For a step-by-step workflow, read Set up your first project. For current provider support, consult Capabilities; this chapter describes reviewed surfaces and their limits.
Chat
The seven-button project bar is Chat, Code, Overview, Plan, Monitor, Review and Terminal. Code opens a menu; Plan, Monitor and Review open their remembered section, with detailed tabs underneath. Overview opens the project home directly. Saved layouts and explicit pane choices retain their destinations.
For the selected Claude, Codex or Gemini agent and Profile, open Profile → Account & sign-in… in the composer’s agent and model selector. An empty chat can offer Set up agent and Later; Later dismisses that invitation without running checks. A generic error offers the neutral Agent setup… action beside Reconnect. Setup uses the current selection, which may differ from a queued switch or the unproven identity of an earlier failed request. Opening or closing it does not reconnect, send or automatically retry a message.
A chat tab’s menu offers its move actions and Close for a closable tab. These actions target the tab whose menu you opened.
Open a code header’s menu for Copy Code, which copies the whole block, or Expand Code for a larger view. The expanded view’s toolbar also offers whole-code copying.
The composer menu includes Paste and Match Style, input recall, Quick Actions… and Attach Image… where available. Clear Draft Text leaves pending images in place. An image chip’s menu offers Remove Attachment and Remove All Attachments for those pending images.
Antigravity has a separate native sign-in flow. After returning, use Check connection & refresh models in setup or Account & sign-in. The returned model names and IDs are shown in setup and shared with the chat picker and Agents & Models. Refresh models is also available from the picker. Opening the Antigravity picker can also refresh catalog metadata; sign-in and completions remain separate actions. Refreshing changes no selected model and sends no completion; catalog availability does not certify account status. An empty, unavailable or unrecognized catalog is explained instead of silently presenting a blank list.
Select a project and open Chat to send prompts, choose agent/account/model and inspect streamed replies, tools and attachments. A split layout can hold independently configured panes. Each Chat strip can contain several chat and file tabs. Use its + button for a new conversation in that pane; Command–T targets the primary pane. A tab can move between panes of the same project while keeping its conversation identity; pending tool confirmations and unsaved file edits must be resolved first. Moving the last tab creates a fresh source chat. Closing the sole tab preserves its conversation and follows Last chat tab fallback. The original chat can be reordered or closed like its siblings without changing its conversation identity during a reorder. Closing any busy chat with siblings asks before stopping its turn; save dirty files and resolve conflicting composer drafts before closing those tabs.
Open the composer’s agent and model selector for Agent, Profile, Model and provider-supported Effort. For Claude, Codex and Gemini, Profile choices remain available while the current account is default; there is no second Profile selector. Effort uses a stepped slider with a provider-default choice, and describes requested configuration rather than provider-reported effective effort. The same selector’s Thinking section offers Show thinking updates. It changes the existing appearance preference for available thinking updates in all chats; it does not turn model reasoning on or off. Permission mode and automatic continuation retain their existing menu actions. The footer groups attachment, / commands and the icon-only Loadout control on the left, with Send or Stop on the right. Session info remains in the command menu. When available, the wide composer shows a labeled Weekly quota value; compact widths hide that separate weekly control. Passive usage appears in the composer’s Weekly indicator when available, without a separate pane notice. Notices for held messages, recovery after a usage limit and model-capacity errors keep their actions. Chat telemetry labels its own measure as Context, Turn spend or Cost. Hover either control to read its visible measure first, followed by the provider's session and weekly quota breakdown where available. Open Chat telemetry for quota details, text size and Find in conversation; unavailable quota evidence is not zero usage. Chat tabs use plain target captions. For example, a tab can show Codex Sol while its composer shows Codex GPT-6.1-Sol · High. The composer names the requested reasoning level; the tab’s hover card retains session detail, while accessibility descriptions preserve the model and requested effort. A requested account switch waits until an active turn ends; the selector retains running and queued Profile information. The Privacy control appears only for a checked preview with a non-default mode or warning; an absent control is not evidence that a privacy check passed. A refused request can show a separate privacy notice with recovery guidance.
