Skip to content

Releases: ac0033/dsh-ctm

Release list

v1.0.0 — Context Transparency Manager (first public release)

v1.0.0 — Context Transparency Manager (first public release) Latest
Latest

Choose a tag to compare

View all tags
@ac0033 ac0033 released this 16 Aug 15:26
v1.0.0
2b25a7e

dsh-ctm (Context Transparency Manager) turns the model's context into a visible, editable, effectiveness-scored surface — a self-contained bundle plugin for DeepSeek Harness. First public release.

Highlights

  • Full context visibility — a turn → segment flow view colored by role / tokens / cache status / effectiveness, with markdown rendering, paging, and live event-driven refresh (no polling).
  • Provider-measured token accounting — MECE KPI buckets (uncached input / cache hit / output + derived hit rate) read from the host's tokenUsage projection with a full-log fallback; per-turn and per-step real usage; per-segment estimates clearly labeled as estimates.
  • Logged, undoable edits (Apply for real) — replace / delete / rollback land in the session log as surface replace events at agent/pre-step, the same mechanism and timing as official compaction, honoring DSH's model-visible ⟺ logged invariant: replay, fork and token accounting always match what the model actually saw. Deletions shadow minimal tool-pair-balanced ranges, so the model never sees a dangling tool call. Everything is undoable.
  • Read-only by default — every edit is view-only until Apply for real is switched on; the initial system prompt is editable via the official system-prompt/assemble hook.
  • Zero runtime dependencies, zero coupling — host↔client over plain HTTP POST /ctm with a Zod-validated contract; bilingual UI (中文 / English).

Install

dsh plugin --profile <name> add github:ac0033/dsh-ctm#v1.0.0

Git installs build from source via the package's prepare script — pnpm will ask you once to add dsh-ctm to allowBuilds in the profile's pnpm-workspace.yaml, then re-run the same command.

Known limitations

  • Undoing a rollback (or a delete involving tool pairs) restores the removed content as user messages — an append-only log cannot re-add assistant/tool roles (role demotion; content is fully preserved).
  • The host's conversation tab is a full transcript view by design: shadowed content stays visible there; CTM's own tab shows the model's actual view.

中文摘要

CTM 把模型上下文变成可见、可编辑、可评分的一等对象：provider 实测的 MECE token 口径（未命中输入 / 缓存命中 / 输出 + 命中率）；替换 / 删除 / 回滚以官方 compaction 同款机制写入会话日志，模型所见与日志严格一致，全部可撤销；默认只读，中英双语界面，零运行时依赖。已知限制：撤销回退时内容以 user 角色恢复（append-only 日志的角色降格），详见 README。

Assets 2
Loading