Console

With the standalone binary or Bun, run 0 with no arguments to open the interactive chat console. 0 console opens it with explicit options. From the prompt, the operator can investigate targets, review source, verify findings and work on candidate fixes with the available tools.

Two front-ends share the same engine session (createConsoleSession from @0/core):

Front-end	Requirement	Features
TUI (full)	Bun runtime + TTY (stdout.isTTY && stdin.isTTY)	All slash commands, visual transcript, sidebars, approval prompts, scope extensions, subagent inspection, command palette, theme picker
readline (Node)	Node.js 24+, --scope <file> required	Text-only REPL; limited command subset; scope extensions denied; no interactive tool-approval surface — see approval limitations

The runtime auto-detects Bun and uses the TUI when both Bun and a TTY are available, falling back to the readline console otherwise. The standalone release binary includes its runtime; Bun is needed separately when running the full terminal UI from source.

Launch

# Interactive chat — requires a configured LLM provider
0 console

# Start with an engagement target
0 console --target https://example.com --scope ./scope.json

# Start with a role (tool set)
0 console --role discovery --target https://example.com --scope ./scope.json

# Start in YOLO mode with an initial engagement scope
0 console --yolo --scope ./scope.json --target https://example.com

# Resume the most recent saved session
0 console --continue

# Open a session picker to resume a specific one
0 console --resume

# One-shot: run a prompt and exit (non-interactive)
0 console --mode recon --print "Summarise findings" --continue

# Resume a specific session by id (or unique prefix)
0 console --resume a1b2c3d4

Key flags

Flag	Description	Default
--target <url>	Engagement target the tools operate against	(optional; set in chat)
--scope <file>	Initial authorization scope file; required under Node	(none)
--role <role>	Tool set: audit, review, discovery, attack, verify, report	audit
--mode <mode>	Autonomy mode: standard, recon, copilot, yolo	yolo
--yolo	Shortcut for --mode yolo	—
--model <id>	Override the LLM model ID	provider default
--max-tool-calls <n>	Safety cap on tool-call rounds per operator message	100
--allow-scanners	Expose scanner wrappers (sqlmap, nikto, …)	off
--finding <id>	Focus the chat on one persisted finding	(none)
--finding-intent <intent>	Finding workflow: investigate, verify, draft_fix, impact (requires --finding)	investigate
--db-path <path>	Persistent findings database, also used by history screens	ZERO_DB_PATH or ~/.0/0.db
--resume [id]	Reopen a saved session; omitting id opens a picker	(none)
--continue	Reopen the most recent session, no picker	(none)
--print [prompt]	One-shot non-interactive; reads from argument or piped stdin	(none)

A --scope file is required for the Node readline fallback. Under the Bun TUI it is optional. YOLO public-network tools accept absolute URLs without a launch target; explicit configured restrictions and exclusions still apply. Outside the TUI, YOLO also requires at least one in_scope entry, including with --print. Resuming a transcript does not supply a scope file for this check. For text-only analysis of saved context, explicitly choose --mode recon; Recon is a restricted tool policy, not a guarantee of no network activity.

Choose your autonomy policy

The console defaults to YOLO. Use --mode standard in the Bun TUI for interactive per-action approval, or --mode recon for the restricted Recon tool policy. Co-pilot does not add per-action approval. Standard also bypasses that gate when no approval callback is wired; see the readline and headless limitation. Explicit restrictions and exclusions still apply. This does not change the ordinary scan command’s requirement for a scope file on live targets.

Setup and navigation

On first launch, Escape goes back one setup decision, including Density → Theme. Within a provider login or search, Escape cancels that local operation first. Connect and Models use Ctrl+N to skip; preferences and sharing use s. Back, Confirm, and Skip also have clickable controls. At Welcome, Escape skips setup and opens chat without marking setup complete. Ctrl+C explicitly quits. Confirmed settings and credentials remain saved; model choices are applied to the current audit when you finish or skip setup. Unconfirmed preference previews are discarded when you go back. /onboard opens setup again.

Outside setup, Alt+Left and Alt+Right move through console route history. Nested popups own input until closed; Escape first closes the current popup or edit before returning to the previous screen. Shift+Tab moves backward through the engagement launcher’s fields.

Long-running work and context

The interactive console has no cumulative turn-token cap by default, including subscription-backed providers. Bare 0, 0 console, and resumed sessions share the 100-tool-round default; --max-tool-calls overrides it explicitly. Provider subscription quotas and explicitly configured engine budgets still apply independently.

With auto-compaction enabled and a known model window, the console maintains context between tool rounds and continues the same task. Summaries retain the opening task, latest instruction, and complete recent tool exchanges. Context recovery is bounded and reports when it cannot reduce the prompt; a provider quota or authentication error is not treated as context overflow.

A “turn token budget” pause identifies a local cumulative budget, not the size of the current context. Older builds imposed a 2m-token default. If that pause appears unexpectedly, check 0 --version and /doctor for the running artifact, then restart after updating; an already-running process retains its loaded code. A separate source checkout or generated bundle may be older than the installed standalone executable.

Review previous work

The console attaches a persistent findings database. query_findings can search all sessions or a particular scan ID within that attached database. Use --db-path to open a scan’s run-local state.db; it works independently of --finding, including with --print. The console’s default remains the local ~/.0/0.db (or ZERO_DB_PATH), whereas fresh scan workflows use run-local databases. Do not assume a global history listing means every run’s findings are loaded into this chat.

Saved conversations are a separate store, shared with /resume. The model can use list_conversations to discover them and read_conversation to retrieve their user/assistant text. Discovery defaults to the current working directory; ask for all projects to widen it, or narrow the results with search text. Transcript reads are paginated and size-limited, with explicit truncation and continuation metadata. Known credentials are redacted; hidden reasoning, raw provider payloads, and tool-result bodies are not returned.

Both conversation tools are read-only and work in Recon mode without approval. Already-running consoles retain their loaded code: restart after updating to make these tools available.

Roles

--role selects the tool group exposed to the session:

Role	Tools
audit	Full tool registry (default)
review	Source-code review tools
discovery	Reconnaissance and enumeration
attack	Offensive/exploit tools
verify	Verification and patch validation
report	Reporting role’s tool set

Autonomy modes

Cycle the mode with Shift+Tab in the TUI, or the /mode command.

Mode	Behavior
Standard	Prompts before each effectful (non-read-only) tool call when an approveTool callback is wired, as in the TUI. Without that callback, this gate is bypassed. Scope-extension requests are separate.
Recon	Allows tools classified read-only plus a conservative passive-network reconnaissance set. Effectful/exploit tools are refused; allowed network tools still use authorization and transport checks. Not an offline mode or OS sandbox.
Co-pilot	Skips Standard’s per-action approval gate. Eligible tools proceed automatically, subject to scope and the other authorization controls.
YOLO	Public-network tools need no launch target or per-discovered-host approval. Explicit operator-configured scope, exclusions and prior refusals remain effective.

Readline and headless approval limitation

Neither Node/readline nor headless --print offers interactive tool-approval prompts. Selecting Standard or Co-pilot does not make these paths deny effectful calls whenever an operator cannot be asked:

• A Standard launch leaves approveTool unset, so the Standard gate falls through rather than denying the call. Other non-Co-pilot launch modes also leave the callback unset.
• A Co-pilot launch wires an always-rejecting callback, but Co-pilot skips the per-action gate, so that callback is not consulted.
• In readline, /mode changes the engine mode without replacing the launch callback. Switching a session launched in Co-pilot to Standard makes the retained callback reject effectful calls, not prompt for them. Switching to Standard from a launch without a callback still bypasses this gate.

Use the Bun TUI in Standard mode for interactive per-action approval. Session-only scope extensions are denied separately on the readline and --print paths. Scope, exclusions, Recon restrictions and workspace-trust checks remain independent controls.

In YOLO, the target is optional task context, not a second permission gate. Search results and discovered URLs do not update the target or configured scope. An explicitly empty configured scope remains deny-all; absent scope remains unconfigured. Private-network access, saved credential forwarding and workspace host-code trust remain separate controls. Public URLs do not grant access to private addresses returned by DNS.

