@theseus-whale/dsh-continuity 0.0.0-stage → 4.12.2

package/bin/relabel-dry-run.mjs

@@ -0,0 +1,381 @@
+#!/usr/bin/env node
+/**
+ * relabel-dry-run.mjs — 历史重打的**只读演练**：只输出"如果重打，会改哪些字段"。
+ *
+ * 不写 bc-index.jsonl、不写 baseline、不写 raw。想真正写盘必须另有写盘脚本，且先备份。
+ *
+ * 用法：
+ *   node bin/relabel-dry-run.mjs --index <bc-index.jsonl> --raw <session.v4.jsonl.zstd>
+ *     [--concepts <concepts.md>] [--baseline-file <file>]
+ *     [--center-min 0.02] [--lead 0.02] [--limit 500] [--json]
+ *
+ * 选项：
+ *   --index <file>         必填，bc-index.jsonl
+ *   --raw <file>           必填，raw 会话（.jsonl 或 .jsonl.zstd，帧式也支持）
+ *   --concepts <file>      A 表（默认 <pkg>/templates/concepts.md）
+ *   --baseline-file <file> 显式 baseline（{topic:mean} 或 {version,window,samples}）。
+ *                          不给则用"被对齐上的那些轮的 leave-one-out 均值"——**只用于演练，
+ *                          绝不能当生产 baseline**；输出里会标明来源。
+ *   --center-min / --lead / --topic-min-score / --min-score / --chunk-chars / --max-chunks
+ *   --limit <n>            只看前 n 条 diff（默认 60）
+ *   --json                 只输出 JSON
+ *
+ * 对齐规则：只用 turn 对齐 index 与 raw；raw 里找不到该轮文本 → 记 unresolved，不猜。
+ */
+import fs from 'node:fs';
+import path from 'node:path';
+import zlib from 'node:zlib';
+import { fileURLToPath } from 'node:url';
+
+const here = path.dirname(fileURLToPath(import.meta.url));
+const pkgRoot = path.resolve(here, '..');
+
+const argv = process.argv.slice(2);
+const opt = (name, fallback = null) => {
+  const i = argv.indexOf(`--${name}`);
+  if (i < 0) return fallback;
+  const next = argv[i + 1];
+  return next === undefined || next.startsWith('--') ? true : next;
+};
+const num = (name, fallback) => {
+  const v = opt(name);
+  if (v === null || v === true) return fallback;
+  const n = Number(v);
+  return Number.isFinite(n) ? n : fallback;
+};
+
+const indexFile = opt('index');
+const rawFile = opt('raw');
+if (!indexFile || indexFile === true || !rawFile || rawFile === true) {
+  console.error('用法：node bin/relabel-dry-run.mjs --index <bc-index.jsonl> --raw <session.jsonl[.zstd]>');
+  process.exit(2);
+}
+
+const conceptsPath = (() => {
+  const explicit = opt('concepts');
+  if (explicit && explicit !== true) return explicit;
+  const candidate = path.join(pkgRoot, 'templates', 'concepts.md');
+  return fs.existsSync(candidate) ? candidate : null;
+})();
+if (!conceptsPath || !fs.existsSync(conceptsPath)) {
+  console.error('找不到 A 表：请用 --concepts 指定');
+  process.exit(2);
+}
+
+const centerMin = num('center-min', 0.02);
+const lead = num('lead', 0.02);
+const topicMinScore = num('topic-min-score', 0.33);
+const minScore = num('min-score', 0.25);
+const chunkChars = num('chunk-chars', 600);
+const maxChunks = num('max-chunks', 12);
+const limit = num('limit', 60);
+const asJson = Boolean(opt('json', false));
+const baselineFile = opt('baseline-file');
+
+const { createBLabeler } = await import('file://' + path.join(pkgRoot, 'src/memory/bc-label.mjs'));
+const { loadConcepts } = await import('file://' + path.join(pkgRoot, 'src/memory/concepts-file.mjs'));
+const { embedTexts } = await import('file://' + path.join(pkgRoot, 'src/facets/embedding.mjs'));
