swordluan8-hash/token-aware-orchestrator
Token-Aware-Orchestrator test: narrow Codex context before execution
In an isolated Python repository, the task, budget, and success criteria were fixed first; Codex completed the change in one attempt and passed the tests.
01 · PROBLEM
What I wanted to solve
Codex may read too many unrelated files and historical details before acting. A control layer should define the task, model, budget, success criteria, and permitted file scope before execution begins.
02 · PROCESS
What I did
We created a calculator repair task in an isolated Python repository, declared the task type, target model, budget, success criteria, and expected files, then let Token-Aware-Orchestrator hand the job to Codex. It ran unit tests and produced a JSON handoff record.
03 · RESULT
What actually happened
The execution succeeded in one attempt with exit code 0 and no local retry or escalation. One unit test passed. The task changed src/calculator.py and tests/test_task.py and recorded execution, testing, and diff results.
04 · PITFALLS
What others should watch for
The project is still v0.4.0-alpha. A budget is an execution guardrail, not proof that every model bill falls by the same percentage. This small test proves the workflow runs; it does not prove universal 50% savings.
05 · VERDICT
My conclusion
As a local control layer for Codex, it can already fix task boundaries and preserve an auditable result. The next step is a same-task comparison on larger real repositories using actual tokens, duration, and correctness.