Option–Tab and Option–Shift–Tab cycle chat tabs in the selected project’s primary pane. New extra chats start fresh; an empty extra chat offers Resume latest for its selected provider/account and refuses a latest session already held by another tab. Open the Chat telemetry popover and choose Find in conversation, or use Command–F, to search the current Chat, Project or All conversation scope. Command–F first uses the most recent click that still identifies a pane in that window, then the pane containing the focused control, then the pane under the pointer. A click in a toolbar or outside the panes does not override valid pane focus. Chat and Roadmap retain their existing search; the other supported project panes focus their local filter or Find. A selected .md or .markdown file tab uses read-only document Find with Previous/Next matches and Reload; other file tabs retain their editor’s Find behavior. See shortcut scope.
The auto-continue message editor saves text after typing pauses for 400 ms. Changing between This chat and All chats, leaving the field or choosing Done flushes pending changes. Reset to automatic removes the selected scope’s override. Editing or closing this message does not send a continuation turn.
The transcript shows captions such as Automatic · operator note or Automatic · usage-limit retry. The caption remains visible when sender names are hidden, so a queued operator note or generated retry is distinguishable from a message you just entered.
An automatic caption explains a message’s origin; it is not an approval, proof of unfinished work or permission for another turn. A saved session or an expired limit alone does not establish pending work. Text you explicitly send remains your message even if it looks like an automatic marker. A recognized marker read from saved history is labeled reported, unverified; that label does not verify the historical sender or provider behavior.
Tab names: hover an open chat tab to inspect its permanent @name. Right-click it for Copy @name or Rename @name…; an unnamed tab can be named there too. Names use lowercase letters, digits and hyphens, starting with a letter. Automatic naming preserves manual names, and a rename keeps the old name reachable for queued notes. commandem announce name ls lists the local name directory. A name identifies a routing target, not a currently running process.
Boundary and recovery: Sending starts provider work and can spend quota. Native resume remains provider/account-owned; changing a model is not changing accounts.
Files
Browse project files and open a file in the configured Commandem tab, separate preview pane or external application. With no stored preference, file links use Commandem tab; existing choices remain in effect. Hidden and ignored files are controlled by Layout preferences. Resolve an attachment's actual path before handing it to another app.
Selected Markdown tabs display a bounded read-only preview from a regular file in the tab’s original registered project. Links and images stay passive; unavailable files and shortened previews are labeled. This does not change the inline Files source editor.
Find: While browsing Files, Command–F focuses the existing file filter. It ranks loaded relative paths and shows up to 200 results; this is the existing fuzzy filter, not a complete archive search. While an inline source editor is open, Command–F retains that editor’s native Find instead. Escape clears the file-filter query and leaves that field unfocused.
The visible Close button on a file or image preview returns to the captured prior secondary-pane layout. Replacement previews retain the first origin. Deliberately hiding the pane, explicitly replacing the secondary pane with a project, or applying a Saved View discards that return destination. This reviewed behavior is limited to the visible preview Close button; existing menu and shortcut whole-pane behavior is unchanged, and it does not certify indefinite chat-session retention.
Boundary and recovery: A link is not permission to execute a file. Notebooks and documents need a compatible installed handler; previews and edit support differ by file type.
Search
Search within the selected project's code and files. Use the project selection and query to scope results, then open the matching file and line.
Boundary and recovery: Project search is not Matrix retrieval or a global chat-history search. Generated and ignored files can change what is available.
Log
Inspect project log output and recorded activity while diagnosing a run or check. Match timestamps and run identity rather than inferring cause from neighboring rows.
Find: Filter the already loaded tail by case-insensitive line text. Earlier matching loaded lines can appear before the display cap; any remaining display omissions are labeled. Copy still takes the whole loaded tail. Find does not restart the reader; Escape or Close clears the filter.
