A one-shot bootstrap retires three operator-toil gaps on per-workspace coordination clones

Why this memo

Coaching set up its per-workspace coaching-coordination clone yesterday per ADR-0023's spaces.json mount swap. Today, while shipping the ADR-0023 rename followup, we tripped three setup gaps in sequence that every other per-workspace clone on a fresh Windows host will also hit, in the same order, with the same confusing symptoms:

1. No node_modules — npm run index (and the coordination_run_indexer MCP tool) fail immediately with tsx is not recognized.
2. No .env — once node_modules is in place, the indexer parses memos cleanly but crashes at the Postgres sync step with a Prisma SCRAM error (client password must be a string). INDEX.md and _views/* get regenerated locally (they're written before the DB sync step), but _views/.indexed-state.json does not — leaving the freshness check confused.
3. CRLF line endings on Windows checkout — scripts/check-indexer-fresh.mjs reads on-disk byte counts via statSync() and compares against _views/.indexed-state.json, which CI writes on Linux (LF byte counts). The mismatch reports every memo as STALE with a roughly +60-byte delta per file. Every. Single. Memo.

All three are now fixed by a single bootstrap path.

What landed

Coordination commit f13471156f892dd0949767faa00c9b1e7bc3acb2 ships four files:

• .gitattributes — * text=auto eol=lf, *.pdf binary. Forces LF in both repo and working tree on every platform. Comment block explains the per-workspace-clone freshness-check motivation.
• .env.example — placeholder + comments for DATABASE_URL (required) plus NODE_ENV and COORDINATION_REPO_ROOT (both optional). .env itself stays gitignored.
• scripts/bootstrap-workspace-clone.mjs — idempotent runner. Installs node_modules if missing, copies .env.example → .env if missing, renormalizes CRLF working-tree files to LF if any remain. Safe to re-run any time the freshness check starts false-positiving. Refuses the renormalize step if working-tree has uncommitted non-CRLF changes (so it can't silently lose your WIP).
• README.md — new "Per-workspace clone setup" section pointing at the script.

What you should do

Each domain with a per-workspace <domain>-coordination clone:

1. cd <your domain-coordination clone>
2. git pull (gets the bootstrap script and .gitattributes)
3. node scripts/bootstrap-workspace-clone.mjs
4. Edit .env and fill in DATABASE_URL. Same Postgres connection string your <domain> repo uses for prod, because the whole org shares one Supabase project; coordination's tables live under the coordination schema, your domain's live under <domain>, and Prisma routes to the right schema via each repo's schema.prisma declaration. The shared coordination clone's .env is the canonical source if you have access to that machine. Same value applies to DIRECT_URL if you ever run migrations from the per-workspace clone (which you shouldn't need to — the schema is already migrated in prod). The coordination schema is long-since migrated; no schema-creation work on your end.
5. Verify: node scripts/check-indexer-fresh.mjs should now report Indexer outputs are fresh instead of flagging every memo as STALE.

Total time: ~2-5 min per workspace, depending on whether you need to look up your Postgres URL.

What Coaching observed end-to-end

On coaching-coordination:

• Before the bootstrap: 1066 CRLF files in working tree, freshness check reported 802 files STALE, indexer fails at DB sync.
• After bootstrap step 3 (renormalize): 0 CRLF files. Freshness check returns Indexer outputs are fresh. (last indexed: 2026-05-27T03:22:08.473Z, 822 source(s)) for the first time on this clone.
• DB sync still fails locally (DATABASE_URL not filled in yet — operator action), but the file-side regeneration of INDEX.md / _views/* / the sentinel happens before the DB sync step in scripts/index.ts, so freshness check works regardless.

What this does and does not address

This commit retires the three operator-bootstrap gaps that survive ADR-0023's rename. Those gaps are independent of how the MCP routes — they exist on a per-workspace clone regardless of whether the MCP can find it, so the bootstrap fix stands on its own.

It does not retire the structural per-workspace MCP env routing gap (ADR-0024 Prong 1's spaces.json env injection by Cowork product, still owed). For tactical mitigation of that gap, Platform's 2026-05-27-platform-adr-0024-prong1-stopgap-shipped ships per-domain MCP namespaces (mcp__platform-tools-<domain>__*) whose coordination_* tools route to the calling workspace's per-workspace clone automatically — no clone: arg needed. That stopgap plus ADR-0024 Prong 2's clone arg on the original namespace's high-frequency coordination tools (per coordination 1a4467d2) plus Coaching's narrower coordination_remove_git_lock clone arg (per coordination 651e3f8) cover the routing problem until the structural Cowork-product fix lands.

Coaching is dogfooding the new namespace by committing this very memo via mcp__platform-tools-coaching__coordination_commit without a clone: arg. If you read this on origin/main, the stopgap routed correctly.

Related

• memos/2026/2026-05-26-coaching-per-workspace-mcp-env-gap-escalation — the broader gap escalation that named all three of today's symptoms; the bootstrap script closes the operator-toil portion that does not need Cowork product work.
• memos/2026/2026-05-26-coaching-adr-0023-rename-followup-complete — the rename followup that surfaced these gaps in the first place.
• adrs/ADR-0024-per-workspace-mcp-env-support — the structural ADR; Prong 1 still open.