Redbook

Redbook — Xiaohongshu CLI

Use the redbook CLI to search notes, read content, analyze creators, automate engagement, and research topics on Xiaohongshu (小红书/RED).

OpenClaw users: Install via clawhub install redbook or npm install -g @lucasygu/redbook.

⚠️ Research discipline (read this first). XHS 风控 throttles reading, not just writing. Any research that reads more than a handful of notes MUST run through the Research Loop — human-paced by default (~20 s/note, one at a time; a typical job finishes in tens of minutes, large ones spread across the day). Hammering read / comments / analyze-viral in a tight loop trips captcha / IP block (300012) within a few dozen hits and degrades the account for hours. ⚡ Fast mode is emergency-only and requires a printed warning + explicit user opt-in. Never fire reads in parallel or with zero delay.

Usage

/redbook search "AI编程"              # Search notes
/redbook read <url>                   # Read a note
/redbook user <profileUrl>            # Creator profile
/redbook account-report --file ids.txt --json  # Batch account metrics
/redbook analyze <userId>             # Full creator analysis (profile + posts)

Quick Reference

Always add --json when parsing output programmatically. Without it, output is human-formatted text.

XHS Platform Signals

XHS is not Twitter or Instagram. These platform-specific engagement ratios reveal content type and audience behavior.

Collect/Like Ratio (collected_count / liked_count)

XHS's "collect" (收藏) is a save-for-later mechanic — users build personal reference libraries. This ratio is the strongest signal of content utility.

Comment/Like Ratio (comment_count / liked_count)

Measures how much a note triggers conversation.

Share/Like Ratio (share_count / liked_count)

Measures social currency — whether users share to signal identity or help others.

Search Sort Semantics

Content Form Dynamics

Analysis Modules

Each module is a composable building block. Combine them for different analysis depths.

Module A: Keyword Engagement Matrix

Answers: Which keywords have the highest engagement ceiling? Which are saturated vs. underserved?

Commands:

redbook search "keyword1" --sort popular --json
redbook search "keyword2" --sort popular --json
# Repeat for each keyword in your list

Fields to extract from each result's items[]:

• items[].note_card.interact_info.liked_count — likes (may use Chinese numbers: "1.5万" = 15,000)
• items[].note_card.interact_info.collected_count — collects
• items[].note_card.interact_info.comment_count — comments
• items[].note_card.user.nickname — author

How to interpret:

• Top1 ceiling = items[0] likes — the best-performing note for this keyword. This is the proven demand signal.
• Top10 average = mean likes across items[0..9] — how well an average top note does.
• A high Top1 but low Top10 avg means one outlier dominates; hard to compete.
• A high Top10 avg means consistent demand; easier to break in.

Output: Keyword × engagement table ranked by Top1 ceiling.

Module B: Cross-Topic Heatmap

Answers: Which topic × scene intersections have demand? Where are the content gaps?

Commands:

# Combine base topic with scene/angle keywords
redbook search "base topic + scene1" --sort popular --json
redbook search "base topic + scene2" --sort popular --json
redbook search "base topic + scene3" --sort popular --json

Fields to extract: Same as Module A — Top1 liked_count for each combination.

How to interpret:

• High Top1 = proven demand for this intersection
• Zero or very low results = content gap (opportunity or no demand — check if the combination makes sense)
• Compare across scenes to find which angles resonate most with the base topic

Output: Base × Scene heatmap.

             scene1    scene2    scene3    scene4
base topic   ████ 8K   ██ 2K     ████ 12K  ░░ 200

Module C: Engagement Signal Analysis

Answers: What type of content is each keyword? Reference, insight, or entertainment?

Commands: Use search results from Module A, or for a single note:

redbook analyze-viral "<noteUrl>" --json

Fields to extract:

• From search results: compute ratios from interact_info fields
• From analyze-viral: use pre-computed engagement.collectToLikeRatio, engagement.commentToLikeRatio, engagement.shareToLikeRatio

How to interpret: Apply the ratio benchmarks from XHS Platform Signals above.

Output: Per-keyword or per-note classification.

