DevFlow: use canonical WebView context IDs across tooling (Unit 3)

[mattleibow]

Summary

Make WebView targeting consistent and unambiguous across HTTP, Client, CLI, MCP and Inspector.

• Give every registration a unique canonical webview-<index> ID, including unnamed and duplicate-AutomationId hosts. Index, AutomationId, elementId and descriptive hostKind remain metadata, not alternate selectors.
• Use contextId throughout the protocol and tooling, with CLI --context-id. Explicit blank, invalid or removed selectors fail without selecting another document. Omission/JSON-null retains the active-host default; query contextId takes precedence over typed-body contextId.
• Emit only ready for readiness. Preserve internal readiness checks and native-owner correlation.
• Reject malformed/non-object Inspector JSON, duplicate contextId and removed webviewId fields before dispatch. Initially select an active-ready host and freeze the selected target before JavaScript confirmation.
• Update human-readable context listing/status, consumer tests, protocol schemas/examples and mirrored guidance.

Compatibility

Obtain IDs from /api/v1/webview/contexts. Numeric indices, AutomationIds and element IDs no longer select contexts. HTTP webview, Inspector/MCP webviewId, CLI --webview/-w and the duplicate isReady field are removed; C# named arguments use contextId.

Update Agent, CLI, Client, WebView and Blazor packages together and rebuild consumers. No selector/readiness compatibility shims remain. Portable Client support, Driver namespaces/type forwards and descriptive hostKind extensibility are preserved.

Validation

Contract tests cover valid/default/null targeting, query/body precedence, duplicate IDs, removed/blank values and zero dispatch for invalid selections. Inspector proxy tests reject 40 invalid requests without downstream calls and dispatch six valid/default/null requests exactly once. Client, CLI/MCP and protocol regressions pass; netstandard2.0 builds.

Actual Chromium Inspector tests verify initial selection and a target frozen through confirmation. Agent/shared-engine/CLI/MCP document-isolation proofs exercise one click/fill event sequence, useful errors, native PNG targeting and failure exits. Mac Catalyst target isolation and original Hybrid/native callbacks are also exercised; no Windows/GTK/WPF or physical-device runtime claim.

Related: #654, #621, #232, #407. No timing, screenshot-flake or native-crash resolution is claimed.

[github-actions]

✅ Expert Code Review: 2 independent reviewers found no issues. Methodology: 2-model adversarial consensus. (1 reviewer failed to complete — reduced coverage, no tiebreaker)

Generated by Expert Code Review (auto) for #662 · copilot · gpt56 · 322.4 AIC · ⌖ 21.4 AIC · ⊞ 37.6K · ◷

[mattleibow]

Current consolidation position: this draft canonical-context layer is directly above the prepared #661, at head 99c89ad77813c93530b6de82dc5d8c2d875b8528, in the unified native stack. Earlier base hashes and original implementation counts in the initial description are historical.

The canonical-only slice maps 31 original files to 35 because regressions were composed into the smaller bug fixes existing test files. Direct single-evaluation CLI JavaScript and CDP Input.insertText remain intact rather than restoring the old bundled typed-CLI migration. Useful failures, native PNGs and aggregate batch exits are preserved. Invalid/removed/blank targets still cannot silently select a default, and Inspector confirmation freezes its selected target. Layout remains the next #664 layer. Every PR remains draft; the complete-graph evidence and retained limits are recorded on #665.