CHANGELOG.md

Changelog

[Unreleased]

[0.12.0] - 2026-09-04

Highlights

• Take photos directly in image questions and attachments.
• Choose a physical or virtual camera, then preview and retake photos before using them.
• Stay in control: the camera starts only when you ask, and no audio is recorded.

Added

• Added browser camera capture for image questions and per-question attachments, with camera selection, a live preview, and photo confirmation. Existing uploads, drag-and-drop, and image paths still work as before. Thanks to @tryingET for #21.

[0.11.0] - 2026-08-11

Added

• Added the interview.launcher setting for choosing Glimpse, browser, or Orca form opening. Thanks to Warren Winter (@w-winter, #19) for the contribution.

[0.10.1] - 2026-08-02

Added

• Added a strict, no-emit TypeScript check with exact optional-property validation.

Fixed

• Response unions now preserve homogeneous choice/image arrays, non-attachment uploads require image questions, and optional saved-interview fields are omitted instead of assigned undefined.
• Remote-session detection no longer blocks Pi while checking active logins, and successful host-side browser opens now explain clearly where remote users should open the interview.
• Moshi and other in-app browsers now restore the post-submit close guidance when the preview returns from the background after ignoring window.close().
• Low-port bind retries now remove the stale listening callback before scanning to the next candidate.

[0.10.0] - 2026-08-01

Added

• One-tap access from remote and Moshi sessions: tokenless top-level navigations from loopback now receive a 200 landing shell that hops to the tokenized form URL, so Moshi's browser preview lands on a working form over its per-session SSH forward. The shell is stateless server-side: it never touches the heartbeat, so a drive-by request cannot arm the abandon watchdog. HEAD / answers 200 (headers only) and the server prefers low ports (8377+, scanning forward on collision, ephemeral fallback) because moshi-hook's discovery ignores ephemeral-range listeners and expects 200 on GET /.
• When the browser launch fails or the session looks remote (SSH_CONNECTION/SSH_TTY, which also catches mosh, or an active remote login in who - covering pi started locally but driven over ssh), the tool prints the form URL with tailored hints: a Moshi preview tip (only when the moshi-hook gateway answers on 127.0.0.1:24543), an exact ssh -L command with the real port, and a mosh can't-forward-ports caveat.

Fixed

• Over SSH on macOS, the Glimpse window no longer opens on the host's own display where closing it silently cancelled the interview; Glimpse is skipped for remote-looking sessions.
• The Glimpse window now closes on every finish path, not just abort: submitting the form from another client (e.g. Moshi's browser) used to leave a dead form window open on the desktop.
• When window.close() is ignored after submit (in-app browsers like Moshi's preview, or tabs the script didn't open), the success overlay now tells the user to close the page and return to the terminal instead of appearing stuck.

Security

• Requests with a non-loopback Host header are rejected with 403, closing DNS-rebinding attacks against the local form server.
• ~/.pi/interview-sessions.json (which stores tokenized URLs) is now written with mode 0600 instead of world-readable.

[0.9.1] - 2026-07-30

Changed

• The session-timeout countdown now schedules updates at second boundaries instead of running a continuous animation-frame loop, and activity events move the expiration deadline instead of recreating timers, cutting idle renderer work by ~96%.

Fixed

• Clicking "Stay" on the expiration warning no longer leaves the previous expiration timer armed, which could expire the session shortly after choosing to stay.

[0.9.0] - 2026-07-25

Added

• Added the opt-in interview.glimpseFloating setting to keep native macOS Glimpse windows above other windows. Thanks to Ben Tang (@0xbentang, #5) for the request.

Changed

• Updated the extension to Pi 0.82.1 APIs, provider-backed model requests, and bundled core-package peer dependencies.

Fixed

• Linux browser auto-open now uses detached launcher fallbacks before an observable Pi exec attempt, while preserving the manual URL fallback. Thanks to @borysow (#6) for the report.
• Browser auto-open failures now keep the interview server available at its manual URL without emitting stale updates after cancellation. Thanks to Yoni Lerner (@yonilerner, #9) for the implementation and @borysow (#6) for the report.

