feat(executors): Fable/Claude-5 model family, ultracode effort level, free-text model entry + effort for preset models

[miguelrisero]

Problem

The Claude executor picker was missing the Fable / Claude-5 family, had no way to enter a custom model id, and could not attach an effort level to preset-injected models (e.g. Miguel's claude-fable-5 default surfaced with empty reasoning options). Two real bugs compounded this: (1) model_supports_effort() excluded fable, so CLI-mode launches of claude-fable-5 silently dropped --effort; (2) there was no ultracode effort level, and any UI-only effort id not in the ClaudeEffort enum is silently discarded by apply_overrides().

What changed

Backend (crates/executors):

• Added fable (alias, "Fable") to the top of the model list in default_discovered_options(), newest-first: Fable, Opus, Opus (1M context), Sonnet, Haiku.
• Fixed model_supports_effort() to include fable, so effort is no longer dropped for Fable models at interactive CLI launch.
• Added a real Ultracode variant to ClaudeEffort (serde/strum lowercase → ultracode on the wire), appended to the discovered effort options, with rank 6 + label "Ultracode" in model_selector.rs.
• Interactive CLI path (installed 2.1.204, which accepts the ultracode alias) passes --effort ultracode verbatim. Headless path maps Ultracode → xhigh via a new ClaudeEffort::headless_arg() because the pinned headless CLI (@2.1.154) hard-rejects ultracode and its dynamic-workflow orchestration is interactive-only.
• Regenerated shared/types.ts and shared/schemas/claude_code.json (flat string enum preserved — the variant note is a plain // comment so the schema shape is unchanged for the settings agent-config editor).

Frontend:

• ModelSelectorPopover / ModelList: free-text model entry — when the trimmed search query matches no model, a Use "<query>" row is offered that selects the arbitrary id. The flat (no-provider) picker now always shows the filter/entry box.
• appendPresetModel(): gated reasoning-options fallback — an injected/custom effort-capable Claude id (opus/sonnet/fable, provider-less) inherits the effort options from the first config model that has them, so Fable presets and custom ids can pick an effort (including ultracode). Provider-scoped models (codex path) are unaffected.
• Added modelSelector.useCustom i18n key across all locales.

How verified

• cargo test -p executors — 84 passed. New tests: fable+xhigh red/green for the model_supports_effort fix, ultracode FromStr/AsRef round-trip, Ultracode.headless_arg() == "xhigh", interactive passthrough of ultracode.
• Discriminating CLI test run: npx -y @anthropic-ai/claude-code@2.1.154 --effort ultracode --version → error: option '--effort <level>' argument 'ultracode' is invalid. It must be one of: low, medium, high, xhigh, max — confirms the headless→xhigh mapping is mandatory.
• pnpm run generate-types → ultracode present in shared/types.ts and shared/schemas/claude_code.json; schema stays a flat enum (one-line diff).
• New Vitest spec modelSelector.test.ts (5 passed) covering the gated reasoning-options fallback.
• pnpm run local-web:check, web-core:check, ui:check pass; ui:lint, local-web:lint, and check-unused-i18n-keys.mjs clean; cargo clippy -p executors clean; pnpm run format run.
• NOT verified: end-to-end live launch (no /opt access, no paid model calls); fable[1m] validity (see Risks); actual workflow-orchestration behavior of ultracode in a live session.

Risks / notes

• fable[1m] was intentionally left out: only opus/sonnet [1m] handling is evidenced in the 2.1.204 binary; not adding an unverified entry.
• default_model / CLI_DEFAULT_MODEL were left as opus per the brief — flipping the repo fallback to fable is a separate conscious call (Miguel's profile already defaults to fable). Possible follow-up.
• Headless CLI pin (@2.1.154) was NOT bumped (blast radius: loop automation + title_gen). Full model ids like claude-fable-5 pass through on 2.1.154 regardless; only the bare fable alias needs ≥2.1.204, which affects headless spawns only.
• ClaudeEffort is serialized in profiles.json / stored ExecutorActions — this only ADDS a variant (backward compatible); nothing renamed or removed.
• Free-text entry lets typos through; a bad id fails only at claude startup inside tmux (visible but easy to miss). normalize_claude_model_id only fixes dotted claude-* ids.

---

ship-it battery

Automated multi-pass review battery run on this PR; all local checks green before each push. Follow-up commits: 3193fd57, 44650e12, c90b75ca, 987df99a, 14a6046f.

Council round 1 (mixed: Codex ×4 personas + Claude cross-check) — all CAUTION, no P0/BLOCK. Applied: corrected the headless_arg doc comment (2.1.154 hard-rejects unknown --effort, not warn-and-ignore); made model_supports_effort case-insensitive to faithfully mirror the TS modelSupportsEffort (closes a mixed-case silent --effort drop); added headless_arg_matches_as_ref_except_ultracode (pins headless == interactive wire value for every tier except the deliberate ultracode→xhigh delta) and effort_wire_strings_are_stable (golden test enforcing the locked ADD-only never-rename policy at CI). Rejected: tolerant #[serde(other)] deserialization (would reintroduce the silent-drop this PR fixes; downgrade isn't a product requirement) and coupling the generic model_selector ranker to the Claude enum (inverts dependency direction).

/simplify (4 cleanup agents) — applied: derive the picker effort list from ClaudeEffort::VARIANTS (strum, verified to honor serialize_all) instead of a hand-maintained array, so a future tier can't be silently missing; drop the redundant reasoning_label("ultracode") arm (Title-case fallback already yields "Ultracode"). Rejected: collapsing headless_arg to _ => self.as_ref() (keeps compile-time exhaustiveness at the version-skew boundary) and extracting a shared row component (pre-existing triplication, out of diff scope).

Codex + /code-review xhigh — Codex: injected the current override model (not just the profile preset) so a free-text custom effort-capable Claude id gets an effort selector, and gated the appendPresetModel effort fallback behind an explicit flag set only for CLAUDE_CODE. xhigh caught the key defect: the free-text custom-entry row + always-on search box had leaked to every flat (no-provider) executor (Codex, Gemini, Copilot, Droid, Qwen, Cursor), not just Claude — a footgun (arbitrary/typo ids for executors whose downstream wasn't verified; Codex customs would lose reasoning options). Fixed by gating both behind a new allowCustomModel prop set only for Claude; non-Claude flat pickers keep exact prior behavior (MODEL_LIST_PAGE_SIZE restored). Skipped (minor, noted): a custom Claude id isn't written to the recently-used LRU (it still persists as the active selection).

Council round 2 (mixed, updated diff) — 4 APPROVE + 1 CAUTION, no P0/P1/BLOCK. Applied: gated the override-model injection on isClaude too, so non-Claude executors get exactly the pre-PR preset-only injection. Rejected: wiring the new Vitest spec into CI (pre-existing repo-wide condition — vitest is intentionally not a devDep and tsc excludes *.test.ts; ran locally, 7 pass).

CodeRabbit — CLI could not authenticate in this non-interactive environment (environment_unsupported: no API key, no localhost-callback/stdin fallback) on both --plain and --agent. Deterministic env limitation, not a transient rate-limit, so per the fallback a Codex xhigh review was substituted on the final diff → no actionable findings.

Verification: cargo test -p executors (all pass, incl. new tests), cargo clippy -p executors clean, pnpm run check (frontend tsc all green), npx vitest run modelSelector.test.ts (7 pass), pnpm run format applied. No new migrations; generated shared/types.ts + shared/schemas/claude_code.json regenerated (not hand-edited). Locked decisions honored: no default_model flip, no headless pin bump, ADD-only enum.