+const { eventText } = await import('file://' + path.join(pkgRoot, 'src/session-log.mjs'));
+
+const INJECTED = /^\[(?:抽取|工作线|标签|规则|习惯|技能|归位|候选|存\?|闸|回正|学习|追忆|记忆|continuity:[a-z-]+)\]/;
+const MACHINE = /^(?:Current runtime context|Current working directory|Current DSH file policy|Approval prompts|Time sampled)/;
+
+const conceptRows = loadConcepts(conceptsPath);
+const labels = Object.fromEntries(conceptRows.map((r) => [r.label, r.description]));
+const kinds = Object.fromEntries(conceptRows.map((r) => [r.label, r.kind]));
+const concepts = conceptRows.map((r) => r.label);
+const topics = conceptRows.filter((r) => r.kind === 'topic').map((r) => r.label);
+const workflowNames = conceptRows.filter((r) => r.kind === 'workflow').map((r) => r.label);
+
+// ---------- raw 读取：支持纯 jsonl 与"一串 zstd 帧" ----------
+function readRawLines(file) {
+  const lines = [];
+  if (/\.zstd$/i.test(file)) {
+    const buf = fs.readFileSync(file);
+    const MAGIC = Buffer.from([0x28, 0xb5, 0x2f, 0xfd]);
+    const starts = [];
+    for (let i = 0; ; ) {
+      const k = buf.indexOf(MAGIC, i);
+      if (k < 0) break;
+      starts.push(k);
+      i = k + 4;
+    }
+    for (const start of starts) {
+      let out;
+      try { out = zlib.zstdDecompressSync(buf.subarray(start)); } catch { continue; }
+      for (const line of out.toString('utf8').split(/\r?\n/)) if (line) lines.push(line);
+    }
+    return lines;
+  }
+  return fs.readFileSync(file, 'utf8').split(/\r?\n/).filter(Boolean);
+}
+
+const rawLines = readRawLines(rawFile);
+const events = [];
+const seenSeq = new Set();
+for (const line of rawLines) {
+  let event;
+  try { event = JSON.parse(line.replace(/^\uFEFF/, '')); } catch { continue; }
+  const seq = Number(event?.seq);
+  if (Number.isInteger(seq)) {
+    if (seenSeq.has(seq)) continue;
+    seenSeq.add(seq);
+  }
+  events.push(event);
+}
+events.sort((a, b) => Number(a.seq) - Number(b.seq));
+
+// ---------- 与 updater 相同的 turnsFromEvents ----------
+const markers = events
+  .map((event) => ({ seq: Number(event?.seq), turn: Number(event?.data?.turn) }))
+  .filter((item) => Number.isInteger(item.seq) && Number.isInteger(item.turn))
+  .sort((a, b) => a.seq - b.seq);
+const turnAt = (seq) => {
+  let found = null;
+  for (const marker of markers) {
+    if (marker.seq > seq) break;
+    found = marker.turn;
+  }
+  if (found !== null) return found;
+  return markers.find((marker) => marker.seq >= seq)?.turn ?? null;
+};
+const isHuman = (event) => {
+  const source = event?.data?.source ?? event?.source;
+  if (source && typeof source.kind === 'string' && source.kind !== 'user') return false;
+  if (source && source.peer === true) return false;
+  const text = eventText(event, { includeReasoning: false }).trim();
+  return Boolean(text) && !INJECTED.test(text) && !MACHINE.test(text);
+};
+
+const turns = new Map();
+for (const event of events) {
+  const type = event?.type;
+  if (type !== 'user/message' && type !== 'assistant/message') continue;
+  if (type === 'user/message' && !isHuman(event)) continue;