[0.8.7] - 2026-05-01

Changed

• Ask results are now saved automatically, so the inline Ask panel no longer shows separate Pin, Move up, Use rewrite, or Add rewrite as option actions.
• Question-to-question keyboard navigation now uses Cmd+Left and Cmd+Right instead of bare left/right arrows, leaving plain arrows available for text selection and caret movement.

Fixed

• Option rows now allow normal text selection and copying without toggling the radio/checkbox, and focused text inputs now paste text explicitly without being intercepted by image clipboard handling.

[0.8.6] - 2026-04-22

Fixed

• Migrated extension tool schemas from @sinclair/typebox to typebox 1.x so packaged installs follow Pi's current extension runtime contract.
• Choice-question rerenders now restore only the question being updated instead of clearing every single/multi selection in the form. This fixes the lingering radio-group bug where picking an option in one question could wipe answers from earlier questions and made submitted payloads look incomplete to the agent.

[0.8.5] - 2026-04-22

Fixed

• Radio selections now draw their checked state directly on the input element instead of relying on a ::before pseudo-element, which makes selected single-choice answers render reliably in the native macOS interview window.

Changed

• Removed the repo package-lock.json because this package ships as a library package and the lockfile was only adding stale dependency metadata without affecting published runtime installs.

[0.8.4] - 2026-04-22

Fixed

• Rich-option clarification fields now appear for selected single/multi options with inline content blocks instead of being incorrectly hidden behind the old plain-string-only gate.
• Option-level recommendation normalization now strips temporary option-level recommendation metadata after canonicalizing to question-level recommended and conviction, so saved snapshots and revalidation do not fail on reload.
• Removed a dead choice-note sync hook and simplified clarification gating to reuse the existing option-support helper, keeping the client logic aligned with the rest of the file.
• README clarification docs now match the shipped behavior and no longer claim the inline note field is limited to plain-string options.

[0.8.3] - 2026-04-21

Fixed

• Option-level recommended: true flags on rich option objects now normalize into the interview's canonical question-level recommendation state, so recommended pills, default selections, saved snapshots, and browser boot data stay consistent.
• Schema validation now rejects conflicting recommendation shapes, including mixed question-level and option-level recommendation sources, invalid option-level conviction usage, and multi-recommendation single-select questions.

[0.8.2] - 2026-04-21

Fixed

• Recommendation badges and preselection now survive review-time option normalization. The server and client reconcile recommended labels against the current option labels instead of silently dropping them on harmless whitespace/formatting changes.
• Completed, cancelled-with-answers, and timed-out interview sessions now send the agent an answered-only result payload keyed by full question text plus a fenced JSON block with the exact structured response data. This avoids losing selected options behind terse internal ids like scope and preserves choice notes, attachments, and image answers.
• Upload encoding failures now preserve the underlying FileReader error message instead of collapsing everything to a generic Failed to read file, which makes submit/save upload problems debuggable again.

[0.8.1] - 2026-04-21

Changed

• Ask model controls refined: The Ask panel now shows a clearer Model badge, uses pill-based provider/model selection instead of a raw dropdown, and adds editable response depth levels (Quick, Standard, Deep).

Fixed

• Ask loading feedback: Asking about an option now shows a visible loading state while the answer is being generated instead of leaving the panel visually idle.
• Ask model list validity: The Ask picker now shows a curated subset of likely-valid models instead of exposing raw registry entries that may be unavailable for the current account.

[0.8.0] - 2026-04-21

Added

• Option clarifications for plain choice questions: Plain string single/multi options can now reveal a separate inline Optional clarification... field when selected. Submitted and saved choice responses now use structured { option, note? } objects instead of bare strings.

Fixed

• Clarification draft retention: Switching away from a selected option now hides its clarification field without discarding the typed draft, so switching back restores the earlier text.
• Clarification preservation during rewrites/review: Use rewrite and question-level Review options now preserve structured choice answers and attached clarification notes instead of degrading them into plain strings or dropping them.

