feat: whale-size route taxonomy for model + thinking-effort picker by encyc · Pull Request #2338 · Hmbown/CodeWhale

encyc · 2026-05-29T02:01:02Z

Refs #2026. Related: #1676.

Summary

Adds a central whale-route taxonomy that maps each (model, reasoning_effort) pair to a friendly whale-species label, sorted from largest/deepest to smallest/fastest.

For DeepSeek providers, the /model picker now shows a single-column whale-route list instead of the two-column Model | Thinking layout. Pass-through providers retain the classic layout.

Routes (largest → fastest)

Route	Model	Effort	Description
Blue Whale	v4-pro	max	Architecture, debugging, security reviews
Fin Whale	v4-pro	high	Complex tasks, multi-file refactors
Sperm Whale	v4-pro	off	Straightforward code generation
Humpback	v4-flash	max	Lightweight analysis, first-pass reviews
Minke Whale	v4-flash	high	Tool execution, read-only scouting
Porpoise	v4-flash	off	Lookups, searches, simple edits

New module: `whale_routes.rs`

WhaleRoute struct with label, model, effort, sort_order, hint, description
WHALE_ROUTES const array (6 routes)
for_model_effort() — look up route by model + effort
by_sort_order() — look up by sort index
8 unit tests covering ordering, lookup, case-insensitivity, uniqueness

Model picker changes

show_whale_routes flag activates for DeepSeek providers
Single-column route list with whale labels and hints
Fallback row for "auto" and custom model ids
Tab key is a no-op in whale-route mode
10 updated/new unit tests

Files

crates/tui/src/tui/whale_routes.rs  | +196  (new)
crates/tui/src/tui/mod.rs           |   +1
crates/tui/src/tui/model_picker.rs  | +287 / -48

3 files, +436 −48

Greptile Summary

This PR introduces a whale-route taxonomy that maps (model, reasoning_effort) pairs to friendly whale-species labels (Blue Whale → Beluga) and wires them into the /model picker for DeepSeek providers, replacing the two-column model/effort layout with a single-column route list. Pass-through providers retain the classic layout.

whale_routes.rs defines the 6-entry WHALE_ROUTES const, WhaleRoute struct, and two lookup helpers (for_model_effort, by_sort_order), all covered by 9 unit tests.
model_picker.rs adds show_whale_routes / selected_route_idx fields, routes selection logic, render helpers, and 10 updated/new tests; fallback rows handle "auto" and unrecognised custom model IDs correctly, and selected_route_idx is initialised from .position() (array index) rather than sort_order, avoiding a future fragility.
The PR description table names the 6th route "Porpoise", but the module docs and code consistently use "Beluga" — the code is self-consistent, the PR description has a stale label.

Confidence Score: 5/5

Safe to merge — the three compile-blocking and logic issues from the previous review round are all resolved.

The impl-block placement, popup-height type cast, and array-index-vs-sort_order initialisation flagged in earlier rounds are all correctly addressed. Fallback-row logic is consistent across new(), resolved_whale_model/effort, whale_route_row_count, and render. The 19 total tests cover the key paths well.

No files require special attention.

Important Files Changed

Filename	Overview
crates/tui/src/tui/whale_routes.rs	New module defining the WhaleRoute taxonomy: 6 canonical (model, effort) routes with labels, hints, descriptions, and utility lookup methods; 9 unit tests covering ordering, case-insensitive lookup, uniqueness, and edge cases.
crates/tui/src/tui/model_picker.rs	Adds show_whale_routes / selected_route_idx fields, splits render into render_whale_routes and render_classic helpers in the correct impl block, uses .position() for index init, and correctly sizes the popup. Navigation, fallback rows, and tests are all consistent.
crates/tui/src/tui/mod.rs	Single-line module registration; no issues.