+  const text = eventText(event, { includeReasoning: false }).trim();
+  if (!text || INJECTED.test(text) || MACHINE.test(text)) continue;
+  const seq = Number(event?.seq);
+  const turn = Number.isInteger(Number(event?.data?.turn)) ? Number(event.data.turn) : turnAt(seq);
+  if (!Number.isInteger(turn)) continue;
+  const item = turns.get(turn) ?? { turn, userTexts: [], assistantTexts: [] };
+  if (type === 'user/message') item.userTexts.push(text);
+  else item.assistantTexts.push(text);
+  turns.set(turn, item);
+}
+
+// ---------- index 读取 ----------
+const indexRows = fs.readFileSync(indexFile, 'utf8')
+  .replace(/^\uFEFF/, '')
+  .split(/\r?\n/)
+  .filter((line) => line.trim() !== '')
+  .map((line) => JSON.parse(line));
+
+const labeler = createBLabeler({ embed: embedTexts, labels, concepts, kinds });
+
+const topicOfRecord = (record) => {
+  if (Array.isArray(record.genericTopic) && record.genericTopic.length > 0) return record.genericTopic[0];
+  const generic = Array.isArray(record.generic) ? record.generic : [];
+  return generic.find((name) => kinds[name] === 'topic') ?? null;
+};
+const workflowsOfRecord = (record) => {
+  if (Array.isArray(record.genericWorkflow)) return record.genericWorkflow;
+  const generic = Array.isArray(record.generic) ? record.generic : [];
+  return generic.filter((name) => kinds[name] === 'workflow');
+};
+const scoreOfRecord = (record, concept) => {
+  const scores = record.conceptScores ?? {};
+  const value = Number(scores[concept]);
+  return Number.isFinite(value) ? value : null;
+};
+
+// ---------- baseline ----------
+// v2 文件形如 {version:2, topics:{...}, workflows:{...}}；v1 只有 samples（当作 topics）
+let explicitBaseline = null;
+let workflowBaseline = null;
+let baselineSource = 'leave-one-out(演练用)';
+if (baselineFile && baselineFile !== true) {
+  const raw = JSON.parse(fs.readFileSync(baselineFile, 'utf8').replace(/^\uFEFF/, ''));
+  const topicSource = raw?.topics && typeof raw.topics === 'object'
+    ? raw.topics
+    : (raw?.samples && typeof raw.samples === 'object' ? raw.samples : raw);
+  const workflowSource = raw?.workflows && typeof raw.workflows === 'object' ? raw.workflows : null;
+
+  const buildMap = (source, names) => Object.fromEntries(names.map((name) => {
+    const value = source?.[name];
+    if (Array.isArray(value)) {
+      const nums = value.map((x) => Number(x)).filter((x) => Number.isFinite(x));
+      return [name, nums.length > 0 ? nums.reduce((a, b) => a + b, 0) / nums.length : Number.NaN];
+    }
+    if (value === undefined || value === null || value === '') return [name, Number.NaN];
+    const n = Number(value);
+    return [name, Number.isFinite(n) ? n : Number.NaN];
+  }));
+
+  const fill = (map, names) => {
+    const known = Object.values(map).filter((v) => Number.isFinite(v));
+    if (known.length === 0) return null;
+    const fallback = known.reduce((a, b) => a + b, 0) / known.length;
+    for (const name of names) if (!Number.isFinite(map[name])) map[name] = fallback;
+    return map;
+  };
+
+  explicitBaseline = fill(buildMap(topicSource, topics), topics);
