Changelog

All notable changes to this project are documented here. The format follows Keep a Changelog, and versions follow Semantic Versioning without a v prefix.

Unreleased

2.0.0 - 2026-10-06

Changed

• Replaced nine standalone skill entries with one claude-in-codex index and nine workflow references loaded on demand.
• Invoke $claude-in-codex followed by a workflow name or request; the cic CLI commands are unchanged.
• The installer links the indexed bundle and removes legacy skill symlinks owned by this checkout while preserving user-owned files and unrelated links.
• Plugin packages include the complete indexed bundle; the listing website is the GitHub repository.

Migration

• Reinstall the plugin or run install.sh, then restart Codex.
• For copied skills.sh installations, remove the old standalone copies through the skills CLI after reviewing ownership, then install the indexed skill. Choose project or global scope explicitly.

1.1.0 - 2026-10-06

Added

• cic improve: an automatic prompt engineer. It drafts two rewrites of a rough brief, scores them on a rubric, grounds the winner in the repository, and prints a ready-to-run cic run command.
• --hint (directional stimulus), --example (few-shot; literal text or @file), and --image (screenshots and diagrams, sent to Claude as image blocks) for cic run. ask gains --hint and --image, and say gains --image.
• Active-prompt pass: a "done" report with confidence below 0.7 gets one targeted turn to resolve the stated uncertainties.
• Same-model self-consistency in cic council (for example --members claude:sonnet,claude:sonnet,claude:sonnet).
• cic setup: writes the Codex exec-policy rule and runs cic doctor, for installs through skills.sh, pip, or uv.
• Codex plugin marketplace manifest (.agents/plugins/marketplace.json).
• OpenAI (ChatGPT / Codex) plugin package:
  • portable plugin.json (Agent Plugins 1.0) with the directory listing
  • icons designed in HTML/SVG (assets/icon.html), exported with scripts/export-icon.sh
  • PRIVACY.md and TERMS.md
  • a submission build that validates the package against the directory rules
• Commit, changelog, and PR-message chores automatically get the repository's recent commit subjects as style examples.
• cic run --dry-run lists the promptingguide.ai techniques each brief applies.
• Context profiles (--profile standard|lean|minimal|full). The default, standard, keeps Claude Code's system prompt cacheable across repos and skips MCP servers. In measurements, a second job pays about 3k new prompt tokens instead of about 10k (about 60% cheaper).
• Token accounting: every report shows prompt and output tokens and the cached share; cic stats totals usage by model with tips. Resumed sessions are billed per job, not as running totals.
• cic result --full. Default reports are length-capped (12 changes, 8 findings, about 6k characters of text) to save orchestrator tokens.
• Docs: docs/how-it-works.md, docs/cli.md, and docs/prompt-audit.md (per-template technique audit and its two improvement iterations).

Changed

• Every task template was upgraded against promptingguide.ai:
  • generated-knowledge steps for implement and refactor
  • an explicit ReAct loop for debug
  • tree-of-thought viability ratings for plan
  • an evidence-then-answer chain and per-option trade-offs for research
  • word-budgeted answer formats with an example for explain and ask
  • run-what-you-document for docs
  • refute-before-reporting for reviews
• Repair messages (verification failures, continue, pair findings) now require a reflection: name the failing check, recall the approaches already tried, and don't repeat a failed one.
• The work contract defines calibrated confidence anchors.
• Council members state the strongest alternative they rejected. The moderator trusts a majority only when it is backed by cited evidence.
• The README was rewritten for users. Architecture and reference material moved to docs/.
• Failing-check output is focused (signal lines plus summary) before it goes back to Claude. Repair turns are capped at 6 KB of check output; council answers at 3.5 KB each; pair diffs at 40 KB.
• Skill descriptions, which Codex keeps in context in every session, were cut from 3.7k to 1.9k characters.
• --lean now means the lean profile (no user plugins or hooks) instead of safe mode; use --profile minimal for safe mode.

1.0.0 - 2026-10-06

Added

• The cic CLI: routed delegation (Haiku, Sonnet, Opus) with a detached worker, a strict success check, independent --verify re-runs, repair turns on the same session, model escalation, steering, model switching, cancel, and reply.
• review, ask, named session and say conversations, --plan (Opus plans, then a fresh run executes), and --worktree isolation.
• Multi-agent modes: pair (driver and navigator; Claude↔Claude or Claude↔Codex), council (Claude, Codex, and Gemini with a moderator), the file-based message bus, and serve bus agents.
• Nine Codex skills, install.sh and uninstall.sh, the Codex exec-policy rule, and cic doctor.
• CI on Linux and macOS (Python 3.10–3.14), and a tag-based GitHub release workflow.