TAUSIK

English

Русский

English

Русский

Appearance

English | Русский

TAUSIK 1.11: less context, fewer rounds, the same evidence ​

TAUSIK 1.11 targets the cost of accepted work in Codex. It reduces repeated framework context and avoids deterministic model returns without weakening QG-0, QG-2, verification evidence or fail-open behaviour.

Measured changes ​

Surface1.11 result
Repeated MCP prefix, same 147 tools63,333 → 42,519 bytes (−32.9%)
Two reviewed low-value test groups87 → 45 nodes (−48.3%)
Codex cache share in the observed baseline97.8%
Default quality lane, public snapshot12,559 passed · 100 skipped · 143 deselected
Slow quality lane, public snapshot129 passed · 14 skipped · 12,659 deselected

The high cache share means the next useful levers are fewer model rounds and economical routing, not more blind context deletion.

What changed ​

  • /start is one bounded call. task start can return its package, and task show --package is the normal bounded retrieval path.
  • Ordinary verify selects affected tests. Ambiguous, unmapped or sensitive changes fail open to the complete applicable lane.
  • Verify output is compact, but full logs and machine counts stay under .tausik/verification/.
  • Codex and Kilo usage are read from native local records without storing conversation content or double-counting cache/reasoning subsets.
  • Routing can choose an economical bounded worker and escalate for named risk; it never claims to switch the active coordinator.
  • User-facing answers default to controlled prose and use tables or diagrams only when they materially reduce reading effort.

Honest limits ​

This release does not claim that every project will use 30% fewer weekly Codex credits. Natural accepted-task rounds have not yet met the release target, Luna is not yet qualified as the routine default, and one verification replay still loses preflight evidence. Those measurements continue on the 1.11.x line.

Attribution ​

The 1.11 work was developed by Andrey Yumashev with OpenAI Codex. Release commits carry Co-authored-by: Codex <codex@openai.com> so GitHub attributes Codex as a contributor.

See the full changelog and the Russian release notes.

GitHub Release links: