fix: batch summary writes to avoid output truncation on large harvests

When a harvest produced 15+ articles, asking the agent to emit the entire
.summaries.json in a single response hit the model's output-token ceiling —
the response was truncated and the tail articles silently got no summary
(their 原文 kept the __SUMMARY_<slug>__ sentinel forever). Reproduced every
time a high-volume source ran; ≤9-article sources never hit it.

Add scripts/summary-add.js to incrementally merge small batches (≤4) into
.summaries.json (keyed by url, overwrite-no-dup), with a --status mode that
diffs manifest vs summaries so interrupted runs can resume. Rewrite SKILL.md
Flow A Step 2 + Guardrails to mandate batching, and surface a batchHint in
harvest.js's manifest so the agent sees the guidance on first contact.

SKILL.md

@@ -15,6 +15,8 @@ Fetch latest articles from configured news sources, archive them to Obsidian vau
 | Preview latest (read-only) | `node scripts/harvest.js <source_id> <count> --preview` |
 | Preview with full body | `node scripts/harvest.js <source_id> <count> --preview --full` |
 | Finalize Chinese summaries | `node scripts/finalize.js "<dayDir>"` |
+| Merge a summary batch | `node scripts/summary-add.js "<dayDir>"` |
+| Check summary coverage | `node scripts/summary-add.js "<dayDir>" --status` |
 | Add source | `node scripts/source-manage.js add` |
 | List sources | `node scripts/source-manage.js list` |
 | Edit source | `node scripts/source-manage.js edit <id>` |
@@ -235,19 +237,20 @@ It prints a **compact manifest** to stdout (no full body — only a `bodyPreview
 
 > **Browser sources and any source with a recipe** are auto-handled: before fetching, `harvest.js` calls `ensure-chromium.js`, which health-checks the CDP port and launches Chromium itself (cross-platform binary discovery, headless by default) if it isn't up — a no-op if already running. No manual `chromium ... &` or `sleep` is needed. To debug launch failures: `node scripts/ensure-chromium.js` (prints binary path, port, and stderr-log tail on failure).
 
-#### Step 2. Generate Chinese summaries → write `.summaries.json`
-From the compact manifest, for each article write one entry to `<dayDir>/.summaries.json`:
+#### Step 2. Generate Chinese summaries → write `.summaries.json` (batched)
+
+**Why batch:** writing all summaries in a *single* response hits the model's output-token ceiling once you have ~10+ articles — the response is truncated and the remaining articles silently get no summary (their 原文 keeps the `__SUMMARY_<slug>__` sentinel forever). This failure is **not rare** — it reproduces every time a source yields many articles. Avoid it by working in small batches that each get merged into `.summaries.json` by `summary-add.js`.
+
+Each summary entry has this shape:
 
 ```json
-[
-  {
-    "url": "https://theconversation.com/...-286430",
-    "titleZh": "特朗普新规或拖慢联邦科研",
-    "category": "科技/AI",
-    "tags": ["科技/AI", "The Conversation"],
-    "summaryZh": "## 事件概述\n...\n\n## 背景\n...\n\n## 关键引语\n> \"...\"\n\n## 影响分析\n..."
-  }
-]
+{
+  "url": "https://theconversation.com/...-286430",
+  "titleZh": "特朗普新规或拖慢联邦科研",
+  "category": "科技/AI",
+  "tags": ["科技/AI", "The Conversation"],
+  "summaryZh": "## 事件概述\n...\n\n## 背景\n...\n\n## 关键引语\n> \"...\"\n\n## 影响分析\n..."
+}
 ```
 
 - `titleZh`: concise Chinese title (≤30 chars)
@@ -255,6 +258,23 @@ From the compact manifest, for each article write one entry to `<dayDir>/.summar
 - `summaryZh`: the 4-section markdown body (事件概述 / 背景 / 关键引语 / 影响分析)
 - The `bodyPreview` in the manifest is enough to write the summary. **Do not re-read the full 原文.md** — if you need more, read a targeted slice.
 
