[barelyworkingcode]

Closes #624

PR 2 of 2 for #624: an "Unsandboxed" chip on every agent row that relay marks unsandboxed: true, on sessions and their drop-in terminals. Targets epic/623-projectless-adopt.

Non-obvious decisions

• The rule lives in AgentBoard.isUnsandboxed. The tab bar and the drop-in marking call it too.
• The chip is role="img" with aria-label="Unsandboxed", and its id joins the row's aria-describedby, so a screen reader hears it on the row (contract amendment).
• Fixed before merge: Drop in on a session adopted in no project opened a tab that the tab bar filtered out, because the terminal has no project. TabManager.render() now also shows a project-less terminal that relay marks unsandboxed. Relay's drop-in answer carries no mark, so AgentBoard.dropIn marks the opened terminal from its session until the next terminal_list (contract amendment for public/tab-manager.js).
• Today no longer lists agents, so criterion 1 names the Agents rail and the phone Agents dialog (amendment).
• The Pi/Codex "Not in a project" line has no hermetic spec: fakerelay worlds hold only Claude rows (fakerelay: outside rows for Pi and Codex agents relay#591). Its proof is relay's TestAdoptNoProjectPiCodexKeepReason, the unchanged ADOPT_REASONS.no_project, and the Verify and close the projectless-adopt epic (relay) relay#587 devbox pass (amendment).

Minor review notes (not fixed)

• The journey counts two chips page-wide during the drop-in; the spec places each chip on its row.
• Long lines in devboxverify/outside-claude.js:69 and the README paragraph.
• Theoretical: during a drop-in the terminal chip may show from eve's own mark before relay's terminal_list arrives.

Proof

• npx playwright test on 0d92544: 843 passed, 33 skipped, 0 failed (4.5m). An earlier run had one failure, G6.37, which fails under CPU load on the base too: Flake: G6.37 jump to the agents that need me loses focus under CPU load #632 (flake).
• --stack on 0d92544 (stack e624): SUMMARY pass=48 fail=0 blocked=1 notrun=10; the one BLOCKED is outside-adopt-codex (no codex daemon on this VM), the same under main's harness.

Fix attempts: one review round (two blockers, eight minors), the drop-in defect (two fixes), one journey audit-read fix. Tokens: about 1.0M across subagents.

Review: no blocking findings (round 2, default model)
Journey: outside-adopt-unsandboxed PASS on stack e624, head 0d92544 (log /tmp/c585/e624-dbv-head.log)
Harness weakening: none; outside-adopt and the other journeys are unchanged, the new journey is additive
Main harness vs head: main's devboxverify (952d1f0) against head 0d92544 on stack e624m: SUMMARY pass=47 fail=0 blocked=1 notrun=10 (blocked: no codex daemon)
Break-and-restore: hiding the chip in AgentBoard turned all five @G6.109 specs red, and on a stack outside-adopt-unsandboxed FAIL "the managed row shows no Unsandboxed chip after the adopt"; green after the restore

🤖 Generated with Claude Code

[barelyworkingcode]

merge-check: merged. approved, target, contract, evidence, owner
Report: 2. The gates themselves: criterion removed A managed session row on the Agents rail, the phone Agents dialog and Today shows an "Unsandboxed" chip when relay lists it with unsandboxed: true. (issue #624)