New framework: TAUSIK

Verification layer for AI coding agents that gates task start and completion on specifications and evidence, with project memory and signed receipts.

Reviewed 2 weeks ago · Sep 30, 2026

[Website](https://tausik.tech/)
 [GitHub · Kibertum/tausik-core](https://github.com/Kibertum/tausik-core)
          
[Compare with other frameworks →](https://sddobservatory.com/compare/?f=tausik,grill-me-skills,spec-kit)

• Stars
• 18
• Contributors
• 1
• Open issues
• 1
• Age
• 6 months
• Last push
• yesterday
• Latest release
• v1.11.4 Oct 12, 2026

View weekly commit data #

Metrics updated Oct 13, 2026 · collected automatically from the GitHub API

Core approach #

Runs over an existing agent through MCP tools, skills, and host-dependent hooks. QG-0 requires a goal and acceptance criteria before task start; QG-2 checks verification evidence before task completion. Scoped verification runs can issue ed25519-signed receipts recording the checked code state, while a local database preserves decisions and dead ends across sessions.

Workflow #

1. Install as a git submodule and run the Python bootstrap for the chosen host
2. /start — open a session and load project state and knowledge
3. /plan — capture requirements and create tasks with goals and acceptance criteria
4. /task <slug> — pass QG-0 and implement within the task scope
5. tausik verify --task <slug> — run scoped checks and issue a receipt when signing is configured
6. /ship — review, test, verify acceptance criteria, pass QG-2, and commit
7. /end — capture knowledge and prepare the session handoff

Supported tools #

• Claude Code
• Codex
• Qwen Code
• Cursor
• Kilo
• OpenCode

Strengths #

• Task gates connect a written definition of done to recorded verification results
• Signed receipts can be exported and checked offline with tausik receipt verify
• Persistent project knowledge records decisions, patterns, and failed approaches
• Integrates with several existing agents rather than requiring a replacement agent loop

Limitations #

• Hooks are strongest in Claude Code, Qwen Code, and trusted Codex projects; other hosts primarily enforce task start and completion
• The signing key is accessible in the working tree: receipts provide integrity evidence, not independent attestation against the agent
• Verification coverage depends on declared scope and configured gates; recorded exceptions and unsigned completion paths exist
• No general token savings or independent production outcomes were established; the maintainer's reported token comparison found no saving

Notes #

TAUSIK (Task Agent Unified Supervision, Inspection & Knowledge) implements a task-specification and verification workflow over existing coding agents. It declares conformance to SENAR Core. The reviewed repository documents bootstrap installation, host integrations, task gates, receipts, and self-use in TAUSIK’s own development.

The submission’s claim that every task needs a passing, signed receipt is broader than the documented behavior. The receipt guide retains a fresh-run completion path without an explicit signed handle, and the task reference documents a recorded acknowledgment when no test gate applies. Receipt scope divergence is recorded and does not always block completion. These distinctions matter when interpreting a green result.

The experimental rating reflects limited independently verifiable adoption and the evolving verification contract. This review checked public documentation and the verify-first gate source; it did not run the software or independently validate the project’s self-reported effectiveness.