wiki-review-queue

name	wiki-review-queue
description	Review candidate queue for wiki-pipeline Phase 1B gate — stores scored article candidates in ~/.wiki-review-queue.json with content dedup, TTL expiry, and reject cache. CLI: review-queue.mjs.
version	2.0.0
author	Hermes Agent
license	MIT
metadata	{"hermes":{"tags":["wiki","review-queue","pipeline","candidates","phase1b"],"category":"wiki","related_skills":["wiki-pipeline","web-content-reviewer","llm-wiki"]}}

Wiki Review Queue v2

Persistent review candidate queue for the wiki-pipeline Phase 1B. When compile --review (or pipeline --review mode) runs, scored articles are added as JSON candidates to ~/.wiki-review-queue.json instead of directly ingesting into wiki/.

Reviewers then approve or reject each candidate via the CLI.

Architecture

wiki-pipeline --review
    │
    ├── Phase 1 (web-content-reviewer): score article
    │       │
    │       └── score ≥ 49 ──► add to ~/.wiki-review-queue.json
    │       │                         │
    │       └── score < 49 ──► reject immediately (no candidate)
    │       │
    └────────────── Phase 2 deferred ──┘
                                  │
                          user reviews candidates
                                  │
                    ┌─────────────┴─────────────┐
                    │                           │
              review approve                  review reject
                    │                           │
        write wiki/ + refresh index         reject cache
        + flush source state               (content_hash stored)
        + register concept deps            candidate removed

Candidate JSON Schema

Each candidate is stored in ~/.wiki-review-queue.json:

{
  "candidates": [
    {
      "id": "k3f9a2b1",
      "content_hash": "a8da29aa0bfa06cd3c6bb588a6703dd8...",
      "file": "raw/rss-inbox/interconnects-article.md",
      "review_value": 8,
      "review_confidence": 7,
      "source_url": "https://...",
      "sources": ["raw/rss-inbox/interconnects-article.md"],
      "strategic_context": "frontier-3",
      "contradicted_by": "",
      "approved": undefined,
      "created_at": "2026-05-18T10:00:00.000Z",
      "reviewed_at": undefined
    }
  ]
}

v2 fields:

• content_hash: body-only SHA-256 — same content never re-queued
• approved: true | false | undefined (pending)
• created_at / reviewed_at: ISO timestamps for TTL tracking

CLI Commands

All commands are executed via node review-queue.mjs:

--list

List all candidates (auto-purges expired entries before display).

node review-queue.mjs --list

Output:

  Wiki Review Queue — 3 candidate(s)

  k3f9a2b1  score=56  ⏳ pending  2026/5/18 [→ frontier-3]
         file=raw/rss-inbox/...  url=https://...

--add

Add a candidate (with automatic dedup and reject cache check):

node review-queue.mjs --add \
  --file raw/rss-inbox/interconnects-article.md \
  --value 8 --confidence 7 \
  --url https://interconnects.ai/... \
  --sources raw/rss-inbox/interconnects-article.md \
  --strategic-context frontier-3

Dedup: If content_hash already exists in pending queue → skip with log.
Reject cache: If content_hash was previously rejected → skip with log.

--approve <id>

Approve a candidate and remove from queue:

node review-queue.mjs --approve k3f9a2b1

--reject <id>

Reject a candidate, cache its content_hash to prevent future re-add:

node review-queue.mjs --reject k3f9a2b1

--pending

Show pending count (auto-purges expired):

node review-queue.mjs --pending

--purge-stale

Force-expire entries older than TTL (default 14d):

node review-queue.mjs --purge-stale

--show-rejects

List rejected content hashes:

node review-queue.mjs --show-rejects

--clear-rejects

Clear the reject cache:

node review-queue.mjs --clear-rejects

--reset

Clear the entire queue (use with caution):

node review-queue.mjs --reset

Low v×c Scanner Integration

The low-vxc-scanner.py script (~/wiki/scripts/low-vxc-scanner.py) identifies entities with low review scores and can enroll them into the review queue:

