English · 한국어 · 简体中文 · 繁體中文 · 日本語 · Español
ScopeWeaver is a standalone derivative of Autumn-27/ARTEX, based on commit 160fe13. It is maintained separately at cskwork/scopeweaver, without a GitHub fork-network relationship.
The ScopeWeaver modifications dated 2026-10-02 cover the product name, English-default/Korean-selectable interface and messages, documentation and bundled agent guidance, setup/build/start/update/password-reset prompts, report and agent language handling, and distribution naming. The executable is scopeweaver (scopeweaver.exe on Windows); archive names follow scopeweaver-<version>-<os>-<arch>.zip. Docker Compose builds the checked-out source locally. The configured release workflow targets this repository and GHCR.
The original LICENSE is retained unchanged. Original authorship, contributor acknowledgments and upstream changelog history remain attributed to ARTEX. The Go module github.com/Autumn-27/artex, source entry point ./cmd/artex, ARTEX_* configuration contracts, artex database defaults and ARTEX login username remain for compatibility.
This modification date is not a release tag. It does not assert that binary releases or container images have been published. Historical validation documents under sidequestion/ describe upstream runs; they are not new ScopeWeaver verification receipts. Translated historical JSON preserves technical identifiers and measurements while labeling translated narrative.
Modifications on 2026-10-15 add an optional native Ollama profile format (/api/chat with num_ctx, keep_alive and num_predict), three new-form Ollama templates, keyless local profiles, per-profile concurrency, Ollama-aware pre-warm, compaction and approval-judge defaults, and an opt-in local_llm_only LLM network policy. SCOPEWEAVER_LLM_* and SCOPEWEAVER_LOCAL_LLM_* variables are added as aliases that take precedence over the retained ARTEX_* contracts. Existing Anthropic, OpenAI, OpenAI Responses, custom URL, proxy and GLM template behavior is preserved, and the policy is off by default.
The follow-up native hardening candidate on 2026-10-15 adds urgent next-slot admission without preemption, actual native-origin judge selection with schema-constrained output and 90-second/ask defaults, and 72% compaction with 16/32 KiB tool-output caps through 64K. It narrows opt-in strict destinations to loopback plus explicit inference CIDRs (including LAN/Docker migration), refuses redirects/proxies, and blocks all agent search and update network actions while retaining local status/rollback. Read-only GPU-memory estimates and shared-model warnings, safe judge-settings loading, and English/Korean/Simplified Chinese interface text accompany the change. These are application controls, not whole-host offline or model-quality guarantees; no dependency, migration, automatic model pull or release is introduced. See the provider guide and candidate verification.
Additional modifications on 2026-10-03 complete the server translation audit and add GLM-5.3 configuration templates plus its required always-enabled reasoning compatibility. Templates use the existing provider adapter and do not create or activate a profile automatically.
The subsequent upstream fixes from ARTEX commits e2c8a23, 86729b6, and
a951e4a are backported here: same-origin production SSE, separate approval-model
token accounting, and fail-closed authentication with atomic initial password
creation and bounded database pooling. New messages and deployment guidance are
available in English and Korean. The original authorship is credited in the
backport commits; the other ScopeWeaver changes remain in place.