dsh-md-convert v0.5.6

v0.5.6 fixes a cleanup safety issue found during the real install/call test.

• Fix: sweepStale() no longer deletes every tmpdir entry sharing the
  "dsh-md-convert-" prefix — only plugin-owned leftovers (old mkdtemp
  format dsh-md-convert-<6 alnum>, or - dirs with a dead pid) are
  swept; user-created dirs with the same prefix are preserved
• New cleanup regression test; suite now 10/10 passing

Install: dsh plugin add github:yakoylp/dsh-md-convert

dsh-md-convert v0.5.5

v0.5.5 fixes the tool result being blocked by the DSH runtime value check.

• Fix: chain in the output schema was declared as an array while
  convert.js returns a string — every successful conversion was rejected
  with "value.chain" must be an array and the result never reached the
  conversation. Schema now declares chain: { type: "string" }, matching
  convert.js / cli.js / README examples
• Render callback prints the chain string directly (no .length/.join)
• Regression test validates execute() return values against the declared
  output schema (verified the old schema fails it with the exact runtime
  error)
• Test suite: 9/9 passing

Install: dsh plugin add github:yakoylp/dsh-md-convert

v0.5.4

关键修复

• 路径解析彻底修复：md_convert 相对路径现基于会话工作区（exec.agent.session.header.cwd）而非进程 cwd，Windows 盘符绝对路径原样保留——修复所有调用报 E_FILE_NOT_FOUND、路径被拼到 DSH 依赖目录的问题。
• OCR 异步化：pip 安装 / 模型下载 / OCR 三处分钟级同步调用改为异步（新增
  unAsync），插件运行于 DSH 宿主时不再冻结整个 web 服务。
• 缺模型明确报错：扫描件首次转换且模型未缓存时返回 E_OCR_DEPS + 指引执行 dsh-md-convert deps，不再报晦涩的 paddle 错误。
• 进程信号钩子不再泄漏到宿主、临时目录并发安全（按 pid 清理）、措辞修正（PP-StructureV3 → 路由 OCR）、入参校验、Python 除零保护。

完整清单见 CHANGELOG.md。
pm test 8/8 通过。

dsh-md-convert v0.5.3

v0.5.3 completes the boot fix: v0.5.2 activated but the tool registration
still threw because it did not match the DSH tools service contract.

• Fix: register the md_convert tool as a DSH ToolDefinition —
  single definition object with name, compiled JSON Schema parameters
  (required array at object root), mandatory output { schema, render },
  and execute() (the runtime never calls run())
• Drop exec.cwd (ToolRunContext has no such field; relative paths now
  resolve from process.cwd())
• Regression test now enforces the same checks as the real runtime:
  definition shape, supported JSON Schema keyword subset, required-array
  placement — so a malformed registration fails tests, not the next boot
• Test suite: 6/6 passing

Install: dsh plugin add github:yakoylp/dsh-md-convert

dsh-md-convert v0.5.2

v0.5.2 fixes the remaining plugin boot failure: after v0.5.1, dsh web still
aborted with cannot get property "config" without inject.

• Fix: plugin config is read from the apply(ctx, config) second argument
  (with {} fallback) instead of ctx.config, matching the official DSH
  plugin pattern
• Add cordis activation regression test (test/cordis-entry.test.mjs) that
  boots the entry through real cordis and asserts the md_convert tool
  registers; verified it reproduces the loader error against the old code
• Test suite: 6/6 passing (4 original + 2 new cordis-entry tests)

Install: dsh plugin add github:yakoylp/dsh-md-convert

dsh-md-convert v0.5.1

v0.5.1 fixes a plugin boot failure when installed into DSH.

• Fix: the plugin no longer injects the non-existent workspace service.
  Previously the entry stayed pending (waiting for service: workspace) forever
  and dsh web aborted with 1 entry did not activate
• inject now declares tools, which ctx.tools.register() actually requires;
  the workspace lookup is read via ctx.get with process.cwd() fallback
• Test suite: 4/4 passing (incl. real conversion smoke of 7 sample files)

Install: dsh plugin add github:yakoylp/dsh-md-convert

dsh-md-convert v0.5.0

OCR pipeline replaced with a modular routing engine — CPU-first, lightweight models, cost/performance oriented.

• New routing pipeline: PP-DocLayout-L layout analysis (threshold 0.3) routes each region:
  • text -> RapidOCR (PP-OCRv6 ONNX, fastest)
  • tables -> SLANet structure + RT-DETR cell localization + OCR fill
  • single-column boxes (text frames) -> recursive inner layout analysis
  • formulas -> FormulaNet-Plus-S (on by default)
  • stamps -> comments
• Heading levels come from the layout model (regex heuristics removed)
• Overlap dedup: containment keeps the larger region, near-overlap keeps higher score
• Removed old monolithic PP-StructureV3 pipeline and its flags (--ocr-fast / --no-table / --ocr-formula / --ocr-seal / --ocr-chart)
• New deps: rapidocr + onnxruntime; new model set (PP-DocLayout-L, SLANeXt_wired/SLANet_plus, RT-DETR-L x2, PP-FormulaNet_plus-S)
• Measured: single-page scanned letter ~24s; 23-page scanned application ~3.3 min
• Offline after one dsh-md-convert deps run; auto-install of pip deps on first use
• Install: dsh plugin add github:yakoylp/dsh-md-convert

dsh-md-convert v0.4.0

Add --ocr-fast mode (PP-OCRv5 mobile models).

• --ocr-fast / ocr.fast config: uses PP-OCRv5_mobile_det/rec instead of server models
• Measured on a single A4 scanned page: 61.7s (server) -> 35.0s (mobile), ~40% faster
• Accuracy comparable on typical scans (per-character differences possible)
• Pre-download models: dsh-md-convert deps --ocr-fast (offline afterwards)
• Install: dsh plugin add github:yakoylp/dsh-md-convert

dsh-md-convert v0.3.0

Configurable OCR pipeline.

• Document orientation classification / unwarping / textline orientation now ON by default (auto-corrects tilted/rotated scans; lightweight models)
• Formula / seal / chart recognition OFF by default; enable per run with --ocr-formula / --ocr-seal / --ocr-chart, or via plugin config ocr.formula/seal/chart
• dsh-md-convert deps --ocr-formula|--ocr-seal|--ocr-chart pre-downloads the feature models (offline afterwards, as before)
• Verified offline: default scan conversion ~51 s; seal recognition works offline once its model is cached
• Install: dsh plugin add github:yakoylp/dsh-md-convert

dsh-md-convert v0.2.1

OCR fix: fully offline capability + ~20x speedup.

• PP-StructureV3 runtime now sets PADDLE_PDX_DISABLE_MODEL_SOURCE_CHECK=true: with models cached (pre-downloaded via dsh-md-convert deps), OCR makes zero network calls — works without internet
• Default-trimmed pipeline (formula/seal/chart recognition off, table kept) + render max-side cap 1600px: scanned A3 page conversion dropped from 20+ min (timeout) to ~53 s
• dsh-md-convert check/deps now verify and pre-download OCR models to ~/.paddlex/official_models/ (one network run at install time; offline afterwards)
• Regenerated scanned reference output (clean table + text OCR)
• Install: dsh plugin add github:yakoylp/dsh-md-convert