@gobing-ai/knowledge-kit 0.1.5 → 0.1.7

package/.claude-plugin/marketplace.json

@@ -6,7 +6,7 @@
     },
     "plugins": [
         {
-            "version": "0.1.5",
+            "version": "0.1.7",
             "name": "kk",
             "source": "./plugins/kk",
             "description": "knowledge-kit — a set of toolkit for knowledge management"

package/dist/index.js

@@ -51661,7 +51661,8 @@ async function createItemDecisionAsker(options = {}) {
   const injected = options.hub ?? await loadInjectedHub();
   const fallback = options.fallback ?? (injected === undefined ? (decisionId, input, maker) => runDecisionSubprocess(decisionId, input, maker, fs) : undefined);
   const hub = injected ?? await (await import("@gobing-ai/ts-ai-decision")).createDecisionHub({
-    catalogs: [catalogPath]
+    catalogs: [catalogPath],
+    builtins: { makerOptions: { fetch: bufferedDecisionFetch } }
   });
   const backendFailures = new Map;
   return {
@@ -51696,6 +51697,24 @@ async function createItemDecisionAsker(options = {}) {
     }
   };
 }
+function createBufferedDecisionFetch(inner = fetch) {
+  return async (input, init) => {
+    const response = await inner(input, init);
+    const text = await response.text();
+    const buffered = {
+      ok: response.ok,
+      status: response.status,
+      statusText: response.statusText,
+      headers: response.headers,
+      body: null,
+      text: async () => text,
+      json: async () => JSON.parse(text),
+      clone: () => buffered
+    };
+    return buffered;
+  };
+}
+var bufferedDecisionFetch = createBufferedDecisionFetch();
 function runDecisionSubprocess(decisionId, input, maker, fs = createNodeFileSystem()) {
   const command = process.env.KK_SPUR_BIN ?? "spur";
   const scratch = join21(tmpdir3(), `kk-decision-${randomUUID2()}`);

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@gobing-ai/knowledge-kit",
-    "version": "0.1.5",
+    "version": "0.1.7",
     "description": "An ingest → create → publish content pipeline CLI (Bun).",
     "repository": {
         "type": "git",

package/plugins/generations/daily-article-gen/dist/index.js

@@ -12399,12 +12399,14 @@ ${items}
 `;
 }
 var AI_NEWS_TAG = "ai-news";
+var AI_DAILY_SUB_CATEGORY_TAG = "sub-category:ai-daily";
 function tagsFor(docs) {
   const existing = docs.flatMap((d) => {
     const meta = d.metadata;
     return Array.isArray(meta?.tags) ? meta.tags.filter((t) => typeof t === "string") : [];
   });
-  return [AI_NEWS_TAG, ...existing.filter((t) => t !== AI_NEWS_TAG)];
+  const mandatory = [AI_NEWS_TAG, AI_DAILY_SUB_CATEGORY_TAG];
+  return [...new Set([...mandatory, ...existing.filter((t) => !mandatory.includes(t))])];
 }
 var LEADING_H1_RE = /^#(?!#)\s*[^\n]+\n+/.source;
 var NOTE_EXCLUDED_RE = /\u672A\u5199\u5165[\uFF1A:](.*)/g;

package/plugins/generations/daily-article-gen/src/index.ts

@@ -122,16 +122,18 @@ export function renderDailyArticle(docs: Doc[], date: string): string {
     return `# ${episodeTitle(date)}\n\n${intro}\n\n${items}\n`;
 }
 
-/** Single-source daily tag: stamped on every Content this generator emits (task 0119). */
+/** Single-source daily tags: stamped on every Content this generator emits (task 0119 + sub-category series tag). */
 const AI_NEWS_TAG = 'ai-news';
+const AI_DAILY_SUB_CATEGORY_TAG = 'sub-category:ai-daily';
 
