Violin v4.0.0

Breaking changes

• Require Hermes 0.21.5 or later. The container pins Hermes v2026.9.24 to commit f97608f178d1ffeca59860195ab7da295f7c8e5f and runs it in a supported Python 3.13 environment.
• Version engagement runtime state and bind loaded skills to their delivered content. Start a new engagement when moving from Violin 3.3; existing engagement state is not migrated.

Fixed

• Distinguish receipt authenticity from demonstrated vulnerability proof and require claim-to-evidence review before reporting.

• Burst responses report pending workflow review from persisted batch state and identify the batch required by PTT updates.

• Target commands now reject paths into another engagement through the normal single-command and burst guard.

• PTT task updates no longer advertise ignored review fields; record execution outcomes with violin_review_batch. The ignored research_attempted batch-review input is removed, while hypothesis research recording remains available.

• Execution accounting tests and smoke checks use the same retry-safe state transition as the runtime; obsolete state helpers are removed.

• Preserve canonical hypothesis and PTT sections during Markdown updates, normalize hypothesis arguments and vulnerability classes, and explain valid skill choices.

• Finalize execution receipts safely on retries, keep captured HTTP response bodies intact, and apply a consistent secret-redaction contract.

• Clarify closeout sequencing, local administrative command admission, and coverage/methodology disposition rules.

• Reorganize guard code by commands, engagement state, evidence, and skills; avoid repeating heavy release checks after CI runs them separately.

• Authenticate saved finding proof with explicitly cited receipts, scope each command receipt to its own outputs, and preserve authentication for identical-byte rewrites.

• Keep burst review and bootstrap state updates atomic, explain denied burst slot accounting, and allow an exact repeat after a failed execution.

• Provision research and discovery tools in the container, resolve the Kali httpx command correctly, and avoid requiring the optional SecLists archive.

• Bound coverage-close diagnostics while retaining actionable status, evidence, and key guidance.

• Explain semantic no-progress warnings and distinguish advisory pivots from active-lock recovery.

• Bound foreground execution to Hermes deadlines and report commands skipped after a burst stops.

• Reject malformed task and hypothesis bindings before mutation, preserve inline hypothesis placeholders, and reset no-progress streaks after productive reviews.

• Keep proof preflight consistent with execution normalization and explain coverage and scope-denial remedies.

• Initialize the documented reconnaissance technology evidence directory before benchmark execution.

Changed

• Open saved evidence before batch review and claim promotion, including relevant sections omitted from a truncated view.

• Pin benchmark source and runtime metadata, pass provider selection explicitly to Hermes, and keep the persisted Hermes model/provider defaults intact.

• Require at least 15 of 20 confirmed findings together with complete coverage, methodology, and a comparable benchmark protocol.

• Defer container browser provisioning until redirect and subresource scope enforcement is available.

• Remove obsolete accounting paths and split runtime, skill-receipt, and test modules into cohesive units.

• Run benchmark validation and calibration in Docker, keep the evaluator outside the agent image, and publish sanitized aggregate summaries.

• Support Violin 4.x; end maintenance of the 3.x release line.