{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreiguxpn323uy3kccgawtsnkyu4c5fsmhfmft63v5beyku47rdzr7ki",
    "uri": "at://did:plc:lk3jfj3zq4k4wxnk474axylu/app.bsky.feed.post/3mfoc25mkkh42"
  },
  "path": "/t/show-preventing-doc-drift-in-agentic-coding-workflows/1375031#post_1",
  "publishedAt": "2026-02-25T07:43:59.000Z",
  "site": "https://community.openai.com",
  "tags": [
    "github.com",
    "GitHub - amsminn/vericontext: Deterministic, hash-based verification for docs..."
  ],
  "textContent": "I’ve been using a “deep init” style setup with coding agents —\neach directory has its own README.md / AGENTS.md describing roles and constraints.\n\nIt works well early on.\n\nBut as the codebase grows, those docs inevitably drift,\nwhile the agent still treats them as ground truth.\n\nThat creates a subtle reliability issue:\nthe model is following instructions — they’re just stale.\n\nTo experiment with this, I built a small tool called VeriContext.\n\nWhen documentation references code, it embeds a SHA-256 hash of the exact snippet (inside an HTML comment, so rendered Markdown stays clean).\nOn verification, either the hash matches or it fails (fail-closed, no fuzzy matching).\n\nIt can run:\n• as an agent skill (verify before finishing a task / committing), or\n• via pre-commit / CI (e.g. npx vericontext verify workspace …)\n\nThe goal isn’t better prompting — it’s constraining documentation to stay aligned with code.\n\nI’m curious how others are handling doc / context drift when using Codex or other coding agents.\n\nWould strict fail-closed verification feel too rigid in practice?\nHas anyone tried AST-aware approaches instead of raw snippet hashing?\n\ngithub.com\n\n### GitHub - amsminn/vericontext: Deterministic, hash-based verification for docs...\n\nDeterministic, hash-based verification for docs that reference code. Fail-closed. Zero fuzzy matching.",
  "title": "Show: Preventing doc drift in agentic coding workflows"
}