Module D: Creator Discovery & Profiling

Answers: Who are the key creators in this niche? What are their strategies?

Commands:

# 1. Collect unique user_ids from search results across keywords
#    Extract from items[].note_card.user.user_id

# 2. For each creator:
redbook user "<userId>" --json
redbook user-posts "<userId>" --json

Fields to extract:

• From user: interactions[] where type === "fans" → follower count
• From user-posts: notes[].interact_info.liked_count for all posts → compute avg, median, max
• From user-posts: notes[].display_title → content patterns, posting frequency

How to interpret:

• Avg vs. Median likes: Large gap means viral outliers inflate the average. Median is the "true" baseline.
• Max / Median ratio: >5× means they've had breakout hits. Study those notes specifically.
• Post frequency: Count notes to estimate posting cadence. Prolific creators (>3/week) vs. quality-focused (<1/week).

Output: Creator comparison table.

Module D2: Known Account Reporting

Answers: Given a known list of account IDs/profile URLs, how often did each account post and how did each post perform?

Use this when the user already has KOS/KOC account IDs and wants clean account-scoped metrics. Do not approximate this with keyword search; keyword search introduces unrelated posts and noisy attribution.

redbook account-report --file accounts.txt --month 2026-07 --json
redbook account-report "<profileUrl1>" "<profileUrl2>" --max-pages 2 --json

Input files are newline-separated user IDs or profile URLs; blank lines and # comments are ignored. Prefer profile URLs containing xsec_token when available. JSON output includes account summary fields (fetchedPosts, postsInMonth, complete, pagesFetched), engagement totals and averages, top posts, and per-note rows with title, type, publish time, metrics, and webUrl.

Default mode fetches only the first page per account. complete: false means more pages exist. Use --all only when the user explicitly wants full history and the account list is small enough to pace safely; otherwise prefer --max-pages <n>.

Module E: Content Form Breakdown

Answers: Do image-text or video notes perform better for this topic?

Commands:

redbook search "keyword" --type image --sort popular --json
redbook search "keyword" --type video --sort popular --json

Fields to extract:

• Compare Top1 and Top10 avg liked_count and collected_count between the two result sets
• Note the type field: "normal" = image-text, "video" = video

Output: Form × engagement table.

Module F: Opportunity Scoring

Answers: Which keywords should I target? Where is the best effort-to-reward ratio?

Input: Keyword matrix from Module A.

Scoring logic:

• Demand = Top1 likes ceiling (proven audience size)
• Competition = density of high-engagement results (how many notes in Top10 have >1K likes)
• Score = Demand × (1 / Competition density)

Tier thresholds (based on Top1 likes):

Output: Tiered keyword list.

Module G: Audience Inference

Answers: Who is the audience for this niche? What do they want?

Input: Engagement ratios from Module C + comment themes from analyze-viral + content patterns.

Fields to extract from analyze-viral JSON:

• comments.themes[] — recurring phrases and keywords from comment section
• comments.questionRate — % of comments that are questions (learning intent)
• engagement.collectToLikeRatio — save behavior signals intent
• hook.hookPatterns[] — what title patterns attract this audience

Inference rules:

• High collect rate + high question rate → learning-oriented audience (students, professionals)
• High comment rate + emotional themes → community-oriented audience (sharing experiences)
• High share rate → aspiration-oriented audience (lifestyle, identity signaling)
• Comment language patterns → age/education signals (formal = older, slang = younger)

Output: Audience persona summary — demographics, intent, content preferences.

Module H: Content Brainstorm

Answers: What specific content should I create, backed by data?

Input: Opportunity scores (Module F) + audience persona (Module G) + heatmap gaps (Module B).

For each content idea, specify:

• Target keyword — from opportunity scoring
• Hook angle — based on hookPatterns that work for this niche
• Content type — 工具型/认知型/娱乐型 based on what the audience wants
• Form — 图文 or 视频 based on Module E
• Engagement target — realistic based on Top10 avg for this keyword
• Competitive reference — specific note URL that proves this angle works

Output: Ranked content ideas with data backing.

Module I: Comment Operations

Answers: Which comments deserve a reply? What is the comment quality distribution?

Commands:

# 1. Fetch all comments
redbook comments "<noteUrl>" --all --json

# 2. Preview reply candidates (dry run)
redbook batch-reply "<noteUrl>" --strategy questions --dry-run --json

# 3. Execute replies with template (5 min delay with ±30% jitter)
redbook batch-reply "<noteUrl>" --strategy questions \
  --template "感谢提问！关于{content}，..." \
  --max 10

Fields to extract from --dry-run JSON:

• candidates[].commentId — target comments
• candidates[].isQuestion — boolean, detected question
• candidates[].likes — engagement signal
• candidates[].hasSubReplies — whether already answered
• skipped — how many comments were filtered out
• totalComments — total fetched

Strategies:

• questions — replies to comments ending with ？ or ? (learning-oriented audience)
• top-engaged — replies to highest-liked comments (maximum visibility)
• all-unanswered — replies to comments with no existing sub-replies (fill gaps)

How to interpret:

• High question rate (>15%) = audience is learning-oriented → reply to build authority
• High top-engaged comments (>100 likes) = reply to visible comments for maximum reach
• Many unanswered comments = engagement gap, opportunity to increase reply rate

Safety: Hard cap 30 replies per batch, minimum 3-minute delay with ±30% jitter (default 5 min), --dry-run by default (no template = preview only), immediate stop on captcha. See Rate Limits & Safety for details.

Output: Reply plan table with candidate comments, strategy match reason, and status.

Module J: Viral Replication

Answers: What structural template can I extract from successful notes to guide new content creation?

Commands:

# 1. Find top notes for a keyword
redbook search "keyword" --sort popular --json

# 2. Extract structural template from 2-3 top performers
redbook viral-template "<url1>" "<url2>" "<url3>" --json

Fields to extract from viral-template JSON:

• dominantHookPatterns[] — hook types appearing in majority of notes
• titleStructure.commonPatterns[] — specific title formula
• titleStructure.avgLength — target title length
• bodyStructure.lengthRange — target word count [min, max]
• bodyStructure.paragraphRange — target paragraph count
• engagementProfile.type — reference/insight/entertainment
• audienceSignals.commonThemes[] — what the audience talks about

How to interpret:

• Consistent hook patterns across notes = proven formula for this niche
• Narrow body length range = audience has clear content length preference
• High collect/like in profile = audience saves content → create reference material
• Common comment themes = topics to address in new content

Composition with other modules:

• Uses Module A results to identify top URLs for template extraction
• Feeds into Module H (Content Brainstorm) as structural constraint
• Uses Module C classification to validate engagement profile

Output: Content template spec — the structural skeleton for content creation. An LLM (via the composed workflow) uses this template to generate actual title, body, hashtags, and cover image prompt.

Module K: Engagement Automation

Answers: How should I manage ongoing engagement with my audience?

This module is a workflow that composes Modules I and J with human oversight.

Workflow:

1. Monitor — redbook comments "<myNoteUrl>" --all --json to fetch recent comments
2. Filter — redbook batch-reply --strategy questions --dry-run to identify reply candidates
3. Review — Human reviews dry-run output (or LLM reviews with persona guidelines)
4. Execute — redbook batch-reply --strategy questions --template "..." --max 10
5. Report — Summary of replies sent, errors encountered, rate limit status

Safety rules:

• Always --dry-run first, human approval before execution
• Maximum 30 replies per session (hard cap)
• Minimum 3-minute delay between replies, default 5 minutes, with ±30% random jitter
• Never reply to the same comment twice (check hasSubReplies)
• Stop immediately on captcha — do not retry
• See Rate Limits & Safety for XHS risk control thresholds

Anti-spam guidelines:

• Vary reply templates across batches
• Limit to 1-2 batch runs per note per day
• Prioritize quality (targeted strategy) over quantity
• Uniform timing patterns trigger bot detection — jitter is applied automatically

Module L: Card Rendering

Answers: How do I turn markdown content into Xiaohongshu-ready image cards?

Commands:

# Render markdown to styled PNG cards
redbook render content.md --style xiaohongshu

# Custom style and output directory
redbook render content.md --style dark --output-dir ./cards

# JSON output (for programmatic use)
redbook render content.md --json

Input: Markdown file with YAML frontmatter:

---
emoji: "🚀"
title: "5个AI效率技巧"
subtitle: "Claude Code 实战"
---

## 技巧一：...
Content here...

---

## 技巧二：...
More content...

Output: cover.png + card_1.png, card_2.png, ... in the same directory.

Card specs:

• Size: 1080×1440 (3:4 ratio, standard XHS image)
• DPR: 2 (retina quality, actual output 2160×2880)
• Styles: purple, xiaohongshu, mint, sunset, ocean, elegant, dark

Pagination modes:

• auto (default) — smart split on heading/paragraph boundaries using character-count heuristic
• separator — manual split on --- in markdown

How to interpret:

• Uses the user's existing Chrome for rendering (via puppeteer-core) — no browser download needed
• Purely offline — no XHS API or cookies required
• Output images are ready for redbook post --images cover.png card_1.png ...

Dependencies: Requires puppeteer-core and marked (optional, install with npm install -g puppeteer-core marked).

Composition with other modules:

• Pairs with Module H (Content Brainstorm) — generate content ideas, write markdown, render to cards
• Pairs with Module J (Viral Replication) — extract template, write content matching the template, render
• Output feeds into redbook post --images for publishing

Module M: Note Health Check (限流检测)

Answers: Are any of my notes being secretly rate-limited by XHS?

XHS assigns a hidden level field to each note in the creator backend API. This level controls recommendation distribution but is never shown in the UI. Your note may look "normal" while secretly receiving zero recommendations.

Commands:

# Check all notes (first page)
redbook health

# Check all pages
redbook health --all

# JSON output for programmatic use
redbook health --all --json

Level definitions:

Additional checks:

• Sensitive word detection — flags titles containing automation/AI keywords (自动化, AI生成, 批量, etc.)
• Tag count warning — flags notes with >5 hashtags (risk factor)

How to interpret:

• Level -1 or below = your note is being throttled. Consider editing or deleting + reposting.
• Level -102 = irreversible. Delete the note and create fresh content.
• Sensitive word hits = risky title keywords that may trigger throttling. Rephrase.
• Excessive tags = potential spam signal. Use 3-5 targeted tags.

Output: Terminal dashboard with color-coded distribution summary, limited notes list, and risk factor warnings.

Discovery credit: @xxx111god — xhs-note-health-checker

Composed Workflows

Combine modules for different analysis depths.

Quick Topic Scan (~5 min)

Modules: A → C → F

Search 3–5 keywords, classify engagement type, rank opportunities. Good for quickly validating whether a niche is worth deeper research.

Content Planning

Modules: A → B → E → F → H

Build keyword matrix, map topic × scene intersections, check content form performance, score opportunities, brainstorm specific content ideas.

Creator Competitive Analysis

Modules: A → D

Find who dominates a niche and study their content strategy, posting frequency, and engagement patterns.

Full Niche Analysis

Modules: A → B → C → D → E → F → G → H

The comprehensive playbook — keyword landscape, cross-topic heatmap, engagement signals, creator profiles, content form analysis, opportunity scoring, audience personas, and data-backed content ideas.

Single Note Deep-Dive

Command: redbook analyze-viral "<url>" --json

No module composition needed — analyze-viral returns hook analysis, engagement ratios, comment themes, author baseline comparison, and a 0-100 viral score in one call.

Viral Pattern Research → Content Template

# 1. Find top notes
redbook search "keyword" --sort popular --json

# 2. Extract template from top 3 notes (replaces manual synthesis)
redbook viral-template "<url1>" "<url2>" "<url3>" --json

viral-template automates what previously required manual synthesis across analyze-viral results. It outputs a ContentTemplate JSON that captures dominant hooks, body structure ranges, engagement profile, and audience signals.

