repost-with-agent

Repost-with-agent (v4.5.8)

GitHub Pages: https://ethansk.github.io/Repost-with-agent/

A skill-only agent/OpenClaw-compatible plugin that drives the running agent through cross-platform reposting. No CLI, no MCP server, no platform SDKs, no Playwright. The plugin ships zero code that does the work — it ships instructions (skills) and the agent's existing toolkit (Read, Edit, Write, Bash, the current harness's browser automation, and the current harness's configured message-delivery tool) does everything.

Works with any website the agent can safely operate through the logged-in browser: sources, destinations, and compose flows are defined by pair config, platform docs, per-pair learnings.md, and custom rules rather than a hardcoded platform list. This repo ships documented/validated example surfaces for LinkedIn, X, Bluesky, Threads, and Facebook.

That generic model still has real limits: no API keys, no stealth, no CAPTCHA / 2FA bypass, no private-inbox scraping unless explicitly authorized, no posting where site policy or account ownership is unclear, and no assumption that a new site's DOM will be stable. New websites need a preview validation pass, account identity checks, destination-specific dedupe, and learnings/custom rules as the agent discovers the safe flow. For OpenClaw runs, browser automation means OpenClaw's own browser profile (openclaw, CDP port 18800), not Ethan's personal browser/profile.

TL;DR

1. Clone this repo.
2. Register this repo as a directory-source plugin in the harness you want to use:
   • OpenClaw: add the repo path to plugins.load.paths and enable plugins.entries["repost-with-agent"].
   • Claude Code / compatible loaders: point the loader at .claude-plugin/.
3. Restart/reload the harness only when first registering the repo, moving the repo, or changing manifests/commands. Plain skill/doc edits are read from this directory by fresh agent runs.
4. In a fresh session: /pair create to set up a source → destination pair.
5. /repost-run <pair-id> to do a manual end-to-end repost.
6. /repost-setup-cron to install the default daily all-enabled-pairs sweep (/repost-run all). If you want something else, say so: per-pair jobs, subset jobs, dry/preview sweeps, custom cron expressions, and manual-only pairs are all valid user-owned configurations.
7. For historical source backfills, use source-item fanout semantics: one source item fans out to every enabled destination pair before the slot moves on.

That's it. The agent does everything else.

Registration / update mechanics

This is a directory-source plugin. Harness config points at this repo instead of copying the skill files elsewhere. There is no repo-owned shell registration helper.

• OpenClaw primary path: ~/.openclaw/openclaw.json includes this repo in plugins.load.paths and enables plugins.entries["repost-with-agent"].
• Claude Code / Claude-compatible path: the loader points at the .claude-plugin/ directory-source metadata when that harness is used.
• After a normal repo edit or git pull: no re-registration is needed as long as the configured path still points at this repo.
• Re-register the repo path: only for first setup, if the repo moved, if the plugin was disabled, or if harness config was reset.
• Restart/reload the harness: needed for first registration and usually safest after manifest or slash-command additions/removals. Existing long-running agent sessions may keep old startup context; start a fresh run/session for newly edited skill text.

Marketplace / packaging sanity check

Before publishing or sharing a bundle, package only the tracked repo files (for example via git archive or a clean checkout). Do not zip the live working directory if it contains ignored local files such as .env, .claude/, browser/cache artefacts, or ~/.repost-with-agent state.

Architecture in one sentence

This plugin is a folder of Markdown skills + slash commands; the running agent reads them and executes the procedure using its native tools. The plugin itself runs zero code at runtime.

The agent maintains a per-pair learnings.md so it doesn't re-figure quirks every run — pagination caps, DOM changes, rate-limit signatures, and account-specific gotchas accumulate across scheduled ticks instead of being rediscovered from scratch each time. v4.2.0 adds a structured entry shape (optional ### Selectors, ### Step playbook, and ### Quirks sub-sections) so each entry doubles as a recipe the next run can follow verbatim — read learnings.md FIRST, fall back to docs/destinations/<platform>.md only when learnings.md is silent or a cached selector misses. (See skills/repost-learnings/SKILL.md.)

Exact-first text fidelity