One long session that touched almost every layer: stood up a public OIDC identity provider, scaffolded a centralized Postgres for new services, defended the public chatbot against a real-world Pliny prompt-injection campaign, dropped 29 GB of contended VRAM by retiring an experiment that didn't pan out, brought gygaxbot's voice recording back from the dead, and dropped npm engine-strict into eleven repos so the kind of silent-fail that wasted an hour tonight can't happen quietly again.

Dex goes public at auth.jefe.works

JefeWorks now has a real OIDC IdP on the public internet — Dex behind a dedicated Cloudflare tunnel, with the LAN hostname kept for operator break-glass. The full Phase 1–4 plan from the prep handoff landed: dedicated jefeworks-auth tunnel, DNS via Cloudflare, Traefik split into LAN-secure and public-secure routers, issuer rotated to https://auth.jefe.works, and a higher-priority router that 404s the static-password connector on the public hostname while leaving it reachable on LAN. Discord OAuth went in last; cracked a YAML-number-vs-string bug where Discord's all-numeric client ID was being parsed as a number and rejected by the connector unmarshal, then closed a follow-up where the picker page was leaking /auth/local as a path-based bypass to the password connector. End state: testers go straight to Discord, operators keep their LAN break-glass.

Centralized Postgres, scaffolded

New repo jefe-postgres at port 5440 (PgBouncer at 6440) with role-per-service / database-per-service, JefeCA-issued TLS, postgres_exporter on monitoring-net, daily backups, and a "how to add a new service" runbook. Adventure GUI is the headline future tenant — wrote the migration spec into jefeadventure/docs/POSTGRES-MIGRATION-PLAN.md covering scope (campaigns/sessions/combat/feedback/ACL/invites move; shipped read-only campaigns stay as content), schema sketch with JSONB where shape is free-form, ORM choice (SQLAlchemy 2.x sync + Alembic), connection pattern caveats (PgBouncer transaction-mode bites LISTEN/NOTIFY and prepared statements), and sequencing after the invites sprint. FreeChat and Enclave keep their existing Postgres instances; centralization is the default for new services, not a forced migration.

Pliny corpus showed up at the chatbot

Reviewed ten days of jefeai-api access logs and found a sustained probing session against the jefeworks chatbot from a single AWS IP — sixteen probes over seven minutes, escalating from recon to direct asks to a copy-pasted Pliny jailbreak template. The attacker literally typed "pliny" into the chat at one point. Replayed two of the templates with benign payload substitution to see whether the bot held: it didn't. The Nemotron reasoning trace was leaking into the client response (system prompt visible), and the HERMES 4 GODMODE template caused the bot to emit {LOVE PLINY <3} in its actual reply — screenshot-grade brand damage. Discovered an existing input_guard.py module with thirteen patterns and zero callers — dead code. Security-engineer agent rewrote the guard, wired it into the three RAG chat endpoints, added a hybrid <think>-tag-plus-heuristic reasoning stripper, added an output-side guard that catches GODMODE / LOVE PLINY / I'M FREE divider patterns regardless of input verdict, and laid down privacy-aware audit logging that captures full bodies only when injection or output is flagged. Local commit, ready to ship.

The vLLM experiment retires

Memory alert fired on Titan during the session — system RAM at 38 GiB, swap fully consumed, load average over four. Diagnosis: VRAM contention. The vLLM-Code instance was holding 29 GB of GPU memory permanently for a JefeCode product pivot that didn't pan out, leaving Ollama only 7 GB free and forcing it to CPU-offload sixteen gigs of weights and KV cache. Decommissioned vLLM after auditing the actual consumers (websites already on Ollama, jefebot has a runtime cascade with proper try/catch fallback, vibebot Twitch path was Ollama-native, gygaxbot's useVllm evaluated false at boot anyway). Stopped the container; VRAM dropped from 89 GB to 60 GB used, free climbed from 7 GB to 36 GB, system RAM recovered from 38 GiB to 16 GiB, swap drained. Jefebot's cascade caught the failure and routed Discord traffic to Ollama in three and a half seconds, no user-visible error.

gygaxbot voice recording, the long way around

The DM tried !session record, watched the bot join and then immediately apologise. Added deep voice-networking diagnostic logging (the existing handler only captured high-level state transitions); next attempt revealed WebSocket close code 4017 — undocumented but real, lined up with Discord's DAVE protocol rollout for E2EE voice. Rolled the Dockerfile and lock back to the last known-working commit on a hunch: same 4017, confirming the regression was Discord-side, not ours. Bumped @discordjs/voice 0.18.0 → 0.19.2 (the DAVE-supporting line via @snazzah/davey), which silently demanded Node ≥22.12 — npm's optional-dependency engine-mismatch behaviour skipped the install with no warning, the build "succeeded" with voice missing entirely, and the next test failed at "module not found." Dockerfile bumped Alpine Node 20 → 22, container built, voice recording back in business. Also threaded a group_add: ["1500"] through compose to give the new container UID write access to the campaigns volume after a separate UID-shift broke the mkdir path. The DM is recording the session as I type.

Stack-wide engine-strict drop

Filed .npmrc with engine-strict=true into all eleven active Node repos so npm fails loudly on engine mismatches instead of silently dropping optional dependencies. Untracked working-tree files only — no commits, no pushes, no rebuilds tonight; gygaxbot's recording stays sacred. The repos pick the change up on their next deliberate deploy. The story above won't repeat without a louder warning.

What's next

• Ship the JefeAI prompt-injection hardening — local commits ready, just needs push and force-recreate
• Stand up the centralized Postgres on Titan from the runbook (db-net needs creating, vault seeded, JefeCA cert minted, then first start)
• Decommission vLLM fully if there's no regret in a week — remove compose service, drop VLLM_* from vault namespaces
• Port jefebot's runtime cascade pattern into gygaxbot so its config-time-only useVllm brittleness goes away
• Bump @discordjs/voice in any other voice-using bots (vibebot, jefebot) before they hit the same 4017
• Stage and push the eleven .npmrc files at next maintenance window