Reply Management

Modules: I

Single-module workflow for managing comment engagement on your notes. Use batch-reply --dry-run to audit, then execute with a template.

Content Replication

Modules: A → J → H → L

Keyword research → viral template extraction → data-backed content brainstorm → render to image cards. The template provides structural constraints that guide Module H's content ideas. Module L renders the final markdown to XHS-ready PNGs.

Content Creation End-to-End

Modules: A → J → H → L → post

The full pipeline: research keywords → extract viral template → brainstorm content → write markdown → render to styled image cards → publish via redbook post --images cover.png card_1.png ...

Account Health Monitoring

Modules: M

Run redbook health --all periodically to catch throttled notes early. If level drops below 1, investigate the note's content for policy violations. Combine with Module I to check if throttled notes still have unanswered comments worth replying to.

Full Operations

Modules: A → C → I → J → K → M

Comprehensive automation playbook — keyword analysis, engagement classification, comment operations, viral replication templates, and engagement automation workflow.

Rate Limits & Safety

XHS enforces aggressive anti-spam (风控) that detects automated behavior through device fingerprinting, activity ratio monitoring, and timing pattern analysis. The CLI applies safe defaults based on platform research.

Safe Thresholds

Anti-Detection Measures

• Timing jitter: ±30% random variation on all batch delays. Uniform intervals are a bot signature.
• Hard caps: Maximum 30 replies per batch (down from 50). No override.
• Rate limit warnings: post, comment, and reply commands display safe interval reminders after each action.
• Captcha circuit breaker: Batch operations stop immediately on captcha (NeedVerify).

What Triggers Risk Control

• Uniform timing — replying at exact 3-second intervals flags bot detection
• High frequency — >50 interactions/minute across any action type
• Activity ratio anomaly — more comments than post views signals inauthentic behavior
• Device fingerprint mismatch — XHS fingerprints 21 hardware parameters

Best Practices for Agents

1. Always --dry-run first, review candidates, then execute
2. Use the default 5-minute delay — do not override --delay below 180000 (3 min)
3. Limit batch runs to 1-2 per note per day
4. Vary reply templates between batches
5. Space post commands 3-4 hours apart (2-3 notes/day maximum)

Research Loop — Rate-Limit-Safe Long-Running Research

The Rate Limits & Safety section above paces writes. This section paces reads — the part research actually hammers. read, comments, and analyze-viral each hit the note-detail API (with xsec_token); fire 30–50 in a tight loop and XHS returns NeedVerify / captcha or IP-blocks you (300012), and a tripped account stays degraded for hours. So research here is paced, not parallel: one note in flight, spaced to look like a human reading rather than a script scraping.

The mechanism is a file-based loop: a queue of work items on disk, one item processed per "tick", every result checkpointed immediately. Because all state lives in files, the loop is resumable and survives the session ending — a fresh session (or a Push heartbeat) picks up exactly where the last left off. Pairs with the /checkpoint-research skill for the synthesis step.

Pace — the safe band (why these numbers)

These are calibrated to XHS's documented read-side safe band, not guesswork. Community crawlers (MediaCrawler) default to ~2 s/request, single-threaded (CRAWLER_MAX_SLEEP_SEC = 2, MAX_CONCURRENCY_NUM = 1); reverse-engineering write-ups put the "looks-human" frequency at 3–8 s/request and the per-IP ceiling at ≤5 requests/min (~12 s average). This skill runs on your real logged-in cookies + residential IP + real Chrome fingerprint — the lowest-detection profile (the thing scrapers spend most of their effort faking). So the binding risk isn't request signature, it's behavioral: a human doesn't open 100 note-detail pages in two minutes. We therefore pace to a human reading cadence — comfortably inside the safe band — rather than to the absolute minimum. (5-minute gaps, an earlier guess, are ~25× slower than the ceiling and buy nothing.)