Browser HTTP requests use the scoped, address-pinned transport while a browser tool action is active. Cancelled or ended actions cannot dispatch delayed HTTP requests or deliver a held response to the page. Underlying held connections may remain until the existing transport deadline; this is not a full browser network sandbox or a WebSocket/WebRTC isolation claim.

Acquiring a public repository in YOLO

Source checkout is separate from permission to test its hosting service. In YOLO, a standalone public HTTPS git clone through bash or run_command can fetch code even when the repository host is not the launch target:

cd /home/dev/coding && git clone --depth=1 https://github.com/golang/go.git golang-go-audit

Run inspection, builds, or other commands in subsequent tool calls. Checkout does not change configured engagement scope. Previously declined hosts and explicit exclusions still apply.

This acquisition path uses standard HTTPS on port 443, public-address DNS validation and a pinned tunnel, isolated Git configuration, no credential helpers, and the existing command timeout/output limits. It does not follow redirects, fetch submodules, accept arbitrary Git configuration, or execute appended shell commands. Private/authenticated repositories need their normal authorized workflow; this is not a blanket network-scope bypass.

Supported runtimes

The console constructs the direct API runtime, including supported API-key, provider-subscription and hosted transports. It does not choose the Claude Code, Codex or Gemini CLI subprocess wrappers, and has no --runtime option. Installing one of those CLIs is not by itself a console connection.

Use /connect and /model, or supply the provider configuration described in API Keys. The --runtime auto|api|claude|codex|gemini options on scan/review commands are a different surface; see runtime configuration.

First interaction

On the first no-argument launch, guided setup walks through connection, model selection, display preferences and analytics consent. The final Done confirmation marks setup complete. Cancelling does not undo choices already saved, but setup appears again on the next launch.

For Cloud sign-in, your own API key or a subscription connection, follow the setup guide. A normal /connect selection prepares the next chat; it does not automatically replace a healthy chat’s current provider. After connecting, reselect the model in /model to apply it to the current conversation.

When the TUI launches:

• Home screen — product mark, engagement panel, composer (text input) centred on the screen.
• Status bar — active model, mode, working directory, cost/token counters (when enabled).
• Header — product name, configured scope, optional objective and clickable sidebar controls.
• Conversation — Messenger framing by default: your messages align right, answers align left. Saved alternative styles remain effective.
• Agents sidebar — visible by default on wide terminals, with worker activity, plan and findings. Hide it without replacing the conversation.

Type a message and press Enter to send it. The engine streams its response token-by-token. Tool calls appear as bordered cards showing the command or edit, output, and exit code (controlled by richToolCards setting). The transcript auto-scrolls to newest content. PageUp / PageDown (or Ctrl+Up / Ctrl+Down) scrolls through history.

The context meter displays unavailable when the runtime does not report a usable context window; it is not a turn-budget percentage. Hosted sessions show the Cloud account’s reported credit state separately from estimated model cost.

Use /copy (aliases /export and /dump) while idle to export the complete public conversation, not just visible transcript rows. It saves private local JSON even when copying fails. An OSC52 notice means the content was sent to the terminal clipboard; it does not verify clipboard contents.

Use /impact <finding-id> to discuss a persisted finding in the current chat. Without an ID, select a finding from the conversation. Its prompt asks the model to distinguish observed impact from conditional chains and missing evidence, and not to execute tools or expand authorization. This is an analysis workflow, not a new tool-permission boundary; the session’s actual mode and gates still apply.

