changelog

Every change, documented.

Keep a Changelog format · semantic versioning · raw on GitHub

unreleased

Provider coverage expanded 12 → 37 rules for LLM-inference parity: OpenRouter, Groq, Together, Fireworks, xAI, Perplexity, Hugging Face, Cerebras, Replicate, Vercel AI Gateway, NVIDIA NIM, Chutes, Requesty, Novita, Upstage, Baseten, Baidu Qianfan, fal.ai, DashScope, Moonshot Kimi, Mistral, Cohere, Scaleway.
Generic fallback rules: unattributed sk- 32-char key shape and generic JWT.
OpenAI rule tightened to the T3BlbkFJ marker — no longer double-fires with the generic sk- rule.
Prefixless and generic-prefix shapes are keyword-gated: the provider name must appear in the scanned text before the regex is tried.
Providers with no publicly documented key format get keyword-gated opaque rules — attribution + rotation URL, not shape confidence. No guessed regexes, ever.

planned next

Silent redaction via updatedToolOutput for PostToolUse hooks — the agent never sees the raw value.
Active verification mode — "is this key actually live?"
--file-path flag on scan --stdin for path-based allowlists.

Initial release.
Detection core: regex patterns (12 built-in providers), bash escape-hatch detector, entropy candidates.
Claude Code plugin with 6 hook scripts covering 8 detection rows.
CLI peer with scan, list, close, watch commands.
Pattern layering: builtin → ~/.token-hunt/patterns.local.toml → ./.token-hunt/patterns.local.toml.
Allowlist engine (path-based + fingerprint-based).
rotations.log JSONL store with 72-hour OVERDUE tracking.
Bash escape-hatch detection — the ${VAR:+SET}${VAR:-UNSET} bug class no regex scanner sees.
TOKEN_HUNT_DISABLE=1 opt-out · TOKEN_HUNT_ALLOW=1 one-shot bypass.
Fail-closed: scanner crashes BLOCK with a bypass hint, never silently allow.