_{Reviews (5): Last reviewed commit: "fix(#2338): prevent known model + Auto e..." | Re-trigger Greptile}

…mbown#2026) Adds a central whale-route taxonomy that maps each (model, reasoning_effort) pair to a friendly whale-species label sorted from largest/deepest to smallest/fastest: Blue Whale — Pro + max thinking Fin Whale — Pro + high thinking Sperm Whale — Pro + no thinking Humpback — Flash + max thinking Minke Whale — Flash + high thinking Porpoise — Flash + no thinking For DeepSeek providers the /model picker now shows a single-column whale-route list instead of the two-column Model|Thinking layout. Each route sets both model and effort at once. Pass-through providers retain the classic layout. New whale_routes module: - WhaleRoute struct with label, model, effort, sort_order, hint, description - WHALE_ROUTES const array (6 routes) - for_model_effort() and by_sort_order() lookups - 8 unit tests Model picker changes: - show_whale_routes flag activates on DeepSeek providers - Selected route maps to model + effort simultaneously - Fallback row for "auto" and custom models - Updated test suite for whale-route behavior (7 new/updated tests) Refs: Hmbown#1676, Hmbown#2026

gemini-code-assist

Code Review

This pull request introduces 'whale routes' for DeepSeek providers in the model picker modal, mapping model and thinking-effort combinations to friendly whale-species labels sorted from largest to fastest. For other providers, the picker falls back to the classic two-column layout. Feedback on the changes highlights a compilation error in model_picker.rs due to a type mismatch in popup_height calculation, as well as a correctness bug in resolved_whale_model() where selecting the fallback row can incorrectly resolve to the initial model instead of 'auto'.

gemini-code-assist · 2026-05-29T02:02:20Z

+    fn render_whale_routes(&self, area: Rect, buf: &mut Buffer) {
+        let popup_width = 62.min(area.width.saturating_sub(4)).max(44);
+        let row_count = self.whale_route_row_count();
+        let popup_height = (row_count + 4).min(area.height.saturating_sub(4)).max(8) as u16;


This line will fail to compile because row_count + 4 is of type usize while area.height.saturating_sub(4) is of type u16. In Rust, the .min() method requires both arguments to be of the same type. Casting the u16 value to usize resolves this type mismatch.

Suggested change

let popup_height = (row_count + 4).min(area.height.saturating_sub(4)).max(8) as u16;

let popup_height = (row_count + 4).min(area.height.saturating_sub(4) as usize).max(8) as u16;

gemini-code-assist · 2026-05-29T02:02:20Z

+    fn resolved_whale_model(&self) -> String {
+        if self.selected_route_idx < WHALE_ROUTES.len() {
+            WHALE_ROUTES[self.selected_route_idx].model.to_string()
+        } else {
+            // Past the last whale route: "auto" or custom.
+            self.initial_model.clone()
+        }
+    }


There is a correctness bug when resolving the model from the fallback row. If initial_model is a standard model (e.g., "deepseek-v4-pro") and the user selects the fallback row (which is rendered as "auto — select per turn" because show_custom_model_row is false), resolved_whale_model() incorrectly returns self.initial_model.clone() (i.e., "deepseek-v4-pro") instead of "auto". Checking self.show_custom_model_row ensures that we correctly resolve to "auto" when the fallback row represents the auto-select option.

Suggested change

fn resolved_whale_model(&self) -> String {

if self.selected_route_idx < WHALE_ROUTES.len() {

WHALE_ROUTES[self.selected_route_idx].model.to_string()

} else {

// Past the last whale route: "auto" or custom.

self.initial_model.clone()

}

}

fn resolved_whale_model(&self) -> String {

if self.selected_route_idx < WHALE_ROUTES.len() {

WHALE_ROUTES[self.selected_route_idx].model.to_string()

} else if self.show_custom_model_row {

self.initial_model.clone()

} else {

"auto".to_string()

}

}

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5ddac40909

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-05-29T02:06:24Z

+        } else {
+            // Past the last whale route: "auto" or custom.
+            self.initial_model.clone()


Return auto for the standard fallback row

When the current model is one of the standard whale routes, the picker appends a fallback row that is rendered as auto — select per turn, but selecting that row reaches this branch and returns initial_model instead of auto. In that scenario, choosing the displayed auto row applies the current model/effort unchanged rather than enabling per-turn auto selection; only the custom-model fallback should preserve initial_model.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-05-29T02:06:24Z

