Settings reference#
Every control in PacketBench's Settings view, what it defaults to, what it actually changes, and when the change takes effect. Settings is reached from the left navigation; deep links from elsewhere in the app (a "Fix this in Settings" button, for example) open the matching section directly.
How Settings is organised#
Groups on the left, sections as chips, scope pills on the right of the section heading.
The left rail lists six groups. Selecting a group opens its first section;
sections appear as chips under the group heading. A search box above the rail
matches group and section labels, descriptions, and a keyword list — searching
fingerprint finds Remote Hosts, push to talk finds Dictation.
| Group | Sections |
|---|---|
| General | Preferences · Date & Time |
| Workspaces & Terminal | Workspace defaults · CLI Clients · CLI Accounts · Remote Hosts · Project Rules |
| Agents & Models | Agent behavior · Providers & Models |
| Automation | Flights & Autonomy · Task Role Defaults · PacketAgent |
| Integrations & Data | Git Hosts · MCP · Issues · Memory · Dictation · Modules |
| Security & Diagnostics | Trust & Diagnostics |
Scope pills#
Each section shows one or more scope pills next to its heading. They are a promise about reach, and this page uses the same vocabulary:
| Pill | Meaning |
|---|---|
| App | One value for the whole installation. Takes effect immediately. |
| Project | Belongs to the currently open project — usually stored in the project's own files. |
| Workspace | Belongs to the active workspace; other workspaces are unaffected. |
| New sessions | Read when a terminal/PTY session starts. Running sessions keep what they started with. |
| New conversations | Read when an agent conversation starts. Running conversations keep what they froze. |
| New Flights | Read when a Flight is created or launched. Existing Flights keep their policy snapshot. |
Most settings persist to localStorage under packetbench:* keys.
Secrets never do — API keys, git-host tokens, SSH passwords, and the
PacketAgent token all live in the OS credential store. Flight autonomy and
commit-trailer settings persist through the Rust state file.
General#
App-wide appearance, notifications, and keyboard behaviour.
Preferences#
Scope: App. Three cards side by side.
Theme#
| Control | Default | Effect |
|---|---|---|
| Dark / Light | Dark | Switches the whole app's theme tokens immediately. |
Notifications#
Desktop notifications. Turning the master switch on requests OS permission first; if permission is refused the switch stays off. The event toggles only appear while notifications are enabled.
| Control | Default | Effect |
|---|---|---|
| Enable notifications | On | Master switch. |
| Only when app unfocused | On | Suppresses notifications while PacketBench has focus. |
| Approval needed | On | Fires when an agent is waiting on you. |
| Session complete | On | Fires when a session finishes. |
| Session error | On | Fires when a session fails to start or dies. |
| Cost threshold alerts | On | Fires when a budget guardrail warning threshold is crossed. |
Keyboard Shortcuts#
This card owns the global dictation accelerators; everything else in PacketBench's shortcut set is in-app and not editable here.
| Control | Default | Effect |
|---|---|---|
| Global dictation shortcuts | Off | Explicit opt-in. Registers the three accelerators with the OS. In-app controls and Escape work regardless. |
| Push to Talk (hold) | Ctrl+Alt+Space (Cmd on macOS) | Hold to record, release to transcribe. Re-bindable; must include a modifier. |
| Toggle Recording | Ctrl+Alt+R | Start/stop. Re-bindable; must include a modifier. |
| Cancel Recording | Escape | Fixed, in-app only. |
| Open Dictation | Ctrl+Shift+D | Fixed. |
Press the pencil to re-capture a binding; the row swallows keystrokes until you press a valid combination, and Escape cancels. A rotate icon appears when a binding differs from its default. All three accelerators must be distinct, and PacketBench only unregisters what it successfully registered — an existing OS or app binding is reported as a conflict, never seized. The status line under the rows reads off, registering, active, or the conflict message.
