Authoring agent-data-cli source

Overview

Use this skill to design and implement an agent-data-cli source with stable behavior and clear project fit.

This skill is deliberately stricter than normal feature work because a weak source design causes protocol drift, command confusion, and unreliable sync behavior.

It is the source-authoring path for RSS feeds, HTTP APIs, HTML scraping, browser-driven sites, finance data, news content, and other remote content systems that must fit the source/channel/content model.

Current core contract to keep in mind:

• - source/channel is still the only core resource model
• INLINECODE3 sources return ContentSyncBatch, not flat per-row persistence instructions
• shared persistence is now content_nodes, content_channel_links, and INLINECODE7
• structural relations in core use abstract parent; source-specific meaning belongs in INLINECODE9

Hard Gate

Do not start implementation immediately.

The required sequence is:

1. 1. research
2. spec
3. plan
4. approval
5. implement
6. verify

If the user explicitly wants to skip a stage, say what risk that creates before proceeding.

When to Use

Use this skill when the user wants to:

• - add a new source
• redesign an existing source
• add source capabilities such as channel search, content search, content update, or INLINECODE13
• add support for RSS, APIs, scraping, browser automation, authentication, cookies, or remote side effects

Do not use this skill for:

• - ordinary content operations against an existing source
• unrelated CLI or store changes with no source work

Install From skills.sh

Install this skill directly from skills.sh:

CODEBLOCK0

Install

If agent-data-cli is not present locally, install it first:

CODEBLOCK1

Then load the bundled skills from this repository's skills/ directory and work from the repo root.

Important boundary:

• - source code belongs in the source workspace repo, typically INLINECODE17
• keep agent-data-cli focused on core/cli/store/protocol work
• do not install source runtime dependencies into the core project with INLINECODE19
• use uv pip install or init.sh inside the source workspace instead

Workflow

1. Research

Identify the source type before making architecture decisions.

Classify it as one or more of:

• - RSS
• API
• HTML scraping
• browser-driven
• auth or session driven
• interact capable

Research must confirm:

• - whether the source has a real channel concept
• whether remote discovery and remote sync are distinct
• how to identify unique content
• whether the source has hierarchical or container-like content that should become INLINECODE23
• what time field is available
• how pagination or incremental fetch works
• what config is required
• whether interact is actually possible

Use available web research, local fetch tools, and the repo's fetchers/ where appropriate.

2. Spec

Write a source-specific spec before implementation.

It must define:

• - source to resource mapping
• supported capabilities
• config fields and mode if needed
• content normalization and dedup strategy
• INLINECODE25 strategy
• whether update returns only direct content, or also context nodes and INLINECODE26
• whether the source needs relation_semantic values such as reply, contains, or INLINECODE30
• storage requirements
• error boundaries
• CLI-visible semantics
• testing scope

For native search/query views:

• - treat column names as a soft compatibility surface because multi-source and multi-channel aggregation merges by column header
• prefer explicit names such as published_at, publisher, author, price, INLINECODE35
• avoid vague names such as time, source, value unless that meaning is genuinely exact
• column order is mainly for readability; header naming is what determines merge behavior

3. Plan

Turn the approved spec into an implementation plan.

The plan must break work into:

• - failing tests to add first
• source code units to implement
• INLINECODE39 construction path
• CLI verification steps
• persistence and audit verification

4. Approval

Wait for user approval after the spec and plan.

Do not jump from research straight to code.

5. Implement

Implement with TDD.

• - write failing tests first
• verify the failure is correct
• write minimal code
• rerun focused tests

6. Verify

Before claiming completion, verify:

• - unit tests
• CLI simulation tests
• help output
• capability and config behavior
• persistence side effects
• INLINECODE40 / content_channel_links / content_relations side effects when update is involved
• interact audit behavior when applicable

Read Next

• - references/source-contract.md for repository rules
• INLINECODE44 for feed-like sources
• INLINECODE45 for JSON or HTTP API sources
• INLINECODE46 for browser-driven sources
• INLINECODE47 for remote side effects
• INLINECODE48 for test matrix
• INLINECODE49 before final verification