    #[must_use]
    pub fn new(app: &App) -> Self {
        let hide_deepseek_models = crate::config::provider_passes_model_through(app.api_provider);
+        let show_whale_routes = !hide_deepseek_models;


Limit whale routes to official DeepSeek providers

For non-pass-through compatible providers such as NVIDIA NIM, OpenRouter, Novita, Fireworks, SGLang, and vLLM, provider_passes_model_through is also false, so this enables whale-route mode even though those providers use provider-specific model IDs elsewhere (for example deepseek-ai/deepseek-v4-pro for NIM). In that context the current provider-specific model is treated as a custom fallback, and selecting any whale route applies the bare DeepSeek ID directly, bypassing the provider remapping used by /model <id> and config normalization.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-05-29T02:06:24Z

+        // Fallback row: "auto" if not set, otherwise the current custom model.
+        let fallback_label = if self.initial_model == "auto" {
+            "auto  —  select per turn".to_string()
+        } else if self.show_custom_model_row {
+            format!("{}  —  custom", self.initial_model)


Keep auto selectable for custom models

When the current model is an unrecognized/custom DeepSeek ID, the single fallback row is rendered as the custom model instead of adding both the custom row and the documented auto option. That means a user on a snapshot/custom model can no longer switch back to per-turn auto from /model, whereas the previous two-column picker always kept the auto row alongside the custom model.

Useful? React with 👍 / 👎.

… array index - Fix usize/u16 type mismatch in popup_height (row_count as u16 + 4) - Fallback rows: always show "auto" first, then custom model if applicable - Limit whale routes to official DeepSeek/DeepSeekCN providers only - Use array position (not sort_order) for selected_route_idx lookup - Update tests for new fallback row indices

These two methods were accidentally placed inside the ModalView trait implementation block, which caused E0407 compilation errors on CI. They are now in a separate impl ModelPickerView block.

…fort - model_picker.rs no longer directly references WhaleRoute (uses WHALE_ROUTES.iter().position() instead of WhaleRoute::for_model_effort) - whale_routes.rs: for_model_effort is only used in tests; add #[allow(dead_code)] to suppress -D warnings in release builds

encyc · 2026-05-29T09:22:26Z

CI status: macOS ✅, ubuntu ✅, lint ✅. The Windows failure is a pre-existing flaky test in composer_history::tests::append_history_dispatched_does_not_block_the_caller — a writer-thread timing race on Windows CI runners, unrelated to this PR's changes (model_picker.rs, whale_routes.rs).

Hmbown · 2026-05-31T01:33:08Z

haha I like this - I am working on something similar with a whale pod mode - will use this to help!

Hmbown · 2026-05-31T04:36:04Z

Still like the whale-route idea. I am leaving this open because review found one real model-picker bug: a known DeepSeek model paired with ReasoningEffort::Auto can fall through to the auto model row and silently replace the explicit model when the picker is applied.

Please key that fallback on whether the model is actually auto, add the regression around known-model + Auto effort, and clean up the test that mutates fields whale-route mode never reads. After that, the whale taxonomy slice should be much closer.

Hmbown · 2026-05-31T05:24:19Z

Thanks @encyc, I still like this direction a lot. I tested the merged PR locally against current origin/main: the whale route and model picker tests pass, and the Windows CI failure is in an unrelated composer-history timing test.

I would hold merge for one correctness fix: a known DeepSeek model paired with ReasoningEffort::Auto currently falls through to the auto row, so applying the picker can silently replace the explicit model with "auto". Please key that fallback on whether the initial model is actually auto, add a regression for known-model + Auto effort, and clean up the test that mutates fields whale-route mode never reads.

One small taxonomy note too: #2016 excludes porpoises from the user-facing whale pool, so the final Flash/off tier should use a small whale name from the shared taxonomy instead of Porpoise. After that, this looks harvestable as the model-picker slice; I would keep #2026 open for receipts/footer/agent-card integration.

