English | Русский

Skill ecosystem — one-page map

Single entry point for vendor / repo skills (not the 13 core slash skills + /brain conditional in harness/skills/ — those ship with TAUSIK and are documented in Skills).

Flow (install path)

flowchart TD
  R[skill repo add URL] --> I[skill install name]
  I --> A[skill activate name]
  A --> U[Ready in agent context]
  I --> V[vendor cache]

CLI steps (mirror of CLI — Skills; run via .tausik/tausik):

1. skill repo add <url> — register a TAUSIK-compatible repo (tausik-skills.json / legacy skills.json).
2. skill install <name> — clone if needed, copy skill files, install declared pip dependencies.
3. skill activate <name> — promote an installed skill into the IDE skill load path (see Vendor skills — three tiers).
4. skill list — see active, vendored, and available skills.

Deactivate / remove: skill deactivate <name> · skill uninstall <name> · skill repo remove <name> (see CLI reference).

Risks

• skill repo add without --force only succeeds for the official Kibertum/tausik-skills URL; third-party repos require an explicit opt-in after review (Vendor skills — Repo trust).
• External repos run arbitrary instructions in SKILL.md and may execute scripts during install. Treat unknown repos like untrusted code. Details: Vendor skills (security & trust).
• Context budget: each active skill consumes prompt space; keep only frequently used skills activated.

Claude-native sub-agents

Beyond slash skills, TAUSIK ships named sub-agents that the Agent tool can invoke directly (Agent(subagent_type="<name>", ...)). They live under harness/claude/subagents/ in the source tree and bootstrap deploys them to .claude/agents/. Cursor / Qwen do not have a named-subagent concept — these are Claude-only.

Sub-agent	Invoked by	What it does
tausik-reviewer	/review lite	Single-pass code review with the SENAR 28-item checklist + security docs. Returns structured JSON {critical[], high[], medium[], low[]} so the main context never sees per-file reads or per-agent prose. ~2.9KB definition; rubric loaded at runtime. Token-economy alternative to the default 6-agent fork — pick it for low-stakes diffs and skip it for security-sensitive code.
tausik-gate-fixer	/debug (auto-helper after a failed tausik verify)	Reads the gate's stderr + project troubleshooting docs, returns a 1-3 step JSON fix plan {gate, family, plan: [{step, action, target, change, why}], meta}. Read-only PLAN agent — never applies edits; the invoker applies the plan and re-runs tausik verify. Action vocabulary is fixed (edit, extract_module, add_test, move_file, delete_dead_code, re_run_gate).

Adding a sub-agent: drop <name>.md (with frontmatter name, description, tools, model) into harness/claude/subagents/, then python bootstrap/bootstrap.py --ide claude re-deploys it. Tools should be minimal — read-only sub-agents (Read, Grep, Bash) keep the trust surface small.

Where to go next

Topic	Document
Repo format, pip deps, tausik-skills.json	Vendor skills
Slash skills /plan, /ship, core 13	Skills
IDE-specific skill paths	Skill adaptation
MCP tool names for agents	MCP — tausik_skill_*