Changed

• CLI: compact finish line summary across API, browser, and session views.
• CLI: token counts now render as ↑in ↓out ↻reasoning Δtotal.

Fixed

• CLI/Browser: ignore duplicate --file inputs (log once) and improve attachment presence detection so re-runs don’t spam “already attached” upload errors.
• Browser: harden session reattach (better conversation targeting, longer prompt-commit wait, avoid closing shared DevTools targets).
• Live tests: add coverage + retries for browser reattach/model selection; tolerate transient OpenRouter free-tier failures.

Fixed

• Packaging: switch tokentally to npm release so Homebrew installs don't trigger git prepare builds.

Changed

• Browser: add --browser-thinking-time <light|standard|extended|heavy> to select thinking-time intensity in ChatGPT.

Fixed

• Browser: throttle attachment upload pokes and pace multi-file uploads to avoid duplicate “already attached” warnings.
• Browser: correct GPT-5.2 variant selection (Auto/Thinking/Instant/Pro) with stricter matching and improved testid scoring; thinking-time selection now supports multiple levels. Original PR #45 by Manish Malhotra (@manmal) — thank you!
• Browser: only reload stalled conversations after an assistant-response failure (and only once), instead of always refreshing after submit.

• Changed: API: streaming answers in a rich TTY now use Markdansi’s live renderer (createLiveRenderer) so we can stream and render Markdown in-place.
• Changed: Dependencies: update markdansi to 0.1.5.
• Changed: Dependencies: update devtools-protocol to 0.0.1561482.
• Fixed: Browser: prevent chrome-launcher from auto-killing Chrome on SIGINT so reattach sessions survive Ctrl+C.
• Fixed: Sessions: running browser sessions now mark as errored when the Chrome PID/port are no longer reachable.
• Fixed: Browser: reattach now recovers even if Chrome was closed by reopening, locating the conversation in the sidebar, and resuming the response.

Fixed

• Browser: stop auto-clicking the “Answer now” gate; wait for the full Pro-thinking response instead of skipping it.
• Browser: reject ?temporary-chat=true URLs when targeting Pro models (Pro picker entries are not available in Temporary Chat); error message now calls this out explicitly.
• Browser: attachment uploads re-trigger the file-input change event until ChatGPT renders the attachment card (avoids hydration races); verify attachments are present on the sent user message before waiting for the assistant.
• Live tests: make the gpt-5.2-instant OpenAI smoke test resilient to transient API stalls/errors.

Changed

• API: default model is now gpt-5.2-pro (and “Pro” label inference prefers GPT‑5.2 Pro).
• Tests: updated fixtures/defaults to use gpt-5.2-pro instead of gpt-5.1-pro.
• API: clarify gpt-5.1-pro as a stable alias that targets gpt-5.2-pro.
• Dependencies: refresh (notably zod 4.2.1 and devtools-protocol 0.0.1559729).
• Browser: browser engine GPT selection now supports ChatGPT 5.2 (gpt-5.2) and ChatGPT 5.2 Pro (gpt-5.2-pro); legacy labels like gpt-5.1 normalize to 5.2, and “Pro” always resolves to 5.2 Pro (ignores Legacy GPT‑5.1 Pro submenu) with a top-bar label confirmation.

Fixed

• Browser: prompt commit verification handles markdown code fences better; prompt-echo recovery is more robust (including remote browser mode); multi-file uploads are less flaky (dynamic timeouts + better filename matching). Original PR #41 by Muly Oved (@mulyoved) — thank you!
• Browser: adapt to ChatGPT DOM changes (data-turn=assistant|user) and “Answer now” gating in Pro thinking so we don’t capture placeholders/truncate answers.
• Gemini web: add abortable timeouts + retries for cookie-based runs so live tests are less likely to hang on transient Gemini web responses.

Added

• Browser: Gemini browser mode via direct Gemini web client (uses Chrome cookies; no API key required; runs fully in Node/TypeScript — no Python/venv). Includes --youtube, --generate-image, --edit-image, --output, --aspect, and --gemini-show-thoughts. Original PR #39 by Nico Bailon (@nicobailon) — thank you!
• Browser: media files passed via --file (images/video/audio/PDF) are treated as upload attachments instead of being inlined into the prompt (enables Gemini file analysis).
• Browser: Gemini image ops follow gg-dl redirects while preserving cookies, so --generate-image/--edit-image actually create output files.
• Browser: Gemini web runs support “Pro” auto-fallback when unavailable and include compatibility init for Gemini web token changes.
• Live tests: add opt-in Gemini web smoke coverage for image generation/editing (cookie-based browser mode).

Changed

• Browser guard now allows Gemini models (browser engine supports GPT + Gemini; other models require --engine api).

Changed

• Browser: default model target now prefers ChatGPT 5.2. Original PR #40 by Muly Oved (@mulyoved) — thank you!
• Browser: remove the “browser fallback” API retry suggestion to avoid accidental billable reruns. Idea from PR #38 by Nico Bailon (@nicobailon) — thank you!

Fixed

• Browser: manual-login runs now reuse an already-running Chrome more reliably (persist DevTools port in the profile; probe with retries; clean up stale port state). Original PR #40 by Muly Oved (@mulyoved) — thank you!
• Browser: response capture is less likely to truncate by mistaking earlier turns as complete; completion detection is scoped to the last assistant turn and requires brief stability before capture. Original PR #40 by Muly Oved (@mulyoved) — thank you!
• Browser: stale profile cleanup avoids deleting lock files when an active Chrome process is using the profile.

Added

• GPT-5.2 model support (gpt-5.2 Thinking, gpt-5.2-instant, gpt-5.2-pro) plus browser --browser-extended-thinking automation. Original PR #37 by Nico Bailon (@nicobailon) — thank you!

Changed

• API: gpt-5.1-pro now targets gpt-5.2-pro instead of older Pro fallbacks.
• Browser: “Thinking time → Extended” selection now reuses centralized menu selectors, normalizes text matching, and ships a best-effort helper for future “auto” mode. Original PR #36 by Victor Vannara (@voctory) — thank you!
• Browser: new --browser-attachments <auto|never|always> (default auto) pastes file contents inline up to ~60k characters, then switches to uploads; if ChatGPT rejects an inline paste as too large, Oracle retries automatically with uploads.
  • Note: the ~60k threshold is based on pasted characters in the ChatGPT composer (not token estimates); on rejection we log the retry and switch to uploads automatically.

Changed

• Browser uploads: after setFileInputFiles we now log the chips + file-input contents and only mark success when the real file input contains the uploaded filename; the generic “Files” pill is no longer treated as proof of attachment.
• Inline prompt commit: verification now matches on a normalized prefix and logs the last user turn + counts when commit fails, reducing false negatives for inline/file-paste runs.

Fixed

• Inline fallback (pasting file contents) now reliably submits and captures the user turn; headful smoke confirms the marker text is echoed back.

(Re-release of 0.5.5)