Reel Reactor: project notes

A fun, one-off Claude Code plugin (for a LinkedIn post and GitHub). It drives Instagram through the Playwright MCP and reacts to reels that whitelisted friends DM the user, plus one DM reply per run. It's educational, probably against Instagram's ToS, and only ever used on a throwaway account.

Layout

• .claude-plugin/plugin.json: plugin manifest (name reel-reactor, version: see "Releasing changes")
• .claude-plugin/marketplace.json: single-plugin marketplace so the repo installs with /plugin
• skills/react/SKILL.md: the core. Fast mode: rr.mjs migrate --quiet → collect.js → Claude judges mood/emoji (RNG via node -e) and writes one reply to the friend about the standout reel → act.js → write state → summary table. Watch mode: the same run step by step with visible clicks (for videos, and as the fallback).
• skills/setup/SKILL.md: Node + browser tools check, rr.mjs migrate (create/upgrade config), friends, manual login in the plugin's own browser profile (says so up front), then the scheduled task (create or update, never duplicate).
• Slash commands = skills with $ARGUMENTS (user-only via disable-model-invocation, except status): skills/status, skills/friends, skills/schedule, skills/dry-run. They call scripts/rr.mjs and print its output.
• scripts/rr.mjs: config helper (no deps): migrate, status, dry-run, friends, schedule, task-spec. Owns config migrations (configVersion, currently 3) and the scheduled-task spec (TASK_PROMPT = just /reel-reactor:react plus guardrails, so tasks never go stale). REEL_REACTOR_HOME overrides ~/.reel-reactor for tests.
• scripts/collect.js (read-only) and scripts/act.js: run through browser_run_code_unsafe (older MCPs: browser_run_code) with filename = absolute path. That sandbox has no fs/require, so input goes through instagram.com localStorage (reelReactor.input, reelReactor.act), set with browser_evaluate first.
  • collect replays the page's own GraphQL (PolarisDirectInboxQuery: a thread is checked if its last_activity_timestamp_ms > state.lastSeen[friend], not by Instagram's unread flag, so opening a chat manually doesn't hide reels; else early exit; IGDThreadDetailQuery for messages, reactions and reel target_url) plus REST /api/v1/media/<pk>/info|comments. The old /api/v1/direct_v2/* REST endpoints return 404 on web.
  • act matches a bubble by its preview image file name, scrolls the message list element itself (it's virtualized), scopes React/Reply to the ancestor row that holds exactly one [role=article], and verifies through the thread data (wrong_message if any other message changed). probe: true (dry run) finds everything but clicks nothing.
• .mcp.json + scripts/playwright-mcp.mjs: bundled Playwright MCP. Flags: persistent --user-data-dir (~/.reel-reactor/playwright-profile), --output-dir ~/.reel-reactor/output (no .playwright-mcp/ in the session folder), --allow-unrestricted-file-access (MCP roots are only the session cwd; without it, loading the scripts from the plugin folder fails with "File access denied"), --codegen none (tool results don't echo the script). Optional config.browser. Tools appear as mcp__plugin_reel-reactor_playwright__*.
• config/reel-reactor.example.json: defaults (new installs are live: dryRun: false); rr.mjs creates the config from it.
• scripts/run-once.ps1, scripts/schedule-windows.ps1: CLI-only scheduling on Windows (Task Scheduler ReelReactor-HHMM).
• Primary user path is the Claude desktop app: /plugin marketplace add atinder-harika/reel-reactor → install → new session → /reel-reactor:setup → /reel-reactor:react; the scheduled task comes from setup / /reel-reactor:schedule.

Runtime files (outside the repo)

~/.reel-reactor/: config.json, state.json (lastRun, processedReelIds capped at 500, recheckFriends, lastSeen {friend: last activity ms}; no time cutoff), playwright-profile/ (the persistent logged-in browser profile), output/ (MCP snapshots/logs), run.log.

Scope

• Emoji reactions (~80% of reels, plus the sticker share while stickers are off), ~10% left alone, 1 reply to the friend in the DM thread per run (not a comment on the reel). dryRun defaults to false.
• Not built yet: stickers (stickersEnabled, stickers are reserved in the config).

Guardrails (keep them in any edit)

Only whitelisted threads. Never enter credentials. Stop on login, checkpoint or CAPTCHA pages. No follows, likes or reel comments; the only DM sent is the one reply per run in a whitelisted thread. Never click React/Reply by position. Every reaction is verified. Skip a reel rather than guess at the UI. Replies: ≤15 words, no mentions, links or hashtags, nothing mean or about looks.

Releasing changes (do this automatically, every time)

Installed copies are cached by version (~/.claude/plugins/cache/reel-reactor-marketplace/reel-reactor/<version>), so a change that keeps the same version never reaches users, even after /plugin marketplace update. Claude Code doesn't read GitHub Releases or tags for this; only version in plugin.json on the default branch.

• Any change to what ships in the plugin (skills/, scripts/, .mcp.json, config/, .claude-plugin/) must bump version in .claude-plugin/plugin.json in the same branch/PR, without being asked. Docs-only changes (README.md, CLAUDE.md) don't need a bump.
• Semver: patch (0.3.0 → 0.3.1) for fixes such as selectors, queries or wording; minor (0.3.x → 0.4.0) for new behaviour, commands or config keys; major only for breaking config/state changes.
• One bump per PR: if the branch already bumped past main's version, don't bump again.
• New or renamed config keys need a migration step in rr.mjs (bump CONFIG_VERSION). If the change alters what gets sent, have the migration turn dryRun on with dryRunReason: "upgrade" so the first run after the update is a rehearsal.
• Keep the description in plugin.json true to current behaviour.
• In the PR description, tell the user how to pick it up: with auto-update on (/plugin → Marketplaces → Enable auto-update) it arrives by itself, then /reload-plugins; otherwise /plugin marketplace update reel-reactor-marketplace, then /plugin → Installed → Update now (or uninstall/reinstall), then a new session and /reel-reactor:setup.

Gotchas

• npx doesn't spawn directly on Windows, which is why the bundled server goes through the playwright-mcp.mjs launcher (shell spawn).
• Plugin MCP servers and skills only load at session start, so after installing or updating, start a new session.
• Instagram's DOM and queries change often. Selector and query fixes go in scripts/collect.js / scripts/act.js (kept short on purpose). Behaviour and judgment changes go in skills/react/SKILL.md. Test act.js changes with probe: true first.
• Emoji picker search is picky: 💀 → "death", 🔥 → "flame" (the SEARCH map in act.js).
• The desktop scheduled-task tool has no model option; the user picks Sonnet/Haiku on the task in the app.
• To test the MCP launcher without touching the real login, set REEL_REACTOR_HOME to a temp dir (separate profile).