+  if (workflowSource) workflowBaseline = fill(buildMap(workflowSource, workflowNames), workflowNames);
+
+  if (!explicitBaseline && !workflowBaseline) {
+    baselineSource = 'raw-bootstrap(文件无样本)';
+  } else if (explicitBaseline) {
+    baselineSource = `file:${baselineFile}`;
+  }
+}
+
+// ---------- 逐轮重打（只读） ----------
+const evaluated = [];
+for (const record of indexRows) {
+  const turn = Number(record.turn);
+  if (!Number.isInteger(turn)) continue;
+  const entry = turns.get(turn);
+  if (!entry) {
+    evaluated.push({ record, turn, unresolved: true });
+    continue;
+  }
+  const text = [...entry.userTexts, ...entry.assistantTexts].join('\n');
+  evaluated.push({ record, turn, text, unresolved: false });
+}
+
+// leave-one-out baseline 需要用每轮 raw 分；没给文件时先粗算一遍
+let looBaseline = null;
+if (!explicitBaseline) {
+  const sums = Object.fromEntries(topics.map((t) => [t, 0]));
+  let count = 0;
+  for (const item of evaluated) {
+    if (item.unresolved) continue;
+    const res = await labeler.labelChunked(item.text, {
+      topK: 3, minScore, topicMinScore, topicCenterMin: -Infinity, topicLead: -Infinity,
+      chunkChars, maxChunks, topicTopK: 1,
+    });
+    const scores = res.topicScores ?? {};
+    const keys = topics.filter((t) => Number.isFinite(Number(scores[t])));
+    if (keys.length === 0) continue;
+    for (const t of keys) sums[t] += Number(scores[t]);
+    count += 1;
+  }
+  if (count > 0) {
+    looBaseline = Object.fromEntries(topics.map((t) => [t, sums[t] / count]));
+    baselineSource = `leave-one-out(演练用,${count} 轮)`;
+  } else {
+    baselineSource = 'raw-bootstrap(无可用轮次)';
+  }
+}
+
+const diffs = [];
+const summary = {
+  indexRows: indexRows.length,
+  evaluated: 0,
+  unresolved: 0,
+  changed: 0,
+  unchanged: 0,
+  emptyToTopic: 0,
+  topicToEmpty: 0,
+  topicChanged: 0,
+  workflowChanged: 0,
+  scoreChanged: 0,
+  baselineSource,
+};
+
+for (const item of evaluated) {
+  if (item.unresolved) { summary.unresolved += 1; continue; }
+  summary.evaluated += 1;
+
+  let baseline = explicitBaseline ?? looBaseline;
+  if (!explicitBaseline && looBaseline) {
+    // 近似 leave-one-out：把本轮 raw 分从均值里扣掉（用粗算的那一轮分不够精确，宁可标来源）
+    baseline = looBaseline;
+  }
+
+  const res = await labeler.labelChunked(item.text, {
+    topK: 3,
+    minScore,
+    topicMinScore,
+    topicCenterMin: centerMin,
+    topicLead: lead,
+    chunkChars,
+    maxChunks,
+    topicTopK: 1,
+    topicBaseline: baseline ?? undefined,
+    workflowBaseline: workflowBaseline ?? undefined,
+  });
+
+  const oldTopic = topicOfRecord(item.record);
+  const oldWorkflows = workflowsOfRecord(item.record);
+  const oldScore = oldTopic ? scoreOfRecord(item.record, oldTopic) : null;
+  const oldConfidence = Number.isFinite(Number(item.record.topicConfidence)) ? Number(item.record.topicConfidence) : null;
+
+  const newTopic = (res.topic ?? [])[0]?.concept ?? null;
+  const newWorkflows = (res.workflow ?? []).map((w) => w.concept);
+  const newScore = newTopic ? Number(res.topicScores?.[newTopic] ?? res.topic?.[0]?.score ?? 0) : null;
+  const newConfidence = newTopic ? Number(res.topicConfidence ?? 0) : null;
+
