This file records the checks used for the initial agentview prototype. They are intentionally split between deterministic tests and disposable runtime probes. Existing user containers and live agent sessions were never used as test targets for lifecycle operations.

For the reusable Linux, macOS, and Windows acceptance matrix, native-versus- simulated test boundaries, clean public-install commands, and release artifact provenance checks, use the cross-platform testing guide. This file is the dated evidence log; that guide is the procedure.

The reproducible commands, synthetic populated fixture, exhaustive key route, visual acceptance criteria, and evidence template live in the real-TTY validation guide. This file records completed checks; the guide also contains release gates that are not yet complete.

• cargo test --locked: domain normalization, provider parsing, App Server JSONL and Unix-WebSocket transport behavior, navigation, responsive rendering, process timeouts, ownership persistence, Claude launch parsing, VT100 log reconstruction, and doctor output.
• A disposable mock App Server test launches an owned Codex thread, drops the first supervisor, reconnects from a second supervisor through the same Unix socket, discovers the still-active thread, interrupts its exact turn, and rejects an external thread ID. The same test covers bounded transcript read, exact active-turn steer, idle reply, archive, and delete. It also asserts mode 0700 for the state directory and 0600 for the ownership record. The mock PID is re-verified, terminated, and reaped by a panic-safe test guard.
• Codex reducer tests require an exact owned turn/completed thread/turn pair before releasing active-turn authority, reject external deletion events, and prove a stale idle snapshot remains interruptible. Shutdown tests hold a stable Linux pidfd, reject a second controller and tampered command line, and remove only the exact stopped server record. Process transports run in a new process group so npm wrappers and their native App Server children are both bounded during teardown.
• A disposable mock Pi RPC executable covers managed launch, live discovery, transcript inspection, active steer, confirmation, structured text input, nonblocking exact stop, stopped-session delete, completed native handoff, exact unowned-ID refusal, modeled launch, and shutdown. A second dashboard client reconnects through the same verified daemon while its Pi child remains live. Compatibility tests prove an old daemon advertises no model feature, refuses replacement while it owns active work, and can be safely replaced only after exact owned sessions are completed. Panic-safe cleanup stops the exact test daemon; separate tests reject symlinked state and permissive/replaced authority records.
• A disposable authenticated OpenCode loopback fixture covers durable managed launch, dashboard reconnect, discovery, transcript inspection, active/idle reply, interrupt, exact unowned-ID refusal, and panic-safe pidfd shutdown. It verifies the private state modes plus exact Linux process and listener identity. A separate credential-empty real OpenCode server probe covers create, accepted async prompt, list, inspect, and exact shutdown without claiming a model-backed turn.
• Disposable provider fixtures cover the additional host adapters without touching user credentials or session stores. Cursor tests own a temporary executable, workspace, registry, logs, and child process while exercising launch/discovery/inspect/interrupt/reply and stale-identity refusal. Copilot tests own an ACP subprocess and exercise list pagination plus the retained new/prompt/update/permission/completion/reply/cancel lifecycle; a separate load fixture proves external list results remain unowned. Copilot restart coverage also proves private exact-ID persistence, split user/assistant history replay, real latest-message summaries, provider timestamps, and visibility when the second process cannot reach ACP. Antigravity tests read only a temporary documented cache and build shell-free native commands.
• Mistral Vibe and Qwen Code have public-controller integration tests backed by isolated executable app-server/CLI fixtures. They exercise model selection, native launch, exact ownership persistence, owned-only discovery, native resume, and unowned interrupt refusal through the same public traits used by the dashboard. An outer real PTY additionally drives foreground launch, native background, exact reattach, verified interrupt, and fresh exact resume for both providers. Vibe also covers bounded delayed correlation and ambiguity refusal; Qwen covers immediate launch-failure rollback. Qwen adapter tests also cover fractional millisecond JSONL timestamps and ownership persisted through an independently loaded private-registry handle.
• tests/managed_cursor_copilot.rs drives the public ProviderController surface rather than private supervisor helpers. It covers managed launch, enrich, inspect, reply, interrupt, exact one-shot approval, duplicate permission cancellation, and controller-side refusal of injected unowned records. Copilot runs on Unix CI; the Cursor process lifecycle is Linux-only.
• Protocol tests interleave client responses with string- and numeric-ID server requests and assert the exact {id,result} response shape. Reconnect tests verify that thread/resume replays an unresolved approval with the same ID, a decision remains resolving until serverRequest/resolved, and a second dashboard cannot acquire the process-held controller lease.
