Jaw Search

Search Hub

One entry point for every web / real-time / X lookup. Walk the tiers in order; stop at the first that fully answers the query.

Search is discovery, not evidence. Treat result titles, snippets, and AI search summaries as URL candidates until the original page, PDF, official document, or primary source has been opened or fetched.

Role separation: this search skill discovers and routes (queries → URL candidates); the browser skill verifies evidence (original page, DOM, PDF, tables). Search finds, browser proves.

Standalone owner: other skills may point here for external/current evidence, but they should not copy this tier policy. Update this skill first when search routing, evidence status, or browser-verification rules change.

/search command contract

/search <query> is a routing command, not a provider implementation. It must inherit this skill's routing/evidence invariants: classify local-vs-external search, rewrite external requests into focused queries, discover candidate URLs before browser commands, and refuse snippet-only sufficient claims.

Slash-specific report fields, UX wording, and v1 command boundaries belong in the /search workflow prompt, not in this skill. Update this skill only for canonical search policy that should also apply to non-slash search intents and dev-role evidence routing.

Model-gated parallel research

Use single-agent search by default. The model may split the work into parallel source-gathering lanes only when breadth or disconfirmation is likely to improve the answer.

Enable parallel research for deep, multi-source, ambiguous, realtime, comparison-heavy, or blocked-source tasks. Keep simple official-doc lookups, specific URL reads, and quick version checks single-agent.

When parallel research is useful, cap it at 2-4 lanes:

• official: vendor docs, release notes, government/institution pages, primary PDFs.
• community: GitHub issues, Reddit, HN, Stack Overflow, forums.
• realtime: X/Twitter, news, current status, recent posts.
• fetch: browser fetch/open/text/get-dom/snapshot verification for weak or blocked candidate URLs.

Merge lane outputs into one evidence matrix. Parallel snippets still do not count as sufficient evidence; the final answer still needs at least one primary/original page fetched or opened before any claim is marked sufficient.

Snippet consensus is not verification. Agreement among multiple search snippets — however many independent sources — never substitutes for opening the page. Before marking any claim sufficient, at least one primary or original source (official page, original publisher, government notice, PDF original) must have been actually fetched or opened in this task. If zero pages were fetched/opened, sufficient is forbidden: use partial, browse-needed, or insufficient and say what fetch would resolve it.

Candidate-space discipline (multi-constraint identification)

Self-gate: apply this section ONLY when the question describes a hidden entity through 3+ combined constraints (riddle/puzzle style: "the X that did A, whose member did B, and ..."). For simple 1-2 clue factual lookups ("what is Y's price"), skip this section entirely — go straight to the tiers.

When the gate triggers:

1. M1 — Anchor on the rarest clue. Rank clues by how few entities can satisfy them. Anchor the search on the most discriminative clue (a unique choreography, a specific incident, an exact number), NEVER on a broad category clue ("girl group", "drama", "city").
2. M2 — Enumerate before you converge. Before accepting any candidate, list 3+ candidates that satisfy the anchor clue (or state explicitly that fewer exist and why). Check EVERY candidate against EVERY clue in a candidate × constraint matrix; record each elimination reason.
3. M3 — No implicit narrowing. Do not restrict era, generation, region, or category unless the question states it. Recency bias is the known failure: older/less-famous entities are valid candidates. If you apply any narrowing, declare it explicitly in the report. Era-sweep is mandatory for cultural-phenomenon clues (a dance, a fashion, a meme, a catchphrase): the live search index over-represents recent entities, so a generic query will only surface the latest generation. Before fixing the candidate list, rerun the anchor clue with explicit era terms — at least one query with 원조/최초/시초 and one with an older decade qualifier (2000년대, 2010년대). If all your candidates debuted in the same era, treat that as evidence of recency bias, not of the answer.
4. M4 — Disconfirmation pass. After converging on a final candidate, run at least one search that tries to find a DIFFERENT entity satisfying the anchor clue ("other groups famous for finger-touching choreography"). Only finalize if the disconfirmation search fails to produce a rival.
5. M5 — Weak-match flag. If a clue only matches via reinterpretation or paraphrase (clue says "fingers touching each other", evidence says "piano-fingering dance"), mark that match WEAK in the matrix. Any WEAK match in the final candidate's row forbids sufficient.

