harness/core
4
0
Fork 1
Code Issues Pull requests 14 Projects Releases 200 Packages 1 Wiki Activity Actions

fix(dispatcher): recover incomplete results and configure validator model fallback #283

Merged
karsten merged 2 commits from fix/sequential-validator-recovery into main 2026-11-06 04:59:52 +00:00
Conversation 0 Commits 2 Files changed 9 +412 -39
karsten commented 2026-11-06 04:50:16 +00:00
Owner
Copy link

Fix

  • Independently validate missing/empty/malformed worker reports instead of bypassing the validator. Bounded same-session report/evidence corrections; no inferred success or blind effect replay.
  • Step-specific codified prompt judges the current step, not completion of the entire mandate.
  • Persist settled worker results, bounded cumulative host tool observations and inference attempts. At most two tool-free assessment attempts per result; restart/resume does not reset the budget or replay settled effects. Retained transcript is reconciled before post-restart correction.
  • Preserve two-correction/repeat guards, manual stops, actual permissions/approval/deadline holds, exact explicit model selection and safe inadequacy escalation.
  • Bound checkpoint schema and size, reserve diagnostic space, refuse oversized fixed state before opening an effect intent. Matrix plan-status delivery failure no longer aborts execution.
  • Configurable validator: modelRoles.roles.validator → explicit review.model → reviewer role → default model. Missing configuration falls through; an explicitly selected unavailable model never silently changes. Applied to sequential validation and supervisor completion assessment. Configure with /model-roles set validator provider/model low.

Validation

Final candidate: npm run typecheck and complete npm test 1542/1542 passed. Full suite uses the publisher native-pool environment and two-CPU affinity (no tests skipped). Unrestricted 24-CPU runs exposed unrelated maintenance CLI/Wasm timing flakes. Local npm blocks dependency install scripts; the upstream Matrix native installer was required to load the pinned dependency.
Independent tool-free model review was critically checked: retained evidence-schema bounds and checkpoint capacity fixed, verdict initialization made explicit. The review's UTF-8-byte-limit claim was refuted against ExecutionStore's string.length limit.

Operations

Normal shared-image publisher/promotion after merge. Update and verify Peter, Severus, Randy and Akira. Historical paused tasks and expired mandates are not automatically reopened. User-facing Recovery notices are a separately identified issue, not represented as fixed by this PR.

## Fix - Independently validate missing/empty/malformed worker reports instead of bypassing the validator. Bounded same-session report/evidence corrections; no inferred success or blind effect replay. - Step-specific codified prompt judges the current step, not completion of the entire mandate. - Persist settled worker results, bounded cumulative host tool observations and inference attempts. At most two tool-free assessment attempts per result; restart/resume does not reset the budget or replay settled effects. Retained transcript is reconciled before post-restart correction. - Preserve two-correction/repeat guards, manual stops, actual permissions/approval/deadline holds, exact explicit model selection and safe inadequacy escalation. - Bound checkpoint schema and size, reserve diagnostic space, refuse oversized fixed state before opening an effect intent. Matrix plan-status delivery failure no longer aborts execution. - Configurable validator: modelRoles.roles.validator → explicit review.model → reviewer role → default model. Missing configuration falls through; an explicitly selected unavailable model never silently changes. Applied to sequential validation and supervisor completion assessment. Configure with /model-roles set validator provider/model low. ## Validation Final candidate: npm run typecheck and complete npm test **1542/1542 passed**. Full suite uses the publisher native-pool environment and two-CPU affinity (no tests skipped). Unrestricted 24-CPU runs exposed unrelated maintenance CLI/Wasm timing flakes. Local npm blocks dependency install scripts; the upstream Matrix native installer was required to load the pinned dependency. Independent tool-free model review was critically checked: retained evidence-schema bounds and checkpoint capacity fixed, verdict initialization made explicit. The review's UTF-8-byte-limit claim was refuted against ExecutionStore's string.length limit. ## Operations Normal shared-image publisher/promotion after merge. Update and verify Peter, Severus, Randy and Akira. Historical paused tasks and expired mandates are not automatically reopened. User-facing Recovery notices are a separately identified issue, not represented as fixed by this PR.
karsten changed title from WIP: fix(dispatcher): recover incomplete reports without repeating worker effects to fix(dispatcher): recover incomplete results and configure validator model fallback 2026-11-06 04:59:36 +00:00
karsten deleted branch fix/sequential-validator-recovery 2026-11-06 04:59:53 +00:00
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
harness/core!283
No description provided.