By tduffield
Portable software-development plugin: general-purpose dev agents and dev-ritual skills for any Claude Code project.
Design advisor for non-trivial architectural or approach decisions. Use before committing to an implementation path when the choice is non-obvious and the cost of the wrong choice is high. Returns a recommendation with tradeoffs — not code. Runs on Sonnet with high effort. Good fits: - "Should this new feature be a new module or extend the existing one?" - "How should we structure the data flow between X and Y?" - "Is it worth extracting this into a shared module, or keep it inline?" - "We're choosing between three libraries for Z — which fits best?" Bad fits: - Implementation details after the approach is decided - Trivial naming or organization questions - Questions about how existing code works (use researcher instead)
Senior code reviewer. Reviews completed work against its plan and against quality standards. Returns findings categorized Critical / Important / Minor — not fixes. Runs on Opus with high effort in an isolated context. Good fits: - "Review Slice N of plan X before we continue" (spec compliance + code quality in one pass) - "Review this PR before I merge it" - "Completed a major feature — check it against the plan" Bad fits: - Running tests (dispatch `test-runner` instead) - Security-focused deep review on auth/crypto/secrets (dispatch `security-auditor` instead; this agent can flag but not deeply audit) - Fixing the issues it finds (caller's job)
Council role — User Advocate lens. Dispatched by a planning skill's mandatory council-lite review step for implementation-planning questions where the decision to build has already been made. Represents the end user: UX clarity, error messaging, accessibility, device/platform behavior, and the moments where a user will get confused, stuck, or frustrated. Returns a single-perspective response, NOT a synthesis. Use only when invoked by a planning skill's council-lite review step.
Council role — Builder lens. Dispatched by a planning skill's mandatory council-lite review step for implementation-planning questions where the decision to build has already been made. Focuses on *how* to build it: architecture, code structure, where it lives, libraries to lean on, prior art in this codebase and in the wild. Returns a single-perspective response (recommendation + evidence + uncertainty), NOT a synthesis. Use only when invoked by a planning skill's council-lite review step. For standalone architecture advice use `architect` instead.
Council role — Reliability lens. Dispatched by a planning skill's mandatory council-lite review step for implementation-planning questions where the decision to build has already been made. Focuses on how we'll know it works and how it behaves when it doesn't: tests, edge cases, failure modes, recovery, graceful degradation, and user abuse patterns. Returns a single-perspective response, NOT a synthesis. Use only when invoked by a planning skill's council-lite review step.
Batch up post-implementation follow-up items, clarify any ambiguities inline, then dispatch `sdd-implementer` to do the actual work in a cheaper subagent context — keeping mechanical iteration off the main session's expensive tokens. TRIGGER when: user invokes `/followup`, OR provides a numbered/bulleted list of post-implementation changes that includes the words "follow up", "follow-up", "followup", "follow ups", "fix-ups", "polish", or "iterate on these". DO NOT TRIGGER when: single small item (just do it inline), planning a new feature (use `planning`), debugging a specific failure (use `systematic-debugging`), or "follow up" appears as a passing reference ("I'll follow up on that later") rather than as the framing of a batch.
Record your current git state and shelve a session note so a future session can resume. Use for /forge:handoff, "hand off this work", "save my progress", "I'm done for now".
Resume a shelved work chain — surface the recorded git state and pickup hints from a prior /forge:handoff. Use for /forge:pickup, /forge:pickup <slug>, "pick up where I left off", "resume the work on X".
Use when starting a new feature, substantial change, or multi-step task that needs design alignment before code. TRIGGER when: user says "feature", "implement", "build", "add support for", "design", "plan", "architecture", or describes a multi-step task. DO NOT TRIGGER when: user says "fix", "bug", "debug", "broken", small single-file changes, or tasks with clear intent needing no design.
Use when completing tasks, implementing major features, or before merging to verify work meets requirements
Uses power tools
Uses Bash, Write, or Edit tools
Own this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimOwn this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimBased on adoption, maintenance, documentation, and repository signals. Not a security audit or endorsement.
A portable software-development plugin for Claude Code: general-purpose dev agents and dev-ritual skills that work in any project, with no app-specific assumptions baked in.
forge is the dev-tooling sibling of lore (portable knowledge management). Where lore owns what you know, forge owns how you build: the reusable agents and rituals a developer reaches for regardless of which codebase they're in.
