feat(6.26): per-role model dropdowns from seed + user override by fstamatelopoulos · Pull Request #31 · fstamatelopoulos/cfcf

fstamatelopoulos · 2026-05-03T05:34:26Z

Summary

Replaces the free-text per-role model field across web Settings,
workspace Config, and CLI init / config-edit with a registry-driven
dropdown. Registry = bundled seed in
packages/core/src/adapters/seed-models.ts (intentionally minimal --
generic aliases like opus, sonnet, haiku for claude-code;
gpt-5-codex, gpt-5, o3 for codex) merged with an optional user
override on CfcfGlobalConfig.agentModels[<adapter>]. Override is
managed via a new "Model registry" section on the Settings page.

Why no remote registry / API query

Researched first; documented in commits + CLAUDE.md:

Neither claude nor codex exposes a list-models subcommand
Anthropic / OpenAI Models APIs require auth (API key); not what
Claude Code's OAuth subscription users have
Hosting our own remote registry shifts maintenance responsibility
onto cfcf maintainers

Seed + user override side-steps all three. Hand-edited config values
not in the registry are preserved as (custom) entries on first
render so back-compat doesn't break.

Single edit surface (review feedback)

Original design had a "(custom model name…)" sentinel in every
picker as an inline escape hatch. Dropped after review: having two
add-paths created a UX question the design didn't answer cleanly
("does this persist?"). Settings → Model registry is now the one
place to add / remove / reset; pickers are read-only dropdowns.

Surfaces touched

Settings → Agent roles + Workspace → Config → Agent roles:
free-text model field becomes <AgentModelSelect> (a <select>
with (adapter default) first option + the resolved registry).
Adapter swap re-targets the model dropdown.
Settings → Model registry: new section. Per-adapter card with
the resolved list as deletable chips + add-model input + reset to
seed.
CLI cfcf init / cfcf config edit: numbered selector
matching the existing pickAgent pattern. Hand-edited per-role
models that aren't in the registry survive as a numbered
"(custom — from existing config)" line so re-running init keeps
the user's pick.

UX / theming polish bundled in (review feedback)

Stripped two plan-item refs that had leaked into UI titles
("Model registry (item 6.26)", "Behaviour flags (item 5.1)").
Saved a feedback memory so this doesn't recur.
Static section titles (FormSection h3, Memory page h3) no longer
look clickable. The architect-review__summary class was reused
on static h3s, inheriting cursor: pointer + hover from its
legitimate use on <details>/<summary> blocks. Split into a
new .section-title class without interactivity styling.
Bumped table label / cell font sizes to match the form-input
scale (var(--text-md) ≈ 0.95rem) -- left-column labels on
Server status + role-name cells in Agent roles read at the same
size as the dropdown text next to them.

Tests

8 new unit tests covering resolveModelsForAdapter (seed
precedence, override precedence, empty-array fallback,
malformed-entry filtering, per-adapter isolation, user-only
adapter)
4 new HTTP tests covering GET /api/agents/models

bun run typecheck clean. bun run test: 579 / 100 / 66 / 9 green.
bun run build:web clean.

Test plan

Build + tests + typecheck
Manual: rebuild + restart server → Settings + workspace Config
dropdowns populated from registry; adapter swap re-targets
model dropdown; Model registry section adds/removes
reflected in pickers; "Reset to seed" works; CLI
cfcf init --force walks the numbered picker; existing
hand-edited model values preserved as "(custom)" rows

