dsh-shejing
Shejing (摄鲸) — DSH 的照片工作流插件。通过自身的引擎和目录驱动 Lightroom Classic，支持检查、筛选、整理、分级、验证、归档和回顾。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:yxj1915/dsh-shejing#15ffcfe5d37ea67cd52b9d03fec1820c404b85ceShejing 以 dsh-shejing 发布，当前版本为 0.1.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
A photo workflow plugin for DeepSeek Harness that drives Adobe Lightroom Classic through its own engine and catalog — checkup, cull, organize, grade, verify, archive, retrospective.
The whole design has one goal: make every mistake become a rule for next time, instead of becoming rework.
中文文档见 README.zh-CN.md。
This workflow grew out of real failures, not a whiteboard. Two of them shaped everything:
(0, y) with y ≠ 0, which made Lightroom's spline diverge across the red/blue channels.
Diagnosing and rolling it back took a dozen rounds.So the hardest rule in this plugin — any new parameter combination must be rendered on a single photo and seen by you before it touches a batch — is not a prompt instruction. It is enforced by a hook that DSH routes to you natively. You can approve it; the model cannot fake your approval.
| Stage | Tool | Gate |
|---|---|---|
| Checkup | shejing_checkup | — (read-only) |
| Cull | shejing_cull | marking/moving gate |
| Organize | shejing_organize | irreversible gate |
| Grade | shejing_grade | new-parameter gate |
| Verify | shejing_verify | — (read-only) |
| Archive | shejing_archive | export confirmation |
| Retrospective | shejing_retro | write confirmation |
Plus shejing_doctor (self-check) and shejing_batch_status (read the ledger).
Checkup groups bursts, measures per-group sharpness and exposure spread, classifies each group (burst / exposure bracket / suspected focus stack / panorama), reports highlight and shadow clipping, and renders a contact sheet. It never touches Lightroom — it runs before import.
Cull builds 可导入/ ("importable") and 非导入/ ("not imported") inside the source folder and
moves files between them — same-volume rename, instantly reversible. Originals are never deleted.
Grade applies one of three built-in looks (A warm cinematic / B airy daylight / C deep dusk) or explicit parameters, creating a snapshot per photo and rendering a preview so you can see it.
dsh webThrough the DSH desktop client's plug-in manager, or:
dsh plugin --profile <profile> add dsh-shejing
The plugin registers itself as a profile layer and ships its own skill, so there is nothing to
symlink and nothing to copy into ~/.dsh/skills.
Restart DSH after installing. Client-module metadata is cached until restart, so a page refresh alone will not pick up the browser half.
First run: ask the model to call shejing_doctor. It reports whether the Lightroom plug-in is in
sync, whether the auth token exists, and whether the bridge port is listening.
非导入/ is kept forever; deleting originals is your call, never the
tool's..lrcat. The Lightroom catalog database is only ever read, and only from a
copy. All mutation goes through Lightroom's own API.ProcessVersion (that downgrades photos).The bundled Lightroom bridge is a fork of @pired/lightroom-mcp
(MIT, upstream Automaat/lightroom-mcp), with two
fixes: localized color-label names, and create_collection actually honouring its parent
parameter. See THIRD-PARTY-NOTICES.md.
node scripts/build-client.mjs # build the browser half (no toolchain needed — it is a wrapper)
node scripts/test.mjs # all six suites (needs DSH_HOME for two of them)
node scripts/regression.mjs <cloned-batch-dir> # end-to-end on a real batch (~7 min for 75 frames)
The suites are:
| Suite | What it covers |
|---|---|
smoke.mjs | tool registration surface, every gate path, skill parsing, panel routes |
smoke.mjs --no-connection | graceful activation when a service is missing |
check-client.mjs | actually renders every panel tab with a fake React (catches things node --check cannot) |
test-bridge.mjs | the LR channel: MCP handshake, image blocks, success:false, reconnect, clear errors — against a fake bridge |
test-tools-lr.mjs | the stage tools end to end against a fake bridge: what actually reaches the wire when grading, the curve-endpoint guard, the gate's ask → approve → whitelist cycle, the real export path (plan → gate → export_photos → count verification → SUMMARY.md), and that protocol-level isError becomes a failed call while a payload success:false is passed through for the model to judge |
test_grouping.py | the burst grouping algorithm must produce a partition, plus 200 randomized rounds |
test-packed-install.mjs | the release gate: packs the tarball, installs it into a fresh profile with dsh plugin add, boots it, and checks that the installed copy activates, that relative paths resolve inside the installed package (not the repo), and that the panel channel works |
scripts/regression.mjs drives the real tool implementations against a cloned batch, so it covers
parameter translation, path resolution and ledger writes — not just the Python scripts. It walks
eight steps: checkup → cull dry-run → cull → verify → archive plan → organize rename (dry/real) →
retrospective (dry/real) → graceful degradation when Lightroom is unavailable. Only "actually
export" and "actually write develop settings" need Lightroom; everything else verifies without it.
Both test.mjs and regression.mjs need DSH_HOME set so they never write into your real ~/.dsh.
Pass --reuse to the regression to skip the slow checkup stage and re-test only the downstream steps.
Design decisions and the reasoning behind them are in docs/DESIGN.md.
MIT. See LICENSE.