English | Русский

Workflow

TAUSIK is designed for pair work: the engineer writes in free form, the AI agent interprets and executes. No special commands to memorize — just describe what you want to do.

Typical Work Day

Morning: Starting Work

Write to the agent:

start working

The agent will open a session, show what was done last time, which tasks are in progress, and suggest what to work on. If there are unfinished tasks — it will offer to continue.

Working on a Task

For simple tasks — just say what needs to be done:

add dark theme to user settings

The agent will create a task, formulate acceptance criteria, and start working.

For complex tasks it's better to plan first:

let's plan the migration from REST to GraphQL

The agent will create a task with a detailed plan, break it into steps, estimate complexity, and offer to begin.

Review and Completion

When the work is done:

done, review and commit

The agent will check the code against a 28-point checklist, run tests and gates, verify that all acceptance criteria are met, and offer to commit.

End of Day

that's all for today

The agent will show the summary: what was done, how many tasks were closed, metrics. It will save context for the next session — tomorrow you can continue from where you left off.

Two Work Modes

Quick (for small tasks)

Engineer: "start working"                → /start (opens session)
Engineer: "fix the JWT bug"              → /plan (creates task, plans)
Engineer: "done"                         → /ship (verifies, commits)

Full (for complex tasks)

Engineer: "start"                        → /start (context, metrics)
Engineer: "plan the API refactoring"     → /plan (task + plan + AC)
Engineer: "go ahead"                     → /task (QG-0, begins work)
  ... work, progress, dead ends ...
Engineer: "review the code"              → /review (28-item checklist)
Engineer: "run tests"                    → /test
Engineer: "close and commit"             → /ship (QG-2, gates, commit)
Engineer: "that's all for today"         → /end (metrics, handoff)

Quality Gates

TAUSIK automatically checks quality at two points:

At task start (QG-0):

• Task goal is formulated
• Acceptance criteria are recorded — and say something: templates (<...>, [UPPER], {{...}}), an unexpanded $(...), TODO/TBD, an empty table and "works as expected" are removed first; under 3 words of goal or 5 of criteria is refused
• Blocks if criteria don't include a negative scenario (error, failure, invalid input)
• Warns for security tasks (auth, payments, PII) without security criteria
• Warns if scope is not defined (what to change / what not to touch)

At task completion (QG-2):

• Each acceptance criterion is verified with evidence
• All plan steps are completed
• Tests pass (pytest, ruff, and other gates per stack)
• Warns if knowledge is not documented

These gates cannot be bypassed — the agent cannot start work without a goal and cannot close a task without verification.

When Gates Block You

QG-0 blocks task start:

• Missing goal → add with task update <slug> --goal "..."
• Missing acceptance criteria → add with task update <slug> --acceptance-criteria "..."
• Goal or AC is a placeholder ("the acceptance criteria have 0 word(s) of substance") → write what is checked and how
• No negative scenario in AC → add a criterion like "Returns error on invalid input"
• A session above the advisory threshold is advice, not a refusal: save state with /checkpoint or hand off with /end
• The host-thread budget is separate from session time. Native Codex response/input/context signals trigger an automatic checkpoint and fresh-window prompt at advisory level; hard level refuses continuation. Reopening TAUSIK inside the same host thread does not reset it. Missing native usage is shown as unavailable and fails open.

QG-2 blocks task completion:

• AC not verified → log evidence: task log <slug> "AC verified: 1. ... ✓ 2. ... ✓"
• Tests failing → fix the code, tests run automatically on next task done
• Plan steps incomplete → mark done with task step <slug> <N> or update the plan

The agent handles most of this automatically. If a gate blocks, it will tell you exactly what's missing and how to fix it.

Hooks — Automatic Control

In addition to Quality Gates, TAUSIK uses Claude Code hooks for real-time control:

• No code without a task — attempting to edit a file without an active task is blocked
• Dangerous command firewall — rm -rf, DROP TABLE, git reset --hard are blocked
• Git push only via /ship — direct git push is blocked
• Auto-format — code is automatically formatted after each change (ruff, prettier, gofmt)

Details: Hooks

Project Memory

TAUSIK saves knowledge between sessions. The agent automatically:

• Records decisions — why bcrypt was chosen over argon2
• Documents dead ends — what was tried and why it didn't work
• Captures patterns — API error format, naming conventions
• Passes context — handoff for the next session

This knowledge is loaded at every /start and /task — the agent doesn't repeat mistakes from previous sessions.

Metrics

TAUSIK automatically tracks:

Metric	What It Shows
Throughput	Tasks per session
FPSR	Percentage of tasks solved on the first attempt
DER	Percentage of tasks where a defect was later found
Dead End Rate	Share of dead ends relative to total tasks
Lead Time	Average time from creation to task closure
Cost per Task	Average time by complexity (simple/medium/complex)

Metrics help understand: is the agent working efficiently, or spending time on retries?

Multi-Agent Work

TAUSIK supports multiple AI agents working on the same project simultaneously:

• Task claiming — task claim <slug> locks a task to a specific agent. Other agents see it as taken and pick a different one. task unclaim <slug> releases the lock.
• No conflicts — each agent works on its own claimed task. task next --agent <id> atomically claims and starts the best available task.
• Concurrent writes — the SQLite database runs in WAL (Write-Ahead Logging) mode, so multiple agents can read and write without blocking each other.
• Shared knowledge — all agents share the same project memory, decisions, and dead ends. What one agent learns, others see immediately.

No special setup is needed. Just run multiple agent sessions in the same project directory.

What's Next

• Hooks — automatic control: blocking, firewall, auto-format
• Skills — full list of what the agent can do
• CLI Commands — if you want to manage TAUSIK from the terminal
• Architecture — how the framework works internally