etanhey/google-drive-archive

/google-drive-archive — Brain Drive as Source of Truth

Pattern: Brain Drive holds forever. docs.local/ holds latest only. BrainLayer indexes both. A _DRIVE-LEDGER.md per docs.local/ subdir maps local → Drive.

This skill is for putting things INTO Drive. To query Drive content via Gemini, use /braindrive. To research the web with Drive sources, use /gemini-research.

---

Source-of-Truth Hierarchy

┌──────────────────────────────────────────────────────────────┐
│  BRAIN DRIVE                                                 │
│  ────────────                                                │
│  • Forever storage (free up to 15GB; user has the quota)     │
│  • All versions, heavy assets, raw research outputs          │
│  • Path: Brain Drive/06_ARCHIVE/<area>/<topic>/<date>/...    │
│  • Source of truth — if it's only in docs.local/, it isn't   │
│    durable                                                   │
└──────────────────────────────────────────────────────────────┘
                         ▲
                         │  upload + update ledger
                         │
┌──────────────────────────────────────────────────────────────┐
│  docs.local/  (git-ignored)                                  │
│  ───────────                                                 │
│  • Local cache — only the LATEST versions                    │
│  • Lightweight markdown only (no >100KB media)               │
│  • Per-dir `_DRIVE-LEDGER.md` lists every Drive copy + URL   │
│  • Safe to delete: contents are recoverable from Drive       │
└──────────────────────────────────────────────────────────────┘
                         ▲
                         │  brain_digest extracts entities/relations
                         │
┌──────────────────────────────────────────────────────────────┐
│  BRAINLAYER                                                  │
│  ──────────                                                  │
│  • Searchable index over text content (transcripts, docs)    │
│  • Chunks link back to Drive URLs                            │
│  • Use brain_search before brain_store — dedupe the index    │
└──────────────────────────────────────────────────────────────┘

One-line rule: Heavy or forever → Drive + ledger entry + (if textual) brain_digest. Local copies are caches, not archives.

---

Triggers

Invoke this skill when:

1. Pre-write hint: about to save anything >100KB to docs.local/research/, docs.local/audits/, or docs.local/plans/ — especially if it contains transcripts, audio, video, or large logs.
2. Explicit user request: "save this to Drive", "archive this", "this should be in Brain Drive", "put this in the archive".
3. Post-digest: a research output has been brain_digest'd into BrainLayer and the raw file is a candidate for archival (BrainLayer indexes it; Drive holds the original).
4. /tmp cleanup: about to rm -rf /tmp/<heavy-stuff> — if any of it is forever-knowledge, route it through this skill first.

Skip this skill when:

• The artifact is genuinely ephemeral (CLI test output, throwaway diff, audit prompt that's already digested).
• The content is already covered by another archive flow (e.g. PR merges live in git, not Drive).
• File is <100KB markdown that's the canonical doc kept in docs.local/ (no need to upload markdown notes).

---

Workflow

1. Decide: ephemeral or forever?

Ask: "Will future-Etan or another agent want to re-read this raw artifact a month from now?"

| Answer | Action | |---|---| | Yes (transcripts, audits, plans, research outputs, media) | Forever → continue to step 2. | | No (intermediate processing files, throwaway logs, regeneratable derivatives) | Skip. Just delete after use. Note the choice in commit message or BrainLayer. | | Mixed (e.g. audio.wav keep, audio-16k.wav derivative skip) | Upload only the originals. Document which derivatives are regeneratable + how. |

2. Resolve the Drive archive path

Default convention (matches Brain Drive/06_ARCHIVE/repos/union-2026-04/ precedent):

Brain Drive/06_ARCHIVE/<area>/<topic>/<date-or-tag>/<artifact-folder>/

| <area> | When | |---|---| | research | Web research, deep research, video gems, audits-as-research | | audits | Code audits, security reviews, compliance reports | | plans | Multi-phase plans (large-plan outputs), architecture proposals | | repos | Whole-repo snapshots (e.g. before a destructive refactor) | | voice | Voice session transcripts + audio |

Use mcp__google-drive__search to find existing parent folders before creating new ones — match precedent if one exists.

3. Create the folder tree

mcp__google-drive__createFolder(name=<level>, parent=<parent-id>)

Capture the returned folder ID at every level. You will need them in the ledger.

4. Upload the artifacts

mcp__google-drive__uploadFile(localPath=<absolute>, parentFolderId=<id>)

• Parallel for files in the same folder is safe.
• Big files (>100MB): the MCP handles them; just call once and wait.
• Capture id, link, and size from each upload return.

5. brain_digest text content

For every transcript.txt, report.md, audit.md, plan.md you upload:

mcp__brainlayer__brain_digest(content=<file content>)

Capture the returned chunk_id. This makes the full text searchable — gem extractions and structured chunks cover the highlights, but brain_digest is what catches phrases the gem extractor missed.

