The scout stack goes live — console-only morning sequence.

1 · Merge gates per-PR go from Sam · green CI · in this order

Three PRs carry the stack. The order matters: #96 is the deploy vehicle (the oll-memory service the other two call), #95 layers the seam's scout verbs on top, #94 adds the scout agent service. All three were CI-green at prep time (2026-07-14). Per the house merge gate, each needs an explicit per-PR go — merging is itself a morning console act.

2 · The topology cut recommendation: minimal — memory + GUI up, seam + scout stay local

Two pieces of the stack are stateless callers (the seam, the OpenCLAW scout loop) and two are the things worth hosting (the corpus, the GUI). The minimal cut moves only the second pair to the VPS. The seam keeps running locally exactly as today — it just points its OLL_MEMORY_BASE at the deployed service instead of the laptop container. One moving piece less on deploy morning; the seam's containerization becomes a calm follow-up instead of a morning blocker (it has no Dockerfile yet — §10).

The follow-up (not tomorrow): Dockerfile for services/oll-mcp → Coolify app → OpenCLAW seam URL flips host.docker.internal:5097 → the deployed seam.

3 · Console: Neon DB + the Ollama embeddings app the two dependencies, first

3a — Neon: new database oll_memory

1. Neon console → the existing oll.am project (house rule: new DB on the existing project, never a new project) → New database: oll_memory.
2. Enable pgvector on it: CREATE EXTENSION IF NOT EXISTS vector; (the migration block in §6 includes this — the service's boot hook also attempts it, but don't rely on lazy creation for a prod deploy).
3. Copy the pooled connection string — it becomes oll-memory's DATABASE_URL.

3b — Ollama: one small app, embeddings only

1. Coolify → new app from image ollama/ollama, attach to network ollam, network alias ollama, persistent volume on /root/.ollama (models survive redeploys). No public domain — internal only.
2. After it's healthy, pull the embedding model once (Coolify terminal or SSH): docker exec <ollama-container> ollama pull nomic-embed-text.

4 · Console: oll-memory the standard Base-Directory pattern

Environment

5 · Console: ollscout separate repo — the Dossier GUI at scout.oll.am

ollscout lives in its own repo (bytesbysamu/ollscout, like ollwrite). The production Dockerfile is committed and verified (939d407): Next.js 15 standalone, multi-stage node:22-alpine, non-root, HEALTHCHECK on 127.0.0.1:$PORT/, 12-factor PORT (default 3000). Build + run were proven locally against the real corpus — HTTP 200 with live rail content, container health green.

Environment (the contract is documented in the Dockerfile itself)

6 · Terminal: corpus migration local poc-pg → Neon · recommended: dump/restore

Recommendation: pg_dump/pg_restore. The dump carries the embeddings (768-dim vectors, same data_memory_chunk table both sides), all 58 stored cases, and — the part re-ingest cannot reproduce — the accumulated rulings history (statuses, verdict notes, tick register). Zero re-embedding, minutes of work. Run everything through the poc-pg-1 container so client/server versions can't disagree:

# 0) the Neon URL for the new DB — export without echoing it anywhere
export NEON_MEM_URL='postgresql://…neon…/oll_memory?sslmode=require'

# 1) dump the local corpus (custom format, portable)
docker exec poc-pg-1 pg_dump -U mem -d oll_memory -Fc --no-owner --no-privileges -f /tmp/oll_memory.dump

# 2) pgvector first — the restore fails without the extension
docker exec -e URL="$NEON_MEM_URL" poc-pg-1 sh -c 'psql "$URL" -c "CREATE EXTENSION IF NOT EXISTS vector;"'

# 3) restore into Neon
docker exec -e URL="$NEON_MEM_URL" poc-pg-1 sh -c 'pg_restore -d "$URL" --no-owner --no-privileges /tmp/oll_memory.dump'

# 4) the retrieval knob that does NOT travel with a single-DB dump —
#    re-apply it or prod retrieval silently degrades vs. the audited local setup
docker exec -e URL="$NEON_MEM_URL" poc-pg-1 sh -c 'psql "$URL" -c "ALTER DATABASE oll_memory SET hnsw.ef_search = 80;"'

# 5) sanity — chunk count should match the local side
docker exec -e URL="$NEON_MEM_URL" poc-pg-1 sh -c 'psql "$URL" -tc "SELECT count(*) FROM data_memory_chunk;"'
docker exec poc-pg-1 psql -U mem -d oll_memory -tc "SELECT count(*) FROM data_memory_chunk;"

7 · Re-point the local seam the only OpenCLAW-side change in this cut

In the minimal cut the seam and OpenCLAW don't move — OpenCLAW keeps calling host.docker.internal:5097 exactly as today. The single change: the seam's environment (launchd plist / deploy/run-http.sh env) flips its memory base to the deployed service, with the fresh token:

OLL_MEMORY_BASE=https://memory.oll.am
OLL_MEMORY_SERVICE_TOKEN=<the fresh §8 secret>   # poc-token dies with the poc stack

Restart the launchd job; the seam's OLL_MCP_MATCH_FLOOR needs no env at all anymore — the measured 0.64 floor is the baked default since the retrieval audit, so a restart without the var can no longer loosen the gate.

8 · Fresh secrets mint on deploy morning — nothing dev-grade goes to prod

9 · Smoke verification per service, in dependency order — green before the next

9a — ollama

docker exec <ollama-container> ollama list        # nomic-embed-text present

9b — oll-memory (liveness, then an authed query against the migrated corpus)

curl -s https://memory.oll.am/api/health
# → {"service":"oll-memory","status":"ok"}

curl -s -X POST https://memory.oll.am/api/memory/query \
  -H "X-Service-Token: $NEW_TOKEN" -H "Content-Type: application/json" \
  -d '{"query":"software engineer","collection":"mem:user:2:jobs","acl":["u:2"],"top_k":3}'
# → non-empty results[] with real job chunks + vector scores
#   (this one call proves the whole chain: token wall → Neon → Ollama embed → retrieval)

9c — ollscout

• ✓ https://scout.oll.am/ → 200 and the rail shows the real cases — the "Cannot reach the scout memory" sheet means the token or the internal base is wrong.
• ✓ /corpus → the ledger renders with real counts.
• ✓ One draft button → oll-write responds (proves OLL_WRITE_JWT).

9d — the re-pointed local seam

launchctl kickstart -k gui/$(id -u)/am.oll.mcp-seam    # restart with the new env
# then one ollam_memory_search / ollam_job_get through OpenCLAW or the MCP Inspector
# → cited chunks from the MIGRATED corpus (same case ids as before the move)

10 · Honest deltas & risks what this runbook does not hand-wave

Related pages: oll-scout service deploy runbook (the #94 service itself, 2026-07-11) · the Memory-spine decision · the full scout review.