+  const changedTopic = (oldTopic ?? null) !== (newTopic ?? null);
+  const changedWorkflow = oldWorkflows.join('|') !== newWorkflows.join('|');
+  const changedScore = oldScore !== null && newScore !== null && Math.abs(oldScore - newScore) > 1e-6;
+  const changed = changedTopic || changedWorkflow || changedScore || (oldConfidence ?? null) !== (newConfidence ?? null);
+
+  if (changed) summary.changed += 1; else summary.unchanged += 1;
+  if (!oldTopic && newTopic) summary.emptyToTopic += 1;
+  if (oldTopic && !newTopic) summary.topicToEmpty += 1;
+  if (oldTopic && newTopic && oldTopic !== newTopic) summary.topicChanged += 1;
+  if (changedWorkflow) summary.workflowChanged += 1;
+  if (changedScore) summary.scoreChanged += 1;
+
+  diffs.push({
+    turn: item.turn,
+    oldTopic,
+    newTopic,
+    oldWorkflows,
+    newWorkflows,
+    oldTopicScore: oldScore,
+    newTopicScore: newScore,
+    oldConfidence,
+    newConfidence,
+    changed,
+  });
+}
+
+const payload = {
+  summary,
+  diffs,
+  note: explicitBaseline
+    ? 'baseline 来自文件；本脚本只读，未写任何文件'
+    : 'baseline 由被对齐轮次自身算出（leave-one-out 近似），仅用于演练，不可当生产 baseline；本脚本只读',
+};
+
+const looRounds = Number((summary.baselineSource.match(/leave-one-out\(演练用,(\d+) 轮\)/) ?? [])[1] ?? 0);
+const lowSample = !explicitBaseline && looRounds > 0 && looRounds < 100;
+if (lowSample) {
+  payload.warning = `baseline 仅 ${looRounds} 轮（<100），结果仅供参考，不能当生产结论`;
+}
+
+if (asJson) {
+  console.log(JSON.stringify(payload, null, 2));
+} else {
+  console.log(`  index ${summary.indexRows} 行 · 可重打 ${summary.evaluated} · unresolved ${summary.unresolved}`);
+  console.log(`  A 表 ${conceptRows.length} 条（topic ${topics.length}）· 判据 centered≥${centerMin} lead≥${lead} · baseline ${summary.baselineSource}`);
+  if (lowSample) console.log(`  ⚠ ${payload.warning}`);
+  console.log(`  变更 ${summary.changed} · 不变 ${summary.unchanged} · 空→topic ${summary.emptyToTopic} · topic→空 ${summary.topicToEmpty}`);
+  console.log(`  换 topic ${summary.topicChanged} · 换 workflow ${summary.workflowChanged} · 分数变化 ${summary.scoreChanged}`);
+  console.log('');
+  const shown = diffs.filter((d) => d.changed).slice(0, limit);
+  if (shown.length === 0) {
+    console.log('  没有会变更的行。');
+  } else {
+    console.log('  turn  oldTopic → newTopic                          score         wf');
+    for (const d of shown) {
+      const score = `${d.oldTopicScore === null ? '-' : d.oldTopicScore.toFixed(3)}→${d.newTopicScore === null ? '-' : d.newTopicScore.toFixed(3)}`;
+      console.log(`  ${String(d.turn).padStart(5)}  ${String(d.oldTopic ?? '(空)').padEnd(12)} → ${String(d.newTopic ?? '(空)').padEnd(14)} ${score.padEnd(13)} ${d.oldWorkflows.join('/') || '-'} → ${d.newWorkflows.join('/') || '-'}`);
+    }
+    if (diffs.filter((d) => d.changed).length > shown.length) {
+      console.log(`  … 还有 ${diffs.filter((d) => d.changed).length - shown.length} 条，用 --limit 或 --json 看全`);
+    }
+  }
+}

package/bin/relik_pipeline.py

@@ -0,0 +1,39 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+"""ReLiK reader-only entity linking. Input JSONL: {id,text,candidates:[...]}."""