[0.7.0] - 2026-04-16

Added

• Per-option Ask flow: Single and multi-select options, including rich options with inline content blocks, can now open an inline Ask panel for option-specific follow-up questions, with prompt chips, freeform prompts, structured analysis responses, and follow-up actions like pinning, replacing text, moving options up, or adding suggested rewrites as new options.
• Advanced model overrides for option analysis: The new Ask panel shows the default model quietly and exposes provider-first then model selection behind an Advanced section for low-clutter per-request overrides.

Changed

• Pinned option analyses persist in drafts and snapshots: Saved interview snapshots and local autosave state now preserve pinned option analyses so option-level reasoning survives reloads and saved HTML revival, while staying out of submitted response payloads by default.
• Server-side option state now covers option mutations: Option rewrites and reordering actions from the Ask flow now update the in-memory interview definition so saves and reopened sessions stay aligned with what the user saw.

[0.6.2] - 2026-04-16

Changed

• Content block API (breaking): Replaced question codeBlock and rich-option code with content blocks using source + lang.
• Legacy field enforcement (breaking): Legacy codeBlock and rich-option code now fail schema validation with explicit migration errors. Use content instead.
• Markdown-by-default rendering: content.lang: "md" | "markdown" now renders Markdown preview by default. Set showSource: true to force raw Markdown source.
• Rendering parity: content rendering rules now match between the live interview UI and saved interview snapshots.

Fixed

• Code snippets in the live interview UI and saved interview snapshots now wrap long lines instead of clipping or forcing horizontal overflow.
• showSource is now enforced as markdown-only in schema validation and docs, matching runtime behavior.
• Removed an unused client-side renderedHtml branch from markdown content rendering.

[0.6.1] - 2026-04-04

Fixed

• Tool discoverability with pi v0.59+: Added promptSnippet to the interview tool registration so it remains eligible for inclusion in pi's default Available tools system prompt.
• Review options question rewrite: Review options now also rewrites the question text for clarity while keeping the existing option-review behavior.
• Generated/reviewed question persistence: Generated options and reviewed question rewrites now persist in the server-side interview state, so saves, recovery, and reopened snapshots stay in sync with what the user saw in the form.
• Review-mode recommendation cleanup: When reviewed options remove or rename previously recommended answers, recommended and conviction are now cleared or narrowed so saved interviews do not fail validation on reload.
• Rich option generate/review support: Generate more and Review options now work on rich-option questions too. Rich review responses keep structured { label, content? } option objects instead of flattening them, save/reload snapshots preserve the reviewed structure, and surviving option labels keep their existing option keys so pinned Ask insights stay attached when the option genuinely survives a review pass.
• Saved interview answer integrity: Reloading saved interview HTML now resolves paths only for image answers and attachments. Text, single-select, and multi-select answers remain literal so forms pre-populate correctly.
• Submit/save error clarity: Client-side save/submit failures now include the original error message instead of collapsing to generic text.

[0.6.0] - 2026-03-30

Added

• Generate & review options: Single and multi-select questions show two LLM-powered buttons: "✦ Generate more" appends new deduplicated choices, and "↻ Review options" validates and rewrites the existing options. Default model is the agent's current model, configurable via generateModel in interview settings.

Fixed

• Generate model fallback: When an explicitly configured generateModel fails at request time, interview now surfaces the provider error and retries once with the current session model when it differs.
• Codex-compatible generation requests: Generate/review requests now include a system prompt, so fallback to openai-codex session models works instead of failing with {"detail":"Instructions are required"}.

[0.5.5] - 2026-03-21

Added

• Native macOS rendering: When glimpseui is installed separately (pi install npm:glimpseui), interviews open in a native WKWebView window instead of a browser tab. Window lifecycle (submit, cancel, timeout) and queue toast session switching work correctly in the native environment. Falls back to browser on other platforms or when Glimpse is not detected.
• Inline JSON: Pass questions as a JSON string directly to the questions parameter instead of writing a temp file.

