lsm / neokai / 25260681900

82%

push

github

Fix context window reporting for non-Codex models and increase chat limits (#1765)

* Fix context window reporting for non-Codex models and increase chat composer limits

Non-Codex models (e.g. DeepSeek V4 Pro via OpenRouter with 1M context) had
their context window incorrectly reported, causing premature auto-compaction
or "prompt too long" errors.

Bridge server: resolve context window from config models instead of Codex-only
lookup, so non-Codex models in the bridge models list get correct context
windows in SSE events and /v1/models responses.

OpenRouter: increase maxContextWindow fallback from 200K to 1M to match
models with large context windows.

Chat limits: increase InputTextarea default from 10K→100K chars,
space-task-message handlers from 10K→100K chars, useChatBase from 200K→500K.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* Address review feedback: revert Claude fallback context windows, add regression tests

- Revert openrouter-sonnet/opus/haiku FALLBACK_MODELS contextWindow from
  1M back to 200K — Claude models have ~200K actual context, and over-
  estimating would cause "prompt too long" errors when API is down.
  Only openrouter-auto and capabilities.maxContextWindow stay at 1M.
- Add test asserting FALLBACK_MODELS context windows per alias family.
- Add test asserting capabilities.maxContextWindow is 1M.
- Add test for resolveContextWindow fallback path: model NOT in config
  falls through to getCodexModelContextWindow() for known Codex models.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>

7802 of 10201 branches covered (76.48%)

Branch coverage included in aggregate %.

36 of 36 new or added lines in 3 files covered. (100.0%)

60879 of 84917 relevant lines covered (71.69%)

255.92 hits per line