Skip to content

fix(dreamer): keep scheduled cutoffs independent of manual apply - #264

Draft
joshyorko wants to merge 1 commit into
fix/issue-260-dreamer-startupfrom
swarm/238-scheduled-cutoff
Draft

joshyorko wants to merge 1 commit into
fix/issue-260-dreamer-startupfrom
swarm/238-scheduled-cutoff

Conversation

@joshyorko

@joshyorko joshyorko commented Oct 2, 2026 •

Copy link
Copy Markdown
Owner

Addresses #238 Packet A and fixture P14. Depends on #261's bounded command scheduler repair.

With a manual apply watermark but no scheduled watermark, the first scheduled command incorrectly inherited the manual cutoff, returned an empty success, and advanced past scheduler-unprocessed evidence. The private scheduled invocation now keeps its own empty or persisted cutoff. Manual omitted-cutoff defaults and explicit manual replay retain their behavior; the scheduler does not gain archive replay authority.

A file-backed service regression establishes the two independent watermarks through real operations, checks a failed command and persisted error receipt, reopens the database, retries the same first record, continues a capped equal-timestamp source tail across another restart and a later manual watermark, and verifies no archived evidence or unchanged-tick model call. Separate compatibility checks retain manual default and explicit-empty behavior.

Base: 4972855, the published #261 head.
Tested local checkpoint: f1cb612bda4769bbd64d2796f2e37b680abcfc9e.
Published head: 5b1da23. The published and tested trees are identical: 73482fc775544b638809e5220969d06c7eaaff82.

Validation:

  • Genuine pre-fix first-scan failure was observed. The unchanged base reported zero selected evidence and advanced its scheduled watermark without invoking the configured failing command.
  • 17 focused scheduler, command policy, no-fallback and configuration tests passed, zero failed or ignored.
  • Formatting and whitespace checks passed.
  • Full canonical release gate completed under the owned Linux child reaper with an exit-zero receipt at the tested head: 583 Rust tests across 47 result groups passed, zero failed, ignored or filtered. Five launcher checks, synthetic doctor/evaluations/performance, dry runs and required-file checks passed. The published source tree is identical. Operator dogfood was excluded.

This bounded repair does not close #238 or claim governed automatic model adoption, semantic validation, real consumer learning acceptance, or live activation. No provider/model/auth ownership, budgets, proposal identity, schema, storage policy, or adoption mode changes are included. Latest-head automatic connector review and required checks remain merge gates; no additional GitHub review was requested.

Independent local Sol review found no actionable bounded finding. The completed full synthetic gate covers the identical published tree 73482fc775544b638809e5220969d06c7eaaff82. This remains a draft stacked on #261. Latest-head GitHub CI and automatic connector review remain pending; complete #238 adoption and consumer gates remain open.

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