Fixed

• Queued interview handoff: When submitting an interview with queued sessions waiting, the Glimpse/browser window now redirects to the next queued interview instead of closing. The submit response includes the oldest queued session's URL, and the client navigates to it directly.
• Queue ordering stability: Next-session handoff now uses a deterministic tie-breaker (startedAt, then session id) so simultaneous starts cannot produce ambiguous promotion order.
• Settings parse failures are now visible: Invalid ~/.pi/agent/settings.json JSON no longer silently falls back to defaults; the tool now throws with parse context so config errors are debuggable.
• Submit response parsing: Client submit flow now preserves JSON parse failures from /submit instead of replacing them with a generic fallback object.
• Queued session staleness: Queued interview sessions now stay alive in the sessions registry via a server-side keep-alive interval, so queued sessions without browser heartbeats no longer go stale.

[0.5.4] - 2026-03-16

Fixed

• Single-select questions now unwrap recommended: ["Option"] to "Option" instead of throwing, matching the multi-select coercion that already wraps strings into arrays.

[0.5.3] - 2026-03-15

Fixed

• LLM JSON repair: When an LLM produces slightly malformed inline JSON (trailing commas, markdown code fences, single-line comments, curly smart quotes), loadQuestions now attempts a sanitized re-parse before failing. Valid JSON is unaffected — the repair only runs as a fallback when JSON.parse already failed. Saves agents a retry round-trip on common generation hiccups.

[0.5.1] - 2026-02-15

Added

• Schema validation unit tests (schema.test.ts) covering all validation rules, field types, media blocks, conviction/weight semantics, codeBlock validation, and edge cases.

Fixed

• Saved HTML snapshots now copy local media images to the images/ subfolder and rewrite src paths. Previously, media: { type: "image", src: "/local/path.png" } embedded the absolute path, which broke when the server wasn't running.
• copyMediaImages now resolves relative paths against the server's explicit cwd, not process.cwd(), matching the /media route's path resolution.
• Badge rendering uses badgeNumber !== null instead of a truthiness check that would silently skip badge 0 if numbering were ever changed to 0-indexed.
• loadProgress fresh-load branch now matches createQuestionCard's pre-selection logic exactly (recs.length > 0 instead of q.recommended truthiness, which incorrectly passed for empty arrays).

[0.5.0] - 2026-02-15

Added

• Visual redesign: New typography system with Google Fonts (Outfit, Space Mono, Instrument Serif, JetBrains Mono). Granular CSS variables for font sizes, card spacing, and a 6-color question palette (--q-color-1 through --q-color-6). Each question card gets a colored left border, numbered badge, and staggered fadeUp entrance animation with prefers-reduced-motion support. Radial gradient background atmosphere. Container widened to 740px.
• Rich media content: New MediaBlock type supporting image, chart (Chart.js), mermaid diagram, table, and HTML content in questions. Optional caption, position (above, below, side two-column grid), and maxHeight. CDN scripts for Chart.js and Mermaid injected only when needed.
• Info question type: Non-interactive content panel for displaying context alongside media. Visually distinct from questions: uniform neutral border, no shadow, muted title. Skipped during keyboard navigation; excluded from responses and persistence.
• Conviction signals: Optional conviction field on questions with recommended. All recommended options show a "Recommended" pill badge. conviction: "slight" opts out of pre-selection; "strong" and default (omitted) pre-select.
• Question weight: Optional weight field. "critical" renders a prominent card (5px accent border, tinted background). "minor" renders a compact card (smaller padding, text, and gaps; no shadow).
• Pre-selection: Recommended options are pre-checked on load unless conviction: "slight". Saved state and savedAnswers override pre-selection.
• Media serving: /media GET route serves local images with path security (resolve normalization, directory-boundary checks against cwd/homedir/tmpdir).
• Saved HTML: Media blocks, context text, info panels, conviction indicators, and weight styling all render in saved interview HTML snapshots.
• test-media.json and test-taste.json fixtures for testing.

