Skill: engram-self-index

Turn engram's code-intelligence on this repository itself, verify what it reports against the real source, and register the surviving findings.

The store

~/.engram/mem-alpha-self (project mem-alpha-self) — dedicated, this repo ONLY, deliberately kept across sessions so a second scan activates whats_changed temporal discrimination (it is flat on a single baseline). Do not wipe it casually; a wipe resets the temporal baseline.

Procedure

1. Re-scan (cheap — manifests make it incremental):

python3 scripts/dev/mcp_driver.py --storage ~/.engram/mem-alpha-self \
  --project mem-alpha-self --no-vector \
  call scan_repo '{"path": "/home/videogamer/projects/mem-alpha"}'

Expect sub-second on unchanged trees; errors: 0 always (empty files skip, minified assets filter — any error is a regression).

2. Run the battery:

python3 scripts/dev/mcp_driver.py --storage ~/.engram/mem-alpha-self \
  --project mem-alpha-self --no-vector \
  call architecture '{"limit": 10}' \
  call code_health '{}' \
  call whats_changed '{}' \
  call change_impact '{"target": "<top-bridge-symbol>", "depth": 2}'

Read it like this:

• Bridges (betweenness) = chokepoints/god-modules. A symbol ~5× the next is a named refactor target; check its blast radius before scoping the split.
• Central (PageRank) = coupling magnets. Short generic names (cn, run, get) are centrality artifacts, not findings.
• Dead code = leads, NEVER verdicts. This repo's list is dominated by test doubles (Failing*, Fake*, Stub*), prototype UI, and public-API surface — all legitimately uncalled in-repo.

3. VERIFY against source before believing anything. The 2026-08-28 run flagged FastEmbedEmbeddingProvider dead — grep proved it called from BOTH backends (qualified-path + self-method calls under-resolve; see the [receiver-resolution-gap] backlog item). For every dead-code candidate:

grep -rn "<SymbolName>" --include="*.rs" --include="*.ts" | grep -v "fn <name>"

If real callers exist, it is a resolution bug — add the reproduction to the backlog, do not "fix" the code.

4. Register survivors in docs/backlog.md under self-index-findings with the measured numbers, and store a summary via the harness engram MCP (workspace agentzero).

Cadence

Run after significant merges to main (the repo convention) or on request. Each run also refreshes the temporal baseline — after ≥2 scans, prioritize whats_changed output: recent-churn × high-blast-radius is where review effort pays best.