The real rule: behave like a human reading (not "go slow")

Detection here is distributional, not a rate threshold. XHS doesn't ask "more than N requests?" — it models what a human session looks like and flags the unlikely ones. Pure speed isn't the tell; the shape of the behavior is. Three things give a script away:

• Timing micro-structure. A human's gaps between actions are high-variance and heavy-tailed (2 s, then 40 s, then 5 min — they read, scroll, get distracted). A script emits a tight, low-variance stream. Three distinct searches in five seconds has near-zero variance and is faster than a person can read a result page — that single fact is a stronger signal than the total count.
• Session macro-structure. Humans browse depth-first: search → open a note → dwell → open a related note → wander. A scrape is breadth-first enumeration: search A, search B, search C… never clicking in, never dwelling. A keyword list run top-to-bottom is the signature.
• Action composition. Real sessions mix likes, collects, dwell, scroll, video-watch. A research run is 100% read with zero dwell and zero engagement — an impossible ratio for a person.

So the rule is not "go slow." It's make the session indistinguishable from a person reading: human-shaped variance (jitter + the occasional long pause, not merely a bigger constant), one note at a time with dwell, depth over breadth, a ceiling because people don't open hundreds of things, and stop at the first sign of friction (a human who hits a captcha backs off; a bot retries — so the circuit breaker is itself human-like behavior).

Safety-first is the more agentic posture, not the timid one. An agent's real advantage isn't speed — it's patience at scale: it can run a human-shaped loop for eight hours, which no person will. Aggression imitates the one human trait that gets you caught (impatience) and wastes the one the agent uniquely has (tireless patience). Lean into the patience.

This is a pattern, not a blacklist. The numbers here were calibrated from a real incident — a beverage-trend sweep that fired 3 distinct searches in ~5 seconds across ~17 unrelated keywords and drew an account-level warning. What's encoded is the behavioral invariant (human-shaped timing, one-in-flight, ceiling, dwell, breaker), not "don't run those queries." It generalizes to any topic and any platform. The incident is the calibration sample that tells us where the human/bot line sits — not a rule about matcha.

Three modes

Invariants (all modes): never zero-delay, never parallel, exactly one note in flight, and — except for warned Fast mode — never exceed ~5 reads/min sustained. At the Steady pace a 100-note queue finishes in ~35–50 min; only very large queues or Deep mode stretch toward end-of-day.

Strict rules (non-negotiable)

1. The loop is mandatory; Steady is the default pace. Any research touching more than ~5 notes runs through the loop at Steady pace. Deep/overnight and especially Fast are explicit opt-ins — Fast additionally requires the printed warning + user confirmation.
2. One note in flight. Never batch or parallelize reads. No for url in …; do redbook read …; done without the inter-item delay. No background & fan-out. (MAX_CONCURRENCY = 1, like the reference crawlers.)
3. Stay under ~5 reads/min, always jittered. The per-IP safe ceiling is 5 requests/min (12 s/note); Steady (20 s) sits comfortably under it. Apply ±50% random jitter to every wait — uniform intervals are a bot signature. Only warned Fast mode may cross the ceiling.
4. Cap comment pagination. Comment / sub-comment reads are the highest-risk endpoint — they trip 风控 first (deep comment crawls fail even at very long sleeps). Keep --comment-pages ≤ 3, count each comment page against the pace + budget like any other read, and never deep-crawl a comment thread.
5. Checkpoint every item. Write each result to disk and update the manifest before the next tick. Nothing in memory only.
6. Respect the daily budget. Default soft cap ~200 detail-reads/day — anomalous volume is its own 风控 trigger, independent of pace. When consumedToday >= dailyBudget, stop until the next calendar day — do not "just finish a few more".
7. Circuit breaker is absolute. On the first 风控 signal (see Circuit breaker), STOP the entire loop, persist state, surface to the user, and do not auto-retry through it. Resume only on an explicit human "resume".
8. Reads, not writes. This loop is for reading/analysis only. Engagement actions (comment/reply/like) keep their own write-side limits and are never folded into a research tick.