Changed

• Replaced --font-ui with explicit --font-body or --font-mono references throughout CSS.
• Tufte themes now use Instrument Serif + JetBrains Mono instead of Cormorant Garamond + IBM Plex Mono.
• All four theme files updated with per-theme --q-color-* palettes.
• Option items have visible borders and accent-colored focus rings instead of background-only hover.
• Active card styling uses box-shadow lift instead of gradient tint + border color.
• Responsive breakpoint changed from 720px to 768px.
• focusQuestion() returns boolean to signal whether a valid card was found.
• Badge numbering skips info panels (not questions) with a separate counter (no gaps).
• Tool description updated with conviction, weight, and pre-filled form guidance.

Fixed

• Paste not working in text inputs: Clipboard data with both text and image representations (common when copying from web pages) caused handlePaste to intercept the image and swallow the text. Text inputs now let native paste through when clipboard has text data.
• Paste handler isTextInput check narrowed to textarea and input[type="text"] only — radio, checkbox, and file inputs no longer falsely trigger the text-paste early return.
• Responsive bottom padding removed at 768px and 480px breakpoints where shortcuts bar is hidden.
• Path traversal in /media route: absolute paths with .. segments bypassed startsWith check. Fixed with resolve() normalization.
• Prefix collision in /media route: /Users/nick matched /Users/nick2/secrets. Fixed with dir + "/" boundary check.
• HTTP/HTTPS/data URI images no longer proxied through /media endpoint (caused silent 404s).
• Saved HTML now includes context text for regular questions (was only rendered for info panels).
• loadProgress fresh-load branch (enabling Done buttons for pre-selected multi questions) was unreachable when localStorage.getItem throws. Restructured with a loaded flag so the branch runs outside the try/catch.

[0.4.5] - 2026-02-01

Fixed

• Adapt execute signature to pi v0.51.0: reorder signal, onUpdate, ctx parameters

[0.4.4] - 2026-01-27

Fixed

• Google API compatibility: Use StringEnum for theme mode instead of Type.String() (thanks @Whamp, PR #3)

[0.4.3] - 2026-01-27

Fixed

• Success overlay now centered in viewport when scrolled down (changed position: absolute to position: fixed)

Changed

• Removed unused startedAt variable from script.js
• Removed unused resp.type property assignment in collectResponses()

[0.4.2] - 2026-01-26

Fixed

• Install script now copies package.json with pi.extensions declaration for auto-discovery

[0.4.1] - 2026-01-26

Changed

• Added pi-package keyword for npm discoverability (pi v0.50.0 package system)

0.4.0 - 2026-01-24

Added

• Save snapshots: Save interview state to HTML files for later review or revival
  • Manual save via Save button (header and footer)
  • Auto-save on submit (enabled by default, configurable via autoSaveOnSubmit)
  • Snapshots saved to ~/.pi/interview-snapshots/ (configurable via snapshotDir)
  • Self-contained HTML with embedded JSON data for revival
  • Images copied to images/ subfolder with relative paths
  • Folder naming: {title}-{project}-{branch}-{timestamp}[-submitted]/
• Revival from saved interviews: Pass saved HTML path to interview() to reopen with pre-populated answers
  • Supports editing and re-submitting saved interviews
  • Image paths resolved relative to snapshot location
• New settings: snapshotDir and autoSaveOnSubmit in interview settings
• Save toast notification shows save location on success/failure

Changed

• loadQuestions() now accepts both JSON and HTML files
• Questions parameter can be a path to a saved interview HTML

---

0.3.2 - 2026-01-18

Changed

• Skip-friendly cancel behavior: When user cancels/dismisses the interview, the tool now returns context-aware messages:
  • No answers provided → "User skipped the interview without providing answers. Proceed with your best judgment - use recommended options where specified, make reasonable choices elsewhere. Don't ask for clarification unless absolutely necessary."
  • Partial answers provided → "User cancelled the interview with partial responses: [responses]. Proceed with these inputs and use your best judgment for unanswered questions."
• Timeout preserves partial answers: If the interview times out with partial responses, they're now included in the result message
• Cancel and timeout requests now include current form responses so the agent can use partial input

