Every agentMọi agent
Skills deploy these specialized subagents, each in its own context with its own tools. Most are advisory — they report and do not mutate plan or code unless tasked. The agents guide explains how they compose (the critique panel); this is the full roster with each one's model and tool count.Skill triển khai các subagent chuyên biệt này, mỗi cái trong context riêng với tool riêng. Phần lớn là tư vấn — báo cáo, không sửa plan/code trừ khi được giao. Hướng dẫn agent giải thích cách chúng kết hợp (hội đồng critique); đây là danh sách đầy đủ kèm model và số tool.
advisor
Runs the hs:advise interview in isolation — scout, then one question at a time to turn a raw idea into exact requirements and honest advice. It relays each question back to the orchestrator.
brainstormer
Explores 2-3 genuinely different technical directions with quantified trade-offs and second-order effects, before any code is written. Use before committing to an approach or a major refactor.
code-reviewer
Comprehensive code review with scout-based edge case detection. Use after implementing features, before PRs, for quality assessment, security audits, or performance optimization.
code-simplifier
Simplifies and refines code for clarity, consistency, and maintainability while preserving all functionality. Focuses on recently modified code unless instructed otherwise.
critique-consolidator
Merges findings from several independent critique lenses into one ranked, de-duplicated verdict: PASS / PASS_WITH_RISK / BLOCKED, with severity and repeat-offense detection.
debugger
Diagnoses a failure to its root cause — errors, bottlenecks, failed deployments, failing tests — by collecting real evidence (logs, traces, database state) and reporting what it proves.
decision-reconciler
Tier-2 reconcile pass over the whole Decision Register: semantic contradictions between rulings, code that no longer honors a DEC, DEC-id leaks in code. It FLAGS flips; it never edits the SSOT.
developer
Implements one scoped slice test-first — in place for a sequential phase, or in an isolated worktree as a parallel-team slice that owns a disjoint file set and touches no other owner's surface.
docs-manager
Writes and maintains technical documentation: keeps docs current with code changes, organizes structure, and holds terminology consistent across the tree.
escalation-consultant
Autonomous counsel from the strongest model in one run — no model switch, no interview. Spawn from a lower tier or a stuck subagent for a hard design call. Returns advice, not code.
git-manager
Stage, commit, and push code changes with conventional commits. Use when user says "commit", "push", or finishes a feature/fix.
grid-axis-selector
Picks which axes and values belong in a grid-planning grid before it is built, and writes one grid-axis-src.json receipt of why each was kept or dropped. Never runs the engine or edits a grid.
grid-filler
Fills a SKELETON grid cell with raw content under the anti-confab contract, writing ONE JSON fill-source file for a replay invoker. Never edits the grid; STUBs a cell it cannot ground.
independent-revalidator
Re-derives a load-bearing conclusion from primary evidence BEFORE reading the first-round reasoning, then surfaces where the two disagree. A sealed-room re-derivation, not a re-read.
lane-relayer
Carries ONE request to a partner lane (gemini, ccs) and returns its envelope verbatim. A courier for context isolation and fan-out: the lane skill has already built the command.
market-fit-critic
Critique lens on whether a product-bearing artifact stands up in its market: defensibility, differentiation, path to value capture — separate from whether users want it or it is buildable.
planner
Researches and writes an implementation plan for a feature, architecture, or migration — before significant work starts, or when the trade-off between approaches is still open.
product-value-critic
Critique lens on whether an artifact builds something users actually need: job-to-be-done, value proposition, riskiest unproven belief — separate from whether the work is technically sound.
project-manager
Tracks progress against a plan, consolidates status from completed agent tasks, and names blockers — reporting with data, not effort or intent.
red-teamer
Adversarially attacks an artifact ABOUT TO EXECUTE — a plan before cook, a migration before it runs — hunting the failure mode the author and the happy path both missed.
researcher
Investigates a technical topic across real sources — docs, packages, open-source projects — and returns a ranked, evidence-weighted report with one concrete recommendation.
spec-craft-critic
Critique lens on the WRITING of a spec-family artifact under docs/product/: plain language, clear/concise/consistent/correct, unmeasurable adjectives, term drift. Never edits the spec.
spec-tech-critic
Critique lens on whether a spec-family artifact under docs/product/ is buildable and testable as written: INVEST, Given-When-Then, hidden dependencies, NFR gaps. Never edits the spec.
tester
Validates a change by running it — unit and integration tests, coverage against the project floor, error handling, build. Call after a feature or a bug fix.
ui-ux-designer
UI/UX work: interface design, wireframes, design systems, responsive layout. Also reviews a newly built UI for accessibility, consistency, and mobile behaviour without being asked.
workflow-orchestrator
Designs a spawn strategy before you delegate — subagents vs Workflow vs Teams, fan-out grouping, sub-count, consolidate cadence, early-write paths — WITHOUT spawning anything itself.