# Scan only (report)
python3 scripts/low-vxc-scanner.py --threshold 49 --limit 30

# Scan + enroll into review queue
python3 scripts/low-vxc-scanner.py --threshold 49 --limit 30 --enroll

# Only entities missing scores
python3 scripts/low-vxc-scanner.py --missing-only

# Output JSON for further processing
python3 scripts/low-vxc-scanner.py --threshold 49 --json /tmp/low-vxc.json

Cron: wiki-low-vxc-review runs weekly Monday 8am with --threshold 49 --limit 30 (report only, no auto-enroll).

Pitfall: The review_value field may contain stray characters (e.g., 7 ^[raw/articles/xxx.md]). The scanner handles this by splitting on whitespace and ^ before parsing to float.

Phase 1B: Review Gate Integration

The review queue integrates with wiki-pipeline as Phase 1B — an optional sub-mode of Phase 1:

Phase 1 (scoring)
    │
    ├── score ≥ 49 ──► Phase 1B decision gate:
    │       │
    │       ├── pipeline mode = default  ──► immediate Phase 2 (standard flow)
    │       │
    │       └── pipeline mode = review   ──► review-queue.mjs --add, skip Phase 2
    │
    └── score < 49 ──► Phase 1 rejection closeout (unchanged)

Trigger: When user triggers pipeline with --review flag, or when:

• Batch scoring produces borderline articles (score 45-55) where human judgment adds value
• Articles that report contradictedBy in LLM extraction — auto-route to review queue

Reject Cache

Rejected candidates store their content_hash in ~/.wiki-review-rejects.json:

{
  "hashes": [
    {
      "hash": "a8da29aa0bfa06cd3c6bb588a6703dd8...",
      "file": "raw/rss-inbox/spam-article.md",
      "rejected_at": "2026-05-18T10:00:00.000Z"
    }
  ]
}

Auto-cleanup: Reject cache entries expire after 90 days. Expired entries are cleaned on --add operations.

TTL: Auto-Expiry

Queue entries have a 14-day TTL. When --list, --pending, or --purge-stale runs, entries older than 14 days with approved === undefined are automatically rejected and removed.

Expired entries are logged to stderr for observability:

[review-queue] Auto-expired candidate k3f9a2b1: ... (created 2026-05-04, >14d)

Integration with SHA-256 Incremental Check

The review queue uses the same body-only SHA-256 algorithm as wiki-source-hasher.mjs (sha256-body-v1). When a candidate is approved and processed through Phase 2:

1. Source file gets sha256 field written by --write-sha256
2. Concept dependencies registered via registerConcepts() (exported from hasher)
3. Future inbox scans skip unchanged sources

Low v×c Scanner Integration

~/wiki/scripts/low-vxc-scanner.py automates the discovery of low-quality entities for re-review. It scans all entities with v×c < threshold (default 49) and optionally enrolls them via review-queue.mjs --add.

# Scan only
python3 scripts/low-vxc-scanner.py --threshold 49 --limit 30

# Scan + enroll into review queue
python3 scripts/low-vxc-scanner.py --threshold 49 --enroll --limit 30

Pitfall — malformed frontmatter: Some entities have review_value: 7 ^[raw/articles/slug.md] (citation garbage appended). The scanner uses robust safe_float() extraction: float(str(field).split()[0].split('^')[0].strip()).

Cron: wiki-low-vxc-review runs weekly Mon 8am (job ID: 250040ab1a4c) to scan and report.

Lock & Concurrency

Both ~/.wiki-review-queue.json and ~/.wiki-review-rejects.json are written atomically (temp file + rename) to prevent corruption during concurrent access.

Reference

• references/review-queue-workflow.md — detailed CLI workflow and troubleshooting
• scripts/review-queue.mjs — actual CLI implementation