Survey found: z.ai provider (id `zai`, 12 models, glm-5.2 default), LAN IP `192.168.3.78`, and **port 4000 is occupied** (kneldevstack homepage container) → will use 4001. Now reading the design doc §v2.2 and inbox conventions.Task source confirmed; `inbox-pmo` = `/home/reachableceo/.coordinate/inbox-pmo/`. Checking protocol conventions, crush config, and docker image availability.Context fully gathered. Executing now: config + secrets first.Port 4001 free. Creating config + secrets.Adding secrets filenames to credentials/.gitignore (repo hard rule), then pulling the LiteLLM database image.Image pulled. Starting container on 192.168.3.78:4001 (4000 occupied by kneldevstack homepage).Container crash-loops: current LiteLLM main-stable **rejects SQLite** — database features (virtual keys, spend tracking) require PostgreSQL. Investigating when the SQLite guard landed to find the newest version that still honors "embedded SQLite".The current main-stable hard-pins Prisma to PostgreSQL. Checking when this guard was introduced to pick the newest SQLite-capable version.The June 2026 fail-fast fix confirms SQLite hasn't truly worked since Prisma was pinned to PostgreSQL — checking when `schema.prisma` got pinned.Checking the Prisma schema provider in older tags directly.SQLite was never actually viable — Prisma is postgresql-pinned across all 2025 tags and current code has zero sqlite paths. Before deciding this is a BLOCKED, checking prior art in the tooling dir for related findings.Prior art doesn't resolve it. Governance rules here are explicit ("never self-assign scope"; "if any step fails, BLOCKED report"), and the SQLite requirement is upstream-infeasible. Before filing BLOCKED, I'll fully validate the recommended alternative with throwaway containers (zero residue) so the unblock is a 5-minute decision — first removing my crash-looping container (mine, created this turn — not a pre-existing one).Test postgres up. Launching temp litellm with the exact final recipe (env-file secrets, config minus sqlite line) on loopback 4199.Prisma is migrating the fresh DB — waiting for startup to finish.My test env bug: postgres user mismatch (`litellm` vs default `postgres`). Recreating test stack with matching user.Old root-owned data dir survived; wiping it via container, then restarting.Found it — my `&&` chain short-circuited: the env rewrite never ran (rm failed first), so the test DB initialized with the default user. Fixing properly.LiteLLM healthy with Postgres. Validating the complete recipe (models → key gen → chat through proxy → usage API), all output redacted.All green — one more pass with higher max_tokens (reasoning ate the budget) and spend check, then teardown.Need to prove spend actually records (usage API is a core deliverable) — generating a fresh key, chatting, then polling spend.Spend rows exist but cost is $0 — LiteLLM doesn't know pricing for custom openai-compat models. Adding per-token pricing from the crush catalog into the config and re-testing.Spend records in dollars now ($0.0002342). The key_info parse just hit a response-shape difference — re-checking with correct parsing.Full validation complete. Tearing down the throwaway stack to zero residue, then updating the staged config to the validated recipe and filing the BLOCKED report (per task constraint: SQLite is upstream-infeasible — architecture decision belongs to PMO/founder, not me).Log.md reveals the PMO already hit this exact snag and pre-approved the fallback: "dedicated litellm-db postgres (approved fallback path)". Reading the snapshot for exact wording, then deploying that path.Confirmed: PMO pre-approved "dedicated litellm-db postgres container (standard minimal deploy, not overbuild)". Deploying that path now — production versions.PGDATA perms issue (18-layout subdir not covered by entrypoint chown). Fixing host dir ownership via container, letting restart policy recover.Production proxy healthy on LAN. Issuing the 8 virtual keys (values only into the 600 file).Backing up and wiring crush to the proxy.Direct proxy auth works (200). The 401 came from somewhere else — checking proxy logs during a fresh crush attempt.Root cause found: crush **regenerates providers.json from its bundled catalog on startup** — my endpoint edit was overwritten, so crush called z.ai directly with the virtual key (hence z.ai's "unauthorized"). The durable wiring point is crush's config file. Investigating.v0.87 predates some of this — checking what config commands the installed binary actually supports.Restoring the state file (my earlier edit there was the wrong layer), then wiring via `~/.config/crush/crush.json` custom provider.Schema says `base_url` (not the catalog's `api_endpoint`). Rewriting config.Provider registered. Running the headless verification through the proxy.PROXY-OK. Confirming it transited the proxy (logs + rceo key spend) and that the default (flagless) path also routes via litellm.Both routes verified through the proxy. Capturing redacted usage samples for the report.All green. Finalizing: checking log.md format, then writing the report, logging, and archiving the TASK.Done. REPORT filed at `~/.coordinate/inbox-pmo/REPORT-litellm.md`; TASK archived; log.md updated.

- **Containers**: `litellm` (main-stable, restart unless-stopped, **192.168.3.78:4001→4000** — 4000 was taken by the kneldevstack homepage) + `litellm-db` (postgres:alpine, docker-internal only). No existing container touched.
- **One deviation**: "embedded SQLite" is upstream-impossible (Prisma is postgres-pinned in every LiteLLM tag; current builds fail fast) — deployed the **pre-approved fallback** from the PMO's 17:20 snapshot (`litellm-db` postgres), after validating the whole recipe on throwaway containers first.
- **Keys**: 8 issued (rceo $50, bod/cco/cto/coo $30, offstage-1/2/3 $15 — all /30d); values only in `credentials/litellm-virtual-keys.md` (600, gitignored). Secrets-scan of REPORT/log: clean.
- **Wiring**: backup `providers.json.pre-litellm.bak` (+ `crush.json.pre-litellm.bak`); live wiring is `~/.config/crush/crush.json` (600) because crush **auto-regenerates providers.json** (its revert caused a transient 401 mid-wiring — root-caused, not papered over).
- **Verified**: headless crush → `PROXY-OK` via proxy (explicit + default model), 200s in proxy logs, `/key/info` shows rceo spend $2.035e-05 against $50/30d budget; fallback = restore the .bak + `rm ~/.config/crush/crush.json`.