State contract (resumable, survives session death)

All state lives under one job directory — default research/xhs/<job-slug>/ in the private content repo (ask if unclear which repo):

Seeding the queue: the cheap list endpoints (search, feed, and user-posts with a fresh profile URL/token) are how you build the queue, not part of the paced loop. Run a search/feed up front, extract each note webUrl, and write one read/analyze-viral item per note into queue.jsonl. Then the loop drains the expensive detail-reads at the Steady pace. Always carry the fresh webUrl (with token) as the item arg — tokens expire, so don't queue bare noteIds (see xsec_token). For creator queues, prefer a profile URL containing xsec_token; bare userIds may return code=-1.

One tick = one note

A tick is the unit both drivers call. It processes exactly one item and stops:

1. Load loop-state.json. If breaker != "ok" → exit immediately (surface the reason; do not process).
2. Roll the daily budget. If budgetDate != today, reset consumedToday = 0, budgetDate = today. If consumedToday >= dailyBudget → exit until tomorrow.
3. Pop the next pending item from queue.jsonl. If none remain → finalize: synthesize SUMMARY.md from results/, mark the loop complete, and tell the driver to stop. Done.
4. Run exactly one read: redbook <kind> "<arg>" --json.
5. Inspect the result for a 风控 signal (empty {}, NeedVerify/captcha, Session expired, IP block 300012). If found → trip the breaker: set breaker:"tripped" + breakerReason, leave the item pending (not consumed), persist, surface to the user, STOP. (See Circuit breaker.)
6. On success: write results/<id>.json, set the item status:"done", check it off in manifest.md, consumedToday += 1, set lastTickAt.
7. Stop. One note only — the driver re-invokes after the (jittered) interval.

Drivers — keeping the loop running in the background

The same tick, driven two ways. Pick by how long the job must outlive the current session:

A) In-session pacing (default — Steady). The simplest driver and the right one for most jobs: run ticks back-to-back inside one session, sleeping intervalSec ± jitter between each read. At Steady pace a 100-note queue finishes in ~35–50 min — comfortably within one session — and every item is checkpointed, so a crash just resumes. No scheduler needed. (Pace per read; don't fake it with one long blocking sleep.)

B) /loop self-paced (Deep / overnight / survives idle gaps). For Deep mode or jobs spanning hours, drive ticks with /loop without an interval (self-paced): do one tick, then ScheduleWakeup for intervalSec + jitter, and re-fire. ScheduleWakeup clamps to ≥60 s, so it fits Deep (60–120 s) — not sub-minute Steady (run that in-session, per A). The loop ends itself when the queue drains (Step 3) or the breaker trips (Step 5).

C) Hook — Push recurring issue (multi-day / true background, survives app restarts). Best for "run this every day" research or corpora too big for one day. Create a recurring issue (recurringIntervalSec — a few minutes per tick, or a daily sweep) assigned to a redbook-capable agent. Each occurrence is a fresh session = one tick against the same job directory — so the issue body must point at the job dir and say "run exactly one tick, then exit." It keeps firing until a human retires it; the agent finalizes + comments when the queue drains, and comments + leaves it for human attention if the breaker trips. (Push "scout" pattern — the schedule lives on the issue, the agent's Auto switch gates it.)

Why file-state, not a long-lived process: a single blocking sleep-loop dies with the terminal and can't be resumed. The queue + loop-state.json + per-item result files mean any future session — /loop wake, Push heartbeat, or a human re-running the tick — continues exactly where the last stopped, and the partial corpus is never lost.

⚡ Fast-mode warning (emergency only)

Before any fast-mode run, print this verbatim and require an explicit "yes, fast mode" from the user. Do not proceed on silence or a vague affirmative:

⚠️  XHS FAST RESEARCH MODE — elevated 风控 risk
    Reading at ~12 notes/min crosses the safe ~5/min ceiling and
    materially raises the chance of:
      • captcha / NeedVerify mid-run (stops the loop)
      • IP block (error 300012) for hours
      • account-level recommendation throttling
    Steady mode (the default, ~20s/note) stays inside the safe band
    and still finishes a typical job in well under an hour.
    Proceed with fast mode for this run only? (yes / no)

Fast mode still obeys the floor (≥5 s ± jitter between notes — never zero), c

…(truncated)