Fix LMTerminal streaming and preserve config reads

[sderev]

What changed

• Use model-aware streamed reasoning_effort for the gpt-5* family: keep minimal for gpt-5/gpt-5-mini/gpt-5-nano, use low for gpt-5.1/gpt-5.2, and omit the argument for gpt-5-chat-latest.
• Detect when streamed output cannot use a refreshable rich.Live surface and fall back to plain flushed chunk writes.
• Force per-chunk Rich redraws with auto_refresh=False and live.update(..., refresh=True) in refreshable terminals.
• Stop rewriting ~/.config/lmt/config.json during theme reads; malformed or missing config now falls back to defaults without modifying disk state.
• Add regression coverage for GPT-5-family streamed request args, raw and Rich streaming presentation, dumb-terminal fallback, and non-destructive config reads.
• Add changelog fragments for the streaming and config-preservation fixes.

Why

• The first hotfix pass still failed for real GPT-5-family usage: gpt-5.1 and gpt-5.2 reject streamed reasoning_effort="minimal", while gpt-5-chat-latest rejects reasoning_effort entirely.
• Visible streaming responsiveness still depends on terminal capabilities, so non-refreshable terminals need a plain flushed fallback.
• LMTerminal also had a config-read bug where theme reads could rewrite ~/.config/lmt/config.json when the file was malformed or missing.

How to test

• Automated:
  • uv run pytest -q tests/lib_test.py tests/gpt_integration_test.py
  • gate
• Manual:
  • uv run lmt prompt "count from 1 to 10" and confirm visible output starts before full completion.
  • uv run lmt prompt --raw "count from 1 to 10" and confirm raw chunks appear progressively.
  • TERM=dumb uv run lmt prompt "count from 1 to 10" and confirm output still streams progressively.
  • If your account has access, run streaming prompts with --model gpt-5.1 and --model gpt-5-chat-latest and confirm no reasoning_effort request errors.
  • Put malformed JSON in ~/.config/lmt/config.json, run uv run lmt prompt --no-stream "hello", and confirm LMTerminal falls back to defaults without rewriting the file.

Risk/comp notes

• Streaming behavior changed in lmterminal/gpt_integration.py and lmterminal/lib.py.
• GPT-5-family streamed requests now branch on model prefix to avoid unsupported reasoning_effort values.
• Rich output in non-refreshable environments degrades to plain streamed text instead of trying to live-refresh.
• Config reads are fail-soft and read-only; only explicit saves write config.json.

Changelog fragment

• yes
• These are user-visible streaming and config-preservation fixes.

[sderev]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Can't wait for the next one!

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".