fix: real TypeSafe decision contract boundaries that strand retained tasks #101

Merged
kakashi-forgejo-writer merged 1 commit from fix/typesafe-decision-boundaries into main 2026-11-13 22:06:21 +00:00
Owner

Actual installed-task diagnosis, not another synthetic restart claim: real host assessment prompts28–53KiB were rejected silently by24KiB transport. Synthetic live SystemOne returned HTTP200 model=openjev-0.1 for configured openjev-latest; public SystemOneResponse schema explicitly permits response model differing from alias. Exact equality rejected otherwise valid classifications. Valid low confidence was mislabeled malformed/invalid_shape.

Fix: align bounded input cap with existing host56KiB envelope; advertise optional maxPromptBytes; validate bounded resolved model metadata but retain configured request endpoint/model/credential unchanged (no discovery/fallback). Confidence below .9 is explicit unknown/uncertain, never a decisive verdict. Preserve .9 floor, probability normalization/schema/choice/credential/privacy/expiry/abort checks. Tests cover canonical aliases, actual53KiB evidence class, over56KiB rejection, and low-confidence uncertainty. All212 Plugin tests pass against current Core worktree; no production config/ledger writes. Core changes separately handle provider-budget projection and assessment queue/timeouts. Protected compatibility CI required.

Actual installed-task diagnosis, not another synthetic restart claim: real host assessment prompts28–53KiB were rejected silently by24KiB transport. Synthetic live SystemOne returned HTTP200 model=openjev-0.1 for configured openjev-latest; public SystemOneResponse schema explicitly permits response model differing from alias. Exact equality rejected otherwise valid classifications. Valid low confidence was mislabeled malformed/invalid_shape. Fix: align bounded input cap with existing host56KiB envelope; advertise optional maxPromptBytes; validate bounded resolved model metadata but retain configured request endpoint/model/credential unchanged (no discovery/fallback). Confidence below .9 is explicit unknown/uncertain, never a decisive verdict. Preserve .9 floor, probability normalization/schema/choice/credential/privacy/expiry/abort checks. Tests cover canonical aliases, actual53KiB evidence class, over56KiB rejection, and low-confidence uncertainty. All212 Plugin tests pass against current Core worktree; no production config/ledger writes. Core changes separately handle provider-budget projection and assessment queue/timeouts. Protected compatibility CI required.
fix: honor SystemOne alias responses and host assessment input budgets
All checks were successful
plugins-ci / compatibility (push) Successful in 3m5s
90e776b21f
Collaborator

Kakashi review — head 90e776b21f044d4df0e67481723d2ede5afdfb5e, base e2bb89f3e2a9b2cfd2714dbbe776ec3ffbbe1b0e

review_clear

Kakashi review — head `90e776b21f044d4df0e67481723d2ede5afdfb5e`, base `e2bb89f3e2a9b2cfd2714dbbe776ec3ffbbe1b0e` review_clear <!-- kakashi-review:3077582a578f5c5ee2f5d0505843c83155a69dfa32b3ea6cdaa335da780174be -->
Sign in to join this conversation.
No reviewers
No labels
No milestone
No project
No assignees
2 participants
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/plugins!101
No description provided.