# Local workers and models

This overlay targets the separately configured `jobjakt-dev` stack in the primary checkout. It does not replace that checkout's standalone compose file. Production services, databases and model caches are excluded.

The local runtime uses Ollama 0.40.11 and `qwen2.5:7b` (Q4_K_M, Apache-2.0), matching the model name in the repository's production configuration. Exact deployed digest parity has not been established. The frozen official registry manifest is in `model-manifests/qwen2.5-7b.json`; its SHA256 is `845dbda0ea48ed749caafd9e6037047aa19acfcfd82e704d7ca97d631a0b697e`. The image digest and resource limits are pinned in `ollama.compose.yml`. CPU inference is limited to four CPUs and 8GB; cloud processing is disabled. Model files use a dedicated development volume; the API binds only `127.0.0.1:11434`.

Initial local setup, explicitly authorised by the owner on 27 October 2026:

```powershell
rtk proxy docker compose -f dev/ollama.compose.yml up -d --wait
rtk proxy docker compose -f dev/ollama.compose.yml exec ollama ollama pull qwen2.5:7b
rtk proxy python dev/verify_ollama.py --manifest dev/model-manifests/qwen2.5-7b.json --output .artifacts/pre-phase10-ux/local-ai/verification.json
```

Verification checks the actual image, dedicated cache and every model/config/template/license blob against the frozen manifest before inference. A future changed upstream tag must fail verification; review the artifact before updating the pin. The download network remains available for explicit model maintenance; no cloud provider is enabled.

The existing approved DistilBART CPU checkpoint is reused through a read-only mount after all files are checked against `docs/redesign/phase-9/p9-sec-02/full-model/ARTIFACTS.json`. Transformers is offline, so this operation cannot fetch replacement weights. The parser-capable sidecar is also checked against its previously qualified local image ID.

Preflight and apply (run from this worktree; replace paths only with the verified development checkout and approved artifact):

```powershell
rtk proxy python dev/enable_local_workers.py --checkout "E:/Storage/Projects/Active/Job tracker" --bart-artifact "E:/Storage/Projects/Active/Job tracker/.artifacts/p9-sec-02/full-model/artifact" --output .artifacts/pre-phase10-ux/local-ai/worker-preflight.json
rtk proxy python dev/enable_local_workers.py --checkout "E:/Storage/Projects/Active/Job tracker" --bart-artifact "E:/Storage/Projects/Active/Job tracker/.artifacts/p9-sec-02/full-model/artifact" --output .artifacts/pre-phase10-ux/local-ai/workers-applied.json --apply
rtk proxy python dev/check_local_ai.py --output .artifacts/pre-phase10-ux/local-ai/sidecar-smoke.json
```

The apply helper validates project, Development environment, loopback backend ports, development database, current backend image, offline model mount and Mailpit settings. It recreates only the development backend/sidecar without a build or pull, preserving the database and volumes. It enables rules, reminders, saved-search alerts, daily exports, job enrichment and queued AI processing. These features still respect account entitlements, AI consent, configured schedules and available input. Email goes only to Mailpit at `http://localhost:8026`; it is not delivered externally.

This configuration helper does not build or publish source changes. Its initial backend image used a30-second limit, which could interrupt CPU generation. The separately qualified source activation below supplies the new240-second development timeout.

OAuth/mail synchronisation, billing and external data providers need their own approved credentials and configuration. Enabling workers cannot provide them. CV document import uses deterministic Docling; generative CV-import assistance stays disabled pending its separate source-grounding/accuracy qualification.

## Evidence, 27 October 2026

The local setup and worker apply passed. Independent model-blob verification passed. A real constrained JSON CPU response passed in 20.546 seconds including model loading. Authenticated sidecar checks passed: trained local job summarisation in 2.518 seconds and an Ollama-backed CV rewrite in 40.997 seconds. These checks assert actual nonempty content, not semantic correctness. No production digest parity or CV extraction accuracy is claimed.

Official sources: [Ollama model](https://ollama.com/library/qwen2.5:7b), [frozen registry source](https://registry.ollama.ai/v2/library/qwen2.5/manifests/7b), [Ollama Docker guidance](https://docs.ollama.com/docker), [local/cloud configuration](https://docs.ollama.com/faq).

## Local engine recovery, 28 October 2026

The resumed session found Docker Desktop4.86.0 stopped. Startup failed with Windows error1920 on orphaned AF_UNIX entries in `Docker/run/dockerInference` and `docker-secrets-engine/engine.sock`. Exact socket rename/delete attempts failed. Directory inspection confirmed only zero-byte reparse-point runtime entries; the Docker WSL distribution was stopped.