Screens

Screen	Command	Description
Chat	/chat	Main conversation transcript and composer
Launcher	/launcher, /run, /home	Engagement control pane (start new scans, browse sessions)
Operations	/ops, /runs	Active and recent operation status
Doctor	/doctor	Runtime and configuration diagnostics
History	/history	Scan history from the database (completed scans, not chat sessions)
Findings	/findings, /finds	Session finding list with filtering
Finding detail	/finding, /finding-detail	Full detail on one finding
Replay	/replay	Event-level turn replay for a completed scan
Settings	/settings, /config, /prefs	Console display settings (persist across sessions)
Theme	/theme, /themes	Colour theme live preview
Model	/model, /models	Select the current audit’s model, worker-role overrides and single-model policy
Resume	/resume, /sessions	Saved chat-session list browser
Herd	/herd, /workers	Active subagent worker overview
Communications	/comms, /messages	Agent activity and messages
Hackstore	/hackstore, /store, /market, /marketplace	Extension marketplace
Connect	/connect, /login, /auth	Cloud sign-in, API-key and subscription connections
Usage	/usage, /cost, /tokens	Token, cost, and context-window usage for this chat session
Provider	/providers	Opens the same connection pane as /connect
Scope	/scope	Current engagement scope view
Audits	/audits	Switch among independent live audits without stopping their work
Onboarding	/onboard	Reopen guided setup without replacing the current audit
Harness	/harness	Live harness controls, workspace trust and rollback
Keybindings	/keybindings, /keys, /keymap	Inspect or rebind supported keyboard shortcuts
Back	/back	Navigate to the previous screen

Slash commands

Every command is available as /command in the composer. Type / to open the command menu. The readline console supports a subset (noted below).

Command	Aliases	Category	Readline?
/help	/?, /commands	info	✓
/capabilities	/caps	info	—
/status	—	info	✓
/tools	—	info	✓
/agents	—	info	—
/clear	—	session	✓
/new-chat	/new	navigation	—
/audits	—	navigation	—
/onboard	—	navigation	—
/harness	—	navigation	—
/stop	—	session	—
/history	—	session	—
/transcript	/review	session	—
/findings	/finds	session	—
/finding	/finding-detail	session	—
/impact	—	session	—
/copy	/export, /dump	session	—
/replay	—	session	—
/resume	/sessions	session	—
/explain	/eli5	session	—
/mode	—	mode	✓
/model	/models	mode	—
/chat	—	navigation	—
/launcher	/run, /home	navigation	—
/ops	/runs	navigation	—
/herd	/workers	navigation	—
/comms	/messages	navigation	—
/hackstore	/store, /market, /marketplace	navigation	—
/connect	/login, /auth	navigation	—
/usage	/cost, /tokens	navigation	—
/back	—	navigation	—
/scope	—	navigation	—
/exit	/quit	system	✓
/feedback	—	system	—
/settings	/config, /prefs	system	—
/theme	/themes	system	—
/keybindings	/keys, /keymap	system	—
/doctor	—	system	—
/providers	—	system	—

Command palette

Open with Ctrl+P (or Ctrl+K) from any screen. Type to filter commands; each entry shows its title, keybinding or category, and description. Press Enter to run.

The palette is available on every screen. On the home screen it lists workspace commands and navigation destinations; on the chat screen it lists session actions, settings toggles, and screen switches.

Model picker

For BYOK connections, /model opens the priced core and includes the active model even when it is a custom deployment. With an empty query, Tab switches between this curated list and the full catalog. Any nonblank model/provider query searches the full catalog, regardless of the Tab setting. ↑ / ↓ moves the highlight, Enter selects, Ctrl+U clears the query, and Esc clears a query before going back.

Models with the same ID remain separate provider rows. Moving between them shows each provider’s own price and context window. Provider labels describe catalog entries; configured credentials determine runtime routing.

Hosted connections show only the account’s model catalog. Tab does not add BYOK models, and a catalog failure does not substitute an offline list.

The detail pane keeps its height while filtering, so a single BYOK result still shows its price estimate, credential source, and setup guidance. A listed model does not guarantee account access; missing prices remain unknown.

Model, role-model and single-model selections apply to the current audit immediately while idle, or after the current turn finishes. They also carry into the next audit. Conversation history, scope and self-extension remain intact; an in-flight turn is never reconfigured.

If the selected provider is not connected, the selection stays staged for the next audit. Connect that provider, then select the model again to apply it live. These rules also apply when opening the picker through Ctrl+P.

To assign different models to workers:

1. Connect a provider route that serves all intended models through /connect.
2. Open /model. Ctrl+Left / Ctrl+Right changes the target from the parent model to a worker role. Search, highlight a model and press Enter.
3. Ctrl+Backspace removes the selected role’s override and restores parent inheritance. Ctrl+S toggles single-model mode; while it is on, role picks remain stored but are inactive.

The parent selection closes the picker; a role selection keeps it open for more assignments. Ctrl+R reloads an available hosted catalog. Role overrides select models for work that actually runs; they do not start workers themselves. Workers inherit the parent’s provider, key and endpoint; selecting a role’s model does not switch it to another account. A gateway route can serve several vendors’ models through that one transport. See multi-model role routing for a concrete example and precedence. The role map and single-model switch are TUI and embedding-API controls, not general CLI flags or environment variables.

Keyboard shortcuts

These are default shortcuts in the main Chat screen unless otherwise noted. /keybindings shows effective bindings and lets you change supported actions; fixed composer and safety keys are not all rebindable.

Global exit

Shortcut	Context	Action
Ctrl+C (once)	Chat screen	Shows exit confirmation with running subagent count
Ctrl+C (twice)	Chat screen	Quits
Ctrl+C	Any modal/overlay	Exits (declines pending action, releases caller)
q	Screens without composer	Quit (Run.tsx screens: Findings, History, Operations, …)

Navigation and overlays

Shortcut	Action
Ctrl+P / Ctrl+K	Open command palette (all screens)
Ctrl+O	Open transcript review; in worker focus, expand/collapse its tool output
Ctrl+R	Toggle collapsed/expanded tool call detail across the entire transcript
Ctrl+G	Jump to agents
Ctrl+T	Open agent communications
Esc	Clear composer / close overlay / go back / interrupt running turn
Esc (with no overlay or draft)	Stop a running turn, or navigate back

Composer

Shortcut	Action
Enter	Send message / execute command
Shift+Enter	Insert newline
Esc	Cancel draft / close command menu
Up	Recall previous submission (readline history)
Down	Walk history forward / enter subagent list
Ctrl+U	Delete to start of line
Ctrl+W	Delete previous word
Alt+Backspace / Ctrl+Backspace	Delete previous word
Tab	Auto-complete slash command
Ctrl+Y	Pull last queued message back into composer for editing

Transcript

Shortcut	Action
PageUp / Ctrl+Up	Scroll transcript up (half page)
PageDown / Ctrl+Down	Scroll transcript down (half page)
Ctrl+Home	Scroll to transcript start (transcript review only)
Ctrl+End	Scroll to transcript end (transcript review only)

Sidebars and mode

Shortcut	Action
Ctrl+B	Toggle left sidebar (recent chat sessions + findings)
Ctrl+L	Toggle right sidebar (active agents + context strip)
Shift+Tab	Cycle autonomy mode

Approval and picker modals

Shortcut	Action
↑ / ↓	Move selection
Enter	Confirm selection / approve
Esc	Cancel / decline
Space	Toggle option (multi-select)
Backspace	Remove last character (filter/input field)
Type	Filter items (in picker) / Enter text (in free-text field)

Subagent focus

Shortcut	Action
Down (idle composer)	Enter subagent list
↑ / ↓ (in list)	Navigate subagent list
Enter (on agent)	Drill into focused subagent
Esc / Left (focused)	Return from subagent focus
↑ / ↓, PageUp / PageDown	Scroll the retained worker transcript
Ctrl+O (focused)	Expand/collapse commands, output, diffs, and tool details
Type + Enter (focused)	Steer a live worker; follow up with Main when a one-shot worker has finished

Transcript review overlay (Ctrl+O)

Shortcut	Action
PageUp / Ctrl+Up	Scroll up
PageDown / Ctrl+Down	Scroll down
Ctrl+Home	Scroll to start
Ctrl+End	Scroll to end
Ctrl+O / Esc	Close review

Modes, approvals, and scope

The TUI presents the requests below as modal prompts. Per-action tool approval is a Standard-mode gate with a wired callback, not a Co-pilot guarantee. Scope requests, exclusions, Recon restrictions and workspace trust are separate controls. Readline and --print have no interactive approval surface; see their callback limitation.

