[lucas-koontz]

User story

As an on-call engineer
I want quiet, correlated diagnostics
So that I can diagnose failures without exposing customer data

Why this matters

Operators currently sift through agent chatter to find a failed turn. JSON diagnostics now default to WARNING and retain safe exception locations and conversation correlation. Cloud-turn JSONL and the separate usage stream keep their existing contracts.

Visual review

A deployed before/after capture remains a rollout gate for this runtime. Local regression and container evidence below establish behavior, not a production volume reduction.

Acceptance criteria

• The packaged resident launcher uses the same safe formatter and WARNING default for setup/fork failures while preserving forked-turn root logging.
• Owned diagnostics default to WARNING with a consistent JSON envelope and safe failure fields.
• Exception values, credentials, provider bodies and arbitrary extras are excluded from diagnostic output.
• Cloud-turn output and usage records remain independent; request context resets after each turn.
• Override/reset and rollback behavior are documented in docs/structured-logging.md.
• Deployed versions, collector mappings, alerts, retained operational evidence and matched rollout windows are verified.

How to test

1. From an initialized Linux checkout, run uv run pytest tests/test_diagnostics.py tests/test_cloud_turn_resident.py tests/test_cloud_turn_provider_logging.py tests/test_image_smoke.py.
2. Exercise an ordinary success, a final refusal and a dependency or streaming failure. Check safe JSON and correlation, and confirm protocol/output contracts remain unchanged.
3. Enable the documented temporary verbosity/access override, repeat the checks, then reset it. Confirm failed health checks or operational failures remain visible at the quiet default.

Notes for the reviewer

• Publish the package before consumers. Source changes do not update installed agent packages or scratchpad images.
• Use the maintained structlog processor. The adapter stays with this independently released runtime; no shared logging package or release coupling is introduced. Lock regeneration normalized platform markers without changing existing dependency versions.
• Rollback uses the prior release and settings together. Do not disable audit, metric or protocol streams to quiet diagnostics. Live collector and alert validation remains a release gate.

Current review validation

The resident entrypoint regression failed in all four JSON/text and default/DEBUG combinations before the fix. The final affected suite passes all 62 cases in the Linux Python 3.12 container, including resident lifecycle and image smoke tests; Ruff and diff checks pass. Local macOS coverage passed 42 cases with 20 Linux-only skips before the Linux rerun.

The combined source API/web arm64 images also rebuilt successfully. Container smoke verified the installed Anton version 2.26.10.21.2rc3.dev2+gb73859e9f, matching this commit, alongside the reviewed server version. It passed health/routing, refusal visibility and private-value exclusion; it does not establish provider-backed agent turns or a released package.

The hosted workflow also passed its test suite and checks on this commit.

Earlier validation

These broader results precede the resident follow-up:

Check	Observed result
Full suite before assertion updates	7,610 passed, 271 skipped; two old text-log assertions failed and were updated.
Final affected suites	43 passed: diagnostics, cloud process and provider-error behavior.
Regression sensitivity	Each new diagnostic test failed with its protected behavior removed; restored tests pass.
Pre-PR checks	Diff integrity and the repository sweep passed; regression mutations were restored.

Ships with

mindsdb/cowork-server#741, mindsdb/cowork#1263, mindsdb/data-vault#5, mindsdb/mindshub#12504.

Merge/release order: #621, then mindsdb/cowork-server#741, then mindsdb/cowork#1263. mindsdb/data-vault#5 can proceed independently. Update reviewed release pins and rebuild mindsdb/mindshub#12504 last.