obsidian-ios-real-device-smoke

Files (2)

• SKILL.md
• openai.yaml · agents/

SKILL.md (first lines)

# Obsidian iOS Real-Device Smoke

## Core Rules

- Select this skill under the mobile environment rule in `AGENTS.md` and name
  the user request or verified iOS-specific capability. Device selection does
  not itself authorize deployment or interaction; an Agent-authored plan or a
  simulator/tool failure cannot create that requirement or authorization.
- Do not claim iOS real-device validation for a plugin-asset change unless the current build was written to the iCloud Obsidian `test` vault, all copied assets matched `dist`, the loaded plugin identity matched that artifact, and the affected behavior was observed on the connected iPhone.
- If only a vault-resident smoke runner changed, sync and hash-check that runner, then re-evaluate it in the existing page. Do not redeploy or reload the plugin when `dist` and plugin runtime assets did not change.
- Treat Safari Web Inspector as DOM/CSS/console/network evidence, not touch automation. Use iPhone Mirroring or the user for real touch interaction.
- Discover and use the available native UI tools for Safari and iPhone Mirroring;
  do not require or install a legacy skill by name. If no tool can operate the
  real target, report the blocked evidence instead of substituting emulation. For
  Safari Console JavaScript, use a real paste/input event and verify the command
  entered Console history or produced its sentinel before counting execution.
  Accessibility `set_value` may change the displayed editor text without creating
  executable Web Inspector input, so it is never submission evidence by itself.
- Treat an explicit user request to run iOS real-device smoke as authorization for plugin-asset deployment through `make deploy-icloud` (or its current-build reuse target), or the existing scoped preparation/sync command for a runner-only change, targeting only the iCloud `test` vault. Do not ask again. If the user requested only planning, review, or inspection without a real-device run, do not deploy or sync.
- Choose the smallest lane that covers the latest runtime delta and stop once
  that lane has sufficient current-build evidence.
- Judge a targeted canary against its current Decision/Spec/SDD contract, not a
  stricter topology borrowed from a deterministic Desktop fixture. In
  particular, a live Recovery canary must accept an intended reranker fail-open
  followed by `recovery_relaxed / skipped / not_eligible` when the standard
  attempt already returned usable evidence; require an actual relaxed retry only
  when the canary explicitly owns a deterministic valid-none/partial fixture.
  Likewise, lexical `unavailable / feature_disabled / generation absent` to
  `ready / generation present` with an advanced maintenance epoch is a valid
  first-generation transition. Preserve a raw runner false-negative, classify
  its product observations independently, and do not spend another device
  attempt solely to satisfy the invalid assertion.
- Before spending a submitted Console execution on a cancellation-only runner,
  require its post-load Memory plan to be `ready / none`. A fresh plugin reload
  with persisted rollout flags off may expose the existing lexical generation as
  `feature_disabled`; temporarily turning flags on then correctly requires
  preparation and is not a runnable cancellation precondition. Do not convert
  that setup-only `BLOCKED` into a product failure or repeat the command.
- Give every submitted attempt a unique result filename and copy/hash its raw
  receipt before the next attempt. Never let a retry overwrite the only device
  evidence. If historical tooling already overwrote a receipt, disclose that
  limitation and distinguish any retained projection from the unavailable raw
  file; do not describe the projection as immutable full evidence.
- Apply slice-scoped invalidation after a diagnostic-only repair. When an immutable
  real-device receipt already proves cancel request/observation, late discard, zero
  accepted-after-cancel, stable index identity and queue release, a successor that
  changes only terminal classification may close through exact new-artifact load
  identity plus a regression using the real wrapped `AbortError` shape. Do not
  rebuild Memory or spend another Provider/device attempt solely to replay the
  unchanged Worker/queue behavior.

## Change Lanes

| Lane | Use when | Required iPhone evidence | Skip |
| --- | --- | --- | --- |
| `geometry-only` | CSS/layout/position/size changed; gesture and action routing did not | Asset match, one reload, one real gesture that exposes the UI, one Mirroring screenshot, one Inspector geometry/hit-test probe, fresh errors | Hold threshold, movement cancel, exactly-once action checks, provider flows, unrelated orientations |
| `interaction` | Touch, pointer, keyboard, focus, dismissal, or action routing changed | Geometry basics plus the affected real gestures, cancel path, and exactly-once behavior | Unrelated Pagelet/provider/cross-surface flows |
| `broad/release` | Shared mobile runtime, multiple surfaces, packaging, or release gate | The declared cross-surface and orientation matrix | Anything explicitly outside the release scope |

Do not promote a geometry-only delta into an interaction or release run merely
because more evidence is available. Record unrequested orientations or devices
as `NOT TESTED` residuals.

## Device Preflight And Retry Budget

Before deployment or GUI control, confirm:

- the user requested real-device smoke;
- the local prerequisite covers the same runtime source state;
- iPhone Mirroring is connected and the device/Mac are unlocked;
- Obsidian is on the iCloud `test` vault and intended note;
- Safari Develop exposes the current Obsidian target.

Allow one unlock/reconnect/reattach sequence, or about three minutes, when a
device or Inspector prerequisite is missing. Then report `BLOCKED`; do not keep
hunting through menus or coordinates. After a reload, allow at most one
Inspector reconnect. If `body.in-progress` persists, preserve the manual path
and stop.

Use one submitted Console probe that returns all required assertions. Allow at
most two execution attempts before reporting partial evidence or `BLOCKED`;
submitted-but-ambiguous commands count as attempts. Do not spend additional
attempts recovering focus, autocomplete, or stale element indices.
Re-read app state after state-changing UI actions or window/DOM changes, and
never reuse stale element indices; batch stable prompt entry, execution, and
result capture where possible.

## Workflow

Prerequisite: complete `obsidian-test-vault-smoke` at least at `app-runtime` tier.

1. Bound the changed surface:

```bash
git status --short --branch
git diff --stat
git diff --name-only
```

2. Reuse automated local checks with verified inputs/results under `AGENTS.md`
   **Validation Planning And Reuse**. Complete missing or invalidated **Local
   Validation Gate** checks, counting checks covered by the selected enclosing
   command and supplementing the runtime `<style>` / `innerHTML` / `outerHTML`
   source scan. Treat `rg` exit code `1` with no output as PASS and inspect every
   match. Reuse observed local app smoke only within the same turn while its
   target/runtime state is unchanged; otherwise renew the affected local app
   proof. For broad UI/runtime changes, retain `make deploy` (or its eligible
   current-build target) and local app smoke before device validation.

3. Sync only the changed artifact after the user requested real-device smoke:

- Runner-only change: use that runner's preparation/sync workflow, compare its
  destination hash with the source, and keep the current plugin instance and
  Inspector session. Skip plugin deployment and reload.
- Plugin runtime or CSS change: deploy the current plugin assets:

```bash
make deploy-icloud
```

From edonyzpc/personal-assistant, AGPL-3.0.

More from edonyzpc/personal-assistant