Scope request (Standard mode)

The engine may propose adding hosts to the session scope:

• “Approve for this session” — the exact hosts are added for this session only. Existing deny rules in the configured scope file still take precedence.
• “Reject” — the tool does not run.

Filesystem access request

A source-audit tool may request access to a local directory:

• “Approve this directory” — grants this subtree for the session only.
• “Decline” — tool does not run.

Nothing is persisted to disk.

Safety gate override

When a source-audit tool is blocked by a safety gate:

• “Enable for this session” — lifts the restriction for the session; scope, exclusions, workspace trust and Standard per-action approval (when wired) still apply.
• “Keep disabled” — tool stays blocked.

Tool approval (Standard mode)

In the Bun TUI, Standard asks before each effectful (non-read-only) tool call:

• “Approve this call” — runs once; the next call asks again.
• “Reject” — the model continues without it.

Read-only calls bypass this gate. Co-pilot and YOLO skip it entirely; a Standard session without an approveTool callback also bypasses it. These exemptions do not remove the separate scope, exclusions, Recon or workspace-trust controls.

Operator question (ask_operator)

The engine may present a structured question (multiple choice, free text, or both). This modal authorizes nothing — Esc resolves null (tool renders as “dismissed”), Enter confirms the collected answer.

Capabilities

The capability registry (/capabilities or /caps) lists every primary surface organised by safety tier:

Tier	Meaning
automatic	Runs without operator confirmation
operator-confirmed	Classified for operator confirmation; actual per-action prompts depend on the mode and wired approval callback, not this label alone
blocked	Disabled for the session (can be lifted per-session)

Categories: engagement, findings, verification, connect, settings, evolution, automation.

Sessions, resume, and non-interactive mode

Session persistence

The TUI saves conversations to ~/.0/console-sessions/<id>.json after turns, including failed turns. Files are owner-only (0600), and the directory is 0700. Metadata includes working directory, model, target, mode, preview, optional summary, timestamp and native-message count.

The payload is the native message array: prompts, replies, tool calls and full tool results are plaintext and are not scrubbed for secrets. It can include source, captured requests, credentials and undisclosed findings. Filesystem permissions are not encryption; review exports and backups before sharing. The redacted conversation-history tools described above are a different view. Readline and --print can load saved context, but do not write this TUI store.

Resume

# Open the session picker
0 console --resume

# Resume a specific session by id (or unique prefix)
0 console --resume a1b2c3d4

# Resume the most recent session
0 console --continue

In the TUI, /resume or /sessions opens the same picker, showing preview text, relative age (12s, 5m, 3h, 2d, 6w), model, and turn count for each saved session.

The browser starts with this project’s sessions. Tab includes all projects without clearing your query. Type or paste to search, Ctrl+U clears the query, and Enter resumes the highlighted conversation directly.

To remove a saved transcript, press Delete twice on the same session; Esc cancels. Typing d searches rather than deleting. A failed deletion leaves the session visible and reports the error. A session that cannot be loaded reports the failure in the browser rather than closing the console.

--continue chooses the most recent saved transcript across projects; it is not the same as the picker’s initial current-project filter. Bare --resume needs the TUI picker; use an explicit ID outside it.

Resume restores conversation context, not running workers or session-only authorization decisions. Supply the intended scope and mode again. Current CLI precedence also differs by front-end: a saved target wins over --target; the TUI accepts --model over the saved model, --print prefers the saved model, and readline uses the supplied/default model. For a different engagement, start a new audit rather than assuming resume flags replace all saved context.

Session management

Command	Action
/clear	Clear the idle audit’s conversation; retain target, scope, mode and prior authorization refusals
/new-chat / /new	Create a separate audit using staged model/connection choices
/audits	Switch live audits without cancelling their workers
/stop audit	Stop the current audit’s work
/stop worker <exact name or id>	Stop an owned worker and its descendants
/resume	Browse saved sessions and pick one to resume
/history	Review scan history from the database

/clear is not /new and is not saved-transcript deletion. Live audits have separate conversation/runtime ownership; selecting another audit does not stop background work. Use /stop deliberately, and /resume for disk history.