• Model-catalog tests cover Claude's installed-help alias parser, cursor-paged visible Codex model/list, Pi's bounded offline table, OpenCode's exact provider/model output, Cursor's terminal-progress output, Copilot's short-lived headless models.list, and Antigravity's model table. Managed Pi, OpenCode, Cursor, Copilot, and Antigravity assertions prove the selected identifier reaches the provider-native launch surface; Shift+Tab preserves the composer draft, catalog workers ignore stale-provider results, and picker search stays separate from task input.
• Request-reducer/UI tests reject unowned or wrong-turn requests, incomplete command context, duplicate/malformed structured questions, stale deadlines, and blind file acceptance. Sequential input tests verify exact option normalization and confirm that answers never enter the supervisor record.
• Reference-fidelity tests cover initial row focus, cyclic header/row navigation, direct escape-to-quit, printable-to-compose behavior, context-sensitive ?, ctrl+f filtering, task slash commands, the visible harness palette, searchable/paged model picker, live completed toggle, direct idle local removal plus active-hide confirmation, direct Enter/Right open, Left-from-Peek return, Claude attach-return guidance, selection reconciliation after filtering, Claude worktree grouping, aggregate review/working counts, capability-aware help, and the narrow footer's retained help affordance. Focused rendering tests sanitize terminal-control characters from every provider-derived or dynamic surface and use terminal display width for CJK/grapheme-aware row truncation, padding, and editable cursor placement. The composer render test asserts the cursor immediately follows its typed cells with no phantom left-border column. Editable fields also cover Unicode-safe Option+Backspace/Ctrl+W word deletion and Cmd+Backspace/Ctrl+U line deletion.
• Safety-focused state tests verify that ready-for-review and needs-input sessions are treated as live (and therefore require interrupt authority), Ctrl+X emits exact stop for an active owned row and exact delete only after the refreshed row is idle, completed sessions use delete language, active groups cannot enter a bulk-delete confirmation path, and observe-only rows use reversible local hiding without mixing it with provider deletion.
• Private hidden-session registry tests cover idempotent hide/unhide, atomic 0600 persistence in a 0700 directory, symlink/wrong-mode/control-character refusal, and stable-order filtering of a 70,000-session snapshot. A separate 70,000-row application regression proves 2,000 navigation/draw queries reuse one grouping cache and only a filter change rebuilds it.
• Terminal-loop tests prove model loads and completed-visibility changes return nonblocking effects, a deliberately slow managed launch does not block typed input, and post-launch selection requires both the exact provider and exact provider session ID.
• Session-alias tests prove local precedence over a changed provider title, empty/reset behavior, cross-process reload, CLI/JSON round trips, 0600/0700 persistence, and refusal of control characters, oversized values, symlinks, wrong owners, and permissive files. The real PTY renames, filters by the new name, resets it, and observes the fixture's canonical provider name again.
• Installer tests execute agentview and av, and prove an unrelated pre-existing av is retained.
• cargo build --release --locked: release-mode compilation against the committed lock file and Rust 1.75 minimum-version dependency set.
• Release validation locally enforces Rust 1.75 rustfmt, warning-free Clippy, all targets, real PTYs, release mode, and the installer. The checked-in workflow encodes the wider native-platform contract but was unavailable for v0.1.33, whose Linux x86-64 artifact was built and verified manually.
• scripts/real-tui-tests.sh runs the serialized real-terminal suite against real Unix PTYs with isolated HOME/XDG_STATE_HOME. At 120×34, 105×30, and 100×28 they exercise populated sections, contextual help, grouping toggle, ctrl+f filter apply/cancel/clear, slash commands, the harness palette, multiline new-task launch/cancellation, peek, rename cancellation/submission, native-open suspend/restore, reply, direct selected-row interrupt/delete, approval y/n, bulk delete and archive confirmation, structured input, and fixture-fenced refusals. A 90×24 case sends real arrow sequences and collapses/expands a group. The 55×18 and 31×7 cases verify the bounded narrow layout and explicit too-small fallback. Every case asserts alternate screen and cursor restoration from the raw PTY stream.
• Dedicated provider-onboarding PTYs begin with no credentials. Cursor and Copilot show an account-catalog authentication error, hand Enter/l to a native login, reload exact model IDs, preserve the task draft, and select an exact model. Cursor then deliberately delays create-chat; the dashboard animates launch progress and the resulting managed command contains the exact model. Separate PTYs prove Claude's provider-allocated background ID is resolved to the exact UUID, immediately attaches full-screen, and returns to the exact row with the native gesture, while Antigravity performs first-run login, exact model selection, --sandbox full-screen launch, cache ownership, and native return without a dangerous bypass flag.
• Actual-binary PTY tests additionally prove Pi and Copilot native-first launch: each receives one exact agentview-generated UUID, the selected model/prompt, opens full-screen before returning, and restores the exact new managed row after the native background gesture.
