cn-model-devtools

v0.2.0: detected model id now sent in probe payloads; malformed endpoint responses classify cleanly (no raw tracebacks); clean exit-2 usage errors for bad --out; new no_tool probe for over-eager tool-call detection (suite cn-tc-v2).

Normalize GLM-style tool-call responses for OpenAI-compatible Python clients, including sync, async and streaming adapters.

本地 CN 模型 tool-call JSON 运行期修复代理，让 Claude Code 接 Qwen3/DeepSeek 不再静默崩溃

国产大模型工具调用契约差分测试器

v0.2.0: non-UTF-8 or unreadable input now exits cleanly with code 2 instead of a raw traceback; new --output-ratio flag projects output-token cost under an explicit output-length assumption; version surfaces bumped in lockstep with a new CHANGELOG and consistency tests.

Per-task reasoning-budget governor for Qwen3.8 coding agents: profile think traces, learn class budgets, and advise via an OpenAI-compatible streaming proxy.

ctxfeed is a local MCP project-context backend that shards a whole repo into GLM-5.2's 1M-token window with cache-aware ingest ordering

Capability-gap diagnostician for locally-served CN open-weights models (DeepSeek/Qwen3/Kimi K3/GLM) — attributes your local model's felt 'dumbness' to a specific misconfiguration layer so you fix one layer, not the whole stack.

v0.2.0: fix — init now captures wired-config keys written as '- ' list bullets (markdown/YAML lists previously parsed to a mangled key and were dropped); feature — check surfaces an amber pricing.unit delta when a successor switches quote currency instead of going silent; version bumped lockstep across all surfaces + CHANGELOG.

Run-level exposure gauge for coding agents on local models — tracks wall-clock, state mutations, and token burn since last human review, forces a pause at thresholds.