+**Procedure** (loop until every article in the manifest has a summary):
+
+1. Let `N` = number of articles in the manifest. If `N ≤ 8`, you may write all entries to `.summaries.json` directly (one Write) and skip to Step 3. Otherwise work in batches of **≤4**.
+2. For each batch: use `Write` to create `<dayDir>/.summaries-batch.json` — a JSON array of ≤4 summary entries — then merge it:
+   ```bash
+   node scripts/summary-add.js "<dayDir>"
+   ```
+   This merges the batch into `.summaries.json` (keyed by `url`, so re-supplying an entry overwrites it — no duplicates), then deletes the batch file. Repeat per batch.
+3. Once all batches are done, verify coverage and fill any gaps:
+   ```bash
+   node scripts/summary-add.js "<dayDir>" --status
+   ```
+   Lists done vs. missing against `.manifest.json`. Write any missing articles as another batch and merge. When `--status` reports `missing: 0`, proceed to Step 3.
+
+> **Interrupted run?** Do not re-run `harvest.js` (it dedups via the registry, so it won't re-fetch). Instead run `node scripts/summary-add.js "<dayDir>" --status` to see which articles still need summaries, then resume from there.
+
+
 #### Step 3. Run finalize (script stitches summaries in)
 ```bash
 node scripts/finalize.js "<dayDir>"
@@ -276,6 +296,7 @@ No files are written. stdout is a compact JSON list (one call, no re-fetching /
 
 - **`count` is the target number of NEW articles**, not raw links. `harvest.js` already accounts for registry dedup: it scrolls to collect extra candidate links (`count + seen + buffer`, capped) so the *new* count can still reach `count` even when most of the homepage was already archived. If the run returns **fewer than `count`**, the stderr prints `⚠️ … stream may be exhausted` — that's a genuine source limit (no fresh articles to fetch), **not** a bug; re-running won't help. Wait for new articles or broaden the source instead. Don't re-run `harvest.js` for the same source/day just to "get more" — it deduplicates via the registry, so a second run only fetches genuinely new articles (and merges them into the pending `.manifest.json`, which is safe).
 - **Never `Read` the 原文.md files** to write summaries — the `bodyPreview` in the manifest is sufficient. Targeted slice reads are OK if truly needed.
+- **Summaries in batches of ≤4, never one big response.** A single response holding 10+ summaries hits the output-token ceiling and is truncated — the tail articles silently get no summary and keep the `__SUMMARY_<slug>__` sentinel. Write each batch to `.summaries-batch.json` and merge via `node scripts/summary-add.js "<dayDir>"`; repeat per batch. `summary-add.js --status` shows what's left. This is mandatory for `N > 8`; recommended for any multi-article run.
 - **One `finalize.js` call** after writing `.summaries.json`. Do not hand-write file I/O — the scripts handle all original/registry/index/summary writes.
 
 ---

scripts/harvest.js

@@ -626,7 +626,11 @@ Examples:
     nextStep: `Read this manifest, write Chinese summaries to ${path.join(dayDir, '.summaries.json')}, then run: node scripts/finalize.js "${dayDir}"`,
     summariesContract: {
       path: path.join(dayDir, '.summaries.json'),
-      format: 'array of { url, titleZh, category, tags[], summaryZh } — one per article url; summaryZh = the 4 sections markdown (事件概述/背景/关键引语/影响分析)'
+      format: 'array of { url, titleZh, category, tags[], summaryZh } — one per article url; summaryZh = the 4 sections markdown (事件概述/背景/关键引语/影响分析)',
+      batchHint: `If ${mergedArticles.length} articles is more than ~8, do NOT write all summaries in one response — ` +
+        `it will hit the output-token ceiling and get truncated. Work in batches of ≤4: write each batch to ` +
+        `${path.join(dayDir, '.summaries-batch.json')} and merge with "node scripts/summary-add.js \\"${dayDir}\\"" ` +
+        `(repeat per batch), then run finalize.js once. Check gaps with "node scripts/summary-add.js \\"${dayDir}\\" --status".`
     },
     articles: mergedArticles
   };