Boundary and recovery: Logs can contain sensitive paths or tool output. Redact before exporting; absence of a row is not proof no action occurred.
Checks
Read the project's timeline of commits, runs and transitions. Use it to locate evidence relevant to a status change and then open the original record.
Find: Filter the loaded event text: displayed day, time and event name, plus project identity while All projects is enabled. Hidden exit codes and raw field names are not search text. The source window remains 200 observations; matching display omissions are labeled. Escape or Close restores the loaded feed without another read.
Boundary and recovery: Activity is observational. A notification dismissal does not erase historical activity or undo a commit.
History
Browse provider-session history available to the current project and account. Open the original session for inspection or choose a supported resume path. Saved-history browsing in Chat uses a bounded older/newer reading window. It can release unprotected records at the opposite edge while preserving the captured reading anchor and exact protected identities; reversing direction can load those records again when available. See the linked history reference for source-record limits and refusal boundaries.
Find: Command–F focuses the existing saved-session filter. It searches session titles, IDs and account labels within the current provider selection, not every saved turn’s body. Escape clears the query and leaves the field unfocused. Provider/archive availability remains a separate limit.
Boundary and recovery: Provider/account roots and working-directory mapping matter. An API endpoint does not automatically inherit native CLI archives.
Matrix
The internal tab identifier is graph; the visible product name is Matrix. Inspect conversation relationships and reviewed candidate context for this project, with source provenance.
Boundary and recovery: Related does not mean safe to inject. Review source scope, privacy posture and destination before a bridge is delivered.
Diff
Inspect repository changes against the relevant base before review or merge. Confirm the branch, base and dirty worktree as well as the rendered patch.
A hunk’s menu offers Stage hunk or Unstage hunk for the displayed side. Unstaged hunks also offer Discard hunk…, which opens the existing confirmation.
Find: Command–F opens raw view and focuses a local case-insensitive changed-path filter. It intersects the active intent lens and can include tracked and untracked paths; it does not search patch bodies. Escape or Close clears the path query while retaining any active intent lens; Show all files clears the lens. No-match is explicit and does not mean the worktree is clean.
Action scope: Stage all and Unstage all affect the full change set while a path filter or lens is active. Hunk actions apply the complete original hunk patch; uncertain path/header identity makes those actions unavailable. A displayed excerpt is not the full patch, and Unicode-path read support does not certify atomic whole-rename unstaging.
Boundary and recovery: A readable diff is not passing tests or authorization to land. Preserve unrelated user changes and never reset them to make a review easier.
Review changes before landing
Replace PROJECT with the exact project name. Confirm that its registered working directory is the worktree you intend to review.
Read-only inspection.
commandem diff PROJECTExpected result: The project's current diff, with likely secrets redacted. An empty diff is not a passing test, and reading it does not stage, commit or merge changes.
Terminal
Open the selected project's terminal surface and choose New Terminal explicitly to create a shell. Confirm the working directory and command before execution. On macOS, saved shell holders keep running when the app closes; returning to the project reattaches running sessions. Use Refresh to reread their state, the session close button to end a shell, and Forget ended sessions to remove ended references. An unavailable or ended session is not silently replaced. Scrollback uses Ctrl-A, then [.
For a shell outside one project, open Fleet terminal from the palette or press Option–Command–T. It uses the account home directory. Demo workspaces disable terminal execution and cloud dispatch; their banner is not permission to run either.
commandem open PROJECT --shell-session list
commandem open PROJECT --shell-session create
commandem open PROJECT --shell-session attach --session-id SESSION_ID
commandem open PROJECT --shell-session close --session-id SESSION_IDListing reads references; create starts a shell, attach joins the exact saved shell, and close ends it. Replace SESSION_ID with a returned identity. Fleet actions use --fleet without a project. Persistent terminals are explicit owner actions, not an agent or plugin execution route.
Boundary and recovery: Shell commands can write files, disclose secrets or spend quota. Do not copy diagnostic commands blindly from untrusted transcripts.
