refactor: centralize runtime feature flag bindings (#2142)

This commit is contained in:
Carlos Florêncio
2026-08-02 15:36:21 +01:00
committed by GitHub
parent 443e9c89cc
commit e077658a61
29 changed files with 1495 additions and 506 deletions
+160 -219
View File
@@ -1,128 +1,95 @@
---
status: draft
created: 2026-06-14
updated: 2026-08-02
owner: tbd
---
# Feature Toggles
Decision: [ADR-2026-08-01-release-toggle-gating-contract](../../decisions/2026-08-01-release-toggle-gating-contract.md)
Implementation plan: [Release Feature Toggles](../../plans/release-feature-toggles/plan.md)
## Why
Users can already opt into runtime features through environment variables and
`profiles.yaml`, but those controls are invisible from the app and require
terminal access. Users need a settings page that explains experimental
features, lets them choose install-level overrides, and clearly handles the
restart required for startup-gated features such as Office mode and Debug mode.
Kandev needs to merge medium and large changes into `main` without making an
unfinished or risky behavior available in the next release. Operators also
need a simple way to enable a gated feature on one installation, test it with
real work, disable it again if necessary, and understand when restart is
required. Adding and later removing a temporary release toggle must not require
duplicating flag identity or config-binding code across many files.
## What
- Kandev adds a **Feature Toggles** settings page at
- Kandev provides install-wide runtime toggles under
**Settings > System > Feature Toggles** at
`/settings/system/feature-toggles`.
- The page shows user-facing runtime toggles, not every profile/env knob.
- V1 exposes three user-facing toggles:
- **Office mode** — enables the autonomous-agent Office surface.
- **App status bar** — shows the general-purpose desktop/tablet status bar and phone Status drawer entry. It is stable and off by default in production; enabling it changes visibility only. Disabling it does not suppress the problem-only [WebSocket connectivity warning](../ui/ws-connectivity-warning.md).
- **Debug mode** — enables local diagnostic/debug behavior, including agent
message debug logs.
- Office mode is marked **Experimental** and includes a concise description of
what it does and the risks of enabling it.
- Debug mode is a single user-facing toggle. Agent message debug logs are part
of Debug mode, not a duplicate top-level toggle.
- Every v1 toggle requires restart before the changed value takes effect.
- The page distinguishes effective value, saved override, default value, and
environment-locked value.
- Explicit environment variables remain authoritative. If an env var controls a
flag, the UI shows the flag as locked and does not allow overriding it.
- Users can reset a saved override back to the profile/env default.
- After saving any restart-required change, the UI shows a persistent restart
banner and offers a restart action when Kandev can safely restart itself.
- If automatic restart is unsupported, the UI shows manual restart guidance.
- `/api/v1/features` continues to return effective feature booleans for SSR
gating and existing `useFeature()` callers.
## Tech stack
- Backend: Go, Gin HTTP handlers, SQLite persistence.
- Frontend: Next.js App Router, React, Zustand, `@kandev/ui` shadcn
components.
- E2E: Playwright under `apps/web/e2e`.
- Runtime defaults: `profiles.yaml` embedded by `apps/backend/internal/profiles`.
## Commands
```bash
make fmt
make typecheck test lint
make -C apps/backend test
cd apps && pnpm --filter @kandev/web typecheck
cd apps && pnpm --filter @kandev/web lint
cd apps/web && pnpm e2e -- settings/feature-toggles.spec.ts settings/mobile-feature-toggles.spec.ts
```
## Project structure
```text
apps/backend/internal/runtimeflags/ Runtime flag registry, store, service, handlers
apps/backend/internal/system/restart/ Restart capability/request support
apps/backend/cmd/kandev/ Startup override application and route wiring
apps/web/app/settings/system/feature-toggles/page.tsx
apps/web/components/settings/system/feature-toggles-settings.tsx
apps/web/lib/api/domains/runtime-flags-api.ts
docs/decisions/0018-runtime-settings-overrides.md
docs/decisions/0019-restart-supervisor.md
```
Planned E2E coverage:
```text
apps/web/e2e/settings/feature-toggles.spec.ts TODO
apps/web/e2e/settings/mobile-feature-toggles.spec.ts TODO
```
## Code style
Use typed contracts at the backend boundary and keep env-var knowledge in the
backend registry rather than duplicating it in React components.
