vector-conflict-check

name	vector-conflict-check
description	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(hash mismatch), and surfaces vectors whose content contradicts current ground-truth (stale claims, closed engagements, outdated targets).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.
disable-model-invocation	false
allowed-tools	Read, Write, Bash(sha256sum *), Bash(test *), Bash(date *), Bash(mkdir *), mcp__gateway__pinecone__describe-index-stats, mcp__gateway__pinecone__search-records

Vector Conflict Check

When to run

• After bulk edits to ground-truth or CRM.
• Periodically (weekly) when running on Pinecone.
• Before relying on semantic-recall for high-stakes calls (offer comparisons, pivot decisions).

Procedure

1. Resolve config + guard

${CAREER_DATA_DIR}/config.json → WORKING_FOLDER, CONTEXT_STORE, PINECONE_INDEX. Bail if CONTEXT_STORE != pinecone.

2. Pull index inventory

Use the Pinecone MCP to enumerate vector IDs + metadata for the configured index. Each vector should carry metadata: source_path, content_hash, embedded_at, namespace (set by pinecone-sync).

If metadata is missing on a meaningful fraction of vectors, bail with: "index lacks source-path metadata — run /career:pinecone-sync --reindex first."

3. File-level checks

For each vector:

• Source-deleted → file path no longer exists in WORKING_FOLDER. Propose vector delete.
• Source-renamed → no exact-path match but a same-hash file exists elsewhere. Propose metadata update (cheap) over delete+re-embed (expensive).
• Content-drift → file exists at same path; sha256 mismatch with content_hash metadata. Propose re-embed (delete + upsert via pinecone-sync).
• Clean → path matches, hash matches. No action.

4. Semantic conflict checks

Read ${WORKING_FOLDER}/ground-truth.md and extract assertion-shaped lines (looking-for, hard-constraints, location, currency, current-stage). For each assertion, search the vector index (search-records) with the assertion as query, top_k=10. For matches with high similarity (≥0.85) but semantic contradiction, flag.

Heuristic contradictions to detect (let the model judge, no regex needed):

• Vector says "looking for full-time" while ground-truth says "consulting only".
• Vector says "active engagement with X" while CRM marks X as closed.
• Vector contains a location/currency that no longer matches ground-truth.

5. Write findings

${WORKING_FOLDER}/reconcile/<YYYY-MM-DD>-vector-conflicts.md. Sections:

# Vector conflicts — {{date}}

## File-level
- DELETE  vec_id=...  source=research/companies/acme/...  reason=source-deleted
- REEMBED vec_id=...  source=ground-truth.md              reason=hash mismatch (last embedded 2026-01-12)
- RENAME  vec_id=...  source=domain-notes/foo.md → bar.md reason=path moved

## Semantic
- vec_id=...  excerpt="…still in talks with Acme…"  conflicts with: crm/companies.md:Acme=closed (2026-02-14)

## Proposed actions
- 12 deletes, 3 re-embeds, 1 rename, 4 semantic flags for user review

6. Apply (with explicit confirm)

AskUserQuestion per action class (delete batch / re-embed batch / rename batch / semantic flags). Semantic flags never auto-resolve — they're for the user to decide whether to update ground-truth, update the source doc, or keep both views.

Re-embeds delegate to pinecone-sync --paths=<list>.

Idempotency

Re-running after a clean apply produces an empty findings file (or skips writing).

Failure modes

• Pinecone MCP missing → bail with: "install via /career:detect-mcps or /career:install-companion-plugins --only=private-misc."
• Index empty → exit cleanly with "no vectors to check."
• Network error mid-scan → write partial findings, mark as incomplete: true in frontmatter.