Adds the data + helpers that drive per-role model dropdowns in the web UI and CLI. Two layers: 1. Bundled seed in packages/core/src/adapters/seed-models.ts. Intentionally minimal -- generic aliases (opus / sonnet / haiku for claude-code; gpt-5-codex / gpt-5 / o3 for codex) over date-bound full names so the seed ages slowly. Aliases stay stable as upstream releases new model versions. 2. Optional user override on CfcfGlobalConfig.agentModels[<adapter>] -- managed via the web Settings → Model registry editor (next commit). When set + non-empty, supersedes the seed. resolveModelsForAdapter(adapter, config) is the single source of truth for "what models to surface in pickers"; resolveAllModels(config) is the bulk version for the API endpoint. Pickers always also offer a "(adapter default)" first option (empty value) and a "(custom model name…)" sentinel last option for the long tail. validateConfig coerces malformed agentModels entries to a no-op (filters non-string + blank model names; drops the field if the whole thing is shapeless) so a hand-edited config can't break startup. 8 new unit tests cover seed lookup, override precedence, empty-array fallback, malformed-entry filtering, per-adapter isolation, and the "user-only adapter with no seed" case. Why no remote registry: hosting one shifts maintenance responsibility onto cfcf maintainers (model-list curation, rate-limiting, availability). Seed-plus-override side-steps that; the (custom) fallback handles the long tail. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Returns { adapters, seed }: `adapters` is the resolved per-adapter list (user override else seed), `seed` is surfaced alongside so the web Settings → Model registry editor can render a "Reset to seed" affordance without having to know seed values out-of-band. Web pickers (Settings agent-roles + workspace Config agent-roles) fetch this on mount and after every save that touches `agentModels`. The seed list is small -- response stays in single-digit-KB range -- so the cost of refreshing on every picker open is negligible. 4 new HTTP tests cover seed-only, user-override, seed-surfaced- alongside, and pre-init (no config file) cases. Each test isolates the user's real ~/.cfcf/ via CFCF_CONFIG_DIR. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Replaces the free-text model field in two places (Settings → Agent roles + workspace Config → Agent roles) with the new AgentModelSelect component. Layout per role: - <select> populated from the resolved per-adapter registry (fetched once on mount + re-fetched after a Model registry save) - "(adapter default)" first option (empty value; current free-text empty semantic preserved) - "(custom model name…)" sentinel last option that swaps to a free- text input -- mirrors the ClioProjectDialog pattern shipped in 6.12 - Hand-edited config values not in the registry render as `<option>{value} (custom)</option>` so we never silently lose them When the role's adapter dropdown changes, the model dropdown's options swap to the new adapter's registry; if the previously-selected model isn't in the new list, custom mode kicks in automatically. New "Model registry" section in Settings: per-adapter card showing the current resolved list as deletable chips + an "Add model name" input + "Reset to seed" button when an override is in effect. Saving routes through the existing global-config save path; an empty list (or one identical to the seed) clears the override so future seed updates flow through automatically. types.ts: agentModels?: Record<string, string[]> on GlobalConfig to mirror the core type. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Replaces the free-text "X agent model" prompts in `cfcf init` (and therefore `cfcf config edit`, which delegates) with a numbered selector mirroring the existing `pickAgent` pattern. Layout per role: Dev model (adapter: claude-code): 0. (use adapter default) 1. opus 2. sonnet 3. haiku 4. (custom -- type a name) Choose [2]: The registry per role is resolved live from resolveModelsForAdapter(adapter, config), so: - editing Settings → Model registry on the web side affects the next CLI prompt - the per-role adapter chosen earlier in the same init run drives which list appears (each role can pick from its own adapter's models) - hand-edited per-role models that aren't in the registry default to the "custom" branch (so re-running init keeps the user's pick on Enter) CLAUDE.md gets a paragraph documenting the seed-vs-override model and the maintenance workflow (when an upstream agent CLI ships a new headline model, update the relevant array in seed-models.ts and ship in the next release; user overrides survive the upgrade). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

… (6.26 review) Two UX/polish fixes from the 6.26 review: 1. Plan-item refs leaking into rendered titles: "Model registry (item 6.26)" → "Model registry"; "Behaviour flags (item 5.1)" → "Behaviour flags". Internal references confuse end users and signal an unpolished product. Code comments still carry the trace, which is the right place for them. (Memory file added on the agent side so this doesn't happen again.) 2. Static section titles looked clickable. The .architect-review__ summary class was reused on FormSection's <h3>s but it has `cursor: pointer` + a hover colour swap because it's also used on real <summary> elements inside <details> in ArchitectReview / JudgeDetail. Split: <summary> elements keep architect-review__summary; <h3> static titles get a new .section-title class with the same typography but no cursor / hover. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Surfaced in review: having two ways to add a model -- the inline "(custom model name…)" sentinel in every picker AND the Settings → Model registry editor -- created a UX question the design didn't answer cleanly ("does this persist?"). The custom sentinel was one-shot per role with no persistence; the registry editor is the real management surface. Resolution: single edit surface. The picker is now a plain dropdown of (adapter default) + registry models. To add a model that's not in the registry, the user goes to Settings → Model registry → Add. To remove a model, click the × on its chip. To reset, click "Reset to seed". One place, three operations, no ambiguity. Back-compat preserved: a hand-edited config value not in the registry still renders as `<option>{value} (custom)</option>` on first paint so we never silently coerce it. The user clears it by picking another option, or makes it stick by adding it to the registry. Same back-compat handling on the CLI side: an unknown existing per-role model shows up as a numbered option labeled "<value> (custom — from existing config)". CLAUDE.md note + Settings page intro updated to reflect the simpler model. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