Date & Time#
Scope: App.
| Control | Default | Effect |
|---|---|---|
| Time zone | System default | The zone every date and time in PacketBench is rendered in. Stored as an IANA zone name, so daylight saving is applied per timestamp rather than as a fixed offset. |
A zone saved by an older IANA release that this system no longer recognises stays selectable and is called out in amber; dates fall back to the host zone until you pick a valid one.
Dictation analytics — streaks, hourly activity, and the daily and weekly totals — are bucketed by UTC days in the backend and do not follow this setting. Away from UTC they can disagree with timestamps shown elsewhere for entries near midnight. See Dictation & analytics.
Workspaces & Terminal#
Workspace defaults#
Scope: App, Workspace. Three cards.
Project#
| Control | Default | Effect |
|---|---|---|
| Path · Browse | — | Rebinds the active workspace's project folder. To change another workspace's path, switch to it first. |
| Projects Folder · Browse / Clear | Not set | A parent folder whose subdirectories are listed in the sidebar's projects list. |
With no workspace open, the card shows the last-used folder and a Create workspace button instead — workspaces own their project folder.
Workspace Pane#
| Control | Default | Effect |
|---|---|---|
| Default new workspaces to bypass permission prompts | Off | Pre-checks "Bypass permission prompts" in the workspace-creation modal. Existing workspaces are unaffected. |
| Auto-detect GitHub repo on workspace creation | On | Runs git remote get-url origin at creation and links the detected repo. Turn off if you don't want that call made. |
These control the Workspace pane only; they do not change how the Agents pane groups or resumes conversations.
Local terminal shell#
Applies to new or restarted local terminal panes. Coding CLIs launch normally, and remote workspaces always use the host's login shell.
| Control | Default | Effect |
|---|---|---|
| App default | Auto | Auto = PowerShell on Windows, Bash on macOS/Linux. Also offers PowerShell 7, cmd, Git Bash, WSL, and Custom, per platform. |
| Workspace override · name | Use app default | Per-workspace override. Disabled unless an active local workspace is selected. |
| WSL distribution | Default distribution | Only shown for the WSL profile. |
| Shell executable / Startup arguments | — | Only shown for Custom. Must be a supported shell (pwsh, powershell, cmd, bash, zsh, fish, nu, xonsh); Auto remains the effective launch until it is valid. |
Detect re-scans for installed shells. Test shell actually launches the resolved command and reports the executable, version, platform, and working directory. Unavailable profiles are marked "not found" and disabled, with an install link where one exists.
CLI Clients#
Scope: App, New sessions.
A catalog card, one entry per known coding CLI, showing whether each is installed and where it was found. Per entry, depending on state:
- Install — runs the entry's install command in a transient PTY.
- Browse — pins an absolute executable path for this machine, shown as a manual override with a Clear button.
- Install docs — opens the vendor's instructions.
- Entries marked coming soon are listed but not installable.
Selecting the PacketCode entry reveals its integration panel (detection, doctor status, data home, version, provider summary).
Under Advanced — custom CLI agents:
| Control | Effect |
|---|---|
| Custom | Adds a user-defined CLI agent: name, command (claude, codex, or an absolute path), description, and one argument per line. |
| Edit / Delete | Manage an existing custom agent. Delete is confirmed. |
| Reset built-ins | Clears command overrides on the built-in entries. Confirmed. |
The PTY command allowlist is enforced in Rust and covers claude,
codex, opencode, packetcode, ssh, and the shells. A custom agent
pointing at anything else is rejected at launch. A pinned absolute path is
matched by its program name, so D:\tools\packetcode.exe is accepted.
CLI Accounts#
Scope: App, New sessions.
Named logins for Claude Code and Codex, so you can keep more than one
subscription available. Each row is a pointer at a config directory —
CLAUDE_CONFIG_DIR for Claude Code, CODEX_HOME for Codex — which is the whole
mechanism.
| Control | Effect |
|---|---|
| Add / Edit | Label, CLI, and config directory. A ~/.claude-work style suggestion is offered. |
| Log in | Opens a PTY running claude login / codex login with that account's env var set. |
| Delete | Removes the record only. The directory and the login inside it are untouched. |
There are no secrets on this screen. A pane picks an account via its own selector; a pane with no account uses the ambient login.