• A dedicated real PTY selects the Terminal target, opens its real searchable shell picker, confirms the configured shell, launches a private interactive shell, backgrounds it with the native gesture, discovers and resumes its exact preserved screen, stops it with Ctrl+X, and deletes the completed row with a second Ctrl+X. Provider login PTYs use the same isolated registry.
• tests/setup_installer.rs covers all fifteen coding-harness installers with an isolated PATH and fake curl/bash/npm. For every provider it proves non-TTY setup refuses before download without --yes, confirmed setup uses only the exact official URL or package, creates the configured executable, and emits the next authentication step. A second Linux case hands every already installed provider's exact login arguments to a real PTY.
• scripts/fresh-provider-setup-tests.sh is the networked, explicitly invoked E2E tier. It starts fifteen independent containers from the pinned node:22-bookworm-slim digest with empty homes and no mounted credentials or workspaces, lets the real agentview binary run each current official installer, verifies the installed executable/version, and requires a native PTY login handoff. Browser/device authorization is deliberately bounded rather than completed. On 2026-08-25 this passed for Claude Code 2.1.245, Codex 0.149.1, Pi 0.73.1, OpenCode 1.18.23, Cursor 2026.08.11-e8db854, Copilot 1.0.80, Antigravity 1.1.20, Mistral Vibe 2.24.3, Muse Code 0.2.1 (0.2.1-R1215.1), Qwen Code 0.22.0, and Kimi Code 0.38.0; all disposable containers were removed. Vibe's real passive app-server RPC, Qwen's real JSONL inventory, Muse's credential-free echo provider, and Kimi's unauthenticated provider catalog are also probed in those containers. On 2026-08-27 the same empty-home gate passed for Oh My Pi 18.0.8, Grok 1.0.5 (5115b46bc9), Kilo Code 7.5.5, and OpenHands SDK 1.16.1. No account state or project workspace was mounted into any of those containers.
• tests/self_update.rs runs --version, -v, and -V, then exercises both update and upgrade with isolated fake curl/bash commands. It verifies the exact repository request, install-directory propagation, successful handoff, and cleanup of the downloaded installer without network access.
• A dedicated Linux isolated real-PTY case enables fake Claude and managed Pi launch controllers and exercises the harness palette through the actual binary. It asserts complete choice visibility, arrow/Tab preview, number selection, Enter confirmation, Escape cancellation, model reset on a real switch, and draft preservation throughout; no provider process receives a prompt.
• A second isolated real-PTY case enables only Oh My Pi, Grok, Kilo Code, OpenHands, and Terminal. It verifies that all five appear in the real harness picker, selects each coding harness with a numbered key, opens its asynchronous model picker, checks the exact native model ID, and returns to the unchanged task draft after every selection.
• A provider-enabled PTY regression returns two Claude rows, deliberately stalls the next provider refresh for two seconds, and verifies that both exact selected-row arrow repaints and dashboard exit still finish within 750 milliseconds. It also asserts that an unchanged frame emits no idle terminal output and that cancellation leaves no fake provider child behind. This catches accidental provider I/O or unconditional redraws on the input thread.
• The canonical fixture and a real 150×36 PTY render all fifteen coding providers plus Terminal together, require every provider-specific row label, and retain collision-free normalized IDs.
• A 500-session real-PTY stress case verifies that only the terminal-sized page (capped at 25 rows) is initially present, selects the matching Show-more row with real arrow keys, reveals the second page, then sends 200 down-arrow events in one burst. It requires the exact destination to appear within 750 ms while emitting less than 24 KiB, guarding bounded rendering, input coalescing, and page-aligned scrolling. It also submits /help and a 200-character typing burst, requiring each within 750 ms and bounding terminal output. A separate startup case makes Claude discovery sleep for two seconds and requires an immediately available Antigravity row within 750 ms, proving that the first screen is not gated on the slowest provider.
• An authenticated Antigravity 1.1.20 regression (isolated workspace plus agentview state/cache) verified a 14-model live catalog, exact selected-model launch, foreground response, Shift+Left return, immediate exact managed row, transcript-derived current summary/time, exact Ctrl+X stop, and rediscovery from a fresh agentview process. The CLI reused its existing login internally; the test did not read or copy authentication material.
• A separate default-mode PTY supplies 1,000 completed fixture records without a visibility flag, requires the bounded Completed page within 750 ms, then sends 208 arrow events in one burst and requires the exact destination within 750 ms while emitting less than 24 KiB. It submits /completed hide, verifies immediate removal, then restores the same bounded page with /completed show. A second real-PTY test points OpenCode at a marker-writing, two-second executable, enables /completed show, and proves the external-history command is still never started without --include-external; the adapter test independently retains an unused runner response.

