sogeisetsu · Daily GitHub activitysogeisetsu 的 GitHub 每日动态
@sogeisetsu·2026-10-04 00:00 – 24:00 · UTC+8
sogeisetsu logged 41 commits on 2026-10-04 across three repositories, with no PRs, issues, or reviews.
In sogeisetsu/sogeisetsu, 24 own commits plus 20 automated commits by github-actions[bot] focused on the daily-report pipeline: fixing the test-render.mjs assertion collector so throwing cases no longer fail, reworking dark-mode readability and the headline contract, rendering inline Markdown in AI summaries and Release notes, adding mock regression tests, and switching to deterministic truncation that lifted the daily success rate from 0% to 100%.
An ai fix also disabled reasoning so max_tokens no longer consumed the CoT and left content empty.
In sogeisetsu/meld-deepresearch, 15 own commits shipped the v0.2.0 release: explicit handling for access-blocked sources, validator hints that explain how to fix a failure, same-publisher origin warnings, mid-run budget uplift, genre templates in SKILL.md, chapter background alternation for the Chinese main, and footer/hero copy updates.
In sogeisetsu/opencode-go-model-picker, 2 own commits added scripts/discover-agents.mjs and scripts/snapshot-summary.mjs and documented vision verification, the Windows refresh path, and a role-overview column.
Two releases were published: sogeisetsu/meld-deepresearch v0.2.0 and sogeisetsu/opencode-go-model-picker v1.3.0.
sogeisetsu 在 2026-10-04 于三个仓库共提交 41 次，无 PR、issue 或 review。
在 sogeisetsu/sogeisetsu 中，24 次本人提交，另有 20 次由 github-actions[bot] 自动提交，围绕日报流水线展开：修复 test-render.mjs 断言收集器使抛异常用例不再失败，整改深色模式可读性与 headline 契约，为 AI 摘要与 Release notes 渲染内联 Markdown，补充 mock 回归测试，并改用确定性截断使日报成功率从 0% 提升至 100%。
另有一次 ai 修复关闭 reasoning，避免 max_tokens 计入 CoT 导致 content 恒为空。
在 sogeisetsu/meld-deepresearch 中，15 次本人提交发布 v0.2.0：明确处理无法访问的来源、validator 给出修复提示、同出版方来源告警、运行中预算上调、SKILL.md 中列出 genre 模板、中文主页章节背景交替、页脚与 hero 文案更新。
在 sogeisetsu/opencode-go-model-picker 中，2 次本人提交新增 scripts/discover-agents.mjs 与 scripts/snapshot-summary.mjs，并记录视觉验证、Windows 刷新路径与 role-overview 列。
当日共发布两个版本：sogeisetsu/meld-deepresearch v0.2.0 与 sogeisetsu/opencode-go-model-picker v1.3.0。
20 automated commits by @github-actions[bot] · not counted in the total20 次自动提交，由 @github-actions[bot] 发起 · 不计入总提交数
This release hardens the research protocol and ships the evaluation work that
came out of comparing this skill against other deep-research skills.
The strongest primary sources are often behind a paywall or a bot wall. Until
now, the right response — record the failure, note the gap, and never pass a
snippet off as the original — was implied but not written down. It now is: a
blocked page is logged as an observation with its HTTP status, the gap goes into
the report as an availability caveat, and a claim whose only support is an
unreachable source is labelled unknown instead of quietly softened.
A failed run used to get a code and a location. It now also gets one
"smallest safe fix" hint per actionable error — whether to re-key a claim, add a
second independent origin, or drop an unreachable reference — so a single
allowed correction is spent on the right change.
If an interpretive claim rests on several URLs that all belong to one publisher
(a wire story re-published across one outlet), the validator now warns withW_SAME_PUBLISHER. It is a hint, not a failure: the check is a heuristic and
never blocks delivery.
A genuinely broad topic can now raise its ow…
AI Hardens the research protocol and ships evaluation work from comparing this skill against other deep-research skills. Access-blocked sources are now logged as observations with HTTP status, gaps become availability caveats, and claims supported only by unreachable sources are labelled unknown. The validator now also returns hints on how to fix a failure, not just a code and location.加固研究协议，并发布与其他 deep-research skill 对比得出的评估工作。无法访问的来源现按规则处理：记录为带 HTTP 状态的观察，缺口写入可用性说明，仅由不可达来源支撑的结论标记为 unknown。validator 除返回错误码与位置外，还给出修复提示。
Two versions of work are rolled up here (v1.2.0 and v1.3.0).
scripts/discover-agents.mjs: a read-only, zero-dependency CLI that lists yourcustom agents (slim and native sources) and flags risky fallback chains — duplicate
model ids, chains outside 2–4 entries, non-Go fallback ids not confirmed in the
local registry, and vision lanes whose model is not marked vision-verified. It
prints compact JSON instead of making you read config files by hand.
scripts/snapshot-summary.mjs: a compact digest of the cached model snapshot(~97% shorter than the raw file), with --ids, --trait, and --json.
plain-language description of what each agent does, so a suggestion can be judged
without knowing the role by heart.
price, throughput and the trait-relevant capability (and strictly better on at
least one), the dominated model is replaced — or must be justified in the "why"
column (for example, a cross-provider fallback).
DeepSeek V4.1 Flash accept images. DeepSeek's Go id, lab version name and API
request name are n…
AI Rolls up v1.2.0 and v1.3.0 with better agent discovery and chain safety. Adds scripts/discover-agents.mjs (read-only, zero-dependency CLI that lists custom agents and flags risky fallback chains) and scripts/snapshot-summary.mjs (a compact digest of the cached model snapshot, ~97% shorter than the raw file). The recommendation table gains a role-overview column.合并 v1.2.0 与 v1.3.0，改进 agent 发现与链路安全。新增 scripts/discover-agents.mjs（只读、零依赖 CLI，列出自定义 agent 并标记风险回退链）与 scripts/snapshot-summary.mjs（缓存模型快照的紧凑摘要，比原始文件短约 97%）。推荐表新增 role-overview 列。