Changelog

0.2.0 — 2026-08-15

Added

• An input-surface ladder, so elicitation uses the richest interaction the session offers rather than defaulting to prose questions: a structured question tool (AskUserQuestion, ask_user) where one exists, the clickable ballot where a skill cannot render its own UI, numbered single-keystroke options otherwise. Capability is detected per session, because it varies by mode within a single host.
• render_ballot.py --solo, turning the existing ballot into a solo elicitation surface. This is the richest interaction available in ChatGPT, where interactive UI is reserved for MCP servers.
• decision_engine.py apply-ballot, merging a returned ballot into the model as ratified user judgements so no answer is retyped.
• references/host-capabilities.md, recording what each host offers a skill.

Fixed

• The ballot exported every rating slider whether or not the user moved it, so untouched cells were merged as the user's own judgement. Untouched ratings are now omitted and raised by the ratification queue instead.
• The ballot failed outright when a browser blocked local storage. It now degrades to in-memory progress with a visible warning.
• fully_ratified reported true for a model too incomplete to rank, where an empty queue meant nothing had been assessed rather than everything.

0.1.0 — 2026-08-15

• Initial backend-free Agent Plugins 1.0 package.
• Added adaptive general-purpose decision workflow.
• Added deterministic AHP/incomplete pairwise engine, anchored ratings, hierarchy support and sensitivity analysis.
• Added critique, prioritisation, group and decision-audit modes.
• Added self-contained interactive decision reports.
• Added self-contained offline group ballots with local progress and JSON export.
• Added AIJ/AIP group aggregation and disagreement diagnostics.
• Added ChatGPT/OpenAI and portable Agent Plugins manifests.
• Added GitHub Copilot installation instructions, tests and package validation.
• Added Claude Code support: .claude-plugin/plugin.json, a rohde-tools marketplace entry and decide, critique and ballot command entry points.
• Made bundled-script invocation host-portable, so the engine runs when the working directory is the user's project rather than the skill directory.
• Prepared for GitHub publication: instructive README with an AHP primer and verified install instructions for all four hosts, plus repository metadata in the portable and Claude manifests.
• Added human-steering support: provenance on comparisons, ratings and direct weights; a ratification audit reporting ratified weight/score shares with an influence-sorted queue of outstanding judgements; an initial_lean field; and a challenge command reporting what would have to change for a user-preferred alternative to lead.
• Added references/steering.md, a reconciliation step in the workflow, override and provenance sections in the decision record, and a provenance panel in the report.
• Warned when a declared hard constraint is never screened against an alternative, instead of silently treating it as passed.