Pruning

After TUI writes, pruning keeps the newest 20 unprotected archived sessions across all projects. Active and pending-resume sessions are protected and do not consume that allowance. DEFAULT_PRUNE_KEEP = 20 is a code constant with no environment override. Back up needed evidence before it ages out; corrupt files that cannot be listed are not automatically pruned.

Non-interactive mode (--print)

# Inline prompt
0 console --mode recon --print "Summarise the saved findings" --continue

# Piped prompt — reads from stdin
echo "Summarise the findings" | 0 console --mode recon --print --continue

--print runs one prompt through the engine and exits. Text, tool traces and outcomes can appear on stdout; this is not a JSON-only or answer-only protocol. Combine it with --continue or --resume <id> to load saved conversation context. The YOLO default requires an explicit nonempty scope on this path; the examples choose Recon, which still permits its restricted tool set.

There is no interactive approval prompt in --print. Standard without an approval callback and Co-pilot both bypass the per-action gate; see readline and headless approval limitations before using this path for tasks that may run tools.

Transcript vs replay

The console has two views into past data:

Aspect	Transcript	Replay
Scope	Current session’s conversation turns	Any persisted scan (by scan ID or database)
Content	Operator + model turns, tool calls, outcomes	Event-level turn timeline: stages, tool calls, model output
Access	/transcript (Ctrl+O)	/replay
Data source	Current in-memory conversation; saved native messages can seed a resumed chat	Scan database (--db-path or ~/.0/0.db)
Use case	Review what was discussed and returned by tools in this chat	Inspect the events that a scan actually persisted

The transcript review (Ctrl+O) is a scrollable, virtualised rendering of the current conversation.

The replay screen (/replay) loads a completed scan’s recorded events. Browse scan runs, select one, and step through its events.

Feedback and secrets

Local feedback

/feedback <message> appends a Markdown entry with timestamp, version, model, and mode metadata to ~/.0/feedback.md. This file is yours — never sent anywhere without explicit action.

Staged submission

/feedback submit This scan found an interesting edge case
/feedback send       ← transmits the staged feedback over HTTPS
/feedback cancel     ← clears the staged message without sending

• /feedback submit <message> writes the entry to the local file AND shows a preview of what would be sent (target URL, body, headers with auth redacted, any warnings).
• /feedback send transmits the most recently submitted message to the configured endpoint.
• /feedback cancel clears the staged message.

Transmission

Submission is disabled by any of: ZERO_OFFLINE=1, ZERO_NO_TELEMETRY=1, DO_NOT_TRACK=1. Transmission goes to the URL in ZERO_FEEDBACK_URL, or to the 0cloud feedback endpoint (/api/cli-feedback) when the CLI is authenticated with a compatible configured 0cloud deployment.

The feedback payload body contains: message, timestamp, version, model, mode. The body is capped at 64 KB; request timeout is 5 seconds. Failure to submit never blocks the session.

Automatic problem reports

Problem reporting defaults to automatic. Tool and runtime failures can produce a diagnostic through the same feedback transport, independently of manually staged messages; it does not upload ~/.0/feedback.md. At analytics levels off or usage, the diagnostic is limited to failure categories and runtime metadata. Opting into commands or full allows bounded error messages, stack traces and captured output after redaction. Redaction is not a guarantee that arbitrary engagement data is safe to share; choose the sharing level appropriate for your target and credentials.

Open /feedback → Problem-report preferences to choose off, ask, or automatic. This global preference cannot be overridden by a project. Explicit saved opt-outs and the environment opt-outs above remain effective. ask requires confirmation before sending. Without Cloud authentication or a configured HTTPS endpoint, automatic reports remain local and the console reports submission as unavailable.

Secret scanning

When entering an API key through the TUI’s credential prompt (/connect or /providers), the entered value is stored directly to ~/.0/credentials.json. The store’s redactSecret function produces a display form showing only a prefix and the last 4 characters (e.g. sk-ant-…a4f2) — the full key is never echoed to the transcript.

