v3.12.2

Antigravity permissions-adapter repair.

• Consumes Antigravity's native nested toolCall payload and emits native top-level allow/deny decisions.
• Normalizes command, file, multi-edit, and MCP calls into Signet's existing canonical policy model.
• Makes native MCP wrapper identity authoritative, rejects spoofed identity fields, and fails closed on malformed wrappers.
• Ships an Antigravity hook configuration plus README and project-site setup documentation.
• Aligns Cargo, Kindex, MCP Registry, and Claude plugin metadata on v3.12.2.
• Passes 231 tests, formatting, lint, crate packaging, and crates.io publish dry-run verification.

v3.12.1

Durable safety-default release.

• Ships the GitHub remote-owner identity guard and embeds its check for atomic first-use installation.
• Restores hard denial of ephemeral Task* state in favor of Kindex tasks.
• Requires same-session Kindex engagement before direct edits; missing distinct host session identity fails closed.
• Reconciles current compiled built-ins over stale machine policy snapshots while preserving distinct host rules and user overrides.
• Passes 218 tests, shellcheck, release build, and crate package verification.

v3.12.0

Add Antigravity and OpenCode hook adapters, and change prefer_persistent_task_store default policy to Ask

signet-eval v3.11.1

Fixed

• Scoped preflight state to real client chat/session identifiers before falling back to SIGNET_SESSION.
• Hook evaluation now extracts session identifiers from client payloads for preflight, pause, and disable checks.
• Session-scoped preflights no longer deactivate global preflights or preflights from other sessions.

Verification

• cargo test: 156 unit, 9 CLI integration, 38 hook integration passed.
• cargo publish --dry-run passed.
• Published signet-eval 3.11.1 to crates.io.

signet-eval v3.11.0

signet-eval v3.11.0

Adds advisory INJECT rules for probabilistic behavior nudges near tool-call time, without changing deterministic authorization semantics.

Added

• INJECT decision variant evaluated in a separate post-auth pass.
• Recency-weighted trigger modes: constant, step, linear, and exponential.
• Payload sources: inline text, text_file under ~/.signet/injections/, and HMAC-signed allowlisted from_command output.
• Template substitutions: {tool_name}, {cwd}, {date}, {matched_param.X}.
• signet-eval injections and signet-eval inject-test <rule>.
• examples/inject_examples.yaml.

Fixed

• block_rm no longer false-positives on words/paths containing rm as a substring, such as drone_swarm or firmware.
• matches(parameters, ...) now evaluates against serialized parameters, matching contains(parameters, ...) semantics.

Verification

• cargo fmt --check
• cargo test (199 tests)
• cargo package --allow-dirty
• cargo publish --dry-run --allow-dirty
• mcp-publisher validate server.json

v3.10.1 — fix dropped user rules; ship Task* deny default

Fixed

• load_merged_policy no longer drops user rules from rules.yaml when the system policy.yaml is missing or malformed. Previously this branch returned default_policy() and silently discarded every user rule, so on hosts without a per-host system policy both signet_test and real hook enforcement were no-op for user-installed rules. Missing/malformed system policy now falls back to the hardcoded baseline (self-protection + system defaults) and still merges user rules on top.
• Three regression tests added covering missing system policy, missing system policy with self-protection preserved, and malformed system policy with user rules preserved.

Added

• New locked default rule prefer_persistent_task_store: denies Anthropic's session-local Task* tool family (TaskCreate / TaskUpdate / TaskList / TaskGet / TaskOutput / TaskStop) and routes the agent to a persistent task store such as kindex's mcp__kindex__task_*. Ships locked so it cannot be silently overridden by an unlocking user rule.

v3.10.0

Added

• Codex hook adapter support via --adapter codex and --adapter codex-permission.
• Codex PreToolUse deny-only enforcement mapping and PermissionRequest allow/deny/defer mapping.
• Codex hooks example at hooks/codex-hooks.json.
• Integration tests for Codex hook response shapes.

Changed

• Documentation and crate metadata now describe agent-agnostic support for Claude Code and Codex.

Verification

• cargo test
• cargo package --allow-dirty

v3.9.0 — Session-scoped preflights, vault protection, broad-rule advisory

What's New

Vault passphrase protection (locked rule)

AI agents can no longer run signet-eval setup or signet-eval unlock. This prevents an agent from setting a vault passphrase and locking the human out — which happened in practice.

Overly broad preflight constraint advisory

Submitting a preflight constraint with tool_pattern: ".*" and no conditions is now rejected with an explanation. An agent previously locked itself out by filing a universal constraint that matched every tool call including its own MCP tools. Use force: true to bypass if genuinely needed.

Session-scoped preflights (SIGNET_SESSION)

When the SIGNET_SESSION env var is set, preflights are scoped to that session. Each Claude Code instance sees only its own preflights plus global (session_id=NULL) ones. Deactivation, lock checks, and overrides are all session-aware. Backward compatible — unset SIGNET_SESSION means identical behavior to v3.8.0.

Files changed

• src/policy.rs — new locked rule protect_vault_passphrase (8th self-protection rule)
• src/vault.rs — session_id field on Preflight, DB migration, session-filtered queries
• src/mcp_server.rs — broad-constraint advisory, session_id capture + display
• src/main.rs — CLI preflight status shows session scope

Tests

138 tests passing (107 unit + 7 integration CLI + 24 integration hook)

v3.8.0

What's New

Added

• validate --fix CLI flag — auto-fixes clampable issues (gate.within, ensure.timeout clamping) and removes broken unlocked rules; writes updated policy and re-signs if vault exists
• validate --fix --dry-run — previews what --fix would change without writing to disk
• status now shows complete enforcement state: global disable, disabled sessions, global pause with expiry, per-rule/per-session pauses with timestamps
• status shows enforcement overrides even without a vault (file-based state is independent)

Fixed

• enable (no flags) now clears both global disable AND all session disables in one invocation; previously required running twice if both existed
• status no longer returns exit code 1 when vault is not set up — enforcement info is still useful without a vault

v3.7.2 — Indefinite pause

What's new

Indefinite pause — pause 0 means "until I say resume", but only with --session:

signet-eval pause 0 --session signet --rule require_plan_before_code
# paused forever until:
signet-eval resume --rule require_plan_before_code --session signet

pause 0 without --session is rejected to prevent forgotten global pauses.