0.2.18

Handle Windows Codex CLI launch differences

Windows users can hit Codex CLI failures when Obsidian launches the npm codex.cmd wrapper through Node spawn, and some Codex versions reject the explicit '-' stdin marker. Codexian now omits that marker, keeps macOS/Linux on the normal executable path, and only redirects Windows codex.cmd to the underlying @openai/codex/bin/codex.js entrypoint when it exists. Model and reasoning selectors also update their visible labels immediately after saving.

Constraint: macOS and Windows Codex CLI launch paths differ

Rejected: Apply the codex.cmd workaround globally | macOS/Linux should keep direct executable launch behavior

Confidence: high

Scope-risk: moderate

Directive: Keep future Windows fixes platform-gated so macOS/Homebrew and Linux paths remain unaffected

Tested: npm run build

0.2.17

Surface visual generation progress in the modal

Visual prompt drafting and asset generation now forward Codex CLI progress events into the visual progress modal. Users can see Codex steps while the prompt is drafted, the image asset is generated, the output is verified, and the embed is inserted.

Constraint: Codex CLI image generation may take long enough that users need visible progress feedback

Rejected: Console-only progress | users running from Obsidian do not reliably watch developer console output

Confidence: high

Scope-risk: narrow

Directive: Keep visual progress visible in the modal, not only in chat or console

Tested: npm run build

0.2.16

Show Codex work as a step timeline

Codexian now accumulates progress events into a visible timeline instead of replacing a single hint line. The provider also recognizes more command, file, and action-like Codex CLI progress lines so users can see more of the agent's working process while waiting for the final response.

Constraint: Codex exec only exposes text progress, not the full interactive TUI state

Rejected: Show only the final answer | users need confidence that Codex is actively reading, running, building, and verifying

Confidence: high

Scope-risk: moderate

Directive: Keep progress output concise; avoid dumping entire stdout transcripts into chat

Tested: npm run build

0.2.15

Add one-click Obsidian Skills setup

Codexian settings can now clone kepano/obsidian-skills and copy its skills into ~/.codex/skills, giving Codex CLI Obsidian Markdown, Bases, Canvas, CLI, and Defuddle skill instructions without requiring users to run commands manually.

Constraint: obsidian-skills is an Agent Skills repository, not an Obsidian community plugin

Constraint: Codex CLI discovers skills from ~/.codex/skills

Rejected: Install into the vault plugins folder | that would treat skills as an Obsidian plugin and not make them available to Codex CLI

Confidence: high

Scope-risk: moderate

Directive: Keep this installer pointed at Codex skills unless Codex changes its skills discovery path

Tested: npm run build

0.2.14

Apply GPT Image prompt recipes to visual generation

Codexian now turns short user directions into structured visual prompts using the GPT Image 2 recipe pattern: category selection, structure selection, and variable filling. Prompt drafts include subject, composition, style, environment, lighting, typography, details, and aspect ratio, with dedicated recipes for thumbnails, infographics, posters, UI mockups, products, e-commerce hero images, avatars, and storyboards.

Constraint: Core visual generation must continue to use Codex CLI without OpenAI API keys

Rejected: Keep five generic visual modes | too weak for thumbnail, product, UI, avatar, and e-commerce prompt quality

Confidence: high

Scope-risk: moderate

Directive: Future image prompt changes should preserve the structured recipe fields

Tested: npm run build

0.2.13

Make slash command suggestions keyboard-scrollable

The slash menu rendered but keyboard navigation did not keep the selected command visible, and wheel scrolling could be intercepted by the Obsidian pane. The menu now expands the command set, contains wheel scrolling, and scrolls the selected command into view on arrow-key navigation.

Constraint: Codexian should preserve the ObsidianCode-like slash command affordance while still sending slash prompts raw to Codex CLI

Rejected: Only add more commands | the underlying visibility bug was missing scrollIntoView behavior

Confidence: high

Scope-risk: narrow

Directive: Keep slash keyboard navigation and dropdown scrolling tested together

Tested: npm run build

0.2.12

Expose the current frontier Codex model

Codexian's sidebar model picker was still pinned to the older hardcoded model set, so users could not select gpt-5.5 from the Obsidian UI even when their Codex CLI supported it. The default model and empty-setting fallback now use gpt-5.5 while preserving manual model entry for future Codex releases.

Constraint: Codex CLI remains the authority for whether a model is available to the user's account

Rejected: Hard-code only older gpt-5.4 options | blocks users from selecting the current frontier model

Confidence: high

Scope-risk: narrow

Directive: Keep model picker current, but continue allowing manual model IDs in settings

Tested: npm run build

0.2.11

Preserve note properties and surface slash commands

Generated visual embeds now detect YAML frontmatter and insert below the closing delimiter, preventing Obsidian properties from being corrupted. The chat composer now renders Codex-style slash command suggestions instead of only passing slash prompts through invisibly.

Constraint: Obsidian note properties must remain valid YAML frontmatter

Rejected: Raw slash pass-through only | it preserved CLI input but gave users no visible command affordance

Confidence: high

Scope-risk: moderate

Directive: Keep embed insertion properties-safe when changing visual asset placement

Tested: npm run build

0.2.10

Keep visual generation aligned with Codex CLI capabilities

PNG generation depends on Codex CLI image_generation being available, so settings now expose a direct update and feature-enable path. Slash-prefixed prompts are passed through raw so Codexian does not break Codex CLI command syntax before exec receives it.

Constraint: Codexian must use authenticated Codex CLI, not OpenAI API keys

Constraint: BRAT users need release assets matching manifest version

Rejected: Claim direct gpt-image-2 model selection | Codex CLI exposes image_generation as the supported surface, not a plugin-level model selector

Confidence: high

Scope-risk: moderate

Directive: Do not reintroduce SVG-only modal copy when PNG output is selected

Tested: npm run build

0.2.9

Support Codex built-in PNG image generation

Codex CLI now supports built-in image generation through the image_generation feature flag, so Codexian can offer real PNG generation without using OpenAI API keys. This adds an output selector for PNG versus SVG, passes the selected output type through prompt drafting, strongly instructs Codex to use the built-in image tool only for PNG, and embeds the generated PNG at the top of the note.

Constraint: Codex built-in image generation saves initially under ~/.codex/generated_images, so the prompt requires copying or moving the PNG into the vault attachment path.

Rejected: Replace SVG generation entirely | SVG remains useful for text-safe Korean labels and deterministic Obsidian embeds.

Confidence: medium

Scope-risk: moderate

Tested: npm run build