You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
[2.12.0] - 2026-10-12
Added
The scorer labels a write into a location an agent host loads again in a later session: skill, subagent, command, rule and hook directories, Claude Code settings, per-host config under .codex/, .gemini/ and .agents/, Claude Code project memory, and the files CLAUDE.md, AGENTS.md, GEMINI.md, SKILL.md, .cursorrules, .windsurfrules, copilot-instructions.md and .mcp.json. A write tool counts by its path parameter. A shell command counts when it redirects, tees, copies, moves, links or edits in place into one. Content that only mentions such a path does not count. The label is agent.persist_write, and sequence steps match it like an action name.
Built-in sequence outbound_then_persist: an outbound call (web fetch or search, mail, a shell network command) followed within eight calls by a persistent write sets the sequence signal to 0.5. At the starting weights, where each of the five signals counts equally, that raises the risk score by 0.1. Content fetched in one session and written into a skill or instruction file steers every session after it, the cross-session instruction propagation measured in arXiv 2610.13227.
A file written by outbound_then_persist is remembered past its session in a lineage store (~/.vaara/lineage.json, or VAARA_LINEAGE_PATH). A later call that names the file, in any session, carries the label agent.tainted_read and says so in its decision reason. Built-in sequence tainted_then_outbound: such a call followed within eight calls by an outbound call sets the sequence signal to 0.6, which raises the risk score by 0.12 at the starting weights. A write is remembered only when it can land. A denied write, including one a deny rule refused, leaves no mark. LineageStore.clear(path) removes a mark. The scorer uses a store only when given one (AdaptiveScorer(lineage=...)); the Claude Code hook gives it one.
Fixed
Sequence patterns never matched in a hook that runs once per tool call, such as vaara hook pre-tool-use. Each call built a new scorer with an empty window, so it never saw the call before it. The pipeline now replays the session's earlier actions from the trail into a scorer that has not seen the session, the same records authority decay reads. The replay fills only the sequence window: track record, burst and denial counts are not counted twice. A long-lived pipeline does not replay a session it has already scored.
In the Claude Code hook without a policy, a built-in tool call that sends data out, writes an agent-persistent file or names a planted file now goes through the enforced pipeline. It used to take the record-only path, where a matched sequence could not hold it. At the default thresholds these calls are allowed as before.