Description
Users need to know how to opt in, what they get and what stays out of telemetry. Also the proof step: deploy each of the three use_cases combinations (or plan them where a live deploy is not possible) and tear down cleanly.
Acceptance Criteria
- #1 Docs cover the use case, opt-in, architecture diagram, demo runbook, telemetry by content_capture mode, cost, security notes, teardown and the terraform and helm references
- #2 just check passes, and plans for all three use_cases combinations are recorded
- #3 A live regulatory-only deployment is verified and torn down, with evidence recorded in the task
- #4 Docs cover the use case, opt-in, architecture diagram, demo runbook, telemetry by content capture mode, cost, security notes, teardown, the terraform and helm references and the manual team span-metrics step if used; AGENTS.md records the new gotchas
- #5 just check passes, including the per-combination mock plans
- #6 A live regulatory-only deployment, using an image tag built from images.yml or a just images-push mirror, is verified and torn down with evidence recorded in the task
Definition of Done
- #1 just check
Implementation Plan
Implement the corrected design and frozen regulatory contracts, with disjoint file ownership for parallel work. Prove local runtime/rendered/mock-plan behaviour, integrate into just check, complete CodeRabbit review before landing code, and record live-only acceptance separately. No live deployment, image publication, new credential or CVE exception is authorized by this local implementation pass.
Continue beyond publication: reconcile recorded consumer and live state, prepare an isolated regulatory-only candidate with existing target selection, perform authorized bounded runtime/evaluation/snapshot checks and teardown. Retain stack/credential and shared cluster safety boundaries. No shared-root apply or unrelated workload changes.
Implementation Notes
Operator documentation, Mermaid architecture, security/capture/cost/teardown guidance, Terraform/chart references, docs nav and GitOps examples implemented. Full local just check passes using isolated PyYAML (plain local Python lacks it); no local private denylist configured, so private scrub must be verified by CI. Live regulatory-only deployment and teardown remain unperformed, blocked by image publication plus an authorized target stack/deployment. Do not call the epic live-verified.
Local docs and gate complete; exact resume is an authorized regulatory-only deployment using a published image, both capture modes, rule and score readback, dashboard snapshots and teardown evidence. No live acceptance is claimed.
Continued owner authorization covers completing the bounded live proof. Existing consumer target and credentials were reconciled without changing stacks. Separate encrypted state is empty, and fresh paused regulatory-only plan has89creates/no deletes/no host/no account-wide logging. Source/image review and final immutable pins remain before apply. Source continuation and independent image jq fixes are under final gate/review.
The live capture-on proof exposed incomplete reviewer completions after the fenced-JSON fix. The failed section experiment is retained as baseline. Prompt v2 bounds role outputs and preserves actual returned usage/stop reason before fail-closed escalation; 20 regulatory tests pass, including a real AWS SDK-to-localhost and Agent Observability transport regression. Both capture modes, stored experiment completion and teardown remain required. Historical count-spacing review finding corrected in the design continuation; new notes use readable spacing.
Final Summary
Live regulatory-only proof and teardown completed. Final config source 7530badaaa684bc7690a27c090ef7ac102a27f82 and signed runtime image from 80431c41ee435ad77f0aa6303b51437fa0d83aa8 were verified with exact-source CI and completed CodeRabbit review. Both capture modes have actual stored-generation/trace readback, new-pod operational metrics and reviewed 14-panel snapshots. Guard suite stored 3 passing final scores. Section experiment completed all 4 trials with 16 evaluator and 4 final scores; 2/4 passed quality, with complete token/cost coverage. This intentionally remains quality-failure evidence, not an all-green claim. Bounded loadgen on/off each sent 2 successful requests; injection escalated with zero model usage. Fresh isolated teardown destroyed all 89 managed resources; state has zero addresses and direct readbacks confirm removed workload/dashboard/IAM/Pod Identity/all 12 profiles, with shared cluster ACTIVE. Shared root untouched. Historical telemetry and experiment results retain existing stack retention. Public design records sanitized evidence; detailed private artifacts remain with the consumer board.