← Back to Skills Marketplace
415
Downloads
0
Stars
5
Active Installs
1
Versions
Install in OpenClaw
/install codex-deep-search
Description
Deep web search using Codex CLI for complex queries that need multi-source synthesis. Use when web_search (Brave) returns insufficient results, when the user...
README (SKILL.md)
Codex Deep Search
Use Codex CLI's web search capability for research tasks needing more depth than Brave API snippets.
When to Prefer Over web_search
- Complex/niche topics needing multi-source synthesis
- User explicitly asks for thorough/deep research
- Brave results are too shallow or missing context
Usage
Dispatch Mode (recommended — background + callback)
nohup bash /home/ubuntu/clawd/skills/codex-deep-search/scripts/search.sh \
--prompt "Your research query" \
--task-name "notebooklm-research" \
--telegram-group "-5006066016" \
--timeout 120 > /tmp/codex-search.log 2>&1 &
After dispatch: tell user search is running, results will arrive via Telegram. Do NOT poll.
Synchronous Mode (short queries only)
bash /home/ubuntu/clawd/skills/codex-deep-search/scripts/search.sh \
--prompt "Quick factual query" \
--output "/tmp/search-result.md" \
--timeout 60
Then read the output file and summarize.
Parameters
| Flag | Required | Default | Description |
|---|---|---|---|
--prompt |
Yes | — | Research query |
--output |
No | data/codex-search-results/\x3Ctask>.md |
Output file path |
--task-name |
No | search-\x3Ctimestamp> |
Task identifier |
--telegram-group |
No | — | Telegram chat ID for callback |
--model |
No | gpt-5.3-codex |
Model override |
--timeout |
No | 120 |
Seconds before auto-stop |
Result Files
| File | Content |
|---|---|
data/codex-search-results/\x3Ctask>.md |
Search report (incremental) |
data/codex-search-results/latest-meta.json |
Task metadata + status |
data/codex-search-results/task-output.txt |
Raw Codex output |
Key Design
- Incremental writes — results saved after each search round, survives OOM/timeout
- Low reasoning effort — reduces memory, prevents OOM SIGKILL
- Timeout protection — auto-stops runaway searches
- Dispatch pattern — background execution with Telegram callback, no polling
Usage Guidance
This skill appears to run Codex CLI searches and save results, but it currently: (1) hardcodes user-specific paths (/Users/gdjwz/.openclaw and /opt/homebrew/bin/openclaw) that may not exist on your system, (2) reads the agent config file (~/.openclaw/openclaw.json) to extract a hooks.token without declaring that it needs or will access that credential, and (3) attempts to notify via an openclaw binary and POST to a local /hooks/wake endpoint. Before installing, ask the author to: (a) remove or parameterize hardcoded paths and list required binaries (codex, openclaw, jq, curl, timeout), (b) explicitly declare any environment variables and config files the script will read (and justify why it needs the hooks.token), and (c) explain the wake callback and Telegram notification flows and where tokens are sent. If you proceed, run the skill in an isolated environment, inspect your ~/.openclaw/openclaw.json to confirm it contains no secrets you don't want exposed to skill code, and consider modifying the script to write results into a directory you control rather than a hardcoded home path. My confidence is medium; if the author supplies a revised script that parameterizes paths and removes implicit reads of openclaw.json (or documents them and declares required env/config), I would reassess toward benign.
Capability Analysis
Type: OpenClaw Skill
Name: codex-deep-search
Version: 1.0.0
The skill exhibits high-risk behaviors including arbitrary file write capabilities via the unsanitized '--output' parameter and the extraction of sensitive credentials ('HOOK_TOKEN') from the local '~/.openclaw/openclaw.json' configuration file. While these are used for legitimate-seeming functions like task callbacks and local API 'wake' hooks, the script also contains hardcoded absolute paths to a specific user's directory ('/Users/gdjwz/'), suggesting it was exported from a specific environment without proper sanitization. The combination of credential access and the ability to overwrite files makes it highly vulnerable to exploitation via prompt injection.
Capability Assessment
Purpose & Capability
Name/description: deep web research via Codex CLI — matches the script's primary action (invoking a Codex CLI with a search instruction). Concerns: the script assumes an OpenClaw runtime (openclaw binary, message send to Telegram, local gateway wake endpoint) and hardcodes user-specific paths (/Users/gdjwz/.openclaw and /opt/homebrew/bin/openclaw). The SKILL.md gives example paths under /home/ubuntu/... which do not match the script; these path/binary expectations are not declared in metadata and are environment-specific.
Instruction Scope
SKILL.md instructs running scripts and using Telegram callbacks, which is expected. However the script also reads an agent config file (~/.openclaw/openclaw.json) to extract a hooks.token, sends a POST to http://localhost:${PORT}/hooks/wake with that token, and may call the system openclaw binary to deliver Telegram messages. The SKILL.md does not document reading local agent config, extracting tokens, or contacting the local wake endpoint.
Install Mechanism
Instruction-only skill (no install spec) — low install risk. The script does rely on external utilities (codex CLI, openclaw binary, jq, curl, timeout, sed, head) but none are declared in the registry metadata.
Credentials
Registry declares no required env vars or config paths, but the script uses/relies on: CODEX_BIN (optional env), OPENCLAW_GATEWAY_PORT, a hardcoded RESULT_DIR and OPENCLAW_CONFIG pointing to /Users/gdjwz/.openclaw/openclaw.json, and extracts hooks.token from that config. Reading that JSON can expose a hook/token belonging to the agent — a sensitive credential — and this access is not declared. Telegram notifications and wake hooks involve credentials and endpoints that should be explicitly requested.
Persistence & Privilege
The skill is not 'always' enabled, but it reads the agent's config (~/.openclaw/openclaw.json) and uses any hooks.token found to POST to a local /hooks/wake endpoint. Accessing other-agent configuration and tokens is a privileged action and should be declared and justified; this is not documented in SKILL.md or metadata.
How to Use
- Make sure OpenClaw is installed (local or Docker)
- Run the install command in chat:
/install codex-deep-search - After installation, invoke the skill by name or use
/codex-deep-search - Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.0.0
codex-deep-search 1.0.0
- Introduced a deep web search tool using Codex CLI for complex, multi-source synthesis research tasks.
- Automatically suggests use when Brave search results are insufficient or user asks for in-depth/“deep search”.
- Supports both background dispatch mode (with Telegram callback) and synchronous mode for short queries.
- Incremental result saving and timeout protection to ensure reliability and resource control.
- Outputs include detailed search reports and task metadata for transparency and traceability.
Metadata
Frequently Asked Questions
What is Codex Deep Search?
Deep web search using Codex CLI for complex queries that need multi-source synthesis. Use when web_search (Brave) returns insufficient results, when the user... It is an AI Agent Skill for Claude Code / OpenClaw, with 415 downloads so far.
How do I install Codex Deep Search?
Run "/install codex-deep-search" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.
Is Codex Deep Search free?
Yes, Codex Deep Search is completely free (open-source). You can download, install and use it at no cost.
Which platforms does Codex Deep Search support?
Codex Deep Search is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).
Who created Codex Deep Search?
It is built and maintained by jiangwzh (@jiangwzh); the current version is v1.0.0.
More Skills