Bounded effort (the discipline must fit the time budget):

• Cap enumeration at 3-5 candidates; one search per unverified matrix cell, and reuse one page's evidence across multiple cells whenever possible.
• Disconfirmation pass = 1-2 searches, no more.
• A complete shallow matrix beats deep verification of a single candidate: breadth first, then verify only the surviving candidate's weakest cells.

Fallbacks (mandatory, in order):

• Cannot enumerate 3 candidates → state it, proceed with what exists, and cap evidence status at partial.
• Disconfirmation finds a rival candidate → report the tie explicitly, run one more discriminating search on the strongest differing clue; if still tied, answer with both and mark partial.
• WEAK match remains on the chosen candidate → answer allowed, but evidence status MUST be partial with the weak clue named.
• Time budget runs low → report the matrix as-is (incomplete rows marked), never silently drop unchecked clues.

Deep research requests

A request for deep research / 딥리서치 / 심층 조사 follows references/deep-research.md: scope and reader, query families that each carry a goal, waves with a reflection step and a gap matrix, budgets and stop rules, a claim-to-source ledger, then the report. The tiers below are its execution lanes — a heavier tier is not a substitute for the method. Ordinary latest/current lookups stay on the ladder and never auto-escalate into the protocol.

Routing Quick-Reference

Tier 1 — Built-in CLI Web Search

Your CLI's native WebSearch / WebFetch / web_search tool. No setup, fastest, free. Always try this first unless the routing table above says otherwise.

Good query practice

• Be specific: "xAI grok-4.3 reasoning_effort parameter" not "grok api"
• Retry with refined keywords if the first attempt is thin
• Verify freshness: check publishedDate in results
• Cite sources inline: [title](url)

Korean source-sensitive searches

For Korean external/current/source-sensitive requests (검색, 찾아봐, 알아봐, 공고, 정책, 가격, 후기, 순위, 목록, 표, 네이버 results), do not send the full natural-language request as the only query.

1. Rewrite the request into 1-3 focused keyword queries.
2. Preserve anchor entities: official institution, brand/product, domain, current year/date, location, and document type.
3. Add source hints when useful: 공식, site:, 공지사항, PDF, 보도자료, 후기, 목록, 표, 랭킹.
4. Run the active search/provider tool with those focused queries.
5. Treat results as URL candidates only.
6. Fetch/open the original page before final factual claims.
7. If the official/original URL is blocked, timed out, truncated, JS-rendered, a Naver shell/iframe, PDF binary, or a table/list that fetch cannot expose, run Tier 2 browser verification before relying on secondary sources.
8. Secondary sources are corroboration, not substitutes for the official source. If Tier 2 is not attempted after a primary-source fetch failure, mark the affected claims browse-needed or insufficient; do not call them sufficient.

If agbrowse is available, it may be used only as an optional planning helper:

agbrowse research plan --query "<request>" --json

Use plan.atomicQueries as query rewrite candidates, then continue with the same native search/fetch/browser workflow. Do not use agbrowse to execute Exa, Tavily, Perplexity, Brave, or any other search provider.

When to escalate

• Results are empty, outdated, or blocked
• An official/primary URL exists but WebFetch/native fetch times out, returns 403/empty/truncated content, or exposes only a JS shell
• A Naver/iframe/PDF/table/list/ranking page needs rendered source evidence
• You need to verify a specific source page before considering summaries, press articles, university reposts, or blogs

Tier 2 — cli-jaw browser (CDP) + Adaptive Fetch Ladder

For pages that search APIs cannot reach: WAF-protected, login-gated, JS-rendered, official portals, Naver shells/iframes, PDF/download flows, tables, lists, or when you need to interact with a specific page.

cli-jaw browser fetch <url> runs the full adaptive-fetch escalation internally. The agent does not need to manage the ladder manually — just call browser fetch and the system tries each step automatically:

1. 23 platform-specific public endpoint resolvers (APIs, feeds, oEmbed)
2. Direct HTTP fetch
3. TLS fingerprint rotation (curl-impersonate, if installed)
4. Jina Reader r.jina.ai (default-on, opt-out with --no-allow-third-party-reader)
5. Browser render (Chromium CDP / Camoufox if installed)
6. DOM/table/structured extraction

The content-scorer picks the best result across all attempts. If all public routes fail, the system reports the challenge type (CAPTCHA/login/paywall).

