Deep Web Research

Deep Web Research with Kimi WebBridge

When to Use

• User asks for "deep research", "comprehensive research", "investigate thoroughly"
• Multi-source investigation needed (GitHub, X/Twitter, Google, docs sites)
• Need 10+ tabs organized in named tab groups
• User says "retry" to re-do with same approach

Core Methodology

1. Session & Tab Group Setup

# First navigate: create session + tab group with human-readable label
curl -s -X POST http://127.0.0.1:10086/command \
  -H 'Content-Type: application/json' \
  -d '{"action":"navigate","args":{"url":"","newTab":true,"group_title":""},"session":""}'

Rules:

• One task = one session = one tab group
• Session name = task identifier (e.g., vsouthvpawv-research, web-search)
• group_title = user-language label shown in browser tab group
• Never switch session names mid-task

2. Opening Additional Sources

# Subsequent navigates: same session, newTab:true, no group_title needed
curl -s -X POST http://127.0.0.1:10086/command \
  -H 'Content-Type: application/json' \
  -d '{"action":"navigate","args":{"url":"","newTab":true},"session":""}'

3. Reading Page Content

# Snapshot returns accessibility tree with @e refs
curl -s -X POST http://127.0.0.1:10086/command \
  -H 'Content-Type: application/json' \
  -d '{"action":"snapshot","args":{},"session":""}'

4. Following Links / Interacting

• Use @e refs from snapshot for click/fill
• Prefer snapshot over CSS selectors
• For navigation, use navigate with same session

Proper Use of @e References

Important: The @e references (e.g., @e58, @e100) returned by the snapshot tool are special identifiers that work ONLY as direct values for the selector parameter in tools like click, fill, etc.

Common Mistake

Attempting to use @e references in evaluate() with DOM query methods will FAIL:

// ❌ WRONG - This will throw "Failed to execute 'querySelector' on 'Document': '@e58' is not a valid selector"
document.querySelector("@e58").href

Correct Usage

Always pass @e references directly as the selector value. The Kimi WebBridge resolves @e references server-side before browser execution, making them valid in WebBridge commands but NOT in raw browser JS.

# ✅ CORRECT - Use @e reference directly as selector
curl -s -X POST http://127.0.0.1:10086/command \
  -d '{"action":"click","args":{"selector":"@e58"},"session":"my-session"}'

# ✅ CORRECT - For getting attributes, use the get_attribute action
curl -s -X POST http://127.0.0.1:10086/command \
  -d '{"action":"get_attribute","args":{"selector":"@e58","attribute":"href"},"session":"my-session"}'

When You Need Attributes

If you need to extract attributes like href, textContent, etc. from an element identified by @e reference:

1. Use the get_attribute action with the @e reference
2. Or, if you need to use evaluate, use a standard CSS selector (not an @e reference)

Example using a proper CSS selector in evaluate:

curl -s -X POST http://127.0.0.1:10086/command \
  -d '{"action":"evaluate","args":{"code":"const el = document.querySelector(\"a.target-link\"); return el ? el.href : null;"},"session":"my-session"}'

Workflow Recommendation

1. Use snapshot to get the accessibility tree with @e references
2. Identify the target element by its @e reference and role/name
3. For actions (click, fill): Pass the @e reference directly as selector
4. For attribute extraction: Use get_attribute with the @e reference
5. Never try to manipulate @e references as strings or use them in CSS selectors outside of the tool parameters

5. Closing

Only call close_session when user explicitly asks ("close those tabs").

Research Patterns

Parallel Source Investigation

Open multiple tabs simultaneously for different source types:

• GitHub repos/orgs
• X/Twitter profiles
• Google searches (quoted, site:, filetype:)
• Documentation sites
• Blog posts / Medium / Substack

Identity Resolution

When researching a handle/person:

1. Search handle + keywords
2. Check GitHub profile → pinned repos → org membership
3. Check X/Twitter → bio → followers/following → interactions
4. Cross-reference usernames across platforms
5. Look for "verified" markers, org affiliations

Technical Project Analysis

For GitHub projects:

1. Main repo → README → topics → stars/forks
2. Owner profile → other repos → contribution graph
3. Related repos (forks, dependencies, forks of forks)
4. Issues/PRs for active development signals
5. Contributors list for team/community size

Key Commands Reference

| Action | Purpose | |--------|---------| | navigate (newTab:true, group_title) | Open first tab, create tab group | | navigate (newTab:true) | Open additional tabs in same group | | snapshot | Read page content (accessibility tree) | | click / fill | Interact using @e refs | | find_tab | Switch to earlier tab in session | | list_tabs | Inspect open tabs | | close_session | Close entire tab group (user-initiated only) |

Daemon Management

macOS / Linux:

~/.kimi-webbridge/bin/kimi-webbridge start

Windows (PowerShell):

& "$env:USERPROFILE\.kimi-webbridge\bin\kimi-webbridge.exe" start

Cross-platform (any shell with HOME/USERPROFILE):

# Using $HOME (Linux/macOS) or %USERPROFILE% (Windows)
${HOME:-$USERPROFILE}/.kimi-webbridge/bin/kimi-webbridge${EXE:-} start

The daemon runs on http://127.0.0.1:10086 regardless of platform.

Pitfalls to Avoid

• ❌ Switching session names mid-task → fragments tab groups
• ❌ Using Chrome CDP / built-in browser tools instead of Kimi WebBridge
• ❌ Not using newTab:true for parallel pages
• ❌ Using bare domains in find_tab (use exact URL)
• ❌ Calling close_session without user request
• ❌ CSS/JS selectors when @e refs available

Example: Multi-Source Person Investigation

# Session: "alice-research"
# 1. Google search
navigate("https://google.com/search?q=%22alice%22+github", newTab:true, group_title:"alice deep research", session:"alice-research")
# 2. GitHub profile
navigate("https://github.com/alice", newTab:true, session:"alice-research")
# 3. X/Twitter
navigate("https://x.com/alice", newTab:true, session:"alice-research")
# 4. Company site
navigate("https://company.com/team/alice", newTab:true, session:"alice-research")
# ... repeat for 10+ sources
# Snapshot each, cross-reference, synthesize

References

• [E-Reference Guide](file:///e:/hermes-skills-bundle/deep-web-research/references/e-reference-guide.md)
• [Search Engine Preference](file:///e:/hermes-skills-bundle/deep-web-research/references/search-engine-preference.md)

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: sloemo01
• Source: sloemo01/hermes-skills-bundle
• License: MIT

Install and usage instructions live in the source repository linked above.