Config
Inspect the project manifest and declared commands/conventions. The manifest connects project behavior to CLI checks and orchestration.
Boundary and recovery: A configuration declaration does not prove the command is installed or safe. Review mutations and validate the resulting manifest.
Inspect a project's declared workflow
Replace PROJECT with the exact registered project name. This reads its manifest without applying a patch.
Read-only inspection.
commandem manifest PROJECT --jsonExpected result: Declared convention fields as JSON. Check command paths and project scope before invoking an action; absent fields are not invented defaults.
Stats
Inspect project-level activity and resource summaries as available from their data sources. Use underlying receipts when exact accounting matters.
Boundary and recovery: Observed, reported and estimated usage are different. Missing provider quota data must stay unknown, not zero.
Control
A short control panel for the project: its process, last check, last deploy, its declared actions, and an "open in" list, with the agent plan, approve and review flow in one collapsed section. It reads existing commands and shows the last known state first; a failed read is shown as failed, not as absent. Actions use the project's declared workflow rather than an independent task system.
Boundary and recovery: An action can mutate or execute. Inspect the command and require the same approvals that an equivalent CLI call needs.
Roadmap
Browse Campaigns, Missions, executable tasks, dependency state and completion evidence in the project roadmap. Use durable Mission IDs in links or dispatch references. In the project toolbar, Missions and Campaigns sit under the Plan button and News under the Review button. Plan explains Campaign, Mission and task scope.
Project Overview
Project Overview: Overview is the project home: its own button in the project toolbar, with no section strip above it. It was formerly a sub-tab inside Roadmap; Roadmap now offers Missions, Campaigns and News. It is a page of cards: a row of vitals (checks, uncommitted files, ahead and behind, Missions done, runs in 7 days, tokens in 30 days), then Needs you, Agents, Missions with task meters, a 30-day commits chart, an activity-by-agent card, agent-job outcomes, What changed and Health. Open Missions and Open News target the existing detailed tabs, and a tile that has a destination opens it (Diff, Missions, Run Reports for the project, Usage, Monitor); the page does not replace those views or promise to show every record.
Overview distinguishes a source's status and observation time from the time the summary was collected. Missing model, account and capacity facts remain Unknown rather than zero. Saved or stale facts show a bounded source explanation; an unavailable section is not an all-clear. Mission progress remains reported roadmap intent, not evidence that an agent is executing it. A run reads Running only when its worker is verified alive; a chat-lane entry never does, and runs finished more than a day ago collapse to one line and are not counted as agents. Verified means a passed outcome gate: a passing plain verify command reads Unverified and says so. The “Agent jobs: did they pass their checks?” card counts queued and unattended jobs, excluding chats. Its chart appears only when jobs span at least three active days and at least one job is from the last 14 days; otherwise it shows a count and last date. Unattributed receipts remain excluded and disclosed. The activity card includes chats and subagents from the shared Usage report, uses Usage’s agent colors, and labels unread sources. Pending or malformed daily data stays unknown rather than drawing a zero graph. Usage links from this project open with the project filter selected; global Usage opens across projects. A token read may cover only supported agents, so missing agents are disclosed. In addition, and a Mission with no tasks has no progress bar. Steer and Stop remain unavailable while an exact native target and authority are unresolved.
Find and Refresh in Overview: Find in Overview keeps only the displayed rows that match and hides the rest; it does not search the figures or charts, undisplayed records, or replace Mission search in Roadmap. Refresh asks for another summary. A failed refresh can retain the previous observation with a visible failure notice; it does not establish that saved facts are live.
