[samuv]

Note

Update (2026-10-06, while Samu is on leave): I merged origin/main into this branch instead of rebasing, so the original history is preserved. There was one conflict: docs/acceptance/studio-chat-approvals-review-escape.md stays deleted, because #2105 removed acceptance plans, ADRs, and the Plan / Interface process. References below to "Stage", #1888, and ADR 0366 are historical. The merged branch also picks up #2133 (Studio on Node 26, image smoke test, and Grype in the studio CI job). Compared with main, the PR now changes apps/, the studio CI job (Playwright moves after the integration step), and the Studio user docs.

Verified locally before pushing (Node 26 and pnpm 12.4.2 for Studio):

• Studio gates: lint, typecheck, generated-check, and audit all pass. Tests: server 173, web 593, and integration 8. Playwright with mecated: 46.
• Image: it builds, the smoke test passes, and Grype finds no fixable critical vulnerabilities.
• Other checks: task docs, task site:build, the SDK suites, and actionlint with shellcheck all pass.

Closes #1848
Relates to #1776

Stage and approved contract

Stage: Implementation. This completes the Studio approval, plan review, and Escape behavior in the merged Plan / Interface PR #1888, approved at merge commit 98dfad2e1bed66809f68b104c4f207df2d2b1dda. The exact daemon and SDK controls from merged #1909 are already on main. The browser uses the published @stacklok-oss/mecatl-sdk@0.4.0 (staged integrity sha512-Ii03oT378bVhwntthEcteyKgoM8QCYL5XBprpEX/+GjNgBlTxCkU8byG54c7JkAOYeAoRl8dlKpQ4K7tP2VT9w==) with released mecated v0.0.41.

Interfaces match approved contract: Yes. No implementation deviation in this completion slice required another plan or ADR amendment. The earlier directing-human amendments to ADR 0366, protobuf fields, call correlation, and lease-loss handling were incorporated into #1888 and #1909 before their merges.

Behavior

• Ordinary permission asks appear beside a tool row only when the event's non-empty callId matches a row in the same run. Unmatched child and older-server asks remain separate inline cards. Cards show scrubbed arguments and a bounded Edit/Write diff when parseable; malformed or missing arguments fail closed. Allow once, Always allow, and Deny retain the event's immutable session/run/ask tuple and reconcile uncertain submissions from activity.
• PresentPlan has an inline plan card in main chat and side threads. It shows the plan, optional note, and raw arguments, with Approve & run, Auto-accept edits, and Iterate. The two approving controls require valid plan text. The BFF uses only the published SDK's exact run-and-ask plan control. After allow, Studio follows durable activity and reattaches transcript and controls to the observed execution run; a matching continuation failure or bounded uncertainty is reported honestly.
• Escape consumes one layer per press. A focused dialog or menu closes first and restores its trigger. Subsequent presses handle selection, an explicit ask Deny/Iterate, the side panel, a live run, then a guarded unsent draft. The draft clears only on a released second non-repeat press within 500 ms. Main chat and the independent side-thread chat each own their focused controls.
• Developer steer trace displays only observed IDs, event kind, and outcome metadata. The Studio guide and shortcut reference now describe the browser workflow and hints.

Verification

• task studio:check passed on the assembled candidate, including web/server tests, lint, typechecks, and generated client check.
• The named Studio chat: inline approvals, plan review, and Escape behavior #1848 Playwright journey passed on desktop and mobile. It covers ordinary allow/deny, plan approve/iterate, reload and second-tab stale/duplicate handling, overlay-first Escape, ask denial, and guarded draft clear. Its authority proof obtains real Write and PresentPlan asks from a test-owned mecated --mock-script; accepted browser verdicts and consumed-ask 409s traverse the real Hono BFF and published SDK. No live provider or external network is used. CI now runs the browser step after the daemon build.
• CGO_ENABLED=0 task docs, task site:build, and task ac-trace-strict passed. Strict tracing resolves all 13 Studio chat: inline approvals, plan review, and Escape behavior #1848 ACs. Native local docs/build encountered this macOS host's linker incompatibility; the CGO-free docs/build passed.
• Keyboard, screen-reader, and light/dark responsive inspection passed at 320, 500, and 1280 px. The checks cover visible focus return, hint eligibility, readable card order, and diff addition/removal labels without color alone; screenshots are retained in the run-local inspection artifacts.
• Panel review: 0 ship blockers, 1 important nonblocking suggestion, 0 advisories, 0 reviewer failures. Spec, test adequacy, and architecture findings were repaired and revalidated. The remaining suggestion is to centralize repeated plan/ordinary card presentation across chat surfaces; verdict submission remains surface-specific because their session lifecycles differ.

Per the directing human's request, PR CI is the broad validation gate for repository-wide lint, tests, race checks, build, and the full Studio/browser flow. This PR is for review; it has not been merged or deployed.

[kantord]

Approving. A multi-axis review (spec, standards, test adequacy, security, architecture, CI, reuse) found no blockers. The non-blocking findings and nitpicks are tracked in #2142.

Fully or partially written by an AI agent.