Local lanes dropped the knob's graded value unless the operator declared
reasoning_effort_values (and, on the template channel, an effort key) —
picking Max sent a bare thinking toggle and the effort select
degenerated into seven positions that all meant 'on'. The user's
setting now always rides:
- openai-compatible: the flat reasoning_effort param carries the knob
verbatim (effort_passthrough on the lane default); declared values
still snap ordinally, and a declared effort_param still claims the
template channel and suppresses the flat param.
- anthropic-compatible: the graded value rides chat_template_kwargs
alongside the toggle whenever reasoning control is engaged — under
the operator's effort_param, else the conventional fallback key
(reasoning_effort); templates that don't reference the kwarg ignore
it. thinking_mode=none still injects nothing.
- Commercial lanes untouched: empty declared values still mean 'no
effort control' (o1-mini) and the ordinal snap is unchanged.
Golden writer now pins ensure_ascii=False: the baselines' literal em
dashes came from a hand edit (03f82521) the default-escaping writer
could never reproduce — regens no longer churn unrelated lines.
The except path for a malformed capabilities column appended the row
without the key, so clients had to null-check a field the happy path
guarantees. Initialize each entry with an empty ladder and let the try
block overwrite it — the response schema is stable per row.
The one that mattered: /v1/api/models passed the capabilities column —
a JSON STRING (sa.Text) — straight into effort_ladder_for_model, whose
field filter calls .items() on it; the per-row guard swallowed the
AttributeError, so effort_ladder was silently absent from every row and
the sklc annotation could never fire. The endpoint now parses the JSON
and splits the namespaced server_compat exactly like the model_registry
loader, and a regression test seeds a string-capabilities row.
Projection fidelity: effort_ladder_for_model threads api_surface (the
responses surface ignores extra_body — flat-param-only ladder, matching
create_provider's request-time divergence); google/xai route through
the chat-lane projection they actually inherit (_finalize_extra_body +
flat param); the native-Anthropic branch reflects that output_config
gates on supports_effort alone, independent of thinking_mode; budget
clamping to per-request max_tokens is documented as out of scope.
Hardening and symmetry: admin endpoint 400s (not 500s) on non-dict JSON
bodies and gained HTTP tests; the admin edit-load strips effort_param
from the raw JSON only for the lanes whose save path re-adds it, so
non-compat rows can't silently lose a stored key; the empty-model early
return bumps the ladder sequence so stale in-flight responses can't
re-annotate; alias labels only reference positions the target select
actually offers (the skill shelf omits none/minimal); the sklc models
cache no longer pins a rejected promise and is invalidated on the
models_changed event; debounces unified at 500ms;
merge_reasoning_template_kwargs now always returns a fresh dict for
non-empty input; the shared budget constants are public.
Copilot review feedback on #500. The original
``list_available_models`` had an implicit cs=None branch where the
placeholder still advertised ``registry.default`` (filtered against
enabled rows) when ``app.state.config_store`` was None but
``coord_registry`` was bound — useful in the rare degraded state
where lifespan wired the registry but the ConfigStore failed to
initialise. The PR #500 refactor accidentally dropped that branch:
the helper requires a config_store, so the cs=None case fell out as
"blank coordinator default".
Add an explicit ``elif coord_registry is not None`` branch that
mirrors the helper's tier 3 with the placeholder's enabled-rows
filter applied. New test exercises this path by passing
``config_store=False`` to the test fixture.
Previously /v1/api/models (home composer placeholder) and
console/session_factory.py walked separate two-/three-tier chains for
the coordinator alias. session_factory was missing the
``model.default_alias`` tier, so admins who set the system default in
the Models tab would see it advertised but new coordinator sessions
would silently keep launching on ``registry.default``.
This commit:
- Extracts the chain into ``turnstone/console/coordinator_alias.py``.
``resolve_coordinator_alias`` returns the effective alias under a
shared three-tier policy: explicit pin → ``model.default_alias`` →
``registry.default``. Tier 2 is validated against
``registry.has_alias`` and falls through to tier 3 with a logged
warning if unknown. Tier 1 is intentionally passed through
unvalidated so an explicit operator pin surfaces as 503 at
``registry.resolve`` rather than being silently swapped out.
- Wires both call sites through the helper. The placeholder supplies
an ``alias_filter`` that restricts every tier to enabled DB rows so
the home composer never advertises a model the workstream picker
can't actually offer; the session factory uses no filter (matches
prior 503-on-typo behaviour for explicit pins).
- Adds direct integration tests for the session factory's chain
(``tests/test_console_session_factory.py``) and updates the
placeholder tests' fixture to provide a stub coord_registry, since
the helper now requires one.
Light-review followup on 389400c8.
