v0.2.18

0.2.18
Claude Code and Codex now verify whether teaching landed exclusively through developer self-report: they ask whether the explanation is understood or needs clarification and trust the answer. The skills explicitly prohibit quiz-style checks such as restating a mechanism, explaining it in the developer's own words, predicting behavior, selecting an answer, or responding to another topic-specific question.

v0.2.15

0.2.15
Fixed
Claude Code and Codex feature skills now distinguish calibration levels from engagement signals. A probe answer such as learning or new sets teaching depth; it cannot be passed to fluencyloop calibration signal. The CLI now explains that error directly.
Claude Code treats "not comfortable" answers as new, gives the substantive explanation, and waits for an AskUserQuestion response before journaling, calibrating, or continuing.
Added a project-scoped Claude Code permissions guide for native Windows that reduces routine FluencyLoop, editing, and read-only Git prompts without granting broad Git or Bash access.

0.2.14
Fixed
Codex no longer substitutes ASCII diagrams when an Artifact is unavailable. It points to the generated Markdown file and recommends opening it in an IDE Markdown preview, such as VS Code.
Codex probes now frame explanation neutrally. learning and new answers require a substantive explanation and a standalone comprehension question before the workflow can continue. Claude Code workflows are unchanged.

0.2.13
Fixed
Codex stage skills now request elevation for their own initialization, including backfill's skipped-loop setup path. Feature design has a visible teaching gate before implementation, and planning and feature gates distinguish a learning pause from a build or merge block.
Codex review now honors the settled feature-handoff: automatic preference instead of asking again before opening the PR, and uses the feature's recorded base when it creates that PR. Claude Code workflows are unchanged.

0.2.12
Fixed
Codex planning now requires a visible architecture teaching turn before task decomposition, roadmap, constitution, or ticket work. Unknown, learning, and new domains pause for the developer's response. Claude Code's planning workflow is unchanged.

0.2.11
Fixed
Codex feature runs now require a user-visible teaching turn before decisions are journaled. Unknown, learning, and new domains pause for the developer's response; calibration signals require actual engagement. Claude Code's workflow is unchanged.

0.2.10
Fixed
For literal fluencyloop init requests, the Codex router now requests sandbox elevation before its first execution. This avoids a denied first attempt when the initializer creates protected Git metadata. Claude Code's workflow is unchanged.

0.2.9
Fixed
Codex stage skills now use the plugin-qualified names $fluencyloop:plan, $fluencyloop:feature, $fluencyloop:review, and $fluencyloop:backfill, without repeating the plugin name in the picker. Claude Code commands remain /fluencyloop:.
Codex skills now invoke the bundled dispatcher without exposing its internal path variable, and fluencyloop init preserves Git's original error when repository initialisation fails.
Literal Codex CLI requests now run without preflight narration, inspection, or an automatic transition into a FluencyLoop stage.
Codex now maintains a managed fluencyloop PATH shim on macOS, Linux, Git Bash, and WSL, so its command transcript shows the stable command name instead of a versioned plugin-cache path.

v0.2.6

0.2.6
fluencyloop slice-context now handles an unborn Git branch without attempting to diff an invalid HEAD; it returns the staged and untracked first-project files as the initial slice.
When a live design Artifact cannot be rendered, feature, plan, and backfill workflows now show an ASCII sketch in chat before pointing to the committed Mermaid document.

0.2.5
Codex startup refresh hooks now use only Codex's PLUGIN_ROOT and no-op safely when it is absent, preventing a session-start failure with exit code 127. The Windows hook follows the same guard.
Plugin package tests now read repository text as UTF-8, keeping them reliable on Windows code-page defaults.

v0.2.4

Claude Code skills now invoke the plugin's bundled launcher explicitly, validate the paths it returns, and refuse to hand-scaffold legacy .fluencyloop session files.
Codex's startup hook now recognises both supported installed-plugin root layouts, so its marketplace refresh reaches the current snapshot layout.
The Claude and Codex plan/feature stages now initialise Git automatically in a project directory that does not already have a repository, without prompting the developer.
The plugin package test now reads repository text as UTF-8, so it runs on Windows code-page defaults.
Claude installation guidance now distinguishes slash commands from Bash-tool commands and documents third-party marketplace update behaviour accurately.