The feedback system’s scanForSecrets is a separate path that inspects feedback messages for credential patterns before preview display. This does not affect provider credential storage.

Storable providers (API-key auth)

API-key connections offered by /connect are stored in ~/.0/credentials.json; the account store also supports provider-specific OAuth records and multiple accounts. ChatGPT Codex is an OAuth connection, not a pasted API-key provider. Explicit nonblank environment credentials take precedence over saved credentials. See API Keys for each provider’s supported methods and the additional Azure settings.

Settings

Display settings are layered: default → global (~/.0/tui-settings.json) → project (.0/tui-settings.json). Use /settings in the TUI to toggle them.

The full settings table lives in Configuration. Key console-specific controls include sidebar visibility (showLeftSidebar, showRightSidebar), transcript density and style, theme, and cost display toggles.

Type to search across groups, or use / before a query beginning with r. Bracketed paste and Unicode backspace work in search. ↑ / ↓ select a setting; ← / → cycle its value in either direction; Enter changes it without leaving the search. Ctrl+U clears the query, and Tab switches groups when no query is active.

The detail pane retains room for the description, current/default values, and a visual preview where one exists. Changed settings carry a dot; the status line reports how many differ from their defaults. Changes save immediately. r resets the selected setting and Shift+R resets all settings, after confirmation. A failed save remains explicitly marked as session-only.

Working feedback and live plans

Working feedback distinguishes connecting, thinking, streaming, tool execution, and waiting for operator input. A failed startup is unavailable. During a turn, entering a follow-up interrupts the current turn and sends the queued message.

Enable Reduce motion in settings for static activity glyphs, logo, and highlights; elapsed time remains visible. Working highlights keep their text stationary and use normal foreground colors rather than failure red.

The right-hand plan sidebar prioritizes active tasks, then pending work, then completed work. Phase labels provide context when space permits; tight layouts favor the task itself. Overflow reports remaining/completed counts, and a fully completed plan collapses to a compact summary.

Monitoring subagents

Open /herd or /workers for the worker overview. s searches worker names, identifiers, tasks, and activity. Enter accepts the search; another Enter opens the selected worker. The header counts agents rather than including group headings in the count.

Worker states have distinct text/glyphs as well as color. Focus view retains live progress and failure details, with ↑ / ↓ scrolling, m for a steering message, and Esc returning to the list. Wide terminals show the overview and detail side by side; narrow terminals use a stacked layout.

Inside chat, /agents opens the retained worker roster without leaving the conversation. Completed and failed workers stay selectable. A focused worker shows its task, assistant replies, tool start/completion state, and final answer. Ctrl+O or a tool card’s disclosure expands the retained output rather than another shortened preview. Execution-level truncation limits still apply.

Messages to a live or parked worker use its mailbox. A follow-up to a finished one-shot worker returns to Main with the worker’s result quoted as untrusted context, instead of disappearing into a dead mailbox. Persistent workers send their result back to the parent; the parent consumes queued results at its next model-request boundary.

The bottom status area distinguishes measured context from the turn spend budget. Worker focus shows that worker’s reported model, input/output/cache tokens, and duration; missing measurements remain unknown. Main shows worker counts, plan progress, queued input, and measured context alongside the existing model, mode, directory, Git, and enabled usage indicators.

Click PLAN in the right sidebar to expand and scroll every task, including full wrapped descriptions; click again to collapse it. The unabridged plan stays in the main transcript when the sidebar is hidden.

TUI crash handling

If the TUI crashes, a crash panel shows the message and a short stack. Options:

Action	Key
Restart	R (on the options panel)
Open crash feedback	F
Quit	Q

Crash text is sanitised — credential-shaped substrings are redacted before they reach the panel or any feedback file. The crash panel’s feedback composer works exactly like /feedback: local file by default, opt-in HTTPS transmission.

Related

• Commands reference — all CLI flags across every command
• Configuration — runtime, mode, and feature settings
• API Keys — provider setup
• Scope & Authorization — scope file format and matching
• Getting Started — install and first scan