• Both interactive dashboards were exercised through real allocated TTYs in separate fresh --rm --network none containers from immutable image ID sha256:8f170f660813ac358f347fa8a3580139972f3ea7a9fb087834f1da44669d9392: claude agents 2.1.209 rendered its empty Needs input/Working/Completed dashboard, opened and closed shortcut help with ?, and restored terminal modes on Escape; the mounted agentview release rendered its empty dashboard, opened and closed contextual help, switched to directory view with ctrl+s, and restored the alternate screen on Escape. The agentview probe additionally ran unprivileged with a read-only root, dropped capabilities, no-new-privileges, a PID limit, and isolated tmpfs home/state.
• The exact fresh-container command shape is documented in the TUI validation guide. It pins the immutable image, uses a different --rm container name for each dashboard, and mounts neither host credentials nor a workspace. The committed populated-sessions.json fixture supplies nine sessions across every normalized state and actionable capability for repeatable populated-screen testing without provider access. Fixture mode fences every provider operation, including native open and launch, so synthetic capabilities cannot mutate a provider.
• The current release binary and nine-session fixture were then rerun in fresh isolated containers at 120×34, 55×18, and 31×7. The wide run verified populated status rendering, help, directory grouping, explicit provider names, and clean exit; 55×18 verified responsive truncation; 31×7 verified the explicit minimum-size fallback. Claude 2.1.209 was separately rerun empty at the same sizes: wide help/exit, narrow startup/exit, and tiny intro-only degradation/exit. The exhaustive action matrix ran against the actual binary in host openpty; it was not redundantly repeated inside Docker.
• The v0.1.13 candidate repeated that fresh-container comparison after the history changes: the populated agentview binary rendered both provider columns, help, directory grouping, and clean alternate-screen restoration; Claude 2.1.209 independently rendered its empty dashboard/help and restored the terminal. Both used the same network-disabled, read-only, non-root, capability-dropped container profile, and both --rm containers were gone afterward.
• The v0.1.14 candidate adds a real slow-subprocess regression for detached Claude launch, canonical executable-resolution tests for legacy Pi supervisors, and an exact read-only probe against the existing managed Codex App Server showing that a newly launched thread remains visible without --include-interactive.
• The v0.1.15 candidate drives Enter on a managed Pi row through the real PTY, asserts alternate-screen suspension/restoration instead of Peek, and opens the exact completed Pi history on this Linux host in a read-only native-TUI probe before returning cleanly to agentview. Space remains independently covered as the inline Peek route.
• The v0.1.16 candidate adds a process-level completed-Pi native handoff and a stop-to-delete lifecycle. The mock proves stop returns within 500 ms, exit is observed before Delete appears, native resume receives the exact ID/session directory, deletion revalidates the managed JSONL header/path, and external IDs remain refused. A real PTY also bounds completed-Pi Ctrl+X repaint below 750 ms.
• The v0.1.17 regression case removes the legacy supervisor socket before its verified process exits and proves completed-Pi native handoff waits through that transition. Render tests also prove an idle-but-attached completed Pi row advertises Ctrl+X stop before deletion.
