Fix over-broad rebuild live-server interlock: lsof-scoped, warn-by-default (#1138)

[jeonghun-jj-lee]

What

Fixes the over-broad running-server interlock (OB12) in the rebuild scripts that refused a rebuild when the chat server — or any unrelated opencode process — was running, even though nothing held this rebuild's DB.

Why

The interlock is new in #1135 — the pre-S9 scripts and the in-app Rebuild path never blocked on a live server. It used pgrep -f 'opencode.*serve', which matches any process whose args contain "opencode" then "serve" (including the Amicode server you're chatting through), so it false-refused. Reported live:

==> Rebuild mode: main ... ERROR: A live opencode server appears to be running — close the Extension Dev Host ...

Change (scripts/rebuild_amicode.sh)

• lsof-scoped: the check now runs lsof -t against the actual target DB files ($DBDIR/opencode*.db + -wal/-shm), excluding this script's own pid+parent (a & job shares the parent pgid, so no pgid filter — a real server runs in its own group). No lsof → silent pass.
• Warn-by-default (restores old-script + in-app parity): a genuine holder prints a one-line warning and continues. db_is_zeroed + WAL-checkpoint + restore-only-if-zeroed already guard a torn backup.
• --strict-live-check opts into the hard refuse; --allow-live-server silences the warning.
• Added a --check-live-only hook for isolated testing.
• Drive-by: fixed a dirty-files=0\n0 double-print in the deploy-transparency line.

Tests / gates (all command-run)

• TDD: new AC10 in rebuild_amicode_check.sh — (a) holder → warn+exit 0; (b) --strict → refuse; (c) unrelated opencode…serve-named process holding no target DB → not tripped (regression guard for the exact false positive); (d) no lsof → pass. RED first (failed on current code), then GREEN.
• Harness: 12/12 pass, 0 fail, 1 skip (AC9 heavy build opt-in).
• bash -n = 0 · shellcheck (all 4 scripts) = 0 findings · fork-refs = 0.
• Both scripts ran end-to-end on this machine: rebuild_amicode_locally.sh --yes (build→deploy→binary copy→done) and rebuild_amicode_remotely.sh --yes on a clean tree (full --mode main sync→build→deploy→done); the dirty-tree refuse still fires correctly when the tree is dirty.

Scope

Scripts only — the in-app path never had this interlock, so no coordinator.ts/chat_bridge.ts change; the two surfaces stay consistent (neither hard-blocks on a live server).

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}