Repository navigation
v1.1.0
Base Context 1.1.0
This release includes the six-task repair improvements and Codex-aligned model defaults.
IPython and native runtime
- Long-running IPython cells return pending without cancelling the cell. Collect the final result with
action: "status"before submitting another cell. SDK callers must handle this changed completion contract. - Status and targeted interruption remain available while Python is busy.
- Native completion and EOF wait for delegated work and physical execution to settle.
- Tool error results retain their error status. Task guidance favors bounded computation and finishing unambiguous work.
Model defaults
| Model | Context window | Usable input window | Auto-compaction trigger |
|---|---|---|---|
deepseek/deepseek-flash (low/high/max) |
1,048,576 | 996,147 | 943,718 |
openai-codex/gpt-6.1-sol (medium/high) |
272,000 | 258,400 | 244,800 |
openai-codex/gpt-6-astra (medium) |
272,000 | 258,400 | 244,800 |
These defaults follow the evidenced Codex 0.160.0 configuration. Compaction triggers at or above 90% of the nominal window. Finished-turn threshold compaction is off by default; pre-request/continuation checks and overflow recovery remain. Normal and summary requests omit implicit client output caps for these routes. Backend output limits can still apply.
Explicit user overrides remain supported. Input accounting does not enable hard request admission. Base retains its own estimator, working-set algorithm, and 20K recent-tail policy; retained content is not claimed to be identical to Codex.
Native benchmark tooling
The repository now includes a reusable installed-CLI single-attempt runner with native completion barriers, shared first-prompt deadlines, candidate-service lifecycle management, a Python alias, and installed-package usage collection. The earlier mixed-build benchmark score does not measure this release's new settings.
Release qualification
The final 1.1.0 artifacts passed installed RPC readiness and one fresh DeepSeek-high/task06 attempt: all three stages completed, 5/5 main checks and the edge check passed, in 1,371.23 seconds under the original 2,400-second deadline. No provider or measurement errors were observed. The same artifacts are published without rebuilding. This single-task smoke does not replace or regrade the earlier mixed-build benchmark results.