+import json
+import os
+import sys
+
+from relik import Relik
+from relik.reader.pytorch_modules.span import RelikReaderForSpanExtraction
+
+reader_path = os.environ["RELIK_READER"]
+reader = RelikReaderForSpanExtraction(reader_path)
+relik = Relik(reader=reader, task="span", window_size=32, window_stride=16)
+
+src, out = sys.argv[1], sys.argv[2]
+rows = []
+for line in open(src, encoding="utf-8"):
+    if not line.strip():
+        continue
+    item = json.loads(line)
+    result = relik(item["text"], candidates=item.get("candidates", []))
+    spans = [
+        {"start": span.start, "end": span.end, "text": span.text, "label": span.label}
+        for span in result.spans
+    ]
+    candidates = result.candidates.span if result.candidates is not None else []
+    serial_candidates = []
+    for span_group in candidates:
+        span_group_out = []
+        for candidate_group in span_group:
+            span_group_out.append([
+                {"text": getattr(doc, "text", None), "id": str(getattr(doc, "id", None))}
+                for doc in candidate_group
+            ])
+        serial_candidates.append(span_group_out)
+    rows.append({"id": item.get("id"), "text": item["text"], "spans": spans, "candidates": serial_candidates})
+with open(out, "w", encoding="utf-8") as f:
+    json.dump({"n": len(rows), "rows": rows}, f, ensure_ascii=False, indent=2)
+print(json.dumps({"n": len(rows), "with_spans": sum(1 for row in rows if row["spans"])}, ensure_ascii=False, indent=2))

package/bin/relik_service.py

@@ -0,0 +1,52 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+"""One-shot ReLiK reader service: stdin {text,candidates}, stdout {ok, links}."""
+import json
+import os
+import sys
+
+sys.stdout.reconfigure(encoding="utf-8")
+sys.stderr.reconfigure(encoding="utf-8")
+
+from relik import Relik  # noqa: E402
+from relik.reader.pytorch_modules.span import RelikReaderForSpanExtraction  # noqa: E402
+
+reader = RelikReaderForSpanExtraction(os.environ["RELIK_READER"])
+relik = Relik(reader=reader, task="span", window_size=32, window_stride=16)
+
+def handle(request):
+    if request.get("health"):
+        return {"ok": True, "health": "ready"}
+    try:
+        result = relik(str(request.get("text", "")), candidates=request.get("candidates") or [])
+        links = []
+        candidate_groups = result.candidates.span if result.candidates is not None else []
+        for index, span in enumerate(result.spans):
+            group = []
+            if candidate_groups and index < len(candidate_groups) and candidate_groups[index]:
+                group = candidate_groups[index][0]
+            elif candidate_groups and candidate_groups[0]:
+                group = candidate_groups[0][0]
+            top = next((candidate for candidate in group if candidate.text in (span.label, span.text)), group[0] if group else None)
+            links.append({
+                "text": span.text,
+                "entityId": str(top.id) if top is not None else None,
+                "score": None,
+                "start": span.start,
+                "end": span.end,
+            })
+        return {"ok": True, "links": links, "spans": [{"text": s.text, "start": s.start, "end": s.end} for s in result.spans]}
+    except Exception as error:
+        return {"ok": False, "reason": type(error).__name__, "message": str(error)[:300]}
+
+
+if "--serve" in sys.argv:
+    for raw in sys.stdin.buffer:
+        line = raw.decode("utf-8", "replace").strip()
+        if not line:
+            continue
+        request = json.loads(line)
+        print(json.dumps({"id": request.get("id"), **handle(request)}, ensure_ascii=False), flush=True)
+else:
+    request = json.loads(sys.stdin.buffer.read().decode("utf-8"))
+    print(json.dumps(handle(request), ensure_ascii=False))

package/bin/srl_pipeline.py

@@ -0,0 +1,55 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+"""SRL pipeline: transformer-srl + local CoNLL2012 archive + NLTK POS fallback."""
