Changelog

Recent documentation and code-mode surface changes.

0.1.17 - 2026-07-11

• Updated the ChatGPT contract for GPT-5.6 families and the flat Intelligence picker, with canonical medium|high|xhigh efforts.
• Added dedicated ChatGPT Work submission through agbrowse web-ai work send --power 1..6 and MCP web_ai_work_send; Chat commands reject the Work surface.
• Added tier-aware long-run deadlines, including 5400 seconds for ChatGPT Pro.
• Added fail-closed schema extraction with agbrowse extract, modular proof-first search guidance, and four installable agent skills.
• Redesigned the GitHub Pages landing page and added Playwright/browser-QA intent routing to agbrowse skills.

2026-07 (July)

• Added agbrowse search command: standalone deep search pipeline with query rewrite, candidate discovery, adaptive-fetch evidence scoring, and --deep web-ai escalation.
• Introduced the adaptive-fetch ladder with TLS-impersonation (curl-impersonate/JA3 fallback), Camoufox browser identity, feed-parser, BM25 relevance scoring, table-extractor, and candidate-discovery phases.
• Added the capability registry for tracking capability status across agbrowse and the cli-jaw mirror surface.
• Implemented annotated screenshots in observe-bundle with bounding box overlays for multimodal benchmark reproducibility.
• Added interstitial detector: WAF profile detection for Cloudflare managed+Turnstile, Akamai Bot Manager, AWS WAF, Imperva/Incapsula, DataDome, and PerimeterX challenge pages.
• Introduced freshness gate: gate:truth-table-fresh enforces that CAPABILITY_TRUTH_TABLE.md is no more than 7 days old at release time.
• Security: SSRF DNS pinning guard on adaptive-fetch to prevent DNS rebinding attacks. Validated redirect chains to block open-redirect exploitation.
• GPT-Pro remediation cycles: web-ai code now validates recovered zip artifacts against the PLAN.md/00_plan.md contract and reports missing plan files as actionable failures.
• Updated the ChatGPT Chat contract for the flat Instant (5.5) / Medium / High / Extra High / Pro Intelligence picker and separate family aliases.
• Added canonical thinking efforts medium|high|xhigh; legacy extended selects High and emits one stderr warning.
• Split long-running timeout tiers into chatgpt-pro=5400 (90 min), grok-heavy=3600, deep-research=3600.
• Added dedicated agbrowse web-ai work send --prompt "..." --power N command and MCP web_ai_work_send tool.
• Added research command family: plan, normalize-results, enrich-fetch, browse-plan for offline Korean/source-sensitive search planning.
• Added runway command family: read-only preflight for Runway Apps/Custom media task-runner with selector contract, status, and queue poll.
• Added skills command family: list, get, path, install for bundled agent skill management.
• Expanded browser primitives: hover, select, check, uncheck, drag, mouse-click, move-mouse, mouse-down, mouse-up, scroll, wait, wait-for-selector, wait-for-text, evaluate, new-tab, tab-close, active-tab, select-tab, reload, resize.
• Added web-ai subcommands: project-sources list|add, context-dry-run, context-render, work send.
• ChatGPT Deep Research: --research deep experimental mode with auto-confirm plan card, report artifact save, and session record.
• ChatGPT image output: --output-image for saving generated images with multi-image sibling naming.
• ChatGPT follow-ups: --follow-up for sequential in-command batch prompts.

2026-06-11

• Added full static developer docs under docs/dev/ with Korean V1 parity.
• Documented web-ai code as the primary new-user artifact flow and code-extract as later recovery.
• Added preflight JSON errors for ChatGPT-only code-mode commands, missing prompts, and multi-zip output conflicts.
• Updated source-of-truth command and capability docs for code-extract and claim-audit.

Support labels

Browser primitives and release gates are ready. Live provider automation, code mode, project sources, and generated images remain beta because provider UIs and account state can change outside this repository. Search and research commands are ready. Adaptive fetch and interstitial detection are experimental.