-/** Merge existing item tags (deduped) with the mandatory `ai-news` tag, first position. */
+/** Merge existing item tags (deduped) with the mandatory `ai-news` and `sub-category:ai-daily` tags, leading positions. */
 function tagsFor(docs: Doc[]): string[] {
     const existing = docs.flatMap((d) => {
         const meta = d.metadata as Record<string, unknown> | undefined;
         return Array.isArray(meta?.tags) ? meta.tags.filter((t): t is string => typeof t === 'string') : [];
     });
-    return [AI_NEWS_TAG, ...existing.filter((t) => t !== AI_NEWS_TAG)];
+    const mandatory = [AI_NEWS_TAG, AI_DAILY_SUB_CATEGORY_TAG];
+    return [...new Set([...mandatory, ...existing.filter((t) => !mandatory.includes(t))])];
 }
 
 /** A leading H1 line — agent bodies used to carry their own title; the canonical title is computed (task 0189 R4). */

package/plugins/generations/voxcpm-tts/README.md

@@ -79,6 +79,24 @@ that evidence belongs to task 0258.
 > Grade any render with `scripts/audit-speaker-consistency.py` (needs this plugin's venv):
 > a per-item F0 spread over 45 Hz means the profile is not anchoring a speaker.
 
+### Known artifact: a 40 ms echo lattice on long renders (2026-10-20)
+
+A `reference`-anchored profile renders short text clean and long text with a 40 ms (≈25 Hz) periodic
+reflection plus a pause floor only 16–22 dB below speech — heard as 「回声感、瓮声瓮气」. Measured on the
+raw model output (no ffmpeg join, no mp3): `kk stage audio-quality` reports `cepstrum 40 ms @ 27–31 dB`
+for renders >30 s, `pause floor −16…−22 dB`, while a 10 s render of the same profile/text measures
+`22 ms @ 19 dB / −22.5 dB` and the same days' `gemini-tts` episodes measure `20–25 ms @ 19–20 dB /
+−34…−37 dB`. `inference_timesteps`, `cfg_value` and `normalize` do not move it (measured 2026-10-20).
+
+```bash
+# reproducible in ~2.5 min, no episode needed:
+plugins/generations/voxcpm-tts/.venv/bin/python \
+  plugins/generations/voxcpm-tts/scripts/probe-long-render.py --out /tmp/voxcpm-probe
+kk stage audio-quality /tmp/voxcpm-probe/long-baseline.wav
+```
+
+Treat a driver/profile change as unproven until a LONG render is measured, not a single sentence.
+
 ## Control policy
 
 `normalize` maps directly; an explicit `seed` calls `torch.manual_seed` before synthesis (the
@@ -90,3 +108,16 @@ one for every segment, derived from its `emotion`, and the anchored clip already
 A per-request `engine`,
 `personality=true` or an active `effects_chain` is rejected. Speed stays in shared postprocessing.
 Transcription is omitted, so ASR-dependent QC is reported unavailable (not passing evidence).
+
+#### API/library claim sources and confidence
+
+The externally changeable claims this file makes about the installed `voxcpm` package, each with the
+artifact it was read from. This block is enforced by `structure/api-library-claim-citation` (the file
+is in that rule's `--require-block` registry), so a HIGH claim without a citation fails the gate.
+
+| Claim | Source | Local verification | Confidence |
+| --- | --- | --- | --- |
+| `reference_wav_path` is accepted only by VoxCPM2 checkpoints; any other model id refuses it, so a `reference` profile is VoxCPM2-only while older checkpoints clone through `prompt_wav_path` + `prompt_text` | installed `voxcpm` package | `plugins/generations/voxcpm-tts/.venv/lib/python3.12/site-packages/voxcpm/core.py:239` (`ValueError("reference_wav_path is only supported with VoxCPM2 models")`); reproduced with `scripts/probe-long-render.py --model openbmb/VoxCPM1.5` | HIGH |
+| `generate()`/`_generate()` exposes `cfg_value` (default 2.0), `inference_timesteps` (10), `normalize`, `denoise` and `max_len`, and this plugin passes none of them — the profile and text are the only inputs it sets | installed `voxcpm` package plus this plugin's worker | `…/site-packages/voxcpm/core.py:180`–`:193` (signature and defaults) and `python/worker.py:252`–`:262` (the three `generate(...)` call shapes) | HIGH |
+| `VOXCPM_MODEL` defaults to `openbmb/VoxCPM2`, the model self-reports its output rate, and that rate is what the WAV header carries (no resampling in this plugin) | this plugin's worker | `python/worker.py:181` (`environ.get("VOXCPM_MODEL", "openbmb/VoxCPM2")`), `:207` (`int(model.tts_model.sample_rate)`) and `:264` (`write_pcm16_wav(dest, pcm, sample_rate)`) | HIGH |
+| A `reference` profile pins one speaker (median-pitch swing 22–26 Hz, sd 9–12 Hz) while `instruct` samples a new speaker per request (141 Hz swing), so multi-segment work needs a reference or continuation profile | measured on this machine, 2026-10-15, three texts | `scripts/audit-speaker-consistency.py` grades a render; the 2026-10-15 figures are the ones its docstring records | MEDIUM |