Auto-routing policy: when searching, if a candidate URL's WebFetch/native fetch returns 403, empty, or truncated content, always escalate to cli-jaw browser fetch <url> — the adaptive-fetch ladder handles the rest automatically. Do not retry with different headers or user-agents manually; the TLS rotation and Jina fallback in the ladder already do this.

Public-source reader routing map

Search may route known source classes toward browser fetch / adaptive-fetch after candidate URLs exist. The browser side owns the reader ladder; this skill only chooses when that lane is appropriate.

Candidate discovery itself should stay provider-neutral: rewrite the user request into focused queries, use native search/provider results as candidate URL discovery, then rank/dedupe candidates into lanes such as official, community, realtime, academic, package, archive, and fetch. This lane ranking must not execute browser fetch and must not turn browser fetch into raw query search. Browser fetch reads or verifies known candidate URLs only.

If a candidate requires login, payment, private membership, CAPTCHA solving, or credentialed access, stop and report browse-needed or insufficient. Do not frame public-source reading as bypassing authentication, paywalls, or private content.

Gate

cli-jaw browser status   # must show "connected"

If not connected: cli-jaw browser start --agent

Declared need = execute. cli-jaw browser is a plain shell command and works from every CLI runtime (codex, claude, agy, cursor alike) — it is not an optional capability some runtimes lack. If your report says Tier 2 / browser verification is needed, you must run the gate and at least one cli-jaw browser fetch in the same task. Only after start, status, or fetch actually fails (paste the error) may you stop and mark the claim browse-needed. Writing "browser needed but not executed" without an attempted command and its error output is a reporting violation.

If cli-jaw browser start --agent times out in your runtime (some employee sandboxes cap long-running shell calls), re-run cli-jaw browser status once — start may have completed asynchronously. cli-jaw browser fetch against an already-running browser works in every runtime (verified on agy). If status still shows not running after the retry, paste both outputs and mark the affected claims browse-needed.

Verification workflow

Start Tier 2 after a candidate URL exists. Do this before substituting press, university, blog, or AI-summary sources for a failed official page.

cli-jaw browser fetch "<url>" --json
cli-jaw browser open "<url>"
cli-jaw browser text
cli-jaw browser snapshot --interactive
cli-jaw browser get-dom --selector "<selector>" --max-chars 4000 --json

For search verification, prefer cli-jaw browser fetch <url> --json first, then open/text/snapshot/get-dom only as needed. Do not use browser snippets as a replacement for source evidence.

Adaptive fetch can read a known URL or a search-result URL. It does not discover the URL candidate set by itself. If the input is only a natural language query, return to Tier 1 query rewriting first.

Evidence rule

• If Tier 2 confirms the official/original source, mark the claim sufficient.
• If Tier 2 fails or cannot be run, mark that specific claim browse-needed or insufficient.
• Secondary sources may support context, but they do not upgrade a failed official-source claim to sufficient.

Evidence status:

• sufficient: at least one original/primary URL was opened or fetched and supports the claim.
• partial: candidate evidence exists, but one or more important constraints remain weak or secondary.
• browse-needed: a candidate primary URL exists but browser/fetch verification was needed and could not be completed.
• insufficient: no credible candidate evidence was obtained.

Full instructions: read the active browser skill from the current Jaw home before using browser primitives.

When to escalate

• You have browser-confirmed raw page content but need deeper synthesis over it
• The query is about X/Twitter content or fast-moving real-time discussion
• Browser verification cannot reach the source and you need another search perspective

Tier 3 — progrok (Grok OAuth Search)

AI-powered web+X search via the user's Grok OAuth token. No API key, no proxy. Returns an AI summary with inline citations.