Extending forge for your project? Forge's planning, code-review, and
subagent-dev skills carry named extension points (feature flags, issue tracker,
observability, test commands). The full adopter guide lives in lore's
docs/EXTENDING.md — that is the entry-point for wiring the extension points
to your own stack. The plugins/forge/docs/DEGRADATION.md in this repo is the
canonical reference for each extension point's visible-skip behavior and re-add
path.
Agent fleet populated (v0.2.0). The 15 general dev agents are present and
genericized — they carry zero app-specific or vault-specific strings and
register as forge:<name>:
architect, researcher, troubleshootercode-reviewer, security-auditordoc-finder, test-runner, log-sifter, pr-summarizersdd-assumption-prover, sdd-implementercouncil-builder, council-reliability, council-security, council-advocateThe four council review-lens agents (council-*) are dispatched as a
parallel quartet by a planning skill's council-lite review step — not
standalone. Each holds a single perspective (Builder=architecture,
Reliability=tests/failure modes, Security=threat model, Advocate=UX) and
returns a focused single-lens response rather than a synthesis. If you need
general architecture advice outside a planning context, use architect instead.
First skills shipped. /forge:handoff and /forge:pickup are the first
dev-ritual skills — a symmetric shelve/resume pair (see What lives here
below). They stood up forge's skill test harness
(tests/test_skills_registrable.py + tests/test_skills_generic.py).
A proof-of-life agent (forge-ping) also ships to confirm plugin agent
registration works. Two agents are intentionally not here: planner moves
in a later phase alongside the /planning skill, and design-mockup-writer
is bound to a specific app's visual aesthetic and lives in that app's own repo.
code-simplifier is a separate plugin and was never part of this fleet.
Nothing app-specific belongs in forge; per-project automation stays in that
project's own repo.
Agents (plugins/forge/agents/) — general dev subagents, dispatchable as
forge:<name> once installed.
Skills (plugins/forge/skills/) — dev-ritual skills, invocable as
/forge:<name>:
handoff — record read-only git state + shelve a session note with pickup
hints so a future session can resume.pickup — resume a shelved work chain (surface recorded git state + hints).lore-optional coupling: the handoff/pickup pair drives the lore
CLI (lore handoff / the session-note finder) when it's available, and
degrades to a local forge handoff file at ~/.forge/handoffs/<slug>.md — out
of any repo — when lore is absent, $LORE_VAULT is unset, or lore stats
fails. This is the same one-directional optional dependency forge already has
on lore (forge → lore is allowed; lore never depends on forge). The git
capture is strictly read-only — these rituals record state, they do not
commit, push, or rebase your code.
.claude-plugin/marketplace.json # local dev marketplace (source: ./plugins/forge)
plugins/forge/
.claude-plugin/plugin.json # plugin manifest
agents/ # dispatchable subagents
skills/ # /forge: ritual skills
tests/ # packaging + registrability invariants
Claude Code rejects source: "." — the plugin must live in a plugins/forge/
subdir referenced by source: "./plugins/forge" in the root marketplace.
/plugin marketplace add /path/to/forge
/plugin install forge@forge-local
Then restart the session and confirm with the forge-ping agent. See
MANUAL-SMOKE.md for the full boundary smoke test.
A generic, denylist-driven pre-publish check that blocks a commit when a private string would ship into a publishable repo. The mechanism ships zero private strings — every forbidden token lives in a machine-local denylist that is never tracked in any repo:
plugins/forge/scripts/leak_gate.py # the gate (denylist-driven, fail-closed)
plugins/forge/scripts/install-hooks.sh # chain-safe pre-commit installer
~/.claude/leak-gate.denylist # machine-local denylist (UNTRACKED)
Run it directly:
npx claudepluginhub tduffield/forge --plugin forgePortable knowledge-management plugin: session lifecycle, capture skills, and vault recall for any Claude Code project.
Access thousands of AI prompts and skills directly in your AI coding assistant. Search prompts, discover skills, save your own, and improve prompts with AI.
Complete developer toolkit for Claude Code
Intelligent draw.io diagramming plugin with AI-powered diagram generation, multi-platform embedding (GitHub, Confluence, Azure DevOps, Notion, Teams, Harness), conditional formatting, live data binding, and MCP server integration for programmatic diagram creation and management.
Feature development with code-architect/explorer/reviewer agents, CLAUDE.md audit and session learnings, and Agent Skills creation with eval benchmarking from Anthropic.
Orchestrate multi-agent teams for parallel code review, hypothesis-driven debugging, and coordinated feature development using Claude Code's Agent Teams
Production-grade engineering skills for AI coding agents — covering the full software development lifecycle from spec to ship.