Skip to content

Preserve safe manager failure details and original cause - #9

Merged
Brian Krabach (bkrabach) merged 3 commits into
mainfrom
fix/preserve-manager-failures
Sep 25, 2026
Merged

Brian Krabach (bkrabach) merged 3 commits into
mainfrom
fix/preserve-manager-failures

Conversation

@bkrabach

Copy link
Copy Markdown
Collaborator

Manager turn failures currently discard their cause, publish a provider error even for a local context check, and end with a generic replay warning. Preserve a bounded error category, stage, safe message, retryability, input/generation identity, and explicit non-rollback/non-replay semantics. Keep the original exception as the private cause; do not publish SDK bodies, prompts, file paths, or credentials.

The base engine remains responsible for provider events. Local context failures are identified from their originating module, not merely from a missing provider name. Ordinary finite execution and cancellation behavior are unchanged.

Validation: complete loop-live suite, 81 passed, 4 skipped, using Core 2.0.1, loop-streaming 4cc86dd4 and context-simple 7b100715. New coverage reproduces a real context-simple budget failure, checks safe metadata and cause preservation, verifies input correlation, and rejects false provider/success events or replay. No paid provider calls or live voice validation.

Class review checked current catalog basic ae0954cc, streaming 4cc86dd4, events 6b2f9d54, Microsoft live 1d7be38, and the distinct bkrabach live fork cbc85d2. The fork has the same manager boundary and receives a separate backport. Other loops have no live manager wrapper; their broader legacy error/event behavior is outside this change.

@bkrabach
Brian Krabach (bkrabach) merged commit eecc605 into main Sep 25, 2026
6 checks passed
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