lsm / neokai / 27251953527
82%

DEFAULT BRANCH: dev
Ran
Jobs 0
Files 0
Run time –
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst
27251953527

push

github

web-flow 
fix: route Kimi through local bridge for 262k context window (#2088)

* fix: route Kimi through local bridge to report 262k context window

The SDK binary has a hardcoded context window table. When it sees
`kimi-for-coding` it falls back to ~200k, rejecting requests that
fit within Kimi's actual 262k window.

Route SDK traffic through a lightweight Anthropic Messages bridge
that intercepts GET /v1/models and returns the correct context
window metadata. All other requests are proxied verbatim to Kimi's
API — no protocol translation needed.

- Extend AnthropicMessagesBridgeConfig with optional `models` field
  for enriched /v1/models responses (context_window, max_tokens, etc.)
- KimiProvider now lazily starts a bridge on first buildSdkConfig() call
  and routes ANTHROPIC_BASE_URL through it
- Add bridge model metadata tests
- Update Kimi provider tests with fake bridge factory

* fix: address review feedback on Kimi bridge

- Per-credentials bridge: key by `{baseUrl}::{apiKey}` so sessions with
  different API keys or base URLs get isolated bridges instead of
  sharing a singleton that leaks credentials across sessions
- Follow Codex bridge auth pattern: set ANTHROPIC_API_KEY to sentinel
  value only, omit ANTHROPIC_AUTH_TOKEN to avoid ProviderService
  env-var leak across session switches
- Report max_tokens: 32768 in model metadata (Kimi's documented limit)

* fix: restore Codex 272k context window when SDK reports Anthropic alias

Root cause: #2086 aliased Codex models to Anthropic SDK IDs
(claude-opus-4-1-20250805 / claude-sonnet-4-20250514) so the SDK would
use a larger context window than its ~200k fallback for unknown Codex
IDs. The SDK then reported these Anthropic IDs in getContextUsage().
ContextFetcher's metadataMatchesResponse check only compared against
`id` and `alias` (gpt-5.5 / codex-latest), so it could no longer
match the SDK-reported name, and the UI fell back to SDK-reported
200k capacity instead of the 272k provider metadata.

Fix:... (continued)
Source Files on build 27251953527
Detailed source file information is not available for this build.