ps4emp/CacheSentry
Audit LLM prompt traces for KV cache-reuse stability.
View on GitHubTrust Signals
- Scorecard Score
- not yet scored
- Maintenance Recency
- Stale
- License
- None
Inputs
| name | description | required | default |
|---|---|---|---|
| trace-path | Path to the JSONL trace file to audit | yes | — |
| model | Hugging Face model name to emulate tokenization and chat templates | yes | — |
| input-format | Input format: cachesentry or openai-chat | no | cachesentry |
| render-mode | How to render inputs for diffing: simple or hf-chat-template | no | simple |
| fail-on-severity | Severity threshold to fail the build: off, low, medium, high, critical | no | high |
| max-total-lost-reusable-tokens | Fail if total lost tokens exceeds this | no | — |
| max-pair-lost-reusable-tokens | Fail if any pair's lost tokens exceeds this | no | — |
| min-average-stable-prefix-ratio | Fail if avg stable ratio falls below this (0.0 to 1.0) | no | — |
| fail-on-culprit-kind | Comma-separated list of culprit kinds to fail on (e.g. uuid_or_request_id,timestamp) | no | — |
| output | Path to write the trace report | no | — |
| redaction-mode | Redaction mode | no | mask |
| hash-salt | Optional salt for hash redaction. Pass via secrets. | no | — |
| github-annotations | Emit GitHub Actions annotations for violations | no | true |
| sarif-output | Optional path to write SARIF report | no | — |
Outputs
no outputs