package/plugins/generations/voxcpm-tts/scripts/probe-long-render.py

@@ -0,0 +1,134 @@
+#!/usr/bin/env python3
+"""Reproduce the long-form voxcpm "echo / hollow" artifact in ~2.5 minutes.
+
+Why this exists: the 2026-10-20 operator report ("回声感、瓮声瓮气") is not reproducible on a short
+sentence — a 10 s render comes out clean — but it IS reproducible on a paragraph of ~450 Chinese
+characters, which is the length the daily pipeline actually renders per section. Measured 2026-10-20
+with `kk stage audio-quality` (real-cepstrum echo probe + pause floor):
+
+    short render (~10 s)      cepstrum 22 ms @ 19 dB   pause floor -22.5 dB   clean
+    long render  (~72 s)      cepstrum 40 ms @ 30 dB   pause floor -16.1 dB   reproduced
+    shipped 2026-10-20 §00    cepstrum 40 ms @ 23 dB   pause floor -15.7 dB   reproduced
+
+So the artifact is duration-dependent, and a knob experiment has to use a LONG text or it measures
+nothing. The same run measured `inference_timesteps=20` (29.6 dB) and `cfg_value=2.5` (31.5 dB) against
+the default's 30 dB: none of the exposed sampler knobs remove the 40 ms periodicity.
+
+Usage (needs the plugin venv):
+
+    plugins/generations/voxcpm-tts/.venv/bin/python \
+        plugins/generations/voxcpm-tts/scripts/probe-long-render.py [--out /tmp/voxcpm-probe]
+
+    kk stage audio-quality /tmp/voxcpm-probe/long-baseline.wav   # then read cepstrum + pauseFloor
+
+Options: `--text-file <path>` (default: the built-in paragraph below), `--ref <wav>` (default: the
+operator's `robin-news` clip), `--steps N`, `--cfg F`, `--normalize`, `--name <label>`.
+
+The verdict vocabulary lives in `plugins/kk/scripts/audio-quality.ts` (`kk stage audio-quality`);
+this script only renders.
+"""
+
+from __future__ import annotations
+
+import argparse
+import contextlib
+import os
+import sys
+import time
+import wave
+
+DEFAULT_TEXT = (
+    "大家好，今天是2026年10月20日，星期二。这里是由Robin为你带来的咸话咸说之每日AI乱炖。"
+    "今天聊两件跟动手有关的事，一位自称几乎不懂游戏开发的作者，靠编程智能体把一款手机端的即时战略游戏"
+    "一路做到能离线运行；另一件是 Google 开源了让 PyTorch 跑上 TPU 的项目。两条线索都指向同一件事："
+    "工具链的边界在移动，真正决定成败的不是模型榜单，而是谁能把验收口径写清楚。"
+    "先说游戏那条，作者的说法很朴素，他不写引擎，只把需求拆成可以验证的小步，让智能体一步一步改，"
+    "每一步都跑一次构建，坏了自己回滚。这种做法听起来笨，但它把不确定性压到了单步之内，"
+    "出问题的时候能立刻定位到底层接口还是生成逻辑。我更在意的是这套流程对工程团队的启示："
+    "当生成成本降到几乎为零，稀缺的就不是代码，而是判断哪一步值得交给机器、哪一步必须留给人验。"
+)
+DEFAULT_REF = os.path.expanduser("~/.config/kk/voices/robin-news.wav")
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description="Render one long-form voxcpm sample for echo triage.")
+    parser.add_argument("--text-file", default=None, help="UTF-8 text to render (default: built-in paragraph)")
+    parser.add_argument("--ref", default=DEFAULT_REF, help="reference clip (default: %(default)s)")
+    parser.add_argument("--out", default="/tmp/voxcpm-probe", help="output directory (default: %(default)s)")
+    parser.add_argument("--name", default="long-baseline", help="output file stem (default: %(default)s)")
+    parser.add_argument("--steps", type=int, default=10, help="inference_timesteps (default 10, the plugin's)")
+    parser.add_argument("--cfg", type=float, default=2.0, help="cfg_value (default 2.0, the plugin's)")
+    parser.add_argument("--normalize", action="store_true", help="pass normalize=True")
+    parser.add_argument(