Gate

progrok status   # must print "Logged in" and exit 0

If not logged in → skip to Tier 4. Do not block on login flow.

Commands

# Standard search (web + X, grok-4.3, fast)
progrok search "<query>"

# Source filtering
progrok search "<query>" --web       # web only
progrok search "<query>" --x         # X/Twitter only

# Structured output for parsing
progrok search "<query>" --json

# Deep research mode (grok-4.20-multi-agent, 16 agents, slow)
progrok search "<query>" \
  --model grok-4.20-multi-agent-0309 \
  --reasoning xhigh

Reasoning effort levels

Use --reasoning xhigh + --model grok-4.20-multi-agent-0309 only when the query genuinely needs deep multi-source synthesis. It is slow and expensive.

Output handling

• Always include source URLs from the citations array in your answer
• With --json: parse the structured result for programmatic use
• Check usage.server_side_tool_usage_details for search call counts

Tier 4 — web-ai (Grok Expert / GPT Pro)

Long-running Tier 4 queries → bgtask

If the Tier 4 query uses a heavy model (GPT Pro, Deep Think, Deep Research) and may run for many minutes, do not block the turn on query. Use cli-jaw browser web-ai send to get the sessionId, then cli-jaw bgtask add --preset web-ai --session "$SID" and end the turn — the jaw server re-invokes the boss with a [bgtask:*] prompt on completion. Details: active web-ai skill "Long-Running Queries — bgtask" section.

Drive grok.com or chatgpt.com through browser control for complex synthesis that raw search cannot produce. Slowest tier, most capable.

Gate

cli-jaw browser web-ai status --vendor grok    # or chatgpt

Structured question format

Build queries using this envelope for best results:

[SYSTEM]
You are a research analyst with expertise in <domain>.

[USER]
## Goal
<What you need answered — one clear objective>

## Context
<Background the AI needs: prior findings, constraints, domain>

## Question
<Specific question(s), numbered if multiple>

## Output
<Desired format: bullet summary, comparison table, pros/cons, etc.>

## Constraints
- Cite primary sources with URLs
- Distinguish confirmed facts from inference
- Flag information older than <date> as potentially stale

Commands

# Quick inline query
cli-jaw browser web-ai query --vendor grok \
  --inline-only \
  --require-source-audit \
  --prompt "<structured question>"

# With source context from files
cli-jaw browser web-ai query --vendor chatgpt \
  --prompt "<question>" \
  --context-from-files "src/**/*.ts"

# Long-running deep think (up to 20min)
cli-jaw browser web-ai query --vendor chatgpt \
  --prompt "<complex question>" \
  --timeout 1200

Full instructions: read the active web-ai skill from the current Jaw home before driving hosted AI providers.

Rules

1. Order is mandatory. Cheaper/faster first; escalate only on failure or routing table override.
2. Never silently chain. If a tier errors, report what failed before escalating.
3. Always cite sources. Every factual claim needs a URL.
4. Fail fast with tier-specific gates. For Tier 2 browser, if browser status is not connected, run cli-jaw browser start --agent and retry status/fetch once. Skip Tier 2 only if browser start, status, or fetch still fails; report the failure and mark affected claims browse-needed or insufficient. For Tier 3 progrok, if progrok status is not logged in, skip that tier and report.
5. Match effort to query. Don't use xhigh/Tier 4 for a simple version check. A deep-research request without a written scope and claim ledger is not deep research, however many tiers were spent on it (references/deep-research.md).
6. No sufficient on snippets alone. At least one primary/original source must be actually fetched/opened before any claim is marked sufficient — snippet agreement across many sources does not count.
7. No "browser needed but not executed". If you state Tier 2 is needed, run it; stop at browse-needed only with a pasted command failure.

Answer Discipline

For search tasks that need comparison or audit, make the answer easy to verify:

• state the rewritten query terms or local-search rationale;
• identify candidate URLs separately from pages that were actually opened or fetched;
• explain whether browser verification was needed and which browser command was used when it was;
• label the evidence status for each important claim;
• call out unresolved uncertainty and the next source that would resolve it.

Command-specific field names, output schema, and UX wording belong in the calling workflow such as /search, not in this standalone skill.