```go
type RuntimeFlagDefinition struct {
Key string
EnvVar string
Kind RuntimeFlagKind
Label string
Description string
Stability RuntimeFlagStability
RiskLevel RuntimeFlagRiskLevel
RiskDescription string
RestartRequired bool
Mutable bool
}
```
Frontend components render the API shape and avoid hardcoding flag-specific
logic except for layout/grouping.
- Runtime toggles have two uses:
- **Release toggles** temporarily protect unfinished or risky behavior while
it bakes on selected installations. They are removed after graduation.
- **Operator configuration toggles** are supported long-lived installation
choices such as Authentication or Debug mode.
- A new release toggle defaults to `false` in the `prod`, `dev`, and `e2e`
profiles. A feature spec may deliberately choose a different non-production
default, but `prod` remains off until the default-on rollout.
- An administrator can save an installation override from Feature Toggles.
Explicit environment variables remain authoritative and lock the UI control.
- Every toggle change requires restart before its effective runtime value
changes. Unsupported supervisors show manual restart guidance.
- `/api/v1/features` exposes the startup-effective feature booleans used by SSR
and `useFeature()` callers. Missing or unavailable frontend flag data fails
closed to `false`.
- Feature UI is absent while its release toggle is off. A guessed route returns
`404` when the feature owns a route subtree.
- The backend is authoritative. A disabled feature cannot be invoked by direct
HTTP, WebSocket, MCP, background-job, or agent-tool entry paths even if a
stale or modified client sends its enabled-only request shape.
- Gates live at the narrowest composition or capability boundary that covers
each entry path. Internal helpers do not independently re-read the flag when
their caller has already enforced the same boundary.
- The disabled path preserves the behavior that existed before the gated PR.
Shared refactors, startup initialization, schema migrations, and data-model
compatibility are not considered protected merely because a later behavior
branch is gated.
- Data written while a feature is enabled remains readable and inert when the
feature is disabled. Each gated feature documents any stronger rollback
behavior in its own spec.
- Runtime flag definitions bind their registry key, environment variable,
config read, and config apply behavior in one backend definition. Frontend
feature names and default-false values have one declaration. Graduated flag
keys and environment variables move to an append-only retired-identity set.
- A release toggle follows this lifecycle:
1. merge with every shipped profile default off;
2. enable selected installations with an override or explicit environment;
3. ship one default-on release while retaining the kill switch;
4. remove the live flag and legacy behavior after the default-on release is
proven, while retaining its retired key and environment identity.
- Removed flag keys and environment variables are never reused. Unknown
persisted override rows are ignored so upgrades and downgrades do not
destructively rewrite operator state.
## Data model
Feature Toggle overrides are install-scoped and stored in SQLite.
`runtime_flag_overrides`
Feature-toggle overrides are installation-scoped and stored in
`runtime_flag_overrides`.
| Field | Type | Constraint |
|---|---|---|
| `key` | `TEXT` | primary key, e.g. `features.office` |
| `key` | `TEXT` | primary key, for example `features.office` |
| `value` | `INTEGER` | required, `0` or `1` |
| `created_at` | `DATETIME` | required |
| `updated_at` | `DATETIME` | required |
Absence of a row means "use the effective default". A `false` row is a real
override and must not be treated as missing.
Absence of a row means “use the effective default.” A `false` row is a real
override. Rows whose keys are absent from the running binary's registry are
ignored and are not deleted automatically.
