krishnenduk95

Deep analytical reasoning agent powered by GLM 5.1 at max reasoning (32k thinking budget). Use for non-code reasoning tasks Opus would otherwise do itself — architecture tradeoff analysis, capacity planning, choosing between libraries/databases/frameworks, ranking options, decomposing ambiguous problems, reviewing proposed designs, building decision matrices. Returns structured analysis with explicit assumptions, quantified tradeoffs where possible, and a recommendation. Use this to offload reasoning-heavy questions from Opus without losing depth.

REST / GraphQL / RPC API design specialist powered by GLM 5.1 at max reasoning. Use when designing or implementing API endpoints — routes, input validation, error responses, authentication gates, rate limiting, versioning. Produces the API layer (route handlers + validation schemas + OpenAPI specs + focused tests). Outputs production-grade code, not sketches. Pair with `glm-schema-designer` for the DB layer underneath.

System architecture specialist powered by GLM 5.1 at max reasoning. Use when designing a new system, feature, or major refactor: decides service boundaries, tech stack, data model shape, integration points, and produces a plan file that other subagents can execute. Outputs a structured `plans/NNN-<slug>.md` with dependency DAG, ownership tags, and acceptance criteria. Not for implementation — dispatch `glm-worker`, `glm-schema-designer`, etc., for that.

Read-only codebase investigation agent powered by GLM 5.1 at max reasoning (32k thinking budget) with Opus-level investigative rigor. Use to answer questions about how code works, find where a feature is implemented, trace call graphs, locate files matching patterns, understand data flow, or gather context before Opus designs a change. Returns hypothesis-driven findings with file:line evidence. Cheaper than the built-in Explore agent; use this first for code questions to preserve Opus's context window.

Diff review agent powered by GLM 5.1 at max reasoning (32k thinking budget) with staff-engineer rigor. Use to review code changes (staged diffs, branch diffs, PR diffs, or specific files) across correctness, style, perf, tests, docs, and basic security. Returns a structured review with severity-tagged findings. Use for routine review passes; Opus handles deep security/auth/billing review itself.

Security review specialist powered by GLM 5.1 at max reasoning. Use to audit code for common security issues — OWASP top 10, injection, auth/authz gaps, secret leakage, unsafe deserialization, weak crypto, SSRF, insecure direct object reference. Read-only — produces a structured security report with severity tags. For anything CRITICAL or involving auth/crypto/billing/PII, auto-escalates back to Opus for human-staff review. Pair with glm-reviewer for routine diffs.

Test-case generator powered by GLM 5.1 at max reasoning. Use to write exhaustive test suites — unit, integration, property-based, and fuzz tests — from a spec or existing code. Generates tests BEFORE implementation when doing TDD, or AFTER to shore up coverage of legacy code. Produces tests that fail for real reasons, not flaky tests that waste CI minutes.

UI component builder powered by GLM 5.1 at max reasoning. Use to build React / Next.js / Vue / Svelte / SwiftUI / React Native components from a design brief. Produces working components with proper state management, loading/error/empty states, accessibility, and tests. Matches existing project style (Tailwind / CSS modules / shadcn / Chakra — whatever the repo uses).

High-volume code execution agent powered by GLM 5.1 at max reasoning (32k thinking budget), engineered to produce Opus-quality output. Use for implementing features from a precise plan, writing CRUD endpoints, repositories, handlers, migrations, boilerplate, scaffolding, tests from spec, UI components from design, dependency bumps, and any mechanical pattern-matching work. NOT for architectural decisions, security review, or hard debugging — delegate those back to Opus (the main session). Every dispatch requires a detailed brief with exact file paths, acceptance criteria, and constraints.

Verifier-gated best-of-N — dispatch glm-worker 3× with varied approaches, score with glm-reviewer, pick winner. Use on hard tasks where correctness dominates cost.

Run maximum-depth reasoning on a question — dispatches glm-analyst with chain-of-debate, synthesizes with Opus.

Kick off the dual-model (Opus plans, GLM executes, Opus reviews) workflow on the current request

Full SaaS feature build orchestration — dispatches architect → schema → api → ui → tests → security → docs subagents in proper dependency order, with test-driven verification at each stage. Produces a plan file and executes it end-to-end.