Task previews from the Go dashboard source: source 7794b1dce6d56dcb8d80c8db666c044527c24c07 retains a UTF-8-safe prefix of at most 256 bytes for descriptive current, recorded and last-completed Task values. A shortened value ends with … [preview; N UTF-8 bytes total], where N is the original stored Task text's UTF-8 byte count. Short Task values and Unknown/null values are unchanged. The complete stored Task and exact lane, run, session, provider, model and account identities remain unchanged. This Go source statement does not qualify native delivery.
macOS Unicode argument boundary: an opposed Foundation Process sample preserved an NFC project argument through assignment and its getter, but the receiver observed NFD. Exact argument preservation for every Unicode spelling is therefore not qualified. The original mounted Host control failed before its expected first project receipt and its cleanup remained unresolved; its records did not independently capture OS argument bytes. Strict returned-project, request, cache and stale-publication fences remain in force, so a mismatch must refuse rather than be treated as another project’s successful result. Separate ASCII receiving checks and injected byte-fence checks do not certify Unicode process arguments. No renaming workaround or identity-policy waiver is implied. The limitation is tracked with the Overview work in Mission m-8ec4342f.
Task text limits: the native display retains each supplied current, stored or last-completed Task value up to 512 UTF-8 bytes, including a supplied preview marker and original-byte count. Existing control-character cleaning still applies. A source preview may itself be only part of the stored description. Task values above that display limit receive an explicit excerpt notice. The source-text budget is 2,560 bytes per tracker block and 1,536 per other section block, with the complete document still bounded to 64 KiB. These are display budgets, not new settings, backend limits or changes to saved text.
Work on it prepares a chat draft containing the Mission reference and remaining task scope. Review and send the draft; preparing it does not start an agent.
When derived evidence is available, Roadmap rows show a five-position work track (Shape, Build, Verify, Land and Done) and a separate trust glyph. The selected Mission’s trust chip shows its recorded explanation on hover. Trust distinguishes verified, unverified, integrity warning, failed, ungated, legacy and unknown evidence separately from authored status. It does not re-run a check, authenticate a receipt or authorize a merge.
Find: Mission search includes loaded titles, notes, linked Campaign titles and rationale, Mission comments, stable IDs and task titles. Search shows bounded pages of at most 300 combined matches; Previous results and Next results reach every match. It switches Roadmap’s kanban to List. Clearing search restores the usual browse limits. Editing and dispatch retain the original Mission identity.
Boundary and recovery: A checked task is a claim, not verification. Completion needs locally verified evidence; a status update alone does not implement a feature.
In List view, Sort offers Workflow, Priority, Due date, Step and Recently updated. The choice is saved per project. Right-click a Mission or Campaign card to Pin or Unpin it; pinned cards display a mark. Pins are local owner view state rather than authored roadmap status. Workflow keeps the existing workflow order; a pin does not promise first position in every section or view.
Project news in Roadmap
Project news shows recent recorded updates for the selected project. In the full Roadmap it is its own News tab and fills the pane, with up to twenty rows. The Glance companion shows a smaller fixed-height strip. Neither has a draggable divider or a saved size. The full Roadmap and its full or Glance companion use the same feed. These are recorded events, not an agent-generated summary. Reading them does not start provider work, mark Board messages read or acknowledge alerts.
Keep your place: after the initial successful load, changed snapshots wait behind Show latest updates while the displayed rows stay in place. Choosing it applies the waiting snapshot and returns only the news view to its top; Mission reading position stays in place. Receiving updates also preserves unsent task drafts in the full Roadmap and new-mission drafts in Glance. Refresh checks for another snapshot; Retry tries again when news could not open. Refreshing pauses when the window or pane is hidden. A visible window can keep reading while another window has keyboard focus.
Select an update to inspect its recorded details. Open recorded source appears when a supported web source was recorded and opens it explicitly in the browser; Done closes the details.
Filter displayed rows: All, Features or Needs attention narrows which recorded events are shown; Show maintenance additionally dims routine maintenance commits while All is selected, or hides them when turned off -- it has no effect under Features or Needs attention, which are each already a single classification. Both choices persist across launches and change what is displayed, never which events are recorded. Source initializers: "commandem.news.significanceFilter" → ActivitySignificanceFilter.features.rawValue and "commandem.news.showMaintenance" → true; stored values may differ.