For files >150KB, an inner agent with brain_digest loaded can do this in the background (large content blows up the orchestrator's context).

6. Write / update the local ledger

Path: docs.local/<area>/<topic>/_DRIVE-LEDGER.md

# Drive Ledger — <topic>

> Source of truth for these artifacts is Brain Drive. Local files are caches.

| File | Local path | Drive folder | Drive ID | Drive URL | Size | BrainLayer chunk | Uploaded |
|---|---|---|---|---|---|---|---|
| transcript.txt | docs.local/.../transcript.txt | <folder-id> | <file-id> | <link> | 30KB | brainbar-... | 2026-04-25 13:01 |
| audio.wav | (deleted from /tmp after upload) | <folder-id> | <file-id> | <link> | 312MB | n/a | 2026-04-25 13:08 |

## Skipped (regeneratable derivatives)
- `audio-16k.wav` — 16kHz mono downsample of `audio.wav` for whisper-cli. Regenerate via `ffmpeg -i audio.wav -ar 16000 -ac 1 audio-16k.wav`.

The ledger is what lets a future agent answer "where is the original?" without spelunking through Drive.

7. (Optional) Delete the local heavy copy

Only after:

• mcp__google-drive__listFolder confirms the file is present and sized correctly.
• Ledger entry is written.

Markdown notes <100KB stay in docs.local/ (they're the latest cache). Heavy media (>10MB audio/video) can be deleted locally — Drive is the canonical copy.

8. brain_store the decision (optional, for non-trivial archives)

brain_store(
  content="Archived <topic> to Brain Drive at <path>. Rationale: <why forever>. Skipped: <derivatives + how to regenerate>.",
  tags=["google-drive-archive", "<area>", "<topic>", "milestone"],
  importance=6-8
)

Skip for routine archives (one transcript, one report). Store for unusual decisions: large batches, format choices, retention policy edges.

---

Anti-patterns

| Anti-pattern | Why it's wrong | Do instead | |---|---|---| | Saving 1.5GB transcripts to docs.local/ and committing | Bloats the working dir, hits git ignore edges, no durable copy | Drive + ledger; docs.local/ keeps the markdown summary only | | rm -rf /tmp/extract/ without checking forever-value | Loses raw audio/video that's expensive to regenerate (yt-dlp re-downloads, whisper re-runs) | This skill: route through "ephemeral or forever?" decision first | | brain_digest a transcript but NOT upload to Drive | BrainLayer has the index but the raw file is gone — citations break, full text unrecoverable | Upload to Drive AND brain_digest. Both. | | Upload to Drive but skip the ledger | Drive becomes a junk drawer; future agents can't find the URL | Always update _DRIVE-LEDGER.md | | Re-uploading something that's already in Drive | Wastes quota, confuses search | mcp__google-drive__search first; match existing folder if precedent exists | | Committing _DRIVE-LEDGER.md to git | Ledger is a per-machine cache; the source of truth is Drive | Keep in docs.local/ (which is git-ignored) |

---

Composability

• /research-lifecycle — when condensing/archiving stale research files, use this skill to push originals to Drive before deleting locally.
• /qa-video — video gems and QA processing produce transcripts, frames, and heavy media; this skill is where they land long-term. The qa-video gems output is in /tmp/ by default; route to Drive on completion.
• /large-plan — when a plan completes, archive the full plan folder to Brain Drive/06_ARCHIVE/plans/<name>-<date>/.
• Nightly sweep — scan docs.local/ for >100KB files lacking ledger entries and prompt to archive.

---

Related Tools (reference, not re-document)

• mcp__google-drive__* — search, createFolder, uploadFile, listFolder. The braindrive MCP is already authenticated; for auth drift run nlm login switch <profile> && mcp__google-drive__refresh_auth (this is the repo's established remediation, used by skills/golem-powers/_shared/research/verify-account.sh). authGetStatus is a status check, not a fix.
• mcp__brainlayer__brain_digest — for textual content >2KB.
• mcp__brainlayer__brain_store — for decisions about archives.
• mcp__brainlayer__brain_search — to find existing chunks before re-digesting.

---

Output Format (when reporting an archive operation)

## Archived: <topic>
- Drive: Brain Drive/06_ARCHIVE/<path> (folder ID: <id>)
- Files: N text + M media (X MB / Y GB total)
- BrainLayer chunks: <chunk_id_1>, <chunk_id_2>
- Ledger: docs.local/<area>/<topic>/_DRIVE-LEDGER.md
- Local cleanup: deleted /tmp/<...> (X GB freed) | retained markdown (<size>)
- Skipped derivatives: <list> (regenerate via <command>)

Keep it ≤10 lines. Detail goes in the ledger.