+import json
+import os
+import sys
+
+import nltk
+import spacy
+
+
+def add_pos(doc):
+    tags = nltk.pos_tag([token.text for token in doc])
+    for token, (_, tag) in zip(doc, tags):
+        token.tag_ = tag
+        if tag.startswith("VB"):
+            token.pos_ = "VERB"
+        elif tag.startswith("NN"):
+            token.pos_ = "NOUN"
+        elif tag.startswith("JJ"):
+            token.pos_ = "ADJ"
+        elif tag.startswith("RB"):
+            token.pos_ = "ADV"
+        elif tag in ("IN", "TO"):
+            token.pos_ = "ADP"
+        elif tag in ("DT", "PRP", "PRP$"):
+            token.pos_ = "DET"
+        else:
+            token.pos_ = "X"
+    return doc
+
+
+blank = spacy.blank("en")
+blank.add_pipe(add_pos)
+spacy.load = lambda *args, **kwargs: blank
+
+import transformer_srl.dataset_readers  # noqa: E402,F401
+import transformer_srl.models  # noqa: E402,F401
+from allennlp.models.archival import load_archive  # noqa: E402
+from transformer_srl import predictors  # noqa: E402
+
+archive = load_archive(os.environ["SRL_ARCHIVE"])
+predictor = predictors.SrlTransformersPredictor.from_archive(archive, "transformer_srl")
+
+src, out = sys.argv[1], sys.argv[2]
+rows = []
+for line in open(src, encoding="utf-8"):
+    if not line.strip():
+        continue
+    item = json.loads(line)
+    result = predictor.predict(sentence=item["text"])
+    rows.append({"id": item.get("id"), "text": item["text"], **result})
+with open(out, "w", encoding="utf-8") as f:
+    json.dump({"n": len(rows), "rows": rows}, f, ensure_ascii=False, indent=2)
+print(json.dumps({"n": len(rows), "with_verbs": sum(1 for row in rows if row["verbs"])}, ensure_ascii=False, indent=2))

package/bin/srl_service.py

@@ -0,0 +1,68 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+"""One-shot SRL service: stdin {text}, stdout {ok, srl}."""
+import json
+import os
+import sys
+
+import nltk
+import spacy
+
+sys.stdout.reconfigure(encoding="utf-8")
+sys.stderr.reconfigure(encoding="utf-8")
+
+
+def add_pos(doc):
+    tags = nltk.pos_tag([token.text for token in doc])
+    for token, (_, tag) in zip(doc, tags):
+        token.tag_ = tag
+        if tag.startswith("VB"):
+            token.pos_ = "VERB"
+        elif tag.startswith("NN"):
+            token.pos_ = "NOUN"
+        elif tag.startswith("JJ"):
+            token.pos_ = "ADJ"
+        elif tag.startswith("RB"):
+            token.pos_ = "ADV"
+        elif tag in ("IN", "TO"):
+            token.pos_ = "ADP"
+        elif tag in ("DT", "PRP", "PRP$"):
+            token.pos_ = "DET"
+        else:
+            token.pos_ = "X"
+    return doc
+
+
+blank = spacy.blank("en")
+blank.add_pipe(add_pos)
+spacy.load = lambda *args, **kwargs: blank
+
+import transformer_srl.dataset_readers  # noqa: E402,F401
+import transformer_srl.models  # noqa: E402,F401
+from allennlp.models.archival import load_archive  # noqa: E402
+from transformer_srl import predictors  # noqa: E402
+
+archive = load_archive(os.environ["SRL_ARCHIVE"])
+predictor = predictors.SrlTransformersPredictor.from_archive(archive, "transformer_srl")
+
+def handle(request):
+    if request.get("health"):
+        return {"ok": True, "health": "ready"}
+    try:
+        result = predictor.predict(sentence=str(request.get("text", "")))
+        verb = result["verbs"][0] if result["verbs"] else None
+        return {"ok": True, "srl": {"words": result["words"], "verb": verb}}
+    except Exception as error:
+        return {"ok": False, "reason": type(error).__name__, "message": str(error)[:300]}
+
+
+if "--serve" in sys.argv:
+    for raw in sys.stdin.buffer:
+        line = raw.decode("utf-8", "replace").strip()
+        if not line:
+            continue
+        request = json.loads(line)
+        print(json.dumps({"id": request.get("id"), **handle(request)}, ensure_ascii=False), flush=True)
+else:
+    request = json.loads(sys.stdin.buffer.read().decode("utf-8"))
+    print(json.dumps(handle(request), ensure_ascii=False))