Remote Hosts#
Scope: App, Workspace. Full detail on SSH remote workspaces.
| Control | Default | Effect |
|---|---|---|
| Name / Host / Port / Username | — / — / 22 / — | Identity of the host record (ServerConfig). |
| Auth method | SSH Agent | agent, key (with a validated key path), or password. |
| Remote path | — | Default project directory offered when creating a workspace here. |
| Verify → Trust | Unpinned | Runs ssh-keyscan, shows every returned key with its SHA-256 fingerprint, and pins the chosen one into ~/.packetbench/ssh/known_hosts. |
| Test connection | — | Authenticates and checks the remote path exists, is a directory, and whether it is a Git repo. |
Saving is gated on verification for a new host or a changed hostname. Password
auth stores the password in the OS credential store under ssh-<id>; the form
only reports whether one exists.
Deleting shows everything currently riding on the host — connections, conversations, Flights, workspaces — then removes the record and its stored password. Nothing on the remote machine is deleted.
A host with no pinned fingerprint falls back to TOFU
(accept-new) for interactive use and is refused outright for async Flight
launches.
Project Rules#
Scope: Project.
One editor that writes both AGENTS.md and CLAUDE.md at the project root,
so the same rules apply whether an agent follows the AGENTS.md convention or
Claude's. The editor seeds from AGENTS.md if present, else CLAUDE.md, else a
starter template.
When the two files have diverged, a warning appears with a Unify affordance to collapse them onto one canonical version. With no project open, the card says so.
Agents & Models#
Agent behavior#
Scope: App, New conversations.
Agents#
| Control | Default | Effect |
|---|---|---|
| Default launch location | Project | Where new conversations start: the selected project path, or a fresh project worktree. The chip in the launch composer overrides it and updates this default; it appears only when launching against a non-SSH project, not in a running conversation. |
| Onboarding | Ready to show | Mark complete hides the Agents-pane onboarding; Show again restores it. |
| Auto-archive done conversations | On | Archives finished conversations after an idle period. |
| After N days | 14 | Idle threshold, 1–365. Disabled when auto-archive is off. |
| Auto-failover on quota or overload | On | Retries an API conversation on a same-provider fallback model after a rate-limit, quota, or overload error. |
Two persisted agent settings have no control on this card: the
global transcript density (summary / normal / verbose, default normal)
is changed from the chat header's overflow menu or by keyboard, and the
worktree cleanup-on-archive policy (default only when safe) has no UI at
all. The hourly auto-archive sweep always keeps worktrees regardless of that
policy, because it cannot prompt.
Agent Profiles#
Reusable launch presets. Three built-ins ship and are re-merged on every app update; they are read-only but cloneable.
| Built-in | System prompt | Tools | Memory brief | Plan mode |
|---|---|---|---|---|
| Default | Full autonomous agent harness | All | On | Off |
| Scout | Read-only investigator | Read-only subset | Per the Scout default | On |
| Reviewer | Diff critic, reports findings only | read_file, list_directory, grep |
Off | On |
A user profile carries: name, description, system prompt, an allowed-tools list (comma-separated; empty means all tools), a memory-brief toggle, a permission mode (Auto, Ask risky, Allow all, Deny risky), plan mode, and an optional pinned model that overrides the launcher's dropdown entirely.
The star marks the default profile. Deleting is confirmed. Changes surface immediately in the composer's profile dropdown.
Providers & Models#
Scope: App, New conversations. Three cards.
API Keys#
Stored in the OS credential store, never in app state or files. The card shows only whether a key exists.
| Provider | Key required? |
|---|---|
| Anthropic — Claude Opus, Sonnet, Haiku | Yes |
| OpenAI — GPT-5.5, GPT-4o, o3 | Yes |
| MiniMax (Token Plan) — M3, M2.5, M2 | Yes |
| OpenRouter — 100+ models, one key | Yes |
| Ollama — local models | No |
| Custom endpoint — OpenAI-compatible | Optional; sent as a Bearer token when set |
Deleting a key is confirmed.
