Slash commands
Type these at the > prompt during a fred session. They never leave the REPL, never burn tokens, and never count against your balance.
All commands
| Command | What it does |
|---|---|
/checkpoints | List the last 20 shadow-git checkpoints with timestamps. |
/clear | Reset the conversation. Keeps the system prompt and rules; drops chat history. |
/compact [focus] | Manually compact context. Optional focus hint biases what the summary keeps. |
/cost | Print today's spend + current balance from the server. |
/exit | Quit the REPL. Same as /quit. |
/flash | Switch the main slot to deepseek-v4-flash (the cheap default). |
/help | Show the slash-command list. |
/hooks | List active hooks per lifecycle event. |
/last [n] | Reprint the Nth most recent tool result (default 1). |
/mcp | List connected MCP servers + the tool count each one exposes. |
/mode plan|act|toggle | Switch between plan and act mode. Shift+Tab also cycles. |
/model [name] | Show the active model trio, or switch the main slot to <name>. |
/perms | List permission rules (always-allow / always-deny). |
/perms add <rule> <list> | Add a rule to either the allow- or deny-list. |
/perms remove <rule> | Remove a rule from whichever list it's on. |
/pro | Switch the main slot to deepseek-v4-pro. Once-per-account confirmation. |
/quit | Quit the REPL. Same as /exit. |
/repomap | Print the rendered repo map (debug/inspection only). |
/restore <sha> | Restore the workspace to a checkpoint. Requires y/N confirmation. |
/rules | List active + available project rules from .dsc/rules/. |
/sandbox [mode|approval m] | Show or set the sandbox mode and approval mode. |
/save [path] | Dump the transcript to a JSONL file (default: ./transcript-<ts>.jsonl). |
/style concise|explanatory | Set the output style starting next turn. |
/subagents | List subagent profiles (explore, plan, general). |
/think on|off|show | Toggle extended-reasoning display, or reprint the last turn's thinking. |
/todos | Show the current todo list. |
/tools | List registered tools (view, str_replace, glob, grep, bash, todo_write, …). |
/yolo | Toggle --yolo mid-session (skip the permission prompt for every tool call). |
Input shortcuts
| Pattern | What happens |
|---|---|
@path/to/file.py | Embeds the file contents into your message (capped at 4 KB). |
!<shell command> | Runs the command (30s timeout, 4 KB output cap), captures stdout/stderr into the prompt. |
@ and ! work anywhere in your message; @-mentions can repeat. Use them for "here, look at this" without copy-pasting.
/flash and /pro
The two model commands. /flash switches the main slot to deepseek-v4-flash — fast, cheap, what you want for most edits. /pro switches it to deepseek-v4-pro — slower, smarter, billed at the full retail rate for that model.
First time you run /pro on a given account, Fred prints the per-token rates and asks for a one-time confirm:
> /pro
deepseek-v4-pro is billed at full retail (see /cost or /docs/pricing).
Switch the main slot to deepseek-v4-pro? [y/N]Confirm once and Fred remembers — the prompt won't come back. The choice is recorded as warned_about_pro: true in ~/.config/fred/preferences.json. Skip the prompt entirely with --no-confirm-pro at the CLI or FRED_NO_CONFIRM_PRO=1 in the environment.
Both commands persist to preferences.json so the next session starts with the slot you picked.
/mode plan|act|toggle
Switch between plan mode (read-only, propose-only) and act mode (writes and shell allowed, gated by your approval mode).
/mode plan— force plan mode./mode act— force act mode./mode toggle— flip whichever you're in. Same as Shift+Tab.
Plan mode is for "think out loud, sketch the diff, ask before doing anything." Act mode is for "just do it, with the usual permission prompts." The current mode shows in the bottom-right of the input prompt.
Shift+Tab is bound to /mode toggle. Tap it to flip; tap it again to flip back. The keybinding works at the prompt and during streaming.
/sandbox
Sets the sandbox mode (what tools can touch on disk) and the approval mode (when Fred asks for permission).
Sandbox modes:
ro— read-only. Nostr_replace, nobashwrites.workspace-write— writes confined to the cwd subtree. Default.full— no path restriction. Use when working across multiple repos in one session.
Approval modes (gates the permission prompt):
untrusted— prompt for everything. Strictest.on_request— prompt only for tools the model marks as side-effecting. Default.never— never prompt. Same as--yolo.
Usage:
> /sandbox # show current
> /sandbox ro # set sandbox
> /sandbox approval untrusted # set approval modeBoth pieces also have CLI flags (--sandbox, --approval) and env-var equivalents (FRED_APPROVAL_MODE) — see configuration.
/perms
Permission rules let you skip the "run this?" prompt for commands you've already vetted. They're stored in ~/.dsc/settings.local.json.
> /perms # list active rules
> /perms add bash "git status" # always allow
> /perms add bash "rm -rf" deny # always deny (overrides --yolo)
> /perms remove bash "git status" # back to promptingRules are matched on the literal command string with shell-style globbing. Add patterns liberally for reads (git status, ls *, cat *) and sparingly for writes. Deny rules win over allow rules and over --yolo.
/compact
Manual context compaction. When the running context bumps up against the model's window, Fred summarizes the early turns into a single recap message and drops the originals. Most of the time this happens automatically, but you can force it:
> /compact
> /compact "keep the migration plan and the current diff"The optional focus hint nudges the summarizer toward what to retain. The prior compact line in your history is replaced with the new one — your prompt history isn't lost, but the model only sees the summary.
/checkpoints and /restore
Fred records a shadow-git checkpoint of your workspace at the start of every turn that modifies files. /checkpoints lists the last 20 with their tool calls and timestamps:
> /checkpoints
1. f3a9c7e 17:42:01 edit src/auth.py + write tests/test_auth.py
2. 2b18a4d 17:39:55 edit src/auth.py
3. 9c0e1f2 17:31:08 apply migration 0003_add_index.sql
…/restore <sha>rolls the workspace back to that checkpoint. Fred prints the diff that's about to be reversed and asks for a y/N confirm before touching the working tree.
Checkpoints are cheap and local — they live in a hidden git repo under your home, never push anywhere, and rotate after 20.
Other commands at a glance
/style concisevs/style explanatory— output verbosity. Takes effect on the next turn./think on|off|show— show or hide extended-reasoning blocks.showreprints the last turn's thinking even if hidden./save [path]— dump the full transcript as JSONL for sharing or replay./last [n]— reprint the Nth most recent tool result, useful when streaming ate the output you wanted to read./repomap,/rules,/todos,/tools,/hooks,/mcp,/subagents— inspection commands. Show what Fred has loaded./yolo— toggle--yolofor the rest of the session. Deny rules still apply./cost— server-authoritative spend + balance. Includes the legacy rows stampedlegacy-flat-pre-2026-05./clear— reset chat history. AGENTS.md, rules, and tool registrations stay loaded./help,/exit,/quit— what they say on the tin.
For non-slash CLI flags