fix: default Codex approval policy to on-request

[artemgetmann]

Summary

• change default CODEX_APPROVAL_POLICY fallback in runtime config from never to on-request
• update .env.example and README examples to match

Why

Threads were defaulting to approval=never, which blocks escalation flow for privileged commands.

Validation

• bun run typecheck
• env -u CODEX_APPROVAL_POLICY bun -e "import { CODEX_APPROVAL_POLICY } from './src/config.ts'; console.log(CODEX_APPROVAL_POLICY);" -> on-request

[linuz90]

A couple of things highlighted by Codex:

1. The Claude reasoning behavior changed from keyword-triggered thinking to effectively always-on by default. That's a real latency/cost change, while the README still describes the old behavior. What's your reasoning here?
2. The setup docs currently tell users to clone your fork instead of the upstream repo.

[artemgetmann]

Thanks — both points are fair.\n\n1) Reasoning default:\nMy original intent was quality-first because output quality drops noticeably without reasoning in some workflows. But I agree default-on changes latency/cost expectations.\nI’ve now restored safe default behavior so reasoning is keyword-triggered by default, while medium/high remain explicit opt-in via config ().\n\n2) Setup docs:\nAgreed. I updated setup instructions to use the upstream repo (not my fork).\n\nFollow-up commit pushed to #6: b5a4863

[artemgetmann]

Quick clarification (my previous comment had CLI formatting artifacts):

• Reasoning is now keyword-triggered by default again.
• Baseline always-on reasoning is opt-in via CLAUDE_REASONING_EFFORT=medium|high.
• Setup docs now point to the upstream repo.

If this looks good now, happy for re-review on #6.