pinecone-sync

name	pinecone-sync
description	Incremental sync of working-folder/ into a Pinecone index for semantic recall. Embeds only changed files (mtime + content hash). Index name from config. Delegates first-time MCP setup to private-misc:personal-context-injection. Idempotent.
disable-model-invocation	false
allowed-tools	Read, Write, Bash(find *), Bash(sha256sum *), Bash(stat *), Bash(mkdir *), Bash(date *)

Pinecone Sync

Embeds the workspace into Pinecone so semantic-recall can pull in relevant context for downstream skills (recommendation, meeting-prep, ground-truth deltas). Skips files that haven't changed since the last sync.

Inputs

$ARGUMENTS:

• Optional: --full — re-embed everything (rare; use after schema or model change).
• Optional: --dry-run — list what would be re-embedded, no writes.

Prerequisites

• private-misc plugin installed (provides Pinecone MCP wiring). If missing, bail with: "install private-misc first via /career:onboard companions".
• Config keys: PINECONE_INDEX, CONTEXT_STORE=pinecone.

Procedure

1. Load state

• WORKING_FOLDER.
• ${WORKING_FOLDER}/.pinecone-sync.json — manifest of {path, sha256, mtime, vector-id}. Create if missing.
• PINECONE_INDEX from config.

2. Walk working-folder

Include:

• ground-truth.md
• crm/*.md
• domain-notes/*.md
• recommendations/*.md
• research-briefs/*.md
• plans/*.md
• meetings/*.md
• inbox/*.md (only status: routed entries — drafts churn too fast)

Exclude: cache/, _salvage/, .pinecone-sync.json, hidden files.

3. Diff against manifest

For each file: compute sha256. If hash differs from manifest (or file is new), mark for re-embed. If --full, mark all.

Mark deletions (in manifest, not on disk) for vector deletion.

4. Chunk + embed (proposed)

Chunking strategy:

• Markdown files: split on ## headers; keep heading + body as a single chunk; chunks > 2000 tokens further split on paragraph.
• One vector per chunk. Vector id: <path>#<chunk-index> (stable across runs).
• Metadata: path, section, mtime, type (where type ∈ ground-truth | crm | domain-note | recommendation | brief | plan | meeting | inbox).

If --dry-run: print the chunk plan and stop.

5. Upsert

Use Pinecone MCP upsert-records (namespace = workspace name) with the chunks. Delete vectors for files removed since last sync.

6. Update manifest

Rewrite .pinecone-sync.json with new {path, sha256, mtime, vector-ids} per file.

7. Print summary

synced: <N> files, <M> chunks
unchanged: <U> files
deleted: <D> vectors
index: <PINECONE_INDEX>
manifest: .pinecone-sync.json

Guardrails

• Never embed inbox/* entries that are still pending-route — they're noisy and short-lived.
• --full re-embeds everything; warn the user about cost before running.
• Manifest is the source of truth for "what's already in the index"; don't query Pinecone to figure that out.

Failure modes

• Pinecone MCP not installed → bail with install instructions.
• Index doesn't exist → offer to create via MCP create-index-for-model, then proceed.
• File unreadable → log and skip; don't abort the whole run.