Boundary and recovery: a failed refresh retains the last good updates and shows the failure state. This card contains recent retained activity, with no older-page control. The displayed row count never proves the end of history. The separate Checks tab keeps its own loaded timeline and Find behavior.
Read updates without moving your Mission reading position
Read-only inspection.
- Open a project's Roadmap, read a Mission and leave some Mission text selected.
- When Show latest updates appears, choose it deliberately; inspect the new news snapshot while keeping your Mission reading position.
- Select an update for its recorded details, then choose Done. Open recorded source is a separate browser action when available.
Expected result: after the initial load, changed news rows wait for Show latest updates. Applying them leaves the Mission reading position in place. If refresh fails, the previous updates remain visible with the failure state.
Find current Missions and durable references
Replace PROJECT with the exact registered name. This inspects structured Missions and any associated roadmap document, without changing status or task checkboxes.
News timestamps use relative time, with the absolute time available on hover. Unrecognized timestamp text remains visible. News significance filters use the current client classification; a decoded server field alone is not proof that the view consumes it.
Read-only inspection.
COMMANDEM_NO_ROADMAP_AUTOSYNC=1 commandem roadmap PROJECT --jsonExpected result: The current roadmap, including durable IDs where present. Reuse an actual ID rather than copying the illustrative m-xxxxxxxx placeholder. The prefix disables linked-issue autosync for this inspection. Normal roadmap reads can initiate configured synchronization; review its conflict policy before enabling it.

Messages
Read and send project board messages. Board messages share unread indicators across Board surfaces; historical messages remain after being marked read. They are audit records, not human notifications or OS banners. Local Comments use a separate discussion thread.
Find: Filter loaded message blocks by their displayed sender, recipient, non-default kind, short timestamp or folded body. Hidden message suffixes cannot match. The History view filters displayed metadata and, only with Excerpts enabled under an admitted standard-privacy report, displayed bodies. Record links retain their original IDs and source selection. Filtering does not send, refetch or mark additional reads; Escape or Close restores loaded records.
Local matches do not change delivery or ACK evidence. Source coverage gaps and omitted matching records remain visible; this is not a complete messaging archive.
Boundary and recovery: Clear/read controls are not deletion. A synchronization adapter may be unavailable or require separate private-repository access.
Notes
Open this project’s private scratchpad to keep notes separate from transient conversation text. Reading, adding and clearing notes are separate actions; Find does not send or refresh them.
Find: Search within all loaded note text, using the match count and Previous/Next to select occurrences. Only a bounded excerpt around the selected match is rendered; longer occurrences are partly shown. Copy all loaded notes remains explicit when excerpted. Escape or Close clears and closes Find while preserving the unsent draft.
Boundary and recovery: Private Notes use the local project scratchpad. Local Comments, Board Messages and external adapter delivery have separate audience and storage contracts. Copying note text is an explicit action.
Links
Keep pinned project URLs discoverable for docs, dashboards and trackers. Pin writes the project’s local manifest; opening a URL uses its intended browser or handler.
Find: Filter complete loaded pinned URLs case-insensitively, including text beyond a shortened label. Results use bounded 32-link pages with Previous/Next; every loaded match remains reachable. Open and Remove retain the exact original URL, not a filtered row position. Escape or Close restores the loaded list; Find does not write the manifest or open a browser.
Boundary and recovery: Remote links leave the local app boundary. Access permissions, network availability and installed file handlers remain external requirements.
Current source boundaries
These are source-level boundaries reviewed against 8f06da03. They do not replace the dated source captions or screenshots elsewhere in this handbook, or certify installed-app, physical-input, device, provider or OS-permission behavior.