+        "--mode",
+        choices=["reference", "continuation", "instruct"],
+        default="reference",
+        help="reference cloning (VoxCPM2 only), prompt continuation, or instruct-only (default reference)",
+    )
+    parser.add_argument("--prompt-text", default=None, help="transcript of --ref, required by --mode continuation")
+    parser.add_argument(
+        "--denoise",
+        action="store_true",
+        help="load the ZipEnhancer denoiser and pass denoise=True (the plugin ships load_denoiser=False)",
+    )
+    parser.add_argument("--model", default="openbmb/VoxCPM2", help="hub id or local snapshot")
+    args = parser.parse_args()
+
+    text = open(args.text_file, encoding="utf-8").read().strip() if args.text_file else DEFAULT_TEXT
+    if not text:
+        print("probe: empty text", file=sys.stderr)
+        return 2
+    if not os.path.isfile(args.ref):
+        print(f"probe: reference clip not found: {args.ref}", file=sys.stderr)
+        return 2
+    os.makedirs(args.out, exist_ok=True)
+
+    # The model prints its own progress on stdout; keep this script's stdout for the summary.
+    with contextlib.redirect_stdout(sys.stderr):
+        import numpy as np
+        import voxcpm
+
+        model = voxcpm.VoxCPM.from_pretrained(
+            args.model, load_denoiser=args.denoise, optimize=False, device=None
+        )
+    sample_rate = int(model.tts_model.sample_rate)
+
+    if args.mode == "continuation" and not args.prompt_text:
+        print("probe: --mode continuation needs --prompt-text (the transcript of --ref)", file=sys.stderr)
+        return 2
+    # Only VoxCPM2 accepts `reference_wav_path`; every other checkpoint clones through a prompt pair.
+    anchoring = {
+        "reference": {"reference_wav_path": args.ref},
+        "continuation": {"prompt_wav_path": args.ref, "prompt_text": args.prompt_text or ""},
+        "instruct": {},
+    }[args.mode]
+
+    started = time.time()
+    with contextlib.redirect_stdout(sys.stderr):
+        waveform = model.generate(
+            text=text,
+            cfg_value=args.cfg,
+            inference_timesteps=args.steps,
+            normalize=args.normalize,
+            denoise=args.denoise,
+            **anchoring,
+        )
+    samples = np.clip(np.asarray(waveform).squeeze(), -1.0, 1.0)
+    path = os.path.join(args.out, f"{args.name}.wav")
+    with wave.open(path, "wb") as handle:
+        handle.setnchannels(1)
+        handle.setsampwidth(2)
+        handle.setframerate(sample_rate)
+        handle.writeframes((samples * 32767).astype(np.int16).tobytes())
+
+    seconds = len(samples) / sample_rate
+    print(
+        f"{path}  {seconds:.1f}s  sr={sample_rate}  chars={len(text)}  "
+        f"mode={args.mode} steps={args.steps} cfg={args.cfg} normalize={args.normalize} denoise={args.denoise}  render={time.time() - started:.1f}s"
+    )
+    print("measure it with: kk stage audio-quality " + path, file=sys.stderr)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

package/plugins/kk/plugin.json

@@ -1,5 +1,5 @@
 {
     "name": "kk",
-    "version": "0.1.5",
+    "version": "0.1.7",
     "description": "knowledge-kit Claude Code plugin: skills, commands, subagents, hooks, and rules"
 }