[JasonW404]

Local context estimates could exhaust maintenance or recall allowance even when the provider could still serve the request. This change makes counts advisory, calibrates subsequent whole-request estimates from completed provider input usage, and preserves original/protected context when maintenance fails. Explicit provider input overflow can force bounded source-backed history recovery even when the local estimate fits; output truncation remains distinct.

Shared CodeAgent/model/context changes only. Native runtime integration follows in #4315 after this verified foundation. Saved model output budgets are unchanged.

Validation on 0241c3c (product revision b697eac; final test-only fixture correction):

• Unit suites: 195 shared context/model/runtime, 182 CoreAgent, 141 legacy OpenAI and 96 monitoring tests passed in isolated processes.
• Maintained transport D3: 65 passed, preparation and cleanup passed; strict formal asset validation passed. Real HTTP assertions cover missing usage, invalid summary preservation and bounded overflow recovery.
• Built main/web images deployed on the user-authorized host with existing data. Real CodeAgent accepted estimated excess, compressed history, executed Python, summarized/offloaded and reloaded an original result.
• Nine further real turns accepted eight checkpoints and retained nine exact facts. Browser reopened persisted results. Phoenix default traces: 25da97c495717406150dc56b11d5e3cd, 46dfd8b8f186def4b01b47b61bb6e8cb, 3298259fdae7c37bc3db595bd3557cc0.

Controlled overflow tests use an actual HTTP peer; hosted-model acceptance uses qwen3.8-flash. Full Nexent feature acceptance belongs to the stacked native branch and is still pending. Existing auxiliary summary billing-metric gaps and missing optional history HTTP500 are outside this change.

CI follow-up: the initial runner lost communication. Its retry exposed an eager model-package import in low-dependency runtime tests and mock serialization recursion. Moved the usage helper into neutral context_runtime and restricted structured serialization to Pydantic models. Reproduced then passed the affected 4 bootstrap and 59 history/observation tests; factory tests also pass. Rebuilt/redeployed corrected main image 5925767071ec and repeated actual Python/Agent execution, inspected in Phoenix trace 72caeea199cb3ff2ed2fe6b27a536aa4. Transport D3 65 cases and strict assets revalidated.

Final CI on 0241c3c: all configured checks pass, including unit tests, both image architectures, copyright and CodeQL. Codecov patch coverage 92.24806% (90% target), total 92.1%. The final correction updates test namespaces/imports only; product source matches the rebuilt and verified b697eac image.

[codecov]

Codecov Report

❌ Patch coverage is 92.24806% with 10 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
sdk/nexent/core/context_runtime/usage.py	89.13%	3 Missing and 2 partials ⚠️
sdk/nexent/core/agents/context/manager.py	93.33%	3 Missing ⚠️
sdk/nexent/core/agents/context/llm_summary.py	87.50%	0 Missing and 1 partial ⚠️
sdk/nexent/monitor/monitoring.py	66.66%	0 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!