Subscriptions#
claude login / codex login for the two subscription CLIs, with live auth
badges that refresh from the credential-file watcher.
| Row | What it is for |
|---|---|
| Anthropic (Claude subscription) | Terminal Claude Code sessions use your Pro/Max plan. |
| OpenAI (ChatGPT Plus/Pro) | Terminal Codex CLI sessions use your plan. |
These credentials serve PTY / terminal CLI sessions only. No agent row in the Agents picker consumes them — every API row authenticates with an API key. Signing in opens a terminal running the vendor's own login command; signing out clears the on-disk credential file.
Provider Endpoints#
| Control | Default | Effect |
|---|---|---|
| Ollama base URL | http://localhost:11434 |
Chat uses {base}/api/chat (the only route that accepts num_ctx/keep_alive); discovery uses {base}/api/tags. An endpoint without /api/chat falls back to {base}/v1, where the context window cannot be set. |
| Ollama context cap | 16384 tokens | Ceiling on the derived num_ctx. Under-sizing is invisible — Ollama silently drops the oldest messages. |
| Ollama keep-alive | 30m | How long the daemon holds the model loaded. A Go duration string. |
| MiniMax base URL | https://api.minimax.io/v1 |
Mainland-China accounts must switch to https://api.minimaxi.com/v1; a key is valid against only one host. |
| Custom OpenAI-compatible base URL | Unset — the row is disabled | Used verbatim as {base}/chat/completions, so include a /v1 prefix if your server has one. |
| Custom endpoint model list | Empty | One model id per line. No discovery route works across these servers, so the picker offers exactly what you type. |
Each row has a save tick and a reset arrow (reset returns to the built-in default; for the custom row it clears both URL and model list).
Automation#
Flights & Autonomy#
Scope: New Flights. Two cards.
Flights — auto-trailer on agent commits#
| Control | Default | Effect |
|---|---|---|
| Append a trailer to every agent commit | On | Installs a prepare-commit-msg hook inside each Flight worktree so commits identify their originating Flight and attempt. |
| Trailer format | Run-By: PacketBench flight F-{flightId} attempt A-{attemptId} |
Placeholders: {flightId}, {attemptId}, {flightTitle}. A live preview shows the substitution against sample values. |
Writes are queued and serialised; a failed save restores the previous value and says so.
YOLO / bounded autonomy default#
This governs only Flights that explicitly choose "Use Settings default". Reviewer failures, integration conflicts, the final base-branch landing, credentials, and work outside the allowlists still stop for you regardless.
| Control | Default | Effect |
|---|---|---|
| Assisted / YOLO default | Assisted | Which posture a new Flight inherits. |
| Auto-recover failed attempts | On | Retry a failed attempt without asking. |
| Auto-remediate reviewer findings | On | Let the agent address review findings unattended. |
| Auto-run cooperative task graph | On | Advance the task graph without a prompt. |
| Allow unattended in-project tools | Off (approval-gated) | On = allow_in_project tool posture. |
| Allow configured draft-PR publishing | Off | Permits publishing an attempt as a draft PR unattended. |
| Cost $ | 25 | Spend ceiling for the Flight. |
| Minutes | 120 | Wall-clock ceiling. |
| Retries | 2 | Per task. |
| Reviews | 2 | Review rounds. |
| Agents | 3 | Maximum concurrent agents. |
| Allowed project roots | Empty | One absolute path per line. |
| Allowed targets | local |
One per line — local or a server id. |
Changes are applied by Save autonomy default, and the confirmation states that existing Flights keep their current policy snapshot. Selecting YOLO with an invalid policy blocks the save and names the first problem.