• After the managed-Codex lifecycle fixes, release binary SHA-256 efb0a8d8f62f878fc5bb09c6b67da73a871c26f30e7a2dd56d10922a186cbec9 was staged mode 0755 and rerun in the same immutable image at a real 80×24 PTY with the populated fixture. It rendered all state sections and explicit Claude/Codex provider columns, opened/closed help, exited through Escape, and restored the alternate screen. The eight-target fixture then rendered its full provider header and representative rows at 150×36. Claude Code 2.1.209 independently rendered its reference empty dashboard at 120×34, opened/closed shortcuts, and exited cleanly under the same network-disabled/read-only/non-root controls. All named --rm containers were absent afterward and the exact staging directories were removed.
• Host Claude discovery was compared with claude agents --json --all.
• On the reported host environment, Claude 2.1.236 returned completed rows even without --all; a current-binary JSON probe centrally reduced that result to the two active Needs-input rows and removed the stale Antigravity cache row. Three warm probes completed in 0.44–0.47 seconds. The Claude CLI process itself peaked near 369 MiB, motivating the 15-second default refresh plus ctrl+l manual refresh.
• The reported mixed store contained more than 1,100 Codex threads and 70,915 OpenCode history rows. Before the bounded adapters, Codex spent about six seconds paging and then discarded all of its rows, while OpenCode admitted the full store. With Codex 0.147.0 and OpenCode 1.17.20 resolved from their conventional user installs, active-only discovery completed in 1.55 seconds without warnings. --all --history-limit 100 returned 197 total rows (81 Codex, 100 OpenCode, 10 Claude, and 6 Pi) in 1.48 seconds with explicit partial-history warnings. The default 100-record budget repeated in 1.42 seconds. No task was launched and no history was changed.
• OpenCode 1.17.20 was also probed directly against 501 read-only database rows. Its JSON-array output truncated at line 1,851 when stdout was a pipe; the bounded json_object-per-TSV-row query returned all 501 rows (502 lines including the header) and parsed successfully. This is why history discovery uses streaming TSV rather than a large JSON array.
• A read-only real-host openpty run then used the release binary with the reported provider home. It automatically found six launch-capable harnesses, moved from Claude to OpenCode with three real Down-arrow events, loaded 467 OpenCode model choices through Shift+Tab, cancelled without losing the draft, enabled bounded completed history, exercised Down/Up and typing, and restored the alternate screen on exit. No task was submitted and no provider session was modified.
• Cursor 2026.03.20-44cb435, GitHub Copilot CLI 1.0.80, and Antigravity 1.1.14 were probed with disposable homes/configuration roots and no copied credentials. Cursor and Antigravity empty-state interfaces were exercised in real PTYs through clean exit and terminal restoration. Copilot's real ACP server completed initialize/list without authentication and returned the documented authentication error for session creation. No model task was dispatched; provider-specific details and exact commands are recorded under docs/exploration/.
• Pi 0.84.2 was exercised with temporary configuration/session directories, offline startup, and no credentials. Strict RPC JSONL, canonical session identity, state/name commands, a built-in bash-tool event path that did not invoke a model, clean EOF, and exact native TUI resume/terminal restoration were verified.
• tests/real_managed_launch.rs adds explicit ignored, credentialed host lifecycles without reusing the ordinary agentview supervisor state. Pi 0.84.2 used a temporary daemon and session directory: it launched an exact-response prompt, inspected the assistant transcript, started a second turn, interrupted it, stopped the daemon, and proved the exact provider PID exited in 3.60 seconds. Codex 0.147.0 used a temporary durable App Server: it launched and inspected an exact-response task, exercised a controlled sleep 30 turn and one-time approval when presented, sent exact-turn interrupt (or accepted only Codex's precise completion race), archived and deleted the exact thread, recovered the idle owner when Codex withheld its delete response, restored no test ownership, and removed the exact native listener. The final isolated run passed in 37.13 seconds.
