菜单
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.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Sweep all (or a filtered subset of) email addresses in the workspace CRM and re-verify them via Hunter Email-Verifier. Use periodically — people change jobs, addresses go stale, accept-all status changes. Updates the `verified` and `last-verified` columns in `crm/contacts.md` and prints a summary of what flipped.
Scan the user's existing Claude Code plugins and MCP servers to detect capabilities the career plugin needs (email send, transcription, calendar/tasks, semantic store, CRM). For each capability, decide whether to reuse an already-installed plugin/MCP, install a recommended companion plugin, or add an MCP server. Writes the resolution map to ${CAREER_DATA_DIR}/config.json under a `capabilities` block. Invoked from /career:onboard before install-companion-plugins so we don't double up on capabilities the user already has.
Enrich an existing CRM contact with Hunter Combined-Enrichment — pulls role, seniority, department, social handles, and the contact's company facts in one call. Use to flesh out a thinly-populated CRM row before drafting outreach or prepping for a meeting. Operates on an email address from the CRM; updates the row in place.
Given a person's name and their company (name or domain), return Hunter's best guess at their email address with a confidence score. Use when the user already knows who they want to reach and just needs the address. Differs from find-contact (which lists candidates at a company) — this is targeted at a single named person.
Validate a single email address via Hunter Email-Verifier. Use when the user has an address (handed to them, scraped, guessed) and wants a deliverability + confidence read before sending outreach. Returns Hunter's verdict (deliverable / risky / undeliverable / unknown), score, and the disposable / role / accept-all flags. Does not write to the CRM unless asked.
For CONTEXT_STORE=pinecone — verify the index against the local workspace. Flags vectors whose source file was deleted, renamed, or content-changed (hash mismatch), and surfaces vectors whose content contradicts current ground-truth (stale claims, closed engagements, outdated targets). Proposes deletes/re-embeds; never silently mutates the index.
| 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 *) |
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.
$ARGUMENTS:
--full — re-embed everything (rare; use after schema or model change).--dry-run — list what would be re-embedded, no writes.private-misc plugin installed (provides Pinecone MCP wiring). If missing, bail with: "install private-misc first via /career:onboard companions".PINECONE_INDEX, CONTEXT_STORE=pinecone.WORKING_FOLDER.${WORKING_FOLDER}/.pinecone-sync.json — manifest of {path, sha256, mtime, vector-id}. Create if missing.PINECONE_INDEX from config.Include:
ground-truth.mdcrm/*.mddomain-notes/*.mdrecommendations/*.mdresearch-briefs/*.mdplans/*.mdmeetings/*.mdinbox/*.md (only status: routed entries — drafts churn too fast)Exclude: cache/, _salvage/, .pinecone-sync.json, hidden files.
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.
Chunking strategy:
## headers; keep heading + body as a single chunk; chunks > 2000 tokens further split on paragraph.<path>#<chunk-index> (stable across runs).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.
Use Pinecone MCP upsert-records (namespace = workspace name) with the chunks. Delete vectors for files removed since last sync.
Rewrite .pinecone-sync.json with new {path, sha256, mtime, vector-ids} per file.
synced: <N> files, <M> chunks
unchanged: <U> files
deleted: <D> vectors
index: <PINECONE_INDEX>
manifest: .pinecone-sync.json
inbox/* entries that are still pending-route — they're noisy and short-lived.--full re-embeds everything; warn the user about cost before running.create-index-for-model, then proceed.