Budget guardrails#
The control surface for cost. It deliberately shows no spend figures — no charts, tables, or running totals. Caps are evaluated before an agent or Flight launches, and again on the background poll while work runs.
| Control | Default | Effect |
|---|---|---|
| Daily cap $ | Off (blank) | All providers, today. |
| Monthly cap $ | Off | All providers, this month. |
| Session cap $ | Off | A single conversation. |
| Warn at % | 80 | Notify at this share of a cap. |
| Hard stop at % | 100 | Block launches at this share. Never lower than the warning threshold. |
Blank disables that cap. Reset to defaults clears all three caps and returns the percentages.
Task Role Defaults#
Scope: App, New conversations, New Flights.
AI Provider Routing#
A preferred agent and model per workflow role. Tasks auto-fill from these, and automatic Flight launches (such as GitHub → Draft patch) use the Implementation role.
| Role | Description |
|---|---|
| Implementation | Writing new code |
| Testing | Writing & running tests |
| Code Review | Reviewing code changes |
| Validation | Verifying correctness |
| Research | Exploring solutions |
| Refactoring | Improving existing code |
| Documentation | Writing docs |
Each row picks from CLI agents (marked "not installed" where applicable) or API
executors, then a model for that agent. Default for every role is
claude-code with the system default model. Reset All restores that.
Only API executors can run a Flight attempt, and subscription- login rows are deliberately excluded from this picker — nothing PacketBench routes automatically may resolve to subscription credentials.
Auxiliary AI tasks#
The short single-shot generation tasks PacketBench runs on your behalf. Each row pins a provider and optionally a model, and a Resolves to column shows what the backend would actually use right now — including the error when nothing is configured.
Default for every class is Auto (cheapest configured). These never use a Claude or ChatGPT subscription login.
| Group | Task classes |
|---|---|
| Spec & issues | Spec import · Spec → flight plan · Spec → tickets |
| Code Quality | Explain diagnostic · Summarize checks |
| GitHub | PR description · PR review · Catch-up digest · Issue triage · Issue investigation |
| Memory & flights | Codebase scan · Session summary · Pattern extraction · Flight retrospective |
| Chat | Agent chat · Side chat |
Pinning Ollama requires an explicit model — the backend refuses a model-less Ollama pin, and the model select turns red until you choose one. Reset returns every class to Auto.
PacketAgent#
Scope: App, New Flights.
Endpoint and credentials for deploying a Flight as an always-on worker. PacketAgent owns execution; PacketBench keeps only deployment references and event cursors.
| Control | Default | Effect |
|---|---|---|
| Endpoint | http://127.0.0.1:8484 |
Base URL of the PacketAgent server. |
| PacketAgent workspace ID | — | Which remote workspace handoffs target. |
| Token | Not set | Stored in the OS credential store. Removal is confirmed. |
| Test | — | Probes health, then contract; reports the HTTP status and, when available, the schema match and permitted operations. A healthy but older server still counts as reachable. |
Integrations & Data#
Git Hosts#
Scope: App, Workspace.
| Control | Default | Effect |
|---|---|---|
| GitHub connection | Disconnected | Connect / Reconnect opens the guided setup wizard, which offers browser sign-in (the GitHub device flow, when the build has an OAuth client id) and token paste on one step, validates whichever you use, and replaces the stored credential in place. Disconnect drops it. |
| Default merge strategy | Squash | Merge / Squash / Rebase — the preselected strategy in the PR action bar. |
| Require confirmation for destructive actions | On | Gates merge, close, and convert-to-draft. |
| Default new PRs to draft | Off | Pre-checks "Open as draft" in the PR modal. |
| Publish Flight attempts as draft PRs by default | Off | Pre-checks the equivalent option in the async Flight launcher. |
| Self-hosted Gitea / Forgejo | None | Add a base URL, an optional label, and an access token per host. Removal is confirmed. |
All git-host tokens live in the OS credential store, never in frontend state or workspace records.
MCP#
Scope: App, Project, New conversations. Full detail on MCP hub. Three stacked cards.
