forge:new

  Start new forge run (greenfield or refactor). Triage gate picks path, runs Phase 1 (Perspective), hands off to /forge:scaffold.

• /forge:recipes

  Failure recipe DB (.forge/recipes.jsonl). Lookup past blocker resolutions, list, stats, manual add.

• /forge:resume

  Re-enter Ralph Loop mid-task. Reads .forge/prd.json, picks next failing task, dispatches ralph-coder.

• /forge:scaffold

  Phase 2 entry. Runs after Stage 5 human gate. Dispatches skeleton-scaffolder, contract-generator, task-sequencer in order. Emits prd.json and locks tests. Sets up Ralph Loop preconditions.

• /forge:status

  Show forge state — phase reached, prd.json task progress, last gate fired, blockers.

Subagents (17)

• anti-cheat

  Scans diffs for cheats and stubs — TODO/FIXME, NotImplementedError, hardcoded fakes, skipped tests, mock leakage into prod paths, demo data references at runtime, env fallback masking missing config. Used by Ralph Loop gate and PostToolUse hook.

• architecture-reviewer

  Architecture-fitness gate. Sister agent to anti-cheat — where anti-cheat catches lies about behavior (NotImplementedError, hardcoded fakes), this catches lies about shape (god objects, shotgun surgery, distributed coupling). Runs as Ralph Loop gate, Phase 3 CI check, and on-demand `/forge:status` he

• blast-radius-mapper

  Phase 1 refactor-path entry point. Scans git history + AST + import graph to map the blast radius of a proposed change. Restricts downstream context to only the files that matter. Replaces full dual-interrogation for /forge:new --fix.

• code-simplifier

  Phase 2 Stage 11. Refactors passing implementation for DRY and readability WITHOUT changing behavior. Runs after all Ralph Loop tasks pass. Tests are immutable — if a simplification breaks a locked test, the simplification is wrong.

• context-researcher

  Stage 2 parallel researcher. Scans existing codebase for reuse opportunities — similar features, established patterns, utilities. Output feeds PRD and technical designer.

• contract-generator

  Phase 2 Stage 8. Writes black-box adversarial validation scripts — Playwright for UI journeys, Postman for API contracts, Docker bash for service-level. These are the outer gate of the Ralph Loop's adversarial sandbox.

• critic

  Cheap critic agent. Reviews a draft diff against task acceptance criteria and locked test names BEFORE expensive gates run. Returns pass/concerns/block + line-specific findings. One haiku call saves multiple expensive gate cycles when the draft is obviously wrong.

• design-interrogator

  Creative/UX lead. Co-equal with pm-interrogator in Stage 1. Owns user journeys, mental models, empty/error/loading states, naming, affordances, accessibility. Forces UX decisions via AskUserQuestion before code.

• general-reviewer

  Inter-stage sanity gate. Fires between every Phase 1 stage. Checks for logical gaps, scope drift, contradictions with earlier artifacts, output completeness. Returns pass/concerns/block.

• lessons-keeper

  Reflexion writer. Runs after ralph-coder completes a task (pass OR block). Distills 1-3 short, generalizable lessons from the run and appends to .forge/lessons.md. Also adds recipes to .forge/recipes.jsonl when a blocker was resolved.

• pattern-researcher

  Stage 2 parallel researcher (UX side). Scans existing UI, design system, components, prior UX decisions. Paired with context-researcher.

• pm-interrogator

  Adversarial PM. Extracts edge cases, acceptance criteria, scope boundaries, deps, deadlines. Forces decisions via AskUserQuestion. Co-equal with design-interrogator in Stage 1.

• prd-writer

  Spec-anchored PRD writer. Consumes PM + Design interviews + Context + Pattern research. Emits .forge/prd.md (the constitution) — testable acceptance criteria, scope locks, UX contract refs.

• ralph-coder

  Inner Ralph Loop agent. Implements ONE task from .forge/prd.json. Reads lessons + recipes before coding, writes plan, codes, runs static + anti-cheat + test + adversarial gates, commits on green. Dispatches lessons-keeper post-task. Self-invokes until prd.json all green or blocked.

• skeleton-scaffolder

  Phase 2 Stage 6. Generates dirs, interface stubs, and failing unit tests from shared.md. TDD enforcer — code must be written against tests, not vice versa.

• task-sequencer

  Phase 2 Stage 9. Translates PRD + scaffold into .forge/prd.json — ordered task array, each tied to specific failing tests and acceptance criteria. All marked passes:false initially.

• technical-designer

  Stage 4. Translates PRD into architectural contract — components, data flow, interfaces, file layout, UX contract section. Emits .forge/shared.md.

Hooks (3)

• PostToolUse

  ${CLAUDE_PLUGIN_ROOT}/scripts/anti-cheat-scan.sh fast

• PostToolUse

  ${CLAUDE_PLUGIN_ROOT}/scripts/architecture-scan.sh fast

• UserPromptSubmit

  ${CLAUDE_PLUGIN_ROOT}/scripts/forge-context.sh

Similar plugins