## API surface
`GET /api/v1/runtime-flags`
### `GET /api/v1/runtime-flags`
Returns all user-facing runtime toggle states and metadata.
Returns the registered operator-facing toggle states and metadata.
```json
{
@@ -134,7 +101,7 @@ Returns all user-facing runtime toggle states and metadata.
"description": "Enables autonomous agent office workflows and related settings.",
"stability": "experimental",
"risk_level": "medium",
"risk_description": "Office mode is still evolving. Workflows, routes, and background automation may change between releases and should be reviewed before relying on them.",
"risk_description": "Office mode is still evolving.",
"effective_value": false,
"default_value": false,
"override_value": true,
@@ -142,159 +109,133 @@ Returns all user-facing runtime toggle states and metadata.
"env_var": "KANDEV_FEATURES_OFFICE",
"env_locked": false,
"restart_required": true,
"requires_restart_to_apply": true
"requires_restart_to_apply": true,
"mutable": true
}
]
}
```
`PATCH /api/v1/runtime-flags/:key`
### `PATCH /api/v1/runtime-flags/:key`
```json
{ "override": true }
```
The request body is `{ "override": true | false | null }`. `null` clears the
saved override. Unknown keys return `404`; environment-locked keys return
`409`. Mutation requires an administrator identity. When authentication is
disabled, the synthetic install identity is an administrator.
`override` can be `true`, `false`, or `null`. `null` clears the saved override.
Unknown keys return `404`. Env-locked keys reject writes with `409`.
### `GET /api/v1/features`
`GET /api/v1/system/restart-capability`
Returns startup-effective feature booleans keyed by their public JSON names.
It remains a public bootstrapping endpoint and contains no risk metadata or
saved override details.
```json
{
"supported": true,
"mode": "cli",
"reason": ""
}
```
`mode` is `cli`, `systemd`, `launchd`, or `unsupported`.
`POST /api/v1/system/restart`
Requests a restart. The response is flushed before shutdown starts so the UI
can enter a waiting state and poll `/health`.
`GET /api/v1/features`
Existing contract remains unchanged: effective feature booleans only.
Restart capability and restart-request contracts remain defined by
[Restart supervisor ownership](../../decisions/0019-restart-supervisor.md).
## State machine
Runtime flag state:
```text
profile default -> override saved -> restart pending -> restarted/effective
\-> override cleared -> restart pending -> restarted/default
default off
-> selected-install override on
-> default on with kill switch retained
-> live flag removed, identity retired, new behavior permanent
```
Restart request state:
```text
idle -> requested -> backend unavailable -> backend healthy -> complete
\-> timeout -> manual recovery shown
```
At any retained-flag stage, an explicit environment value has precedence over
a saved override. A saved value that differs from the current process value is
`restart pending` until a new backend boot applies it.
## Permissions
Kandev is single-user in v1. Every user with settings access can view and change
Feature Toggles. Future multi-user support must restrict Feature Toggles,
especially Debug mode, to admins.
- Authenticated administrators and members can read runtime-toggle state and
risk metadata; only administrators can mutate installation-wide overrides.
- `/api/v1/features` remains public because the SPA needs it before authenticated
route composition; it exposes booleans only.
- MCP and agent callers do not receive a generic flag-mutation capability.
## Failure modes
- **Environment-locked flag** — UI disables the switch and shows the controlling
env var. API rejects override writes with `409`.
- **Override store unavailable** — runtime flags API fails with a recoverable
error; existing `/api/v1/features` continues to use startup-effective config.
- **Restart unsupported** — restart banner shows manual restart instructions
and no dead action button.
- **Restart requested while one is pending** — API rejects or idempotently
reports the pending restart.
- **Backend does not return after restart** — frontend times out and shows
manual recovery guidance.
- **Debug mode enabled** — UI warns that local diagnostic endpoints and agent
message logs may include sensitive prompt, file, and tool content.
- **Environment lock** — the UI disables mutation and identifies the controlling
environment variable; the API returns `409`.
- **Override store unavailable** — the runtime-flags API reports a recoverable
error; startup-effective behavior remains unchanged.
- **Frontend flag fetch fails** — client and SSR gates use their all-false
defaults and do not expose the feature.
- **Disabled feature request** — the authoritative backend boundary rejects the
request without mutation or side effects. Route-owned surfaces use `404`;
extensions to an existing endpoint use a stable client error defined by that
feature.
- **Restart unsupported or unsuccessful** — manual restart/recovery guidance is
shown and the running process retains its old effective values.
- **Unknown persisted override** — the row is ignored and hidden, but retained
for downgrade compatibility. Its key cannot be assigned to a future feature.
- **Feature disabled after enabled data exists** — the data remains readable but
cannot trigger enabled-only execution. The feature's own spec defines any
feature-specific resume or cleanup behavior.
## Persistence guarantees
- Saved overrides survive Kandev restarts.
- Effective runtime values are applied at backend startup.
- V1 toggles are startup-gated; changing a value does not affect already
registered routes, constructed services, or running agentctl processes until
restart.
- Environment variables always win over saved overrides.
- Saved overrides survive restart and version upgrades.
- Effective values are resolved at backend startup with precedence:
## Testing strategy
```text
explicit environment > saved install override > active profile > Go zero value
```
- Backend unit tests cover registry metadata, override persistence, precedence,
env locking, and pending-restart computation.
- Backend handler tests cover runtime flag APIs and restart capability/request
behavior.
- Frontend component tests cover Office experimental copy, Debug mode copy,
env-locked rows, save/reset actions, and restart banner states.
- Desktop and mobile Playwright tests cover the user workflow from settings
navigation through saving a restart-required toggle.
## Boundaries
- Always: keep `profiles.yaml` as immutable shipped defaults.
- Always: keep env vars authoritative over DB overrides.
- Always: show restart-required state for v1 toggles.
- Always: include mobile coverage for the settings page.
- Ask first: adding a new external flag service or changing the auth model.
- Ask first: exposing mocks/e2e tuning as user-facing toggles.
- Never: make the Go backend fork an unmanaged replacement process for restart.
- Never: expose agent message debug logs as a separate top-level toggle in v1.
- Applying or clearing a saved override does not live-mutate already registered
routes, constructed services, or agent processes.
- Unknown override rows are non-authoritative and non-destructive.
## Scenarios
- **GIVEN** Office mode is using the production default, **WHEN** the user opens
`/settings/system/feature-toggles`, **THEN** Office mode appears off with a
`Default` source, an `Experimental` badge, and risk text.
- **GIVEN** Office mode is off, **WHEN** the user enables it, **THEN** the
override is saved and the page shows that restart is required before Office
mode takes effect.
- **GIVEN** `KANDEV_FEATURES_OFFICE=true` is set in the environment, **WHEN** the
user opens Feature Toggles, **THEN** Office mode appears on, env-locked, and
cannot be changed from the UI.
- **GIVEN** Debug mode is off, **WHEN** the user enables it, **THEN** the page
explains that local diagnostic endpoints and agent message debug logs will be
enabled after restart.
- **GIVEN** App status bar is using its default, **WHEN** the user opens Feature
Toggles, **THEN** it appears off and explains that enabling it adds both the
desktop/tablet bar and ordinary phone Status entry after restart, while urgent
WebSocket connectivity warnings remain available when it is off.
- **GIVEN** a toggle override exists, **WHEN** the user clicks `Reset to default`,
**THEN** the override is removed and the page shows restart-required state if
the effective runtime value differs from the restored default.
- **GIVEN** restart capability is supported, **WHEN** the user clicks
`Restart Kandev`, **THEN** the UI enters `Restarting...`, polls `/health`, and
refreshes the page after the backend returns healthy.
- **GIVEN** restart capability is unsupported, **WHEN** a restart-required change
is saved, **THEN** the UI shows manual restart instructions instead of a
clickable restart button.
- **GIVEN** a mobile viewport, **WHEN** the user opens Feature Toggles, **THEN**
all descriptions, badges, switches, reset buttons, and restart actions are
reachable without horizontal scrolling.
- **GIVEN** a new release toggle, **WHEN** Kandev boots in `prod`, `dev`, or
`e2e` without an explicit override, **THEN** the feature is off.
- **GIVEN** a release toggle is off, **WHEN** an administrator saves an on
override and restarts Kandev, **THEN** the feature is enabled only on that
installation and the page reports `Saved override` as its source.
- **GIVEN** a release toggle is off, **WHEN** a direct client sends an
enabled-only HTTP, WebSocket, MCP, or agent-tool request, **THEN** the backend
rejects it without persisting or dispatching enabled-only work.
- **GIVEN** a release toggle is off, **WHEN** a user opens the application,
**THEN** the new UI is absent and the pre-feature workflow remains usable on
desktop and mobile.
- **GIVEN** `KANDEV_FEATURES_<NAME>=true`, **WHEN** the Feature Toggles page
opens, **THEN** the toggle is on, environment-locked, and cannot be changed
from the UI.
- **GIVEN** the backend cannot provide feature data during SSR, **WHEN** the SPA
renders, **THEN** every unknown or unavailable feature is treated as off.
- **GIVEN** a release toggle has baked on selected installations, **WHEN** its
production default changes to on, **THEN** operators can still override it
off for that release.
- **GIVEN** one default-on release has succeeded, **WHEN** the toggle graduates,
**THEN** the legacy behavior and live flag declarations/gates are removed,
the identity is retired, and the new behavior remains enabled.
- **GIVEN** a removed key still has a persisted override, **WHEN** Kandev boots,
**THEN** the row is ignored and no unrelated feature adopts its value.
- **GIVEN** a phone viewport, **WHEN** an administrator opens Feature Toggles and
changes an override, **THEN** the same save, reset, lock, risk, and restart
information is reachable without horizontal scrolling.
## Success criteria
- A user can enable or disable Office mode from Feature Toggles and understand
that restart is required.
- A user can enable or hide the App status bar from Feature Toggles and
understand that it applies after restart on every breakpoint.
- Office mode clearly communicates experimental status and risk.
- A user can enable or disable Debug mode without seeing duplicate agent-message
debug toggles.
- Env-controlled flags are accurately shown as locked.
- Restart-required changes have a clear supported or manual restart path.
- Existing SSR feature gating keeps working through `/api/v1/features`.
- Desktop and mobile tests cover the core workflow.
- A risky feature can merge to `main` without changing default production,
development, or E2E behavior.
- An operator can enable that feature on one installation and turn it off again.
- Direct and stale clients cannot bypass a disabled flag.
- Adding or removing a frontend-visible toggle changes one backend definition,
one typed config field, one profile entry, and at most one frontend feature
declaration before its actual gates and tests.
- CI detects incomplete or malformed registry/config/profile/frontend
contracts in either direction.
- A graduated toggle leaves no live legacy branch or reusable stale key or
environment variable.
## Out of scope
- Live-applying Office mode or Debug mode without restart.
- Per-user, workspace-scoped, or percentage rollout flags.
- Exposing mock providers, e2e tuning, or arbitrary env vars as user toggles.
- A third-party feature flag service.
- Multi-user admin enforcement in v1.
- Per-user, workspace-scoped, percentage, or cohort rollouts.
- Live-applying runtime toggles without restart.
- A third-party feature-flag service.
- Protecting incompatible migrations, unsafe shared initialization, or
unconditionally executed refactors with a late behavior check.
- Exposing mocks, E2E tuning, or arbitrary environment variables as user
settings.