Local-first MCP Hub#
| Control | Effect |
|---|---|
| Search | Filters the catalog, configured servers, and capability text together. |
| Curated catalog → Review | Opens a review sheet (source, exact command, file that changes, capabilities, required secrets, network use, removal) before writing anything. Installing writes config only — nothing is executed. |
| Config scope (in the review sheet) | project (default) writes <project>/.mcp.json; global writes ~/.claude/settings.json. |
| Diagnose | Spawns the server, handshakes, calls tools/list, shuts down. Reports connected / degraded / failed with latency and the tool list. Probes stdio only. |
| Read / Write / Network transport | Per-server trust. Read defaults on, Write off, Network on for stdio. Turning Read off also revokes Write. |
| Per-tool chips | Grant or revoke individual tools. A newly diagnosed server pre-grants tools whose names do not look mutating. |
| Reconnect selected | Closes the selected API conversation's backend so its next turn picks up the current trust snapshot. |
Credential, outside-workspace, and protected-publish operations are blocked regardless of these toggles. Every change is written to the local trust audit.
MCP Servers#
| Control | Default | Effect |
|---|---|---|
| Add / Edit | — | Name (immutable after creation), command, space-separated args, scope, and environment pairs. Writes a stdio entry. |
| Delete | — | Confirmed. Removes the entry from the scope's config file. |
| On for agent sessions | All non-disabled servers | Which servers newly started agent conversations begin with. Reset to all returns to the default. |
| Refresh | — | Re-reads both config files. |
MCP Provider#
| Control | Default | Effect |
|---|---|---|
| Enable MCP Provider | Off | Starts a Streamable HTTP MCP server on 127.0.0.1 and shows the URL and a freshly minted bearer token. |
| Port | 3100 | Editable only while stopped. 0 lets the OS choose; the card reports the port actually bound. |
| Allow writes | Off | Off = strictly read-only. On enables append-only coordination notes and confined project-memory writes. Editable only while stopped. |
Live activity (tool calls and resource reads) streams while running and is cleared when the server stops.
The provider's stored config also carries allowedTools and a
scope field that are not sent to the backend and not enforced, and the
card no longer surfaces them.
Issues#
Scope: Project. Three small cards.
| Control | Default | Effect |
|---|---|---|
| Ticket Prefix | PKT | Prefix for generated ticket ids. Upper-cased, max 6 characters. |
| Epics | Empty | Add an epic name; existing ones render as purple chips. |
| Labels | Empty | Add a label; existing ones render as neutral chips. |
Epics and labels are add-only here — there is no remove control on these two cards.
Memory#
Scope: Project, New conversations. Full detail on Memory. The card opens with counts for events, patterns, and the brief source cap, and a button that jumps to the Memory pane. The rotate icon resets every value below.
Capture#
| Control | Default | Effect |
|---|---|---|
| Capture terminal sessions | On | Record PTY sessions longer than 10 seconds when they end, however they end. |
| Capture completed flights | On | Record a Flight when it settles as done. |
Learning#
| Control | Default | Effect |
|---|---|---|
| Summarize sessions on completion | On | Best-effort aux-LLM enrichment after the session is already recorded. |
| Auto-extract learned patterns | On | Run pattern extraction automatically once enough summaries accumulate. |
| Refresh after summaries | 3 | How many new summaries trigger an automatic extraction. Range 1–20. |
Retention#
| Control | Default | Effect |
|---|---|---|
| Expire events by age | Off | Enables age-based pruning. |
| Keep days | 30 | Only meaningful when expiry is on. Range 1–3650. |
| Max stored events | 200 | Oldest beyond the cap are dropped. Range 20–2000. |
| Max learned patterns | 20 | Range 1–100. |
Memory brief budget#
| Control | Default | Effect |
|---|---|---|
| Inject brief into Flight prompts | On | Prepend the composed brief to async Flight prompts at launch. |
| Patterns | 10 | Cap on learned patterns in a brief. Range 0–50. |
| Recent sessions | 5 | Cap on session summaries (last 48 h). Range 0–50. |
| Flight lessons | 5 | Cap on Flight lessons (last 7 days). Range 0–50. |
Project notes contribute up to 5 items, which is fixed and not configurable here.
