Skip to content

feat(lmi): add the lifecycle scenario coordinator - #195

Open
zhongkechen wants to merge 8 commits into
mainfrom
feat/lmi-2-scheduler-squashed
Open

zhongkechen wants to merge 8 commits into
mainfrom
feat/lmi-2-scheduler-squashed

Conversation

@zhongkechen

@zhongkechen zhongkechen commented Nov 18, 2026 •

Copy link
Copy Markdown
Contributor

The Lambda Managed Instances (LMI) conformance runner previously validated evidence collected elsewhere. It can now also schedule lifecycle scenarios against persistent targets through an SDK-owned adapter. Pass --manifest <path> --adapter <module>:<factory> --output <empty-dir> for a live run; omit --adapter to validate saved evidence offline. Both paths apply the same manifest rules, and cases marked NOT_IMPLEMENTED are reported without being dispatched.

How a live run works

  • Schedule against shared targets. The coordinator uses the suite catalog to determine each case’s roles, minimum native concurrency, and isolation requirements. Compatible cases share a target within its capacity; exclusive cases wait for other owners to finish. A suspended case can release native capacity while retaining ownership, so other compatible work can run without letting a deployment replace its target.
  • Record attempts before dispatch. The runner saves the admitted manifest and writes launch attempts to attempts.jsonl before dispatch. If the audit write fails, the launch is aborted. Case outcomes are written to results.jsonl. Evidence is checked against the admitted case, profile, and validation budgets.
  • Keep cleanup separate from the verdict. The coordinator tracks case-owned work through cancellation and retirement. Cleanup evidence can help account for outstanding work, but it cannot turn a failing scenario into a pass. If the coordinator cannot verify that the work has stopped, it quarantines the affected target. Reuse requires positive recovery verification and no remaining live owner.

Validation

Python CI passed all 4,605 tests. The new tests use fake adapters to exercise shared and exclusive scheduling, suspension and resumption, audit writes before dispatch, quarantine and recovery, cleanup on failure, and preservation of retry ordering and correlation evidence. Runner E2E also passed its fixture deployment and validation. An end-to-end run with an SDK adapter against live Lambda Managed Instances remains unverified.

Refs #161.

@zhongkechen
zhongkechen force-pushed the feat/lmi-2-scheduler-squashed branch from fee6746 to 0e667b1 Compare November 18, 2026 22:33
@zhongkechen
zhongkechen deployed to ai-pr-review-runtime November 18, 2026 22:34 — with GitHub Actions Active
Base automatically changed from feat/lmi-1-contract-squashed to main November 19, 2026 18:07
@zhongkechen
zhongkechen deployed to ai-pr-review-runtime November 19, 2026 18:23 — with GitHub Actions Active
@zhongkechen
zhongkechen deployed to ai-pr-review-runtime November 19, 2026 22:45 — with GitHub Actions Active
@zhongkechen
zhongkechen deployed to ai-pr-review-runtime November 20, 2026 02:08 — with GitHub Actions Active

This branch was successfully deployed

1 active deployment
ai-pr-review-runtime — 08aec250 Deployed Nov 20, 2026 by zhongkechen via ai-pr-review / Codex review / Generate Codex review #1256
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant