socialmesh

SocialMesh

Media Diet v2 is an algorithm intelligence platform. A browser extension scans social feeds via direct DOM access and writes structured data to the local filesystem. Claude Code reads that data, synthesizes AlgorithmProfiles, generates AgenticHarnesses, and produces reports. There is no Python backend, no Playwright, and no vault.

Extension communication

The extension exposes a WebSocket server on localhost:9242 and writes a file fallback for environments where WebSocket is unavailable.

WebSocket (primary)

ws://localhost:9242

Send JSON messages:

{"type": "scan", "platform": "instagram", "options": {"depth": 30}}
{"type": "status"}
{"type": "get_profile", "platform": "instagram"}

Receive JSON responses:

{"type": "scan_complete", "platform": "instagram", "scan_id": "abc123", "item_count": 47}
{"type": "status", "connected_extension": true, "available_platforms": ["instagram", "youtube"]}
{"type": "profile", "platform": "instagram", "data": {...}}

File fallback

When WebSocket is unavailable, write a command file and poll for the response:

• Command file: data/extension-bridge/commands/<uuid>.json
• Response file: data/extension-bridge/responses/<uuid>.json

The extension polls for command files every 2 seconds. Poll the response file until it appears (max 60 seconds), then delete both files.

Command file shape:

{"id": "<uuid>", "type": "scan", "platform": "instagram", "options": {"depth": 30}}

Response file shape:

{"id": "<uuid>", "type": "scan_complete", "platform": "instagram", "scan_id": "abc123", "item_count": 47}

Data storage layout

All data is relative to the repository root. These directories are git-ignored.

data/
  extension-bridge/
    commands/          # outbound requests to the extension
    responses/         # inbound responses from the extension
  scans/
    <platform>/
      <scan_id>.json   # raw scan output from the extension
  profiles/
    <platform>/
      latest.json      # most recent synthesized AlgorithmProfile
      <timestamp>.json # historical profiles
  harnesses/
    <platform>/
      latest.json      # most recent AgenticHarness
      <timestamp>.json # historical harnesses
  reports/
    <timestamp>.md     # rendered editorial reports

Scan output shape

Each data/scans/<platform>/<scan_id>.json file contains:

{
  "scan_id": "abc123",
  "platform": "instagram",
  "scanned_at": "2026-04-16T12:00:00Z",
  "item_count": 47,
  "items": [
    {
      "id": "...",
      "author": "...",
      "author_handle": "...",
      "content": "...",
      "url": "...",
      "media_type": "image|video|text",
      "engagement": {"likes": 0, "comments": 0, "shares": 0},
      "topics": [],
      "ts": "..."
    }
  ],
  "feed_metadata": {
    "scroll_depth": 30,
    "duration_ms": 12000
  }
}

AlgorithmProfile shape

data/profiles/<platform>/latest.json:

{
  "platform": "instagram",
  "synthesized_at": "2026-04-16T12:00:00Z",
  "source_scan_id": "abc123",
  "lean_score": 0.3,
  "echo_index": 0.7,
  "dominant_topics": ["cooking", "travel", "politics"],
  "top_authors": [{"handle": "...", "weight": 0.12, "bias_contribution": 0.04}],
  "mirrors": ["accounts amplifying existing views"],
  "doors": ["accounts that would broaden the feed"],
  "algorithm_signals": {
    "recency_bias": 0.6,
    "engagement_bias": 0.8,
    "topic_concentration": 0.5
  }
}

AgenticHarness shape

data/harnesses/<platform>/latest.json:

{
  "platform": "instagram",
  "generated_at": "2026-04-16T12:00:00Z",
  "source_profile_synthesized_at": "2026-04-16T12:00:00Z",
  "tools": [
    {
      "name": "analyze_feed_bias",
      "description": "...",
      "input_schema": {...},
      "analysis_prompt": "Given this profile context: <embedded_profile>, analyze..."
    }
  ],
  "system_prompt": "You are analyzing an Instagram feed with the following algorithmic profile: ..."
}

The system_prompt field is the pre-embedded context loaded into a mesh-analyst sub-agent.

Scan -> profile -> harness -> analysis pipeline

1. Scan: send a scan command to the extension. Wait for scan_complete. The raw scan lands in data/scans/<platform>/.
2. Profile: read the scan file. Use Sonnet to synthesize a structured AlgorithmProfile. Write it to data/profiles/<platform>/latest.json and a timestamped copy.
3. Harness: from the profile, generate an AgenticHarness — a set of tool definitions with embedded profile context as analysis prompts. Write to data/harnesses/<platform>/latest.json.
4. Analysis: spawn one or more mesh-analyst sub-agents. Pass the harness system_prompt as their system context. Each analyst answers a specific question (bias, topic mix, echo index, etc.).
5. Report: the orchestrator receives analyst outputs and synthesizes an editorial report. Write to data/reports/<timestamp>.md.

Model routing

• Haiku: classification tasks within the orchestrator (routing, intent detection, format validation)
• Sonnet: profile synthesis, harness generation, individual analysis steps
• Opus: final editorial report synthesis when depth is requested

Supported platforms

instagram, youtube, x, facebook, github

Adding a new platform

1. The extension must have a DOM scraper registered for the platform. That is outside Claude Code's scope.
2. Once the extension supports the platform, the scan -> profile -> harness pipeline works without changes.
3. Add the platform name to the supported list in this skill.
4. Run /mesh-scan <platform> to generate its first profile and harness.

Slash commands

• /mesh — natural-language entry point and router
• /mesh-scan <platform> — trigger extension scan, synthesize profile and harness
• /mesh-bias — load latest profile, run analysis, produce bias report
• /mesh-status — list available profiles per platform and their ages
• /mesh-tools — list generated harness tools for all platforms
• /mesh-compare — compare profiles across platforms or across time