Project scope#
| Control | Default | Effect |
|---|---|---|
| Match memory by project path | Exact | Exact / Parent directory / Global. Governs filesystem paths only — a remote ssh: scope key always matches by exact identity. |
| Pinned patterns survive cap eviction | On | Off demotes pinned patterns into the same LRU as everything else. |
Dictation#
Scope: App. Full detail on Dictation & analytics.
| Control | Default | Effect |
|---|---|---|
| Whisper Models | small selected |
Download / Verify / Use for tiny, base, small, medium, large-v3. A model counts as Ready only when its checksum and byte length match. |
| Microphone | Default | Stable CPAL device identity. Re-scan and Test buttons alongside. A saved-but-absent device is flagged in amber and recording falls back to the system default. |
| Maximum recording | 5 minutes | Auto-stop and transcribe at this length. Presets 30 s / 1 / 5 / 10 / 30 min; backend clamps to 10 s – 30 min. |
| Language | Auto-detect | Or English, Spanish, French, German, Italian, Portuguese, Japanese, Chinese. |
| Auto-paste after transcription | Off | Enables automatic delivery into the field you were focused on. |
| Paste into other Windows apps | Off, disabled until auto-paste is on | Permits a synthetic Ctrl+V into the foreground app. Re-checked in Rust, and fails closed. |
| Custom Dictionary | Empty | Terms fed to Whisper as an initial prompt. Capped at 100 terms / 1,024 characters. |
Test opens the device for ~1.5 s and reports sample rate, channels, format, peak level, and frames captured — and warns when the probe opened a different device than the one saved.
Modules#
Scope: App.
| Module | Category | Default | Effect when off |
|---|---|---|---|
| Code Quality — lint, type-check, complexity, and test metrics | Analysis | On | Its view is removed from the navigation; if it is open you are returned to Settings. |
| Dictation — voice-to-text with local Whisper | Integration | On | Same. |
Security & Diagnostics#
Trust & Diagnostics#
Scope: App. Five stacked cards.
Workspace/Agents migration evidence#
Local-only, content-free counters for the Workspace/Agents migration — handoff counts and an average attention time. Nothing is uploaded, and no prompts, transcripts, paths, files, diffs, repository URLs, tool arguments, or ids are persisted. Copy puts the evidence on the clipboard; Reset clears it, behind a confirmation.
Release Trust#
Read-only status, not settings:
| Row | Current status |
|---|---|
| Install channel | Manual GitHub Releases |
| Code signing | Not configured for beta builds |
| Auto-updater | Runbook drafted, not enabled |
| Local release gates | lint, build, cargo check, tauri build |
Signing certificates and the Tauri updater are planned release-trust gates, not active guarantees in the current repository. Beta builds are installed manually.
Trust & Provenance#
A bounded local audit of trust decisions — decision metadata only, never transcript or tool output.
| Control | Default | Effect |
|---|---|---|
| Retention | 7 days / 200 events | Or 30 days / 200 events. |
| Source chips | On | Whether provenance chips are shown on evidence in the UI. |
| Copy | — | Copies a redacted audit export to the clipboard. |
| Clear | — | Clears the local audit, behind a confirmation. |
This card is marked data-dictation="off", so dictation will never deliver text
into it.
Crash Reports#
Lists crash files with timestamps; view the contents inline or delete one (confirmed). Refresh re-reads the directory.
Open in dedicated view#
Jump links to two surfaces that used to be Settings tabs:
| Row | Actions |
|---|---|
| History — past terminal sessions and API agent conversations | Open switches to the History view; Preview mounts it inline. |
Prompt Templates — reusable prompts, expanded by typing / in the agent chat |
Edit mounts the inline card, which creates and deletes templates (general / debugging / review / feature / custom) and opens the full manager for in-place editing. There is no dedicated top-level view to jump to. |
Related#
- Memory · Dictation & analytics · MCP hub · SSH remote workspaces
- Agents & conversations — the picker and per-conversation overrides.
- Flight Deck — how the autonomy policy is used at launch.