v0.2.0.0 — Speed, reliability & adaptive red-team loop

Mirage v0.2.0.0

Reliability and agent-quality release from live red-team runs.

Added

• Pace modes (fast / balanced / thorough) with stop-on-first-win
• Full attack transcript (every probe input/output) in JSON and HTML reports
• Strategy hints: pivot after thrashing one attack class; enumerate before name-guessing
• Auto-continue when target replies look truncated (mid-JSON, open braces, cut mid-flag)
• Provider HTTP retry helpers and clearer network errors

Changed

• Agent loop runs in the side panel (stable LLM API access)
• Faster default attacker models (haiku / gpt-4o-mini / grok-4-fast)
• Host permissions allow any http(s) page under test
• Content script waits for real assistant replies; strips user-bubble echoes
• Goal-judge gated; reports show Matched: values not regex sources

Fixed

• False wins from echoed canaries / placeholder Flag{...}
• Bare Failed to fetch from ephemeral service-worker LLM calls
• Host-permission errors no longer burn the full iteration budget

Install

1. Download mirage-0.2.0.0-chrome.zip below
2. Unzip → chrome://extensions → Developer mode → Load unpacked → select the unzipped folder
3. Accept site-access if prompted, open the target tab, Launch Mirage

Full changelog: see CHANGELOG.md

PR: #1