Before work: Define the result
A goal, scope and acceptance criteria exist before implementation begins.
QG-0Keep the agent you already use.Give it an engineering process.
TAUSIK adds tasks, durable project memory and verification gates to Claude Code, Codex, Cursor and other coding agents. It runs locally inside your repository.
One project record connects the request, the decisions and the checks that make “done” credible.
- Open source
- Local project state
- Model-agnostic
Start with one request.
No new interface to learn. Open the repository and tell your coding agent:
Prefer to install it yourself? Show manual commands
- 1
Before the first edit
The requested outcome becomes a task with scope and acceptance criteria.
- 2
While the agent works
Decisions, constraints and dead ends survive the session and the model.
- 3
Before “done”
The task closes only against fresh verification and criterion evidence.
Two required checks keep the work inside its boundaries.
Between them, the agent works as usual. TAUSIK intervenes before work begins and before it closes—where a mistake costs the most.
Before closure: Show the proof
Affected checks and criterion evidence belong to the current state of the code.
QG-2AGENTS.md asks. TAUSIK checks.
“Run the tests before you finish.”
That instruction is useful and stays in your repository. But text alone cannot know whether the tests ran for this commit. TAUSIK adds local state, allowed transitions and a record the next session can verify.
- Instructions
- Describe the behaviour you want.
- TAUSIK
- Records whether the required transition actually happened.
Three layers of accountable agent work.
SENAR defines delivery discipline. RENAR keeps the reasoning trace intact. TAUSIK turns both standards into a working system inside the repository.
SENAR
Delivery standard. Defines the delivery discipline: context before code, explicit acceptance, verification evidence and recoverable history.
Read the standard ↗RENAR
Decision protocol. Preserves the link between a requirement, its interpretation and the resulting decision. It guides reasoning without blocking the work.
Read the standard ↗TAUSIK
Working system. Runs that contract locally through tasks, memory, MCP, skills, hooks and signed receipts.
Install TAUSIK →A new session does not mean a new project.
TAUSIK restores the active task, the last handoff, relevant decisions and approaches that already failed. The agent receives the context it needs instead of re-reading an entire repository and repeating the same exploration.
What returns with the work
- Goal and acceptance criteria
- Decisions and constraints
- Known dead ends
- Latest verification state
Less framework repetition. The same quality contract.
TAUSIK separates a small set of permanent rules from task-specific context and loads detail only when the work requires it.
Measured in the TAUSIK repository on 1–2 October 2026. These are reductions in repeated context and payload, not a promise of lower weekly credits or accepted-task cost.
Read the measurement boundary →Straight answers
What is TAUSIK?
A local open-source governance and verification framework that works on top of an AI coding agent.
Does it replace Claude Code, Codex or Cursor?
No. It gives the agent you already use tasks, durable state, lifecycle gates and verification evidence.
Is it just an alternative to AGENTS.md?
No. Keep your instruction files. TAUSIK adds state and enforceable task transitions that text cannot provide alone.
Does it block every edit in every host?
No. Real-time edit hooks depend on host support and trust. QG-0 and QG-2 still govern task start and closure across supported MCP hosts.
Will it reduce my token bill?
It reduces measured repeated context. It does not promise a fixed saving in weekly credits or task cost.