The "mirrors session_factory.py:109-110" claim was inaccurate —
session_factory's chain is two tiers (coordinator.model_alias →
registry.default) and skips model.default_alias entirely. The
placeholder handler extends that chain with model.default_alias as
tier 2 so admins who set the default in the Models tab see it
advertised in the home composer. Comment now lists the three tiers
explicitly and flags the session_factory-vs-placeholder drift case
(where model.default_alias ≠ registry.default) as a separate issue
to track.
Also lifts the ``from types import SimpleNamespace`` import in the
test fixture to module level — minor readability cleanup.
Two Copilot-review followups on /v1/api/models default resolution.
- console/server.py: coordinator_default_alias now mirrors the full
fallback chain in console/session_factory.py:109-110 — explicit
coordinator.model_alias → model.default_alias → registry.default.
The registry tier was missing, so the home composer placeholder went
blank whenever an operator never set model.default_alias in the admin
UI even though new coordinator sessions still launch on
registry.default (loaded from config.toml [model].default by
load_model_registry). Two new tests cover the registry-default
branch and the disabled-alias guard.
- console/static/app.js: _resolveModelLabel returns "" (not the bare
alias) when the alias isn't found in the dropdown's model list, so
callers can rely on the documented "fall back to neutral placeholder"
contract. Matches the existing doc comment.
Bundles the click-around polish on the console admin UX.
Home composer + schedule modals
- /v1/api/models now exposes coordinator_default_alias + judge_default_alias,
resolved through the same chain console/session_factory.py uses. Both the
home composer's MODEL / JUDGE MODEL placeholders and the schedule create /
edit modal model placeholders rewrite to "Default — alias (model)" once
the API responds. The `models_changed` SSE refresh keeps placeholders
current as operators edit per-role assignments.
- Composer.setOptionPlaceholder added so callers can update just the first
option's text without disturbing the rest of the choice list.
Admin → Models → Roles
- Channel adapter row added (channels.default_model_alias) — the migration
to the Roles sub-tab missed it. Key added to
_MODEL_AFFECTING_SETTING_KEYS so edits fire the SSE refresh, and to the
settings-tab roleKeys skip-list so it only renders in one place.
- Plan/Task agent rows now display "(inherit)" instead of the misleading
"(default — <alias>)" — those roles cascade through plan_model →
agent_model → session model, not a single concrete default.
- coordinator.reasoning_effort accepts "" (inherit), matching
model.plan_effort / model.task_effort.
- Blank options in each role's MODEL select now match the "alias (model)"
shape used by the other rows.
Toggle-switch component
- New .toggle-switch component (visually-hidden native checkbox + styled
track + label). 40×22 hit target meets WCAG 2.5.5 (AAA), inset ring on
the off state for ≥1.5:1 contrast against the modal surface.
- .toggle-stack groups toggles in a column with .toggle-group-divider for
conceptual grouping (used in the Add Model modal between "Active" and the
paired Reasoning toggles).
- .toggle--flush modifier zeroes the default top margin for toggles that
sit flush against a heading or a dynamically-rendered row.
Sweep — every admin-modal boolean checkbox is now a toggle:
schedule (cs/es-autoapprove, es-enabled), policy (ep/epp-enabled),
tool-mode (ctm/etm-default), skill (csk/esk-auto-approve, csk/esk-enabled),
MCP (mcp-auto-approve, mcp-enabled), Add Model (Active, surface-persisted-
reasoning, replay-reasoning), judge bool settings (cancel_on_approval et
al.), and the user-roles-modal role assignment list. The two
ogp-cred / eogp-cred inline credential checkboxes stay as compact inline
boxes since they sit beside text inputs in tight horizontal rows.
Add Model modal — the "Enabled" toggle promoted to "Active" and moved to
the very top of the form. Tooltip explains it gates dropdown visibility
without removing the definition.
MCP authorization — the three radio buttons replaced with a vertical
.segmented-control option list. Selected row paints --accent-dim plus a
filled .segmented-indicator; focus ring uses --accent so it stays visible
on the currently-selected option.
Role permissions modal — the 19 permission checkboxes are now
.toggle-switch.perm-toggle (monospace lowercase identifiers preserved).
The permissions are split into Scopes / Admin / Workstreams & Tools
sections under caps-styled section headers so the row-flow grid no longer
slices `admin.*` mid-column.
Judge bool toggles use a static "Enabled" caption rather than flipping
text on `.checked`; flipping lagged 50–300 ms behind the slider position
because the caption was sourced from the post-save reload.
CSS cleanup — dead `.admin-checkbox` / `.perm-checkbox` rules removed.
Specificity audit (scripts/css_specificity_audit.py) returns no conflicts
on any new component class.
Tests — 525 pass on the affected slices; new tests/test_console_available_
models.py pins each branch of the resolution chain in /v1/api/models so the
home composer placeholder stays correct as precedence rules evolve.
Patrick Buckley