/wiki-news

Search for latest news related to the LLM Wiki's key topics.

Usage: `/wiki-news [cluster|keyword | --gap [<slug>] [--no-filter]]` — argument optional

**If `$ARGUMENTS` is empty**: show the usage below, then proceed with an all-cluster search (do not stop).

Usage: /wiki-news [cluster|keyword | --gap [<slug>] [--no-filter]]

Examples:
  /wiki-news                          # search all clusters (default)
  /wiki-news open-source-ai-definition  # a specific cluster (slug)
  /wiki-news open weights             # a specific cluster (name)
  /wiki-news DeepSeek open model      # free keyword
  /wiki-news --gap                    # gap fill: deterministic hub crawl → _inbox.md (Track A)
  /wiki-news --gap single-source      # enrich a single gap type
  /wiki-news --gap --no-filter        # disable the domain filter

The `--gap` slug is Track A only: `single-source` · `stale-hub`.

Traversal Pattern

Reading domain — the Reporter does an external search per cluster via parallel spawn, then (optionally) an ingest chain. It runs independent cluster queries concurrently in breadth-first parallel mode.

| Phase | Cycle | Owner | |---|---|---| | Read wiki context | GROUND (own context) | Editor-in-Chief (read overview·index·`graph/_clusters.json`) | | Build cluster queries | — | Editor-in-Chief | | External WebSearch | GROUND breadth-first | Reporter multi-spawn (1 instance per cluster) | | Duplicate check | — | Editor-in-Chief (`_source_map.json` matching) | | Write report | — | Editor-in-Chief | | (optional) inbox queue append | — | Editor-in-Chief (`_inbox.md`) — fetch·ingest is the separate `/wiki-ingest inbox` |

Cluster slug list SoT: `graph/_clusters.json::clusters[].slug` (single SoT, varies with ingest accumulation).

Procedure (Editor-in-Chief Orchestration)

1. **Read wiki context** — `wiki/overview.md` (key themes·entities), `wiki/index.md` (cluster breakdown), `graph/_clusters.json` (clusters[].members·top_tags·source_assignments) 2. **Build search queries**:

• If a cluster argument is given — match slug/name against `_clusters.json::clusters[]`, extract 3-5 member hubs with clear labels, combine the hub names (e.g. `open-source-ai-definition` → "OpenSourceInitiative Meta open weights AI definition 2026")
• If no argument — the top 5 clusters by `_clusters.json` size, combining 2-3 hubs each
• Add specific event·person keywords from the overview.md narrative (e.g. "DeepSeek open model release")
• Append the current year for recency
• Queries default to English; build Korean queries (for Korean news sources) only under `WIKI_LANG=ko`

3. **Reporter parallel spawn** — perform WebSearch per cluster (independent cluster queries run concurrently) 4. **Duplicate check** — match each result's title·URL against `wiki/sources/_source_map.json` (or grep `wiki/sources/` for similar keywords) 5. **Write report** — markdown report of category·title·source·date·ingest status (new / already ingested) 6. **Recommend top articles** — suggest ingest + wait for the user's decision 7. **After the user's decision** — append the chosen article URLs to `_inbox.md` as `# source=interactive` (fetch·ingest is the separate `/wiki-ingest inbox`), or save the report to `wiki/syntheses/`

Gap Mode (`--gap [<slug>] [--no-filter]`)

Every Track A enrichment is hub-level: a **deterministic crawl** (the Editor-in-Chief calls `crawl.py` internally) follows the adjacent pages cited by existing sources into `_inbox.md`, and ingest is delegated to `/wiki-ingest inbox`. The cluster-level `sparse-cluster` gap was removed (rationale in the `NON_BACKLOG_BUCKETS` comment in [`tools/_lint/graph_gaps.py`](../../tools/_lint/graph_gaps.py)), and with it the automatic WebSearch path — operator-run WebSearch reinforcement stays valid through `tools/_news/gap_queries.py` (procedure: [`operations/gap-detection-rollout.md`](../operations/gap-detection-rollout.md) § Enrichment channels).

Definitions, thresholds, and domain set have their SoT in [`.claude/operations/gap-detection-rollout.md`](../operations/gap-detection-rollout.md) + [`tools/_news/domains.py`](../../tools/_news/domains.py). Crawl seed derivation, relevance lexicon, and cap have their SoT in [`tools/_news/crawl.py`](../../tools/_news/crawl.py).

Procedure

1. Call `python tools/_news/crawl.py --gap-seed --append-inbox [--gap-type <slug>] [--no-filter]` internally. **Do not call `lint graph gaps` separately** — `crawl.py` runs that diagnosis inside itself. Seeds are auto-derived hub→backlinks→`source_url`; the adjacent pages cited by each hub's existing sources are appended to `_inbox.md` as `source=auto-crawl` after passing the domain allowlist·`by_url` dedup check (no external search needed — the seeds are already trusted sources). 2. Report the crawl hub/candidate counts + `_inbox.md` queue length. Ingest is the explicit `/wiki-ingest inbox`.

Hard Cap

| Item | Value | |---|---| | `--gap-limit` (hubs per gap type) | 5 | | (crawl) `--max-pages`·`--max-depth`·`--per-domain-cap`·`--min-score` | `tools/_news/crawl.py` defaults SoT | | `_inbox.md` queue length alarm | 30 |

Fetch·Ingest Delegation

This command is responsible only up to WebSearch·report·`_inbox.md` append. The fetch (2-stage: deterministic + WebFetch fallback)·raw storage·source authoring are owned solely by [`wiki-ingest.md`](wiki-ingest.md) Inbox Mode — appending the chosen article URLs to `_inbox.md` lets `/wiki-ingest inbox` carry out everything from fetch to ingest (because the raw is the dedup ground truth, source pages are created only after the raw lands).

Human Reviewer Gate

The global gate is SoT at [`CLAUDE.md` "Human Reviewer Gate"](../../CLAUDE.md#human-reviewer-gate). Gates specific to this command:

• Ingest decision (chain into `/wiki-ingest` — explicit approval)
• **`--gap` mode** — no per-query gate: the automatic channel is a link crawl with no queries to approve. Ingest is still the explicit `/wiki-ingest inbox`