…ugh to auto row The whale-route fallback in the picker constructor used show_custom_model_row as the gate for selecting the 'auto' vs custom row, but a known DeepSeek model (e.g. v4-pro) paired with ReasoningEffort::Auto would not match any whale route yet still have show_custom_model_row=false — silently landing on the auto row and replacing the explicit model with 'auto' on apply. Key the fallback on whether the initial model is actually 'auto' instead. When a whale-route fallback selects the custom row, ensure show_custom_model_row is set to true so the row is visible in the picker UI. Also: - Add regression test: known-model + Auto effort must not fall to auto row. - Clean up picker_auto_model_forces_auto_effort_on_apply: remove manual mutations of selected_model_idx / selected_effort_idx which whale-route mode never reads. - Rename Porpoise → Beluga per Hmbown#2016, which excludes porpoises from the user-facing whale pool.

encyc · 2026-05-31T05:44:21Z

@Hmbown thanks for the careful review. Pushed fixes for all three points:

1. Auto effort fallthrough bug (the real one)
The fallback in new() was keyed on show_custom_model_row, but a known DeepSeek model (e.g. v4-pro) paired with ReasoningEffort::Auto would have show_custom_model_row = false (the model IS in PICKER_MODELS) yet not match any whale route. It silently landed on the "auto" row and replaced the explicit model.

Fixed by keying the fallback on initial_model.eq_ignore_ascii_case("auto") — only when the model is actually "auto" does it fall to the auto row. When the whale-route fallback selects the custom row, show_custom_model_row is now also set to true so the row is visible in the picker.

Regression test added: whale_routes_known_model_auto_effort_does_not_fall_to_auto.

2. Cleaned up test
picker_auto_model_forces_auto_effort_on_apply was mutating selected_model_idx / selected_effort_idx — fields whale-route mode never reads. Removed those lines; the test still passes because the constructor selects the correct route.

3. Porpoise → Beluga
Per #2016 excluding porpoises from the user-facing whale pool, renamed to Beluga at all sites (constant, tests, doc comments). Beluga is a proper toothed whale in the shared taxonomy.

All 28 model_picker / whale_routes tests pass locally.

gemini-code-assist Bot reviewed May 29, 2026

View reviewed changes

chatgpt-codex-connector Bot reviewed May 29, 2026

View reviewed changes

greptile-apps Bot reviewed May 29, 2026

View reviewed changes

Comment thread crates/tui/src/tui/model_picker.rs Outdated

Comment thread crates/tui/src/tui/model_picker.rs Outdated

Comment thread crates/tui/src/tui/model_picker.rs

greptile-apps Bot reviewed May 29, 2026

View reviewed changes

Comment thread crates/tui/src/tui/model_picker.rs

encyc added 2 commits May 29, 2026 10:22

fix: move render_whale_routes/render_classic out of ModalView trait impl

bf30aa3

These two methods were accidentally placed inside the ModalView trait implementation block, which caused E0407 compilation errors on CI. They are now in a separate impl ModelPickerView block.

Hmbown mentioned this pull request May 31, 2026

Model picker: whale-size names for Pro/Flash thinking tiers #2026

Open

Hmbown merged commit eb55cfe into Hmbown:main May 31, 2026
9 checks passed

	let popup_height = (row_count + 4).min(area.height.saturating_sub(4)).max(8) as u16;
	let popup_height = (row_count + 4).min(area.height.saturating_sub(4) as usize).max(8) as u16;

Conversation

encyc commented May 29, 2026 • edited by Hmbown Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Routes (largest → fastest)

New module: whale_routes.rs

Model picker changes

Files

Greptile Summary

Confidence Score: 5/5

Important Files Changed

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot May 29, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot May 29, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot May 29, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot May 29, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot May 29, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

encyc commented May 29, 2026

Uh oh!

Hmbown commented May 31, 2026

Uh oh!

Hmbown commented May 31, 2026

Uh oh!

Hmbown commented May 31, 2026

Uh oh!

encyc commented May 31, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

encyc commented May 29, 2026 •

edited by Hmbown

Loading

New module: `whale_routes.rs`