• Failed Codex probe iterations were not left as mystery jobs. Nine exact agentview-CODEX-SMOKE ordinary/archived IDs were enumerated, deleted through independent App Servers, and re-listed as absent; the final passing probe deleted its own ID. No prompt used a project workspace or modified a file, and post-test process scans found no temporary App Server, stdio observer, Pi daemon, or Pi RPC child.
• Pi 0.80.6 was additionally resumed from the reported host's real recursive per-workspace JSONL directory by exact UUID and exact file-parent --session-dir; it opened under a PTY and exited with Ctrl+D without the former No session found matching error. No prompt was sent.
• The official Cursor, Copilot, and Antigravity install/version/help paths were then repeated in three new --rm containers using pinned Debian/Node image digests, tmpfs homes, and no host mounts. Cursor's current installer returned 2026.08.11-e8db854; Copilot's real ACP server again negotiated and listed an empty store; Antigravity again returned 1.1.14. Exact image digests, commands, and outputs are in docs/exploration/fresh-container-provider-validation.md.
• The canonical fifteen-agent-plus-Terminal fixture passed all_supported_providers_coexist_in_one_real_terminal, including provider labels, contextual help, alternate-screen entry, and terminal restoration. The same real-PTY test now opens a managed Pi reply composer, a managed Cursor interrupt confirmation, and a managed Copilot approval card, then proves every action is rejected by the fixture I/O fence.
• The TUI was exercised in a 120-by-30 pseudo-terminal, including navigation, grouping, help, and clean shutdown.
• Claude peek was checked against a real host session using read-only logs; the VT100 reconstruction surfaced the final assistant screen without escape-code leakage.
• Provider-native handoff is covered by a nested real-PTY regression: a plain arrow reaches the provider editor, an unchanged cursor produces the timed bottom-line hint, repeated Left and Right each background the exact frontend, Shift+Right returns immediately, and every reopen restores the retained VT screen. Main-account read-only probes repeated the cycle against a managed OpenCode session and a managed Codex thread; the OpenCode server and Codex App Server remained alive afterward.
• Four opt-in real-host openpty regressions now preserve those reproductions: nested Pi resume/return, Claude 2.1.243 attach with a real plain-Left hint and second-Left return, and a mutation-free composer route through Pi/Claude provider cycling, draft-preserving Shift+Tab Pi/Claude model selection, the dedicated filter, manual refresh, plus the six-harness/OpenCode/bounded-history route described above. They are ignored in ordinary CI because they require installed provider binaries or private host history paths.
• Codex App Server discovery was refreshed repeatedly inside a disposable, network-disabled container. Separate protocol probes established that two servers sharing CODEX_HOME cannot control one another's live turns, motivating the shared durable endpoint.
• The release binary was mounted read-only into a disposable container from immutable image ID sha256:8f170f660813ac358f347fa8a3580139972f3ea7a9fb087834f1da44669d9392 (Codex 0.144.4). It ran as an unprivileged user with no network, credentials, or workspace mount; started the durable Unix listener, completed its WebSocket handshake, listed an empty store, and exited with no warnings. The container and isolated state were then removed.
• Claude and Codex discovery were exercised in disposable, network-disabled containers selected explicitly by immutable Docker ID. Each probe container was removed afterward.