After stopping only the failed local Desktop/backend processes, the runtime directories were moved aside and recreated. Backups remain in `C:/Users/Connor/AppData/Local/Docker/run.pre-p10-stale-20261028`, `run.pre-p10-stale-20261028-02` and `C:/Users/Connor/AppData/Local/docker-secrets-engine.pre-p10-stale-20261028`. They contain only the inspected stale IPC entries. The Linux engine then responded successfully as `docker-desktop`. Model files, database volumes and Docker settings were not reset. This was local development recovery; production was not accessed.

The same failure pattern is reported in [Docker's Windows issue tracker](https://github.com/docker/for-win/issues/15064). The recovery above is supported by current local inspection and a successful engine response; it is not a general instruction to move Docker data directories or reset an installation.

## Qualified local source activation, 28 October 2026

29 October recovery addendum: the same stale IPC startup failure recurred after Desktop had stopped. Exact inspection again found only zero-byte runtime sockets. Failed local Desktop processes were stopped and the bounded socket directories were moved to `Docker/run.stale-20261029-0622`, `Docker/run.stale-20261029-0624` and `docker-secrets-engine.stale-20261029-0624` under the same local AppData parent. No Docker data/settings/volume/model directories were moved. The engine and existing local development services recovered; a fresh isolated MariaDB lane then passed 90 tests with cleanup exit zero. This does not change the source activation identity below.

The existing development stack now runs source checkpoint `3f4db8ca8183704945915ae4d2053e641bafa5b2`. Local URL: **http://jobbjakt.local**. The frontend deliberately rejects an unmatched `localhost` Host header; the canonical host was preserved. Mailpit remains at `http://localhost:8026`.

Backend image: `sha256:462b1d14fe871d64b16a2db28fda30a433657dd4d8ed5f61587a6b80685fb0cf`. Frontend image: `sha256:a59b8313233a20f6a2f0a8cc23785d224b58678e5758c3985fa55f238402886f`. All qualified source-file hashes were rechecked before activation. The backend is the same immutable image used by the passing Linux qualification. The frontend was rebuilt from that source with `/api` routing and the existing local nginx configuration. The qualified static test export was archived separately before this build.

Raw activation/rollback identities, source-image compose override, build logs and actual HTTP checks are in `.artifacts/pre-phase10-ux/local-ai/source-3f4db8c/`. The one-off activation script is retained alongside that directory as execution evidence. Backend/frontend were recreated with no dependency recreation, builds, pulls or volume removal during the compose apply. Previous image IDs remain recorded. Existing database, sidecar, model and email services were retained.

Confirmed: both images healthy; served index matches the built bytes; backend health200; anonymous proxied account request401; actual backend environment has all seven authorised workers enabled and the240-second AI timeout. Ollama blobs were reverified after engine recovery. These checks establish the running source/configuration; model-quality and broader feature qualifications remain separately stated in B_ACCEPTANCE.md.

For subsequent compose operations, retain the generated `source-images.override.json` **after** `dev/local-workers.override.yml`, the primary standalone development compose file and its `.env.dev`. Supply the same read-only `JOBTRACKER_DEV_BART_ARTIFACT` path. Running only the primary compose file can revert these local image/configuration selections. Do not substitute production configuration.

### Cumulative A–D local activation — 2026-10-29

The latest local source is `caa0c6a2efea5a6e9d8910f55fb00fd982cd48fd`, including the dashboard/admin monitor and the strategy JSON correction. Use the newer override `.artifacts/pre-phase10-ux/local-ai/source-caa0c6a/source-images.override.json` last; the B override above is historical. Backend image `sha256:14adb1e05757575e7c9d50c8f504f8209ebbf7f9b5f870daef0113d9dfe7dd49`, unchanged `/api` frontend `sha256:fd4092aa2919463f8cf55b81941df1cd4c2db261ed6dc092b8aaf74730d9d618`, sidecar `sha256:e1dc2dfce83caa4fe785ee9485b86011bb3ec0470511a9dfbdb03206ee1c40b3`. All were healthy after guarded recreation of only the backend/sidecar; the previous images remain available for rollback. No dependencies, model blobs, production services or MariaDB volumes were replaced.

The sidecar smoke helper's immutable image guard now names this qualified source image. Full backend tests pass 1,448; offline Python contracts pass 83. Actual local extraction, model and scheduled enrichment observations, synthetic-account erasure and their evidence limits are in [LOCAL_RUNTIME_ACCEPTANCE.md](../docs/pre-phase10-ux/LOCAL_RUNTIME_ACCEPTANCE.md). Canonical local URL remains **http://jobbjakt.local**. Account AI consent and plan requirements still apply; scheduled enrichment runs every ten minutes. External services requiring provider credentials remain independently unqualified.