Deep Current

Persistent research thread manager with a CLI for tracking topics, notes, sources, and findings. Pair with a nightly cron job to build a personal research digest over time. The shipped code is a local Python CLI for thread management — research is performed by the agent using its standard web_search and web_fetch tools.

A research thread manager for agents. Track topics you care about, accumulate notes and sources over time, and pair with a scheduled cron job to produce regular research digests.

Architecture

This skill ships one component: a Python CLI (scripts/deep-current.py) that manages research threads as local JSON data. It handles:

• Creating, listing, and updating research threads
• Storing notes, sources, and findings per thread
• Thread lifecycle (active/paused/resolved) and decay

What this skill does NOT ship: web search, link following, or report generation. Those capabilities come from the agent's built-in tools (web_search, web_fetch). The cron job prompt instructs the agent to use those tools to research threads, then write findings to a report file.

In short: the CLI manages what to research. The agent's existing tools do the how.

How It Works

1. Threads — Long-running research topics stored in deep-current/currents.json
2. Nightly job — A cron job tells the agent which threads to research (agent uses its own web_search/web_fetch tools)
3. Reports — Each night's findings are written to deep-current-reports/YYYY-MM-DD.md (one file per run)
4. Thread CLI — Manage threads between sessions (add, note, source, finding, status)

Setup

1. Create data directory

mkdir -p deep-current

2. Initialize currents.json

{
  "threads": []
}

3. Schedule the cron job

Create an isolated cron job that runs nightly. The agent will use its own web_search and web_fetch tools to research each thread, then use the CLI to record findings. Example prompt:

You are running a Deep Current research session.

1. Run `python3 scripts/deep-current.py list` to see all active threads.
2. Run `python3 scripts/deep-current.py covered` to see topics and URLs already covered in recent reports. AVOID repeating these.
3. Pick TWO threads based on current relevance — check recent context to decide.
4. For each thread, use web_search and web_fetch to research the topic. Follow interesting links and cross-reference claims. Find NEW angles, developments, or sources not already covered.
5. Update each thread with notes/sources/findings using the deep-current.py CLI.

## Output Format
Create a new file in deep-current-reports/ named YYYY-MM-DD.md:

# Deep Current — [tonight's date]
## [catchy title for thread 1]
[findings with inline source links]
## [catchy title for thread 2]
[findings with inline source links]

Keep it dense and interesting. No fluff. Link to sources. Flag anything actionable.

Recommended: run at 1-3am, use a capable model, 30min timeout.

Thread CLI

Manage research threads with scripts/deep-current.py:

Command	Purpose
list	Show all threads with status
show <id>	Full thread details
add <title>	Create new thread
note <id> <text>	Add dated research note
source <id> <url> [desc]	Add source/reference
finding <id> <text>	Record key finding
status <id> <active|paused|resolved>	Change thread status
digest	Summary of all active threads
decay	Prune stale threads (>90 days inactive + no recent notes)
covered [days]	Show topics & URLs from recent reports (default 14 days) to avoid duplication

Thread IDs are auto-generated slugs from the title. Prefix matching works for short IDs.

Report Format

Each run creates a standalone file in deep-current-reports/YYYY-MM-DD.md. Each report contains:

• Date header
• 2+ research threads with catchy titles
• Dense findings with inline source links
• Actionable flags for anything the user should act on

One file per run — easy to browse, search, or archive.

Research Quality Guidelines

When running a research session (nightly or manual), the agent should:

• Use web_search to find sources, web_fetch to read them
• Cross-reference claims across multiple sources
• Cite sources inline with markdown links
• Flag actionable items explicitly
• Write for a smart reader — dense, no filler
• Use catchy thread titles (this is morning reading, make it engaging)
• Distinguish speculation from sourced facts