Offer the lighter model for a small change, right where you press send

[debtn]

"I don't need Opus to change the name of the title, but I will forget to switch the model, especially for these small things."

Self-contained: one commit, three files (one new). Independent of #8 and #10.

The problem is forgetting, not choosing

The model menu is the last thing on a person's mind mid-thought, and it stays wherever the last big change left it. So the heavy model ran every tiny edit, quietly, for everyone who never thought to look at the picker. Any fix has to arrive at the moment of forgetting — right before send — or it's just one more thing to remember.

Ways to nudge, considered

Approach	What it is	Why / why not
Suggest at send (shipped)	Read the draft; on a small change on a heavy model, offer the lighter one for this message, one tap	Meets people exactly where they forget. Keeps the decision theirs. Cheap to be wrong.
Auto-route small edits	Send anything that looks small to Sonnet automatically, with a note	Zero effort — but silently misclassifying a real feature request costs a bad build, and the code carries a note that a Sonnet-for-edits default was tried and moved back. Worth revisiting with data from this nudge: tap rate tells you whether people agree with the judgment.
Show the footprint per reply	A quiet "~0.4 Wh · Opus 5" under each reply, using the existing energy module	Honest and educational, but a rear-view mirror — it tells you after, when the ask was to catch you before. Good companion later.
Pin per change, not per project	Picking a model pins it for the whole project today, which is exactly how a big-change Opus pick becomes every small edit's model	Real cause of the problem, but a semantics change to the picker; deferred.
Model picks its own tier	A cheap pre-pass classifies the ask, then routes	Extra latency and a call per message to save a call; over-engineered for now.

What shipped

When a draft on a heavy model reads like a small change, one quiet line appears above the composer:

🌿 Small change? Sonnet 5 can do this for about two-fifths of the cost.   [Use Sonnet 5 for this one]   Keep Opus 5

• A nudge, not a switch. The heavy model stays the default unless they tap. The one-shot ride never pins or moves the picker; the very next message runs on whatever it would have.
• Armed, it says exactly what will happen — "This message runs on Sonnet 5 — the next one goes back to Opus 5" — with an undo.
• "Keep Opus 5" hushes it for the sitting. In memory on purpose: a preference that quietly persisted forever would be the nudge you can never get back.
• Never fires on a plan, a first build, the home composer, or a model that's already light.
• Fable → Sonnet reads "about a fifth of the cost"; the fraction is derived, not hand-written.

The judgment is deliberately conservative

"Small" means the preview's own mechanical messages (Copy change…, Swap this photo…), or a short single ask that talks like a tweak (rename, color, typo, bigger, spacing…) and never names anything with shape (a page, a feature, login, calendar, search, "everything", "also"). A miss in either direction is cheap — a false small is a chip you ignore; a false big is no chip — which is what lets the heuristic stay simple instead of clever.

The cost fraction uses first-party output prices (Fable $50 · Opus $25 · Sonnet $10 · Haiku $5 per MTok) — the same proxy the Steward view prices with — rounded to a fraction a person can hold, never a percentage with false precision. Sonnet 5 is the one-tier step for everything heavier; Haiku's shorter output cap and the SEARCH/REPLACE edit format make it a worse bet for real edits.

Files

• src/components/Chat/lighter-model.ts — new: the small-change judgment, the lighter-model choice, the cost fraction, the session hush
• src/components/Chat/MessageInput.tsx — the chip; onSend gains an optional per-send { model }
• src/components/Chat/ChatPanel.tsx — honors the one-shot model on the send path (after the community default logic, never on fix sends); the build log's gen_start records what actually ran

Verification

Driven in Chromium against the dev server, with every model request captured off the wire — 21 checks, zero page errors:

14 drafts judged as intended (6 small, 8 not)	✅
chip appears on a small draft, gone on a big one	✅
armed send goes out as claude-sonnet-5; picker stays Opus, unpinned	✅
the next send goes back to claude-opus-5	✅
undo returns to the offer; keep hushes the sitting; the send stays on Opus	✅
no chip in plan mode, on a first build, or already on Sonnet	✅
on Fable the fraction reads "a fifth"	✅

tsc -b and vite build clean. eslint sits at the same three pre-existing MessageInput problems as an untouched tree — none in the new code.

What I'd do next, if this lands well

1. Log the tap. One friction/build-log event for shown / used / kept gives real data on whether the judgment matches people — and that data is what would justify (or rule out) auto-routing.
2. Footprint under each reply using src/lib/energy.ts, as the companion rear-view mirror.
3. Pin scope. Reconsider whether a picker choice should pin for the project or for the next big change.

One thing noticed in passing, not touched: the community monitor's price table prices /sonnet/ at $3/$15, which is Sonnet 4.6's rate — Sonnet 5 is $2/$10. Since the regex also matches Sonnet 5, the Steward view overstates Sonnet spend by half. Separate fix, owner's call.

🤖 Generated with Claude Code

https://claude.ai/code/session_01RFnFvBZPsVQjc5Xwj7H9nZ

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
relational-builder	Ready	Preview	Sep 3, 2026 8:35pm UTC