English | Русский
AI development framework — plan, build, ship with quality control.
Three messages. Full engineering cycle. Quality gates that the agent can't skip.
Works with Claude Code, Cursor, Qwen Code (GigaCode), Windsurf.
Tell your AI agent:
Add https://github.com/Kibertum/tausik-core as a git submodule in .tausik-lib,
run python .tausik-lib/bootstrap/bootstrap.py --init,
add .tausik/ to .gitignore
The agent will execute all three steps. Restart your IDE after — done.
Then just three messages:
start working
fix the bug — button doesn't work on mobile
ship it
That's it. The agent opens a session, creates a task with acceptance criteria, writes the code, runs tests and code review, verifies each criterion with evidence, commits, and offers to push. Full engineering cycle — you just describe what you want.
| Without TAUSIK | With TAUSIK |
|---|---|
| Agent starts coding immediately | Must define goal + acceptance criteria first |
| Claims "done" without proof | Completion blocked until every criterion has evidence |
| Context lost between sessions | Decisions, patterns, and dead ends persist across sessions |
| Same mistake repeated 3 times | Failed approaches recorded — agent sees what didn't work |
| No tests, no linting | 16 checks auto-run for your stack (pytest, ruff, tsc, eslint, cargo, go vet...) |
| No visibility into process | 6 metrics tracked automatically — throughput, defect rate, lead time |
The key difference: TAUSIK quality gates are enforcement. The agent physically cannot skip steps — no prompts, no hoping, no "please remember to run tests."
If you prefer to set things up yourself:
cd your-project
git submodule add https://github.com/Kibertum/tausik-core .tausik-lib
python .tausik-lib/bootstrap/bootstrap.py --initBootstrap auto-detects your tech stack and enables matching quality gates. Project name is derived from the directory.
After bootstrap, restart your IDE so MCP servers load. Without a restart the agent falls back to CLI mode.
Plan before code. /plan starts with an interview — the agent asks clarifying questions about behavior, edge cases, and constraints. Then it creates tasks with acceptance criteria. No coding until the goal is defined.
Ship with confidence. /ship runs code review (5 parallel agents), tests, verifies each acceptance criterion, commits, and suggests updating docs. One command — full quality pipeline.
Remember everything. Decisions, patterns, conventions, and failed approaches are stored in a local SQLite database with full-text search. New session? The agent picks up right where you left off.
Enforce, not suggest. Quality gates block the agent at two checkpoints: QG-0 (can't start without goal + criteria) and QG-2 (can't close without evidence + passing tests). Hooks block code edits without a task and dangerous shell commands in real time.
Agents in long sessions "drift" — ignore the framework, skip task creation, forget conventions. TAUSIK 1.2 adds real-time drift guards:
- SessionStart hook auto-injects state (active tasks, Memory Block, conventions) into every new session — no manual
/start. - UserPromptSubmit hook detects coding-intent phrases (EN+RU) and nudges the agent if no task is active.
- Stop hook keyword detector catches "I'll implement" / "сейчас напишу" announcements and blocks the stop if no task exists.
- PostToolUse verify-fix-loop audits every
task_doneagainst 5 rule-based checks (file paths, ✓ markers, test counts, file refs, lint status). - Memory Block re-injection — recent decisions + conventions + dead ends pushed back into context on
/startand/checkpoint. - Adversarial critic — 6th parallel
/reviewagent finds 3 weaknesses the others miss.
Plus: /interview (Socratic Q&A before complex tasks), tausik hud (one-screen live dashboard), tausik suggest-model (Haiku/Sonnet/Opus routing), webhook notifications (Slack/Discord/Telegram).
- 35 skills —
/plan,/ship,/review,/audit,/debug,/test,/interview, and more - 82 MCP tools — full programmatic access to the project database
- 16 quality checks — pytest, ruff, tsc, eslint, cargo check, go vet, and more for your stack
- 6 automatic metrics — throughput, first-pass success rate, defect rate, lead time
- Project memory — SQLite + FTS5, graph relations, dead-end tracking, Memory Block re-injection
- 7 Claude Code hooks — task gate, bash firewall, push gate, auto-format, SessionStart, UserPromptSubmit, Stop × 2, PostToolUse verify
- Batch execution —
/run plan.mdexecutes multi-task plans autonomously - Zero dependencies — Python 3.11+ stdlib only; MCP deps in isolated
.tausik/venv/
| IDE | MCP Tools | Skills | Hooks | Rules |
|---|---|---|---|---|
| Claude Code | 80 tools | 34 skills | task gate, bash firewall, push gate, auto-format | CLAUDE.md |
| Qwen Code | 80 tools | 34 skills | task gate, bash firewall, push gate, auto-format | QWEN.md |
| Cursor | 80 tools | 34 skills | — | .cursorrules |
| Windsurf | 80 tools | 34 skills | — | .windsurfrules |
| Codex | — | — | — | AGENTS.md only |
Hooks block code edits without a task, dangerous shell commands, and direct push to main — in real time. Available in Claude Code and Qwen Code. Cursor and Windsurf get the same MCP tools and skills, with quality gates at task start and task done.
TAUSIK was developed using itself. Real numbers:
| Metric | Value |
|---|---|
| Tasks completed | 291 |
| Sessions | 22 |
| Throughput | ~13 tasks/session |
| Test count | 918 |
| Dependencies | 0 core |
Every feature, every refactor, every bug fix went through the same quality gates that ship with the framework.
TAUSIK implements SENAR (GitHub) — an open engineering standard for AI-assisted development. Quality gates, session management, metrics, verification checklists — it's all defined in SENAR.
| Document | Description |
|---|---|
| Quick Start | First setup — 10-15 minutes |
| What is SENAR? | The methodology behind TAUSIK |
| Workflow | A typical day with TAUSIK |
| Skills | 34 structured agent workflows |
| Hooks | Real-time enforcement |
| CLI Commands | Terminal command reference |
| MCP Tools | 80 tools for the AI agent |
| Architecture | How the framework works inside |