Chat recovery, selection and tab limits
In the composer, Up recalls your newest available input; further Up presses move to older inputs. Up enters recall from the first line, preserving ordinary multiline caret navigation. Down moves toward newer inputs and restores your draft past the newest; Escape restores the draft you were writing. In a resumed Claude or Codex conversation, an explicit Up can read up to eight bounded local history pages for older complete inputs. Page targets count conversational messages; raw-record, byte and time limits still apply. The separate recall cache holds at most 200 older inputs and 512 KiB of input text. Truncated previews and unsupported history records are not recalled; a status explains limits or unavailable reads and when to press Up again. Paged-history inputs send as new messages because their full-session rewind position is unproven. Up also recalls prior input while a turn is running. Recall prepares text for review and never sends it automatically. Sending a busy recall creates a follow-up through the normal send or queue path; it does not rewind or stop the active turn. Recall only prepares text for review; explicitly Send to start a model turn. Transcript browsing retains its 60-source-record and 12,000-cost window limits and its existing follow behavior.
A conservative interruption notice can follow observation of a surviving owned process from a prior app session; it is not proof that a turn crashed. Continue and Dismiss are manual choices. A transient reconnecting notice offers Retry now or Cancel separately from a quota hold; the notice alone does not prove the provider recovered.
Transcript options place text in the composer for review. In a multi-select block, checking or unchecking a box updates the current selection in the draft immediately, one option per line in displayed order, while preserving text outside that selection. Unchecking the last box removes the inserted selection. If an update cannot safely locate the previous selection, it keeps your text and shows a notice; a later selection can append a new block. The filled Send N selected control confirms the same draft selection, focuses the composer and clears the checks; review and send the message from the composer. A pending selection pauses automatic following; the configured scroll policy still applies afterward. Reader anchors and bounded scroll convergence reduce particular jumps, not every possible layout shift. A neutral tool-exit label still reports a failure, not success; blocked, denied and timeout outcomes retain their separate meanings. Thinking previews remain abbreviated, not a complete reasoning transcript.
Saved-history paging protects the captured source-message anchor, selected source records, Find targets and held sends within the unchanged source-record and rendering-cost limits. A protected span that cannot fit can stop a page change. Folded activity remains a summary of its source records; only the installed renderer's exact membership mapping admits shared group geometry. An automatic handoff qualified by that mapping keeps its fold policy, and retained expanded run members stay expanded when an older page changes the run head. An admitted reading session captures a fixed trailing-space claim bounded by one viewport. Native space is installed only on a later source publication, preserving the admitted layout while its page is read; incoming output retains reading and the same claim or installed allowance. Latest removes it, as does qualified owned downward arrival at the document bottom after a successful restore. Live and latest views have no allowance. These protections do not promise that every archive or viewport can advance. Current qualification covers an explicit older-page request, source-anchor restoration, a focused 42-method native sample and a separate synthetic owned-window capture. Genuine wheel input, user bottom navigation after resize, active Find and native selected-span UI remain unqualified. Full combined-build and delivery qualification remain separate.
Cross-pane tab dragging targets the receiving Chat tab strip of the same project, not any empty area of a pane. Protected drafts and pending confirmations can refuse a move. Closing a tab may succeed even when its preferences are too large to retain an undo entry: Undo Close Tab is bounded and best-effort. A secondary pane whose project is no longer available offers a close/reselect path rather than silently becoming a new project.
Read stored Notes is a separate reader
The Notes pane also offers Read stored Notes, a separate, explicit reader bound to the exact selected project. It does not turn the scratchpad into a Local Comment thread or grant Matrix, edit, synchronization or outbound-delivery authority. Incognito clears and blocks persistent Note reads; returning to an allowed mode requires Refresh. A stale or unavailable selector is not a verified-empty inventory.
Display limits and diagnostics
Roadmap task titles and comment previews have display limits (120 and 400 characters respectively), with bounded tooltips. Those limits do not truncate saved text or replace the search and pagination limits above. Optional COMMANDEM_TRACE_LINKS diagnostics can include URLs, selected text and drafts; keep such traces private. No automatic redaction or universal scrolling-performance guarantee is implied.