Claude Code plugin marketplace — forge, foundry, and verification tooling
Agent
Hook-invoked fast-tier critic for adhoc methodical-mode iterative gate-dialog. Spawned by the Stop hook on rounds 1-2 of a substantive response. Runs a 10-item rule compliance audit using a cheap small-model pass. Returns per-rule pass/fail/N/A plus an overall verdict (CONCUR | CONCUR WITH CAVEATS | PUSH BACK | WRONG…
Agent
Hook-invoked deep-tier critic for adhoc methodical-mode iterative gate-dialog. Spawned by the Stop hook on rounds 3-5 of a substantive response (escalation tier — the cheap Haiku fast-critic ran rounds 1-2 and either flagged repeated failures or the response is long enough to warrant deep review). Runs the same…
Agent
Independent fresh-context critique. Spawn this agent when you want a take from a Claude that has NOT seen the conversation — useful for sanity-checking a recommendation, reviewing a plan before execution, or breaking out of a reasoning rut. Pass the full problem statement, the proposed approach, and any constraints in…
Agent
Crew critic. Adversarial verifier. Reads the original intent, the plan (if any), and the current state of code/systems. Runs real commands to check the work — does NOT read the worker's journal or trust the worker's self-report. Produces a falsifiable pass/fail verdict with an evidence ledger. Orchestrator rejects…
Agent
Crew fresh-eyes. Blind drift checker. Reads ONLY the original user intent and the current state of the codebase/systems — no brief, no plan, no journal, no critic verdict. Pretends to be a colleague who just walked up and was told what the user wanted. Checks whether the user got what they asked for. Mandatory final…
Agent
Crew router. Classifies a free-form intent into initial worker mode (investigating | planning | executing), scope (light | deep | paranoid), side-effect surface (none | local | remote | production), and flag interpretation (--paranoid, --yolo). One-shot classifier. Returns a JSON config the orchestrator uses to…
Agent
You verify that built code actually delivers what the spec promised. Your power is order: you read the spec and form hard expectations before you look at the code, so you judge the code against the requirement — not the requirement against the code. The default assumption is that the code is broken; VERIFIED is…
Agent
You are an expert engineer given unlimited time to produce correct work, not fast work. You implement exactly one unit of a larger spec. Your unit prompt is authoritative — it carries the spec requirements, acceptance criteria, your file boundary, and (in GRIND) a defect list. Do not re-scope it, soften it, or invent…
Agent
Isolated MCP-driven Playwright spec authoring. Spawn this agent when /e2e:write needs to handle a multi-step flow — the MCP browser snapshots are large and clutter the main context. Pass the flow description, baseURL, target spec path, and any constraints (auth role, starting route). The agent drives the browser…
Agent
V3 brownfield R2 methodology reference. IMPORTANT — this document is a METHODOLOGY REFERENCE, NOT a spawned subagent. V3 R2 runs in the main Claude thread (see forge/commands/plan.md §R2 V3 override) because subagents cannot call AskUserQuestion. Do NOT spawn this file as a subagent — without AskUserQuestion the…
Agent
Verifies stale-knowledge claims in reality.md and orients on ecosystem shape/gotchas for a feature category. Produces a research file consumed by the R2 interviewer. Spawned during R1.5 RESEARCH phase of /forge:plan.
Agent
Adversarial R3.5 spec reviewer. Reads transcript.md FIRST then the draft spec.md, emits spec-review.json with up to 5 A-NNN-cited ambiguity flags, and a binary block/pass verdict. Spawned during R3.5 SPEC REVIEW phase of /forge:plan after the draft spec is written and before the R4 Verbatim-Fidelity Gate.
Agent
This file is a Phase 3 test fixture only. It is NEVER read by a real Foundry run — F0.5's roster is hardcoded paths in plugins/foundry/commands/start.md, never a glob (per RESEARCH.md Pitfall 4). The file's only purpose is to give Plan 03-04 a synthetic agent whose minspecformatversion: v2.1 exceeds a legacy v2.0…
Agent
Maps an existing codebase's stack, architecture, structure, conventions, integrations, and concerns before decomposition. Produces six structured files in foundry-archive/{run}/codebase/ consumed by F0.5 DECOMPOSE and casting teammates. Spawned by the Foundry Lead during F0 when the target codebase is unfamiliar or…
Agent
F2 INSPECT stream for MIGRATION spec type. Runs a deterministic diff between each casting's coveragelist (source items to port) and the actual destination files, flagging missing destinations as COVERAGEINCOMPLETE defects.
Agent
Produces a grounded flow graph (flow-graph.json) of a brownfield codebase's data and control flows. Every node is anchored to a real file:symbol via LSP or grep — nothing invented. Spawned during V3 Forge R0 in brownfield mode to provide the attention anchor that prevents endpoint-anchored plumbing hallucination…
Agent
V3 INSPECT stream. Walks flow-delta.json forward from origin to sink, verifying each packet's produces exists in code AND actually consumes its declared upstream. Mirror of tracer.md (which verifies upstream callers). Paired with tracer, they cover both directions.