---

0.3.1 - 2026-01-17

Changed

• Other input: Changed from single-line text input to auto-growing textarea with line wrapping

Fixed

• Overflow layout bug: Code blocks no longer expand beyond container and break page layout
  • Added min-width: 0 to flex containers to allow proper shrinking
  • Fixed .code-block-lines-container to use min-width: 100% instead of width: 100%
• Done button alignment: Added missing display: flex to .done-item (flex properties were being ignored)
• Session bar responsive margin: Fixed 4px gap at 720px breakpoint where margin didn't match container padding

---

0.3.0 - 2026-01-17

Added

• Code blocks: Display code snippets in questions and options
  • Question-level codeBlock field shown below question text, above options
  • Rich options: options can be { label, code? } objects instead of plain strings
  • Syntax highlighting for diff (lang: "diff") with green/red line coloring
  • Optional file path and line range display in header
  • Line highlighting via highlights array
  • Line numbers shown when file or lines specified
• Light markdown in questions: Question titles and context now render **bold**, `code`, and auto-break numbered lists
• Default theme toggle hotkey: Cmd+Shift+L (Mac) / Ctrl+Shift+L (Windows/Linux) now works out of the box
• Fixed port setting: Configure port in settings to use a consistent port across sessions
• Shared settings module (settings.ts) for consistent settings access across tool and server

Removed

• Voice mode: Removed ElevenLabs voice interview integration entirely
  • Deleted elevenlabs.ts, form/voice.js, form/settings.js
  • Removed voice toggle button, voice indicator, settings modal, API key modal from HTML
  • Removed all voice-related CSS styles and CSS variables
  • Removed v keyboard shortcut for voice toggle
  • Simplified settings.ts (removed voice settings and updateVoiceSettings)
  • Removed transcript handling from server and responses

Changed

• Migrated from ~/.pi/agent/tools/ to ~/.pi/agent/extensions/ folder structure (pi-mono v0.35.0)
• Updated to new extension API: CustomToolFactory -> ExtensionAPI with pi.registerTool()
• Options can now be strings OR objects with { label, code? } structure

Fixed

• Radio/checkbox alignment on multi-line option text (now aligns to top)
• fileInput is not defined error in keyboard handler
• pi.cwd changed to ctx.cwd in tool execute function
• Paste handling: Regular text no longer intercepted as image attachment; only paths ending with image extensions (.png, .jpg, .jpeg, .gif, .webp) are treated as attachments
• Image limit enforcement: MAX_IMAGES limit now consistently enforced for both question images and attachments (was only checking question images)

---

2026-01-02

Added

• Multi-agent queue detection: When another interview is active, new interviews print URL instead of auto-opening browser, preventing focus stealing
• Session heartbeat system: Browser sends heartbeat every 5s; server tracks active sessions
• Abandoned interview recovery: Questions saved to ~/.pi/interview-recovery/ on timeout or stale detection
• Server watchdog: Detects lost heartbeats (60s grace) and saves recovery before closing
• Tab close detection: Best-effort cancel via pagehide + sendBeacon API
• Reload protection: Cmd+R / F5 detected to prevent false cancel on refresh
• Queued interview toast: Active interviews show a top-right toast with a dropdown to open queued sessions
• Queued tool panel output: Queued interview details render in the tool result panel with a single-line transcript summary
• Sessions endpoint: GET /sessions returns active/waiting sessions for in-form queue UI
• "Other..." text input option for single/multi select questions
  • Keyboard selection (Enter/Space) auto-focuses the text input
  • Value restoration from localStorage
• Session status bar at top of form
  • Shows cwd path with ~ home directory normalization (cross-platform)
  • Git branch detection via git rev-parse
  • Short session ID for identification
• Dynamic document title: projectName (branch) | sessionId for tab identification
• --bg-active-tint CSS variable for theme-aware active question styling
• Recovery file auto-cleanup (files older than 7 days)

