[tylerbea]

Asking GAIA for a recipe, a workout, a repair guide, a packing list, a trip plan, a tip split or a loan payment used to return a wall of text, with numbers the small local model often got wrong. Now the answer arrives as an interactive card in the Agent UI. Recipes rescale with servings, switch US/metric and run step timers in a full-screen Start cooking mode. Calculators compute in the app and update as you edit. A Try row on the new chat screen and a one-time tip on the first card show people these exist. Closes #6085.

Every card type was exercised live against Gemma-4-E4B. That run caught and fixed amounts the model writes last ("Olive oil (1 tablespoon)"), "165°F/74°C" pairs, its own step numbering, re-typed recipes under the card, a double-subtracted down payment and currency carried over from an earlier turn.

• Card rendering: the SSE handler now names the card itself when it lifts one, so in-process turns draw cards and a failed call no longer draws an "Invalid payload" card. Reloaded chats rebuild cards from the saved tool steps.
• TUI: draws both cards too: every line of a guide, and a calculator's results computed in the client.
• Evals: new tool docstrings change the schema sent to Lemonade, so re-run gaia_tool_selection, tool_selection, gaia_core (itinerary and workout turns) and gaia_voice (the packing-list scenario expects a bulleted reply) on the next RC eval. No local run was possible here.

Before and after

Before	After

Start cooking (full screen, one step at a time) · Loan calculator · Try row on a phone

Test plan

• npx vitest run in src/gaia/apps/webui: 2099 passed. npx tsc --noEmit -p . is clean.
• python -m pytest tests/unit/test_guide_tools.py tests/unit/test_calculator_tools.py tests/unit/test_sse_narration.py hub/agents/gaia/python/tests/test_capability_matrix.py: 187 passed.
• python util/lint.py --black --isort --flake8 --tool-descriptions --agents passes.
• Live with Gemma-4-E4B in the Agent UI: recipe, workout, how-to, checklist, itinerary, tip and loan cards all draw. Servings, units, timers, guided mode, ticks and calculator edits work, and the cards survive a reload. Checked at desktop and 375 px, in light and dark.
• Reviewer: on a new chat, tap Try → Recipe, send it, then press Start cooking.

[github-actions]

Verdict: Approve with suggestions

Recipes, workouts, how-tos, checklists, itineraries and the tip and loan calculators now show up as interactive cards. The calculators do the math in the app, so the small model's arithmetic no longer reaches the answer. Each tool checks its arguments and returns an error that says how to fix the call. Cards rebuild from the saved steps after a reload, and a failed call no longer draws an "Invalid payload" card. Both have tests. The PR includes before/after screenshots and names the eval categories to watch on the next RC.

🟡 One changelog line is now wrong for integrators. It says the stdio transport "still never sends" the card type. This PR makes stdio send it too. Stdio uses the same output handler and translator as HTTP, and the handler now attaches the card type itself. Please correct that sentence so integrators on stdio know to expect render.

🔍 Technical details

• hub/agents/gaia/npm/CHANGELOG.md, "For integrators" bullet: "The stdio transport still never sends it."
• stdio.py:1396-1439 builds an SSEOutputHandler and feeds its events through CanonicalTranslator.translate. sse_handler.py:985 now sets event["render"] whenever it lifts a card, and sse_translation.py:431 forwards it. So both transports carry render for every tool in _RENDER_TOOL_TO_LANG, not only the new ones.
• Suggested wording: "It now carries the card type on both the HTTP and stdio transports (guide or calculator for the tools above)."
• Removing the render_tool_map kwarg is safe. On main only one test passed it, and that test is updated here.

[kovtcharov-amd]

Thanks @tylerbea — this is a lovely piece of work. Moving the arithmetic out of the model and into the card is exactly the right call for a small local model, and the live Gemma-4-E4B pass clearly paid off (the "Olive oil (1 tablespoon)" and double-subtracted down payment fixes are the kind of thing only a real run finds).

One gap before this can merge: outside the Agent UI, a recipe or calculator answer now loses its content. The tools tell the model "the card is on the user's screen, don't repeat the list / don't state the results", but the TUI has no guide or calculator renderer, so a TUI user asking for a banana bread recipe gets a one-line tip plus a 12-line truncated JSON dump, and a loan question gets no numbers at all. Today the same question gets a full text answer, so this would land as a regression for TUI users once it's on main.

Either of these would close it: a TUI renderer for the two keys (a readable list for guide, the computed results for calculator), or keeping the content in the reply for clients that can't draw the card. The maintainers may pick that up as a companion PR — no need to rework the Agent UI side, which checks out well.

I merged main in and fixed one changelog line (tool count is 105, not 102). Everything else passes locally: vitest 2151/2151, tsc clean, 3755 pytest passing across the touched agent/UI suites, lint clean. Timing-wise this is queued for right after v0.25.3 ships, for v0.26.0.

🔍 Technical details

• TUI: tui/internal/ui/cards/cards.go RenderDeduped sends guide/calculator to renderUnsupported, which caps the raw dump at rawDumpLines = 12. The sidecar now sets tool_result.render on both HTTP and stdio, so the TUI receives these keys.
• Eval watch beyond gaia_tool_selection / tool_selection: gaia_voice/voice_t3_bullets_new_session asks for a camping packing list and fails if the reply isn't a bulleted list, which show_checklist + "do not repeat the list" will likely trip. gaia_core (itinerary and workout turns) is worth a re-run too.
• Verified: card content renders as React text nodes only; print path escapes & < > " '; calculators take typed numbers and never evaluate expressions; the guided view closes on Escape / X / Finish and uses the shared focus-trap and Escape-layer hooks; invalid payloads fall back to the visible "Invalid … payload" card; ChatView.tsx is 97,276 bytes LF.

[kovtcharov-amd]

Thanks @tylerbea, this is ready for a final look. I merged main in and closed the TUI gap from the earlier review: a recipe or loan answer in the TUI now draws as a readable card instead of a cut-off JSON dump. I also fixed how the cards read in an Arabic chat, where English ingredient lines came out reordered ("cups flour 2") with bullets on the wrong side.

One thing left for a follow-up: in Arabic, the card's labels (Ingredients, Start cooking, servings, the calculator's field and result names, the first-card tip) still show in English.

🔍 Technical details

• TUI: new guide and calculator renderers in tui/internal/ui/cards/. A guide draws every line with no 22-row cap, because the tool tells the model not to repeat the list. The calculator recomputes tip and amortization with the same formulas as calculatorDoc.ts. A bad payload shows the existing "Invalid … payload" card. Covered by guide_test.go and calculator_test.go, including escape-sequence and range checks.
• RTL: guide.css uses logical properties for the bullet, the item padding and the Start button. Model-text lines use unicode-bidi: plaintext with alignment set per direction, which is the pattern the sidebar uses. A new rtl.visual.ts test renders both cards in ar and en and passes the layout audit.
• Earlier "Test Gaia Agent" failure: the fixed prefix was 15,144 tokens against a 15,000 budget. After merging main it measures 14,883 locally, so CI on the new head should confirm it.
• Local checks: vitest card suites, tsc --noEmit, the Electron byte-cap suite (63/63), npm run test:rtl (442 + 2 new), PR pytest suites (299), the prefix-stability suite (36), and go test for cards and chat.