flyingtimes/dsh-trajectory-reader
Track interpretation tab page: Interpret assistant behavior (requirements/ideas/execution/results) according to user turns, rule engine + optional LLM narrative, files/commands/errors are clear at a glance, user messages are retained as they are; dsh.bundle.patch can be installed with one click
Adds a 「轨迹解读」 (Trajectory Reader) tab to the DSH Web GUI conversation view ring beside 对话 / 轨迹. The session is segmented by user round, and each round card shows the distilled user need, the action summary of how the assistant fulfilled it (plan / research / implement / verify / delegate), errors and notes, and a reply digest — with the original user message kept verbatim and expandable and the per-tool-call ledger folded away. An optional ✨ button requests an LLM process narrative for a round (need / thinking / execution / result), generated server-side through the host llm service and cached by material hash.
Install
dsh plugin --profile web add @clarkchan/trajectory-readerREADME install Option 1 (npm, recommended). Option 2 installs straight from GitHub: dsh plugin --profile web add "github:flyingtimes/dsh-trajectory-reader#v0.2.3". The package declares dsh.bundle.patch, so dsh plugin add automatically appends it to dsh.profile.bundles — no manual cordis.patch.yml editing. After that, restart dsh web and the conversation tab bar shows 对话 / 轨迹 / 轨迹解读. npm package @clarkchan/trajectory-reader 0.2.5 (registry-verified 2026-09-10; registry repository field -> github.com/flyingtimes/dsh-trajectory-reader).
Compatibility
DSH Web client plugin mounted as a profile bundle (dsh.bundle.patch), loaded by the web profile Loader row with inject: ["llm", "webServer"]. The rules-based round summary is instant and dependency-free; the AI narrative requires the host llm service and is generated on demand with a 120s timeout per round, maxTokens 1200 and at most 12 rounds per request.
Details
- Repo: flyingtimes/dsh-trajectory-reader
- Category: Agent Capabilities
- Stars: 1
- Version: npm @clarkchan/trajectory-reader 0.2.5 (registry-verified 2026-09-10; README also documents the github #v0.2.3 channel)
- Last push: 2026-08-15
- First seen: 2026-08-15
Recent updates
README documents the summarizer's design constraints: the four fixed sections mirror need–thinking–execution–result, the prompt forbids inventing files, commands or causes absent from the material and forbids guessing truncated content, causal chain is emphasized over tool lists, and output is capped in length. Each round's material is JSON-framed with recursively length-capped strings so user text cannot break the structural delimiters.
FAQ
- How do I install dsh-trajectory-reader?
- Run: dsh plugin --profile web add @clarkchan/trajectory-reader (npm, recommended) or dsh plugin --profile web add "github:flyingtimes/dsh-trajectory-reader#v0.2.3", then restart dsh web — the 轨迹解读 tab appears in the conversation tab bar.
- Do I need to edit cordis.patch.yml?
- No. The package declares dsh.bundle.patch, so dsh plugin add automatically appends it to dsh.profile.bundles.
- Is the AI interpretation always available?
- The rules-based per-round summary is instant and dependency-free; the ✨ AI narrative needs the host llm service and shows or hides its button based on an availability probe (GET on the plugin's summarize route). It is generated on demand and cached, so unchanged material is not re-requested.