Changed

• Active question focus styling uses gradient background tint instead of border-only
• Path normalization moved server-side using os.homedir() for cross-platform support
• Session registration uses upsert pattern (handles re-registration after prune)
• Cancel endpoint accepts reason field: "timeout", "user", or "stale"
• Queue toast position moved to top-right with compact layout

Fixed

• "Other" option keyboard selection now focuses text input instead of advancing to next question
• "Other" option accepts typing immediately when focused via keyboard
• Light mode active question gradient visibility (increased tint opacity)
• Question focus scroll uses nearest positioning to avoid jarring jumps
• Server-side timeout only starts when browser auto-opens (not for queued interviews)
• formatTimeAgo handles negative timestamps (clock skew)
• Race conditions prevented via completed flag on server
• Duplicate cancel requests prevented via cancelSent flag on client

---

2026-01-01

Added

• Voice interview mode: Natural voice-based interviewing powered by ElevenLabs Conversational AI
  • Questions read aloud, answers captured via speech
  • Bidirectional sync: click/keyboard navigate to any question, AI adapts
  • Intelligent cycling through unanswered questions
  • Hybrid mode: type/click anytime during voice session
  • Visual indicators: voice-focus styling, status indicator with progress
  • Full transcript returned with responses
  • Activation via URL param (?voice=true), toggle button, or schema config
• Voice controller state machine with WebRTC connection management
• window.__INTERVIEW_API__ bridge for cross-module communication
• getAnsweredQuestionIds() and getAllUnanswered() helper functions
• focusQuestion() now accepts source parameter ('user' | 'voice')
• Voice-specific CSS variables in all theme files
• ElevenLabs agent auto-creation from interview questions
• API key input UI with localStorage persistence

Changed

• InterviewServerOptions extended with voiceApiKey
• InterviewServerCallbacks.onSubmit now accepts optional transcript
• InterviewDetails extended with transcript field
• buildPayload() includes transcript when voice mode used

---

2026-01-02

Added

• Theme system with light/dark mode support
  • Built-in themes: default (monospace, IDE-style) and tufte (serif, book-style)
  • Mode options: dark (default), light, or auto (follows OS preference)
  • Custom theme CSS paths via lightPath / darkPath config
  • Optional toggle hotkey (e.g., mod+shift+l) with localStorage persistence
  • OS theme change detection in auto mode
  • Theme toggle appears in the shortcuts bar when configured
• Paste to attach: Cmd+V pastes clipboard image or file path to current question
• Drag & drop anywhere on question card to attach images
• Path normalization for shell-escaped paths and macOS screenshot filenames
• Per-question image attachments for non-image questions
  • Subtle "+ attach" button at bottom-right of each question
  • Tab navigation within attach area, Esc to close
• Keyboard shortcuts bar showing all available shortcuts
• Session timeout with countdown badge and activity-based refresh
• Progress persistence via localStorage
• Image upload via drag-drop, file picker, or path/URL input

Removed

• "A" keyboard shortcut for attach (conflicted with typing in text areas)

Fixed

• Space/Enter in attach area no longer triggers option selection
• Duplicate response entries for image questions
• ArrowLeft/Right navigation in textarea and path inputs
• Focus management when closing attach panel
• Hover feedback and tick loop race conditions
• Paste attaching to wrong question when clicking options across questions

Changed

• MAX_IMAGES increased from 2 to 12
• Timeout default is 600 seconds (10 minutes)
• Replaced TypeBox with plain TypeScript interfaces in schema.ts
• Consolidated code with reusable helpers (handleFileChange, setupDropzone, setupEdgeNavigation, getQuestionValue)

Initial Release

Features

• Single-select, multi-select, text, and image question types
• Recommended option indicator (*)
• Full keyboard navigation (arrows, Tab, Enter/Space)
• Question-centric navigation (left/right between questions, up/down between options)
• "Done" button for multi-select questions
• Submit with Cmd+Enter
• Session expiration overlay with Stay Here / Close Now options
• Dark IDE-inspired theme