These fresh-container TTY checks validate empty-state rendering and basic interaction, not an authenticated provider lifecycle. The containers had no credentials and no network, so no real Claude or Codex task was dispatched inside them. Populated session parsing, request replay, lifecycle authority, approval/input handling, and destructive-action gates are covered by the deterministic fixtures and disposable mock App Server tests above. A full live task inside a fresh container remains a separate opt-in credentialed test.

“Verified manually” records observed behavior but is not a pixel-diff or automated golden-screen assertion. ANSI pane capture also cannot establish exact color fidelity; screenshots must be reviewed as external, redacted test artifacts.

Validate the canonical fixture through the same parser used by the application:

jq empty fixtures/populated-sessions.json
target/release/agentview \
  --fixture fixtures/populated-sessions.json \
  --no-host-claude \
  --no-host-codex \
  --json --all |
  jq -e '.sessions | length > 0'

The following dependency-free Ruby check resolves every relative Markdown link from the document that contains it. External URLs are intentionally outside its scope:

ruby <<'RUBY'
failed = false
Dir.glob('{*.md,docs/**/*.md}').sort.each do |file|
  File.read(file).scan(/\[[^\]]*\]\(([^)]+)\)/).flatten.each do |target|
 '').sub(/>$/, '')
    next if target.match?(%r{^(?:https?://|mailto:|#)})
    path = target.split('#', 2).first
    next if path.empty?
    resolved = File.expand_path(path, File.dirname(file))
    unless File.exist?(resolved)
      warn "broken: #{file}: #{target}"
      failed = true
    end
  end
end
exit(failed ? 1 : 0)
RUBY

Also run git diff --check so whitespace errors do not undermine rendered examples.

• Docker commands use argument arrays and exact inspected container IDs, not a shell command assembled from user input.
• An existing session has no stop capability unless its provider/runtime key is present in the local ownership registry written at launch time.
• The durable Codex record stores PID start time and exact command-line bytes. Linux matches both through /proc; macOS matches proc_pidinfo start time, KERN_PROCARGS2 argv, and the exact owner of the private Unix socket. Explicit Linux idle recovery opens a pidfd and revalidates the complete identity before signaling that exact process; macOS safely refuses that exceptional restart. Normal dashboard exit never signals the server, and stale sockets are not unlinked automatically. Shared/exclusive private recovery locking keeps other dashboards from attaching during the bounded replacement window.
• Managed Codex launches use on-request approvals and workspace-write, with no danger-full-access or approval-bypass fallback.
• Managed Pi reconnect requires an exact Linux PID start token and command line, with private state/socket/record modes and symlink refusal. Only the daemon's in-memory canonical session IDs receive mutations; external JSONL history remains inspect/open-only.
• Managed OpenCode reconnect requires the exact Linux process start token and command line, ownership of the recorded loopback listener, an authenticated health response, and a canonical session ID in the private 0600 record. External CLI history never gains server authority.
• Managed-Docker tests use an injected command runner: no test contacts the Docker daemon. They cover locked/atomic private ownership persistence, random instance IDs, stopped-only creation, exact start/remove argv, immutable-ID revalidation, and record removal ordering.
• User-supplied Docker targets remain observe-only. The separately tested managed-container API requires immutable identity, matching labels, and an external owner record before lifecycle operations.
• Authentication values were not read into the repository, fixtures, or test logs.
• Provider/runtime/session/group/summary/warning/notice/confirmation text is sanitized before terminal rendering, and dynamic transcript tails are bounded; provider text cannot intentionally inject raw terminal controls through those surfaces.

• Codex file-change acceptance, permission grants, MCP form/URL acceptance, secret structured input, supervisor status/stop, and log rotation.
• Managed-container session launch/control remains separate from container lifecycle; enter the started container through ordinary Docker tooling or observe it with --docker-container.
• Claude inline reply and provider-native rename, for which the explored CLI exposes no safe background-agent command. agentview-local aliases remain available for every normalized session without mutating Claude. Enter opens its native attach and the native return gesture backgrounds that frontend; owned Codex threads support inline idle reply and active steer.
• Managed OpenCode permission and structured-input requests are not yet exposed inline. Durable supervision requires Linux; other platforms retain external history inspection and native resume.
• Cursor has no documented global machine-readable inventory. Only sessions created through its agentview-owned Linux supervisor are shown and controlled; macOS has no managed Cursor discovery/control until an equally race-safe process identity primitive is implemented. Cursor's native TTY picker remains available outside agentview.
• Copilot provider-wide list results remain observe/native-open in the dashboard. agentview-created IDs and bounded latest-message metadata survive a restart, while live prompt/permission authority still applies only to the retained process-local connection and is never reconstructed from disk. Antigravity exposes only its documented last-conversation-per-workspace cache; complete history and inline controls are not claimed.
• Durable Pi supervision currently requires Linux /proc identity. macOS retains documented history discovery, inspection, and native resume.