Files
dsh-context/AGENTS.md
T

12 KiB
Raw Blame History

dsh-context

A DeepSeek Harness plugin for context insight, actions, and management.

Background

  • DeepSeek Harness (dsh):

    • an open-source agent harness developed by DeepSeek AI.
    • Github: https://github.com/deepseek-ai/deepseek-harness
    • NPM: @deepseek-ai/dsh
    • MUST ensure the full clone of the DeepSeek Harness repository is available locally, before any work.
    • MUST always dive deep into the details of dsh source code and dependencies, for its mechanisms, lifecycles, and modules. Ensure every decision is based on the full and actual truth of the dsh source code.
    • Local git clone of dsh:
      • may be found in the ~/dev/deepseek-harness directory
      • git pull on the main branch to update
      • commits and version tags are available for reference and diff
      • run pnpm install to update dependencies after a git pull or switching commit/tag
  • DeepSeek Harness Plugin:

Coding

  • Always consider the minimal change and the most performance efficient implementation.
  • Try best to use the existing classes, utilities, styles, style tokens, events, presets and lifecycles provided by DeepSeek Harness.ess.
  • Use English in code comments, documentation, Pull Request description, and commit messages.
  • Smaller, less-coupling and modulized code and tests are preferred for better maintainability and testability.
  • Avoid adding unnecessary code comments (unless for the pinned major decision or for those provide significant value) and code duplication.
  • Update or remove the outdated or unhelpful code comments when modifying the code.
  • Before any commit, MUST ALWAYS do ALL the following checks:
    • Check the to-do list, and ensure all the items are properly completed or closed.
    • Carefully independently review and simplify all the diffs and all code changes, to ensure they are necessary, correct and not over-engineered.
    • Cleanup the generated temporary files. Cleanup temporary or unhelpful comments.
    • MUST Run pnpm run lint:fix && pnpm run test && pnpm run build in single command and capture FULL output, to ensure:
      • passing all the linting and test
      • the per-file code coverage MUST BE literally 100%.
        • Example output:
          • -------------------------|---------|----------|---------|---------|-------------------
            File % Stmts % Branch % Funcs % Lines Uncovered Line #s
            All files 100 100 100 100

Parsing resilience (log data must never crash or hang a view)

The plugin lives off data it does not own: the durable session log (event shapes vary across dsh versions, producers, and hand-edited replays), the conversation snapshot behind the client join, projection payloads on the wire, history RPC pages, and persisted stores. Treat all of it as untrusted input at every layer. The two failures to design out: any client- or host-side parsing that blanks the page (the error card), and anything that leaves the page stuck on "loading".

  • Never let one bad record take down a view. A malformed node, event, tool entry, or file op degrades to zero rows for that item — the card, the tab, and the session keep working.
  • Host-side projection folds must be TOTAL. The harness projection registry drives apply straight off the session/event bus with no error boundary of its own: one throwing fold stalls that unit's cells and its session/projection push feed, and the browser then waits on "loading" forever. So: unknown event types return the state unchanged; per-event processing is isolated so a malformed event is dropped whole (all-or-nothing — no partial state); and never materialize an undefined-valued property into persisted state, because the plain-JSON precondition makes one such property fail EVERY projection-cache write for the session (sessions then break in unrelated, far-away places).
  • Client-side parsing degrades visibly. Sanitize delivered projection payloads at the boundary (the timelineOf pattern: collections re-proved, scalars zeroed, whole-value absence stays null → loading screen); isolate per-item work in any fold over join/log data (per-item guards, or a bounded catch when a hostile object may throw on property access); every async fetch must resolve to data or a visible retryable state, never an unhandled rejection that leaves a spinner.
  • Re-prove every field at runtime. Structural narrowing over blind casts; optional chaining over non-null assertions; skip elements that fail the shape instead of throwing.
  • Every parser carries hostile fixtures next to its happy path: wrong types, null/missing fields, null or primitive elements inside arrays, unpaired references, and objects that throw on property access. The 100% coverage bar applies to every guard branch — an untested guard is an unverified promise.

Layout & responsive

  • The plugin pane is the shell's center column, not the viewport — its width follows the sidebar's collapse and drags. Drive responsive layout with container queries (the harness's own idiom), not viewport media queries.
  • Fold, never crush: under width pressure, wrap rows or fold grid columns instead of truncating into unreadability. No horizontal scrollbars, no overlapping content, and every field stays reachable on the narrowest pane.
  • Mind container-query side effects: size containment changes containing-block behavior, so fullscreen overlays portal to document.body. Point-of-use specifics live in the CSS comments next to the rules they guard.
  • Verify layout changes visually before shipping: a spread of widths from desktop down to phone (~390px), in both locales (English labels are the truncation stress case), on a long session, including the /context modal.

