Commit Graph
3 Commits
Author SHA1 Message Date
felixxia-oai 4b1cf9a4c9 Use conversations for all Guardian v2 async classification (#56556)
## What changed

- Always use the retained conversation backend for asynchronous classification; keep snapshot sampling test-only.
- Accept legacy `async_classifier_mode` values in configuration and model metadata but ignore them. Emit a startup warning when `[features.guardianv2].async_classifier_mode` is configured.
- Ignore legacy classifier mode differences when checking model-switch compatibility, and fail closed if the conversation backend is missing.

## Testing

Update coverage for legacy configuration warnings, model-switch compatibility, retained approval-review context, and conversation-based scoring.

GitOrigin-RevId: 5e8bb7e4e6e7d1589572f38bfe2498c9fa6825a0
2026-10-23 18:48:10 +00:00
felixxia-oai 47a8bd7321 Avoid duplicate sync reviews in Guardian classifier continuations (#49795)
## Why

Retained Guardian classifier conversations already contain previously delivered synchronous reviews. Appending those reviews again repeats evidence and consumes input budget.

## What changed

Track each review completion with a stable identity in host-only history metadata. Filter already delivered reviews before composing and budgeting continuation input, while preserving distinct completions even when their text is identical.

Fresh conversations receive the review evidence again, and shortened messages no longer count as complete delivery. Delivery metadata stays out of model requests.

## Testing

Add coverage for delivering retained reviews once, appending newly completed reviews, restoring evidence after failed streams, and retaining reviews when history lacks complete delivery metadata.

GitOrigin-RevId: b8d58ffdd7d9c484676af707c65fa50020aae73c
2026-10-01 01:13:29 +00:00
felixxia-oai 726f1492db Add conversation mode to Guardian v2 async classification (#49793)
## What changed

- Add `features.guardianv2.async_classifier_mode` with `snapshot` and `conversation` modes. Local configuration overrides model defaults; snapshots remain the fallback.
- Retain completed classifier requests and responses in conversation mode, appending transcript deltas without repeating previously delivered user instructions.
- Rebuild retained history when authorization, classifier instructions, or parent compaction changes, after an incomplete response, or when the next request exceeds the configured token limit. Add `async_classifier_conversation_token_limit`, defaulting to 100,000 tokens, while preserving the model's input limit for fresh requests.
- Serialize conversation requests in observation order, bound outstanding work, and reject stale work after authorization changes or thread shutdown. Record retained and new input token costs.

## Testing

Add tests for configuration precedence, history reuse and resets, ordered admission, queue overflow, and stale or stopped requests. Exercise the approval matrix in both modes, including resume, fork, mandatory review, authorization changes, and stream failure after an early score.

GitOrigin-RevId: 35b9800104648cde4f467086e678dee12458c3e5
2026-10-01 01:11:33 +00:00