Autosearch autosearch:tikhub-fallback

Decision tree for when to escalate from free native Chinese channels (bilibili / weibo / xiaohongshu / douyin / zhihu) to the paid TikHub fallback. ~$0.0036 per request; 5 platforms covered. Tells the runtime AI when the cost is worth paying vs. when the free paths are enough.

install
source · Clone the upstream repo
git clone https://github.com/0xmariowu/Autosearch
Claude Code · Install into ~/.claude/skills/
T=$(mktemp -d) && git clone --depth=1 https://github.com/0xmariowu/Autosearch "$T" && mkdir -p ~/.claude/skills && cp -r "$T/autosearch/skills/meta/tikhub-fallback" ~/.claude/skills/0xmariowu-autosearch-autosearch-tikhub-fallback && rm -rf "$T"
manifest: autosearch/skills/meta/tikhub-fallback/SKILL.md
source content

TikHub Fallback Decision Tree — Advisory

Cost anchor: TikHub averages $0.0036 per request (per

docs/mcp-channels-playbook.md
measurement). A research session with 1000 TikHub calls costs about $3.60/month. Cheap per call, but accumulates fast if every query defaults to paid.

Boss rule: Free paths first; TikHub is the paid cushion, not the default.

Covered Platforms (5)

PlatformFree native skillTikHub endpointFree path reliability
Xiaohongshu
search-xiaohongshu
(requires cookie)
xiaohongshu/web/search_notes
Low — reliably gated; TikHub is usually required
Weibo
search-weibo
(free API)
web/fetch_search
Medium — free works but TikHub upstream is flaky
Douyin
search-douyin
(free native)
TikHub +
mcporter
-routed free MCP
Medium — free works, TikHub more complete
Zhihu
search-zhihu
(free native)
TikHub endpointLow — Zhihu blocks aggressive scraping
Twitter / X
search-twitter-exa
search-twitter-xreach
(XReach)
Medium — Exa covers public, XReach covers deeper signal

Other hard Chinese platforms not covered by TikHub (do not escalate, degrade gracefully):

  • WeChat Mini Programs
  • Private Xiaohongshu accounts
  • Facebook / Instagram private pages
  • Any account-gated content

Decision Tree

Apply per query, not per session:

1. Try the free native channel skill first.
   ├─ Returns usable results? → DONE (do not escalate).
   └─ Blocked / empty / flaky? → step 2.

2. If `TIKHUB_API_KEY` is not set in env:
   ├─ Try the mcporter free MCP if applicable (Weibo, Douyin).
   └─ Otherwise report to user: "no free path, no TikHub key — gracefully degraded."

3. If `TIKHUB_API_KEY` is set AND the platform is one of the 5 covered:
   ├─ If query is research-critical (user explicitly asked or Best-tier step) → escalate to TikHub.
   ├─ If query is exploratory / fan-out → skip TikHub, try a different channel instead.
   └─ After 3 failed TikHub requests on the same session / same query shape → stop, do not retry a 4th time (rate-limit or upstream regression signal).

4. Bundle call patterns to reduce cost:
   ├─ Dedupe queries across channels before calling TikHub.
   └─ Prefer single calls that return rich payloads over multiple narrow calls.

Per-Platform Notes

  • Xiaohongshu: Free scraping needs cookie that expires; TikHub is usually the primary practical path.
  • Weibo: Free path works for recent hot topics; TikHub upstream has returned
    ok=1 but cards empty
    occasionally — ~20% flake. Pair with
    search-wechat
    for long-form verification.
  • Douyin: Free
    mcporter
    MCP usable if user has cookie configured; TikHub as paid fallback when MCP 403s.
  • Zhihu: Free path rate-limits after ~30 queries/session. Escalate to TikHub on rate-limit.
  • Twitter/X: Prefer
    search-twitter-exa
    (free, Exa semantic site: query); only reach
    search-twitter-xreach
    (XReach paid) when a fresh timeline read is mandatory.

What This Skill Is Not

  • Not a wrapper around the TikHub client (that's
    autosearch/lib/tikhub_client.py
    ).
  • Not a replacement for the channel skill's own
    fallback_chain
    . The channel skills carry
    fallback_chain: [via_tikhub, api_search, api_video_detail]
    in frontmatter — that is the per-channel sequence. This skill is the meta policy deciding whether the runtime AI should even reach into the fallback chain for the TikHub hop.

Boss Rules

  • Do not default to TikHub. Free paths first, always.
  • Do not flip global policy from a single failure. If a free channel failed once, retry with a different query shape before escalating. Only a repeated pattern (3+ failures on same query shape) or a known upstream block (Xiaohongshu) justifies escalating immediately.
  • Budget-aware runtime AI should warn the user when TikHub spend crosses a session threshold (~$1 equivalent = ~280 requests).

Related Skills

  • autosearch:router
    — picks the channel group first.
  • autosearch:model-routing
    — picks model tier (orthogonal to cost).
  • mcporter
    — the free-MCP fallback path (Weibo / Douyin alternatives without TikHub).
  • Individual channel SKILL.md frontmatters already list
    via_tikhub
    in their
    fallback_chain
    .

Quality Bar

  • Evidence items have non-empty title and url.
  • No crash on empty or malformed API response.
  • Source channel field matches the channel name.