Building

  • Run pnpm run build after code changes applied.
  • Run pnpm run watch to keep hot-reloaded on dsh with local plugin installed. It also helps developer to see the code changes in the browser.

DSH web server

  • The dsh web server may be already running, and accessible at http://127.0.0.1:3080/ by browser.
  • Run pnpm run web to restart the dsh web server (kills the running dsh web first, then starts dsh web --no-open).
  • If the auth token issue encounters with browser, kill the dsh process and restart it by running dsh web --no-open.

Dependency

  • Consider updating the dependencies to the latest version if possible, as the deepseek-harness is evolving rapidly.

Compatibility - Important!

  • MUST be able to install and work correctly on @deepseek-ai/dsh 0.1.2-rc.1+, with no regressions in runtime dependencies, message parsing, or any user-visible behavior. docs/compatibility.md is the single source of truth for compatibility — the supported releases, the per-generation event vocabulary, the client seams each line serves, and the runtime baseline gate — so keep that page current and reference it instead of restating its rules here.
  • The supported range spans three durable-log generations that must all keep working from ONE code path. Fold by SHAPE, never by detected version: a log carries exactly one generation and the spellings are mutually exclusive, and src/host/logShapes.ts is the ONE place those spellings meet (firstTokenTimeOfStream, replaceRangeOf, isTokenChunk) — applyTimeline keeps a single code path over all of them. Never branch fold behavior on detectHarnessVersion.
  • Capabilities that only NEWER lines serve are OPTIONAL seams: reach them through a deferred ctx.inject([...], …) — never a hard module inject — so an older harness leaves the plugin fully functional with the capability absent (no pending fiber, no throw). Re-prove the service shape before use, guard the registration, and declare the module in dsh.client.inject.
  • Base every compatibility decision on the ACTUAL dsh source, never on assumption: keep the full clone current (git pull on main; pnpm install after switching commit/tag) and dive into the mechanisms, lifecycles, and modules involved.
  • Low-level logic (token counting, first-token timing, …) MUST mirror the newest supported dsh implementation: the estimators here match dsh-token-meter's estimateSystemMessage (estimateSystemTokens / estimateSystemContent) and the first-token rule matches dsh-llm's assistantStreamFirstTokenTime. The fold's statistics are verified against the harness's OWN projection values on real logs.

Compatibility verification mechanism

  • tests/baselines.ts is the ONE source of truth for the supported baselines and their per-version seam faces. To add a future dsh version: add one Baseline entry there (tag, vendored cordis, faces), add its tag to the release workflow fetch step, then run pnpm test — re-fit or refactor the seam a failing probe names.
  • The per-baseline suites run as part of pnpm test; docs/compatibility.md documents what each check covers. The compat project additionally boots the BUILT plugin (pnpm run build first) against the ACTUAL harness sources at each baseline tag, and needs the dsh checkout (env DSH_REPO, default ~/dev/deepseek-harness).

I18n

  • Chinese (Simplified) and English are supported for UI elements.
  • Update all the supported languages translations when adding or modifying the UI elements.
  • Do not keep the deprecated or unused language keys.

Docs

  • docs directory contains only end-user faced documents.
  • docs/social-preview.png (GitHub social preview) must be exactly 1280 × 640 pixels.
  • README.md
    • Images:
      • Only embed external links in the README.md, in order to help the readers on both GitHub and NPM to access the images
        • For example, putting the image in the docs directory and embedding it in the README.md with links:
          • some image

Temp files

  • Generate one-time temp files in the .tmp directory, and properly clean them up right after use.

Git

  • When asked to commit, please commit the possibly mixed changes separately for each task or purpose.
  • gh cli is installed and logged in.
  • Run git push after create commits.

Workflow

  • To-do list
    • ALWAYS keep the coding agent's to-do list up to date throughout starting or finishing every step/task of planning, investigation and implementation.
    • Before closing any task, review all pending to-do items and ensure each is completed, cleaned up or explicitly closed.

Tool Usage

  • Always read the file first using the read tool before using the edit tool, which prevents errors like "Error: edit requires reading '/path/file' first — read the file, then retry."

Releasing

  • Version X.Y.Z, 大版本.次版本.小版本。
  • Releases are cut by tagging: git tag vX.Y.Z && gh release create vX.Y.Z.
  • A GitHub Actions workflow then builds, tests, and publishes the package to npm automatically by github workflow. Agent don't have to do or check it manually.
  • Write the release notes from the release template

CodeGraph

This repo is indexed by CodeGraph (a .codegraph/ directory exists at the repo root, if not run codegraph init to initialize it).

  • MUST reach for it BEFORE any grep/find or reading files when you need to understand or locate code:
  • Shell (always works): ALWAYS run and collect ALL output of codegraph sync -q && codegraph explore --path /some-path "<symbol names or question>" without truncating text