… review) Spotted in the screenshot: left-column labels in the Server section (Status / Version / Port) and the Role-name cells in Agent roles (Dev / Judge / Solution Architect) read noticeably smaller than the right-hand values + the input dropdowns next to them. The values use the new --text-md (0.95rem) input styling; the labels were stuck on 0.8rem from before the typography pass. Three rules bumped: - .config-display__table th: 0.8rem → var(--text-md) (Server labels) - .project-history__time: 0.8rem → var(--text-md) (Role cells) - .project-history__table: 0.875rem → var(--text-md) (table body) - .project-history__agent: 0.8rem → var(--text-sm) (history rows) Column headers (.project-history__table th, the uppercase ROLE / ADAPTER / MODEL row) bumped a smaller notch -- 0.75rem → var(--text-xs) 0.825rem -- since they're already uppercase + letterspaced and the contrast against body text is part of their visual signal. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Caught by the user during 6.26 dogfooding. I fixed the same string in ConfigDisplay (workspace Config tab) earlier but missed the parallel "Behaviour flags (item 5.1)" in ServerInfo (the Settings page) -- both forms have a Behaviour flags section and they drifted. Audit of every user-facing string in packages/web/src/ + packages/cli/src/ now comes back clean (regex: `"[^"]*item [0-9]\.[0-9][^"]*"|>[^<]*item [0-9]\.[0-9][^<]*<`). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…cale The previous bump only landed on the four table classes I'd identified on the Settings page (config-display__table th, project-history__time, etc.). Other views still had small text from two sources I missed: 1. ~25 inline `fontSize: "0.7Xrem"` style literals scattered across 12 components (ConfigDisplay, WorkspaceHistory, PaSessionDetail, LoopControls, ...). These bypass any CSS rule via specificity, so no amount of class-targeted bumping reaches them. 2. ~50 small `font-size: 0.7Xrem` declarations in app.css for component-specific classes (phase-step__label, judge-assessment__ tests, architect-review__counts, log-viewer__content, modal__title, status-badge, …). Sweep replaces every literal with the corresponding token: 0.75/0.78rem → var(--text-xs) (~13.2px) 0.8 /0.85rem → var(--text-sm) (~14.4px) 0.875rem → var(--text-md) (~15.2px) So all small text now lives at three coordinated sizes that move together when the scale changes. Net effect: every view (workspace detail, history, logs, memory, help, judge/architect/reflection detail panels, modals) gets a ~10-15% size bump on its small-text elements, matching the bump that landed on Settings + workspace Config tabs in the previous commit. PaSessionDetail had three non-standard literals (0.72rem, 0.9rem, 0.92rem) cleaned up to the nearest token in the same pass. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Following the same pattern as 6.20 + 6.12: PR #31 landed on main (per-role model dropdowns + Model registry editor + typography polish + remaining plan-item ref stripping). The v0.18.0 tag was placed after 6.12 but never released (release.yml is workflow_dispatch only and hasn't been triggered), so re-targeting it at this commit captures everything that ships in the actual release artefact. CHANGELOG header now lists 6.20 + 6.12 + 6.26 with a unified rationale; 6.20, Security, and 6.12 sections kept verbatim; new 6.26 section + Typography polish section inserted before the v0.17.1 boundary. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Plan housekeeping after the v0.18.0 release went out (PR #30 + #31 merged, tag pushed, npm publish completed): - Row 6.12 ❌ → ✅ with shipped-state notes. Original audit-only scope expanded into "audit + close every gap surfaced + theming overhaul" in the same PR; preserved the original description for reference. - Row 6.26 ❌ → ✅ with the resolved design (seed + user override + no remote registry; single edit surface in Settings → Model registry). Resolves the open questions from the original draft. - Iter-6 active-set + headline summary lines updated: 6.20, 6.12, 6.26 marked shipped; remaining set is now 6.9, 6.11, 6.13, 6.18, 6.24 (+ 6.19 partial). cli-usage.md: rewrote step 4 of "What `cfcf init` does" to describe the new numbered model picker driven by the per-adapter registry, with a pointer to Settings → Model registry for augmenting the list. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

fstamatelopoulos and others added 9 commits May 2, 2026 22:05

fstamatelopoulos merged commit 26ad7ae into main May 3, 2026
2 checks passed

fstamatelopoulos deleted the iteration-6/per-role-model-dropdowns branch May 3, 2026 05:34

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat(6.26): per-role model dropdowns from seed + user override#31

feat(6.26): per-role model dropdowns from seed + user override#31
fstamatelopoulos merged 9 commits intomainfrom
iteration-6/per-role-model-dropdowns

fstamatelopoulos commented May 3, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

fstamatelopoulos commented May 3, 2026

Summary

Why no remote registry / API query

Single edit surface (review feedback)

Surfaces touched

UX / theming polish bundled in (review feedback)

Tests

Test plan

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant