v5.8.0 — The Trust Edition

Claude Code-native context that just works.

Five pillars, one theme — the experience an agent can parse and a human can read, the proof a third party can verify, and the integration that earns "native":

Added

• Native SessionStart hook (--session-refresh) — every Claude Code session opens with fresh context and a one-line heartbeat: faf: context ✪ 100% — fresh. Freshness-gated (no write, no churn when current), never breaks a session; non-faf directories stay silent.
• faf_setup (tool #35) — the explicit hook installer. Preview first (shows the exact settings JSON + path, writes nothing); confirm: true merges non-destructively into project .claude/settings.json; remove: true takes out exactly the faf hook.
• Determinism parity hash — faf_score/faf_trust carry an engine-agnostic parity receipt: any conformant engine reproduces it; a third party verifies sha256(projection) === parityHash. Falsifiable, not self-asserted.
• The ✪ trust receipt — a render-identical, self-verifying score+parity artifact.
• Typed structured output across the data tools (structuredContent + outputSchema, text retained).

Changed

• Quiet output, everywhere, always — no emoji in tool results; tier glyphs on the quiet ladder ♡ ○ ● ◇ ◆ ★ ✪. One predictable format, no toggle.
• The generated CLAUDE.md block is quiet too — canonical format, matching faf-cli's export.

Fixed

• Enhanced interop — your files are enhanced, never replaced. Interop tools overwrote AGENTS.md / GEMINI.md / .cursorrules / CLAUDE.md wholesale; now a non-destructive injector owns a marked block and preserves everything you've written. A build-failing write-guard makes the old behavior impossible to reintroduce.
• faf_trust works again — self-contained local attestation.

35 tools · 547 tests, 0 fail · IANA application/vnd.faf+yaml

FAF defines. MD instructs. AI codes.