Field round 11 (see BACKLOG.md) opened with a finding that was not
about a tool at all: a reporter drove a device for an hour, worked several things
out the hard way, and read zero android://guide/* resources until a human
told them to. Resources are a pull channel and agents do not pull. So this
release moves the guidance that is expensive to miss into channels that land on
their own, and fixes the three tool gaps that were still producing confident
wrong answers.
Added: the server now returns MCP instructions. They arrive on
initialize and land in the model's system prompt before any tool is called —
the only channel here that does not depend on the agent choosing to read
something. Deliberately about fifteen lines, because they sit in every session's
context forever: adb_reverse before launching a dev client, app_state before
believing anything, absence-is-not-evidence, viewport-scoped reads,
run_sequence for timing, and the guide URIs. Everything else stays in tool
descriptions (read at the moment of use) and error messages (read at the moment
of being wrong).
Fixed: app_state could not tell the RIGHT Metro from a WRONG one. A dev
server left running from another checkout still holds port 8081, so a dev client
connects to it and every signal reads healthy — bundle_source: metro, an
established socket, no warning anywhere — while the JavaScript executing belongs
to a different branch. That is worse than the embedded-bundle case, which at
least looks wrong the moment you check it. app_state now identifies which
dev server, resolved host-side: port, URL, the listening pid and command, and
the project root that process is running in. The live-socket probe also runs
whenever the process could still be on a dev server rather than only when logcat
came back inconclusive, since it is the only thing that reports which port.
Where the host process cannot be identified (no lsof, or a server on another
machine) the response says so instead of implying certainty.
Fixed: source_path staleness detection could silently produce nothing.
Passed in exactly its documented scenario, it returned source_mtime and then
no marker, no comparison and no verdict, while the running JavaScript was
several commits behind. There is now a stale_verdict (stale / current /
undetermined) with a stale_reason on every path that takes a
source_path — not installed, not running, embedded bundle, wrong-checkout
Metro, no HMR marker logged yet, or a real timestamp comparison. undetermined
names why and what to do about it. Silently skipping a check a caller asked for
is the failure mode the whole field-feedback room exists to catch.
Fixed: text matching missed on typographic punctuation.
tap_on_text("Don't allow") found nothing on the runtime permission dialog,
whose label is Don’t allow with U+2019, and the error read as "the element is
absent". Curly quotes, en/em dashes, the ellipsis, the non-breaking-space family
and zero-width characters are now folded to ASCII on both sides of every
text match (tap_on_text, tap_element, wait_for_text, enter_pin,
describe_ui {query}), so the fold can never make a match narrower than a
literal comparison would.
Added: render_stats (tool 79). describe_ui is viewport-scoped at every
filter, including all — Android omits a ScrollView's off-screen children
from the accessibility tree — which makes it the right instrument for "is X on
screen?" and the wrong one for "how many views did this list actually mount?".
render_stats reads dumpsys gfxinfo: total views, render-node kB, a
per-ViewRootImpl breakdown, jank percentage and frame-time percentiles, plus
the Choreographer skipped-frame tally scoped to the app's own pid. Those
numbers together are what argues an unvirtualised list, and they were three
separate shell-outs before. reset: true zeroes the frame counters after
reading so the next call measures one interaction. describe_ui now says out
loud that it stays viewport-scoped, and points here for counting.
Added: android://guide/rn-expo (guide 6). RN/Expo has produced the most
expensive papercuts in this project's history and that knowledge was spread
across two bullets in the driving guide and several tool descriptions. The new
guide leads with the ordered recipe, because the ordering is the part people get
wrong, and collects the traps that had no home: a stray Metro from another
checkout holding the port; expo run:android skipping prebuild when android/
already exists, so config-plugin manifest changes never apply and a runtime
permission request fails for no visible reason; force-stop plus monkey
landing on the Dev Launcher list; a non-default port having to move in two
places at once. Both plugins gained the matching thin skill.
Changed: a failed tap_on_text/tap_element now teaches. An agent reads an
error at the exact moment it is wrong, with full attention and a concrete goal —
no other channel has that property. The miss now names the focused window, lists
what IS on screen closest-first, and states the viewport rule, instead of "no
element matching X", which reads as "the element is absent" and sends people
debugging the app rather than their selector. doctor closes with an rn-expo
pointer and list_devices carries one, since both are near-universal session
openers; wait and press_key now name run_sequence, which the reporter only
discovered after the session despite it fitting their exact problem.
Changed: two more versions stopped being typed by hand. scripts/docsync
now writes VERSION into both plugin manifests, which nothing in the release
gate looks at and which would otherwise have stayed on whatever release first
wrote them. It also fails when a guide resource has no plugin skill wrapping it
— the drift that nearly shipped six guides with five skills.
Verified live on emulator-5556 (API 37) over real stdio JSON-RPC:
instructions returned from initialize; render_stats against two packages,
including the reset path; app_state returning a stale_verdict where it
previously returned silence; an ASCII-hyphen "Mobile, Wi-Fi, hotspot"
exact-matching and tapping Settings' U+2011 Wi‑Fi label; the new miss error;
resources/read on the new guide; and the host-side dev-server resolution
against a real listener on 8081. Not verified: the full device-side Metro
identification, which needs an installed Expo dev build — no third-party package
was available on the test emulator. Its parsers are unit-tested and the host
half was exercised live; flagged in BACKLOG.md as unverified rather than done.
Full Changelog: v0.22.2...v0.23.0