[jthingelstad]

Refs #630 (does not close it).

Jamie, 2026-10-21: "Approved: add an ops command to elixir-mcp and rebuild the battle table's indexes concurrently in production after the 10-23 maintenance".

What

A migrate op, {reindex} (services/migrate/src/ops-reindex.mjs). REINDEX CONCURRENTLY cannot run inside the migrate runner's transaction, so it is an op, as {vacuum} is.

• Report (read, no lock): {"reindex": {"table": "battle"}}. Returns each index's size, scans, validity, the constraint it backs, the foreign keys that use it, and a bloat estimate. Also the heap size, the row estimate, and every invalid index, with _ccnew/_ccold leftovers marked.
• Rebuild (WRITE, production lock): {"reindex": {"table": "battle", "index": "<name>", "apply": true}}. Runs REINDEX INDEX CONCURRENTLY on exactly one of the table's indexes, by quoted name. statement_timeout is 240 s (the Lambda times out at 300 s) and lock_timeout is 60 s. Returns the size before and after, ms, and validity after. It is refused while any index on the table is invalid.
• Leftovers (WRITE): {"reindex": {"table": "battle", "drop_invalid": "<name>"}}. Runs DROP INDEX CONCURRENTLY on one INVALID index. It is refused for a valid index or for one a constraint needs.
• battle is the only table on the allowlist. npm run op treats only the bare {table} form as a read.

The bloat estimate needs no extension. It uses reltuples, the pg_stats widths and the fillfactor. On test data it came within 7% of every freshly rebuilt index.

Docs

• A row in ops.md beside {vacuum}, and a line in the ops skill's "Choosing the read".
• docs/notes/2026-10-21-battle-reindex-op.md.
• Two DECISIONS lines, both [Jamie, 2026-10-21]: shared_buffers follows RDS's own setting and nothing pins it (Operations and AWS; RDS warned of 88 MB at a restart, but after the 10-21 restarts it was still 180 MB); battle's indexes are rebuilt concurrently by {reindex} (Schema and migrations).
• infra/template.yaml is unchanged: its comment that shared_buffers is the engine default is still true.

Tests

services/migrate/test/ops-reindex.test.mjs, on a scratch database:

• a bloated battle_pkey is rebuilt valid and smaller, and both foreign keys follow it;
• a real _ccnew leftover comes from a rebuild cancelled behind an open writer; rebuilds are refused while it stands, drop_invalid refuses the valid index and drops the leftover, and the rebuild then succeeds;
• every refusal is covered;
• the timeouts are held under the template's Lambda timeout.

npm run verify is green on Node 24 (32/32 units). e2e-needed.sh says false.

Deploy

No tool changed, so there is no acceptance family. The rebuild itself is not part of this PR. It runs after the 10-21 maintenance, one index per invocation, smallest first, when Jamie says.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-21T11:42:57.143979Z	00ac016	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[jthingelstad]

Deployed to production 2026-10-21T11:57:35Z (2026-10-21 06:57:35 CDT).

• Commit: 2dab9ab6 (platform lane); this PR's: bfe8793c, 2dab9ab6
• MCP contract: 11.13.3
• JSON API: 3.2.1
• Acceptance: not run
• Timing: started 2026-10-21T11:56:36Z (2026-10-21 06:56:36 CDT); the CI gate waited 2 s, then 58 s to this record: build 2 s, migrate-push 4 s, stack 31 s, grants 2 s, public-stats 1 s, card-art 3 s, site 4 s, smoke 5 s