Steward runtime text-input capability packet - 2026-06-10

Status: adoption run passed; capability candidate pending held-out repeat · Audience: maintainers and agents

Claim to prove

A fresh agent can use Flutter MCP Toolkit against the running flutter_test_app to discover the greeting field, enter text through the interaction tool, and prove runtime state changed.

This is the next harder Skill Steward adoption slice after the hosted dependency cutover gate because it exercises a live MCP/runtime path:

launch app -> discover VM service -> semantic_snapshot -> enter_text -> evaluate_dart_expression

It does not claim all interaction tools, fill_form, web/mobile parity, production app behavior, WebMCP dogfood, or broad repo H5 maturity.

Result - 2026-06-10

The adoption slice passed after bounded repairs to the toolkit runtime path. The proof selected the text field by semantic_snapshot identifier, called enter_text, and verified the live app state through evaluate_dart_expression.

This is not a full H5 proof. It is one successful adoption run plus a capability candidate because the run happened on a dirty local tree after the required repairs were applied.

Evidence artifact:

• docs/evidence/generated/mcp_flutter.runtime-enter-text-greeting.redacted.json

Reproduction entrypoint:

WS_URI='ws://127.0.0.1:<port>/<token>/ws' \
  tool/evals/run_runtime_enter_text_greeting.sh

Runtime target:

• Platform: macOS debug app launched by make showcase.
• VM service URI provenance: local ephemeral URI emitted by flutter run (ws://127.0.0.1:<port>/<token>/ws, token redacted in evidence).
• Subject: commit 87c493c8fb880c58572aa7b59d74e368a68b5144 plus local adoption-run repairs in this change set.

Acceptance facts:

• doctor --json: 10 checks, 10 pass, 0 fail.
• First semantic_snapshot: exposed Semantics identifiers, proving the agent can select by greeting_input_field rather than by first ref.
• scroll(direction: "down", distance: 900): moved scroll position from 0.0 to 480.0, revealing the Type section and field ref s_14.
• Before enter_text: AgentState.instance.greeting == "".
• enter_text(ref: "s_14", snapshotId: 2, text: "steward runtime proof"): returned success: true, via: "editable_state".
• After enter_text: AgentState.instance.greeting == "steward runtime proof".

Caveats found during the run:

• The app service-extension transport included VM metadata key isolateId; the toolkit now strips that before app-side schema validation.
• semantic_snapshot previously omitted SemanticsNode.identifier; the snapshot now exposes identifiers so adoption proof does not depend on brittle ref order.
• Scroll could previously report success while leaving scroll position unchanged. The scroll service now records scrollBefore/scrollAfter and returns no_scroll_movement when it cannot prove movement.
• An exploratory wait_for probe for the echo text exceeded its own timeout and was killed. This is not part of the acceptance check, but it is a follow-up caveat for the interaction harness.

Local repeat - 2026-06-13

The runtime path passed again in the current checkout:

make showcase
WS_URI='ws://127.0.0.1:<port>/<token>/ws' \
  tool/evals/run_runtime_enter_text_greeting.sh

Evidence artifact:

• docs/evidence/generated/mcp_flutter.runtime-enter-text-greeting.redacted.json

Acceptance facts from the 2026-06-13 artifact:

• doctor --json: 10 checks, 10 pass, 0 fail.
• AgentState.instance.greeting before entry was not the target text.
• enter_text submitted steward runtime proof.
• AgentState.instance.greeting after entry returned exactly steward runtime proof.

This repeat is useful freshness evidence for the local runtime path, but it is not the clean held-out proof required for H5 promotion. The checkout still had uncommitted adoption-run repairs and this proof script itself was part of the local change set.

Held-out repeat attempt - 2026-06-13

A second-agent held-out attempt prepared the same proof path but did not reach VM service startup: make showcase failed before the runtime eval could run because macOS denied loading FlutterMacOS.framework during code-signing/system-policy validation. Treat this as blocked input for the next promotion attempt, not as a failed product interaction proof.

Until that launch blocker is cleared in a clean owner checkout, the ceiling remains: helper implemented, focused and local runtime proof passed, held-out repeat pending.

Adoption-run/v2 classification

Target surface

Known target:

• flutter_test_app/lib/showcase_screen.dart wraps the field with semantics identifier greeting_input_field.
• semantic_snapshot should return a ref for that field.
• enter_text should submit steward runtime proof.
• evaluate_dart_expression("AgentState.instance.greeting") should return the submitted text.
• flutter_test_app/lib/agent_state.dart owns the assertion state.

Use enter_text for this one-field proof. Reserve fill_form for a later batch-form capability proof.

Acceptance sequence

Run from this checkout:

cd flutter_test_app
flutter run --debug --machine --host-vmservice-port=8181 -d macos

Use the emitted app.debugPort.wsUri as $WS.

flutter-mcp-toolkit doctor --vm-service-uri "$WS" --json
flutter-mcp-toolkit exec --name semantic_snapshot \
  --args "{\"connection\":{\"targetId\":\"$WS\"}}"

Find the node with semantics identifier greeting_input_field, capture its ref and snapshotId, then run:

flutter-mcp-toolkit exec --name enter_text \
  --args "{\"ref\":\"<ref>\",\"snapshotId\":<snapshotId>,\"text\":\"steward runtime proof\",\"connection\":{\"targetId\":\"$WS\"}}"

flutter-mcp-toolkit exec --name evaluate_dart_expression \
  --args "{\"expression\":\"AgentState.instance.greeting\",\"connection\":{\"targetId\":\"$WS\"}}"

The proof passes only when the final expression returns exactly steward runtime proof.

Detour budget

• If the field ref is stale, call semantic_snapshot again and retry once.
• If app launch or VM service connection fails after two setup attempts, stop restoration and record an unknown case.
• Do not add a new MCP tool in this adoption run unless the existing toolkit path is proven unable to satisfy the acceptance check.
• If direct Dart expression can set the state but enter_text fails, record fallback evidence only; that does not prove text-input interaction.

Promotion requirements

Promote to capability-level H5 only after:

• A clean runtime run proves doctor, semantic_snapshot, enter_text, and evaluate_dart_expression.
• Evidence records platform, VM service URI provenance, subject commit, and dirty state.
• The proof selects the field by semantics identifier, not by whichever ref appears first.
• A falsifier exists, such as asserting the expression is not the target value before entry or proving a wrong ref does not satisfy the acceptance check.
• A future-agent or held-out run repeats the path without hidden local context.

Current interpretation:

• The first four bullets are satisfied for this bounded capability.
• The held-out future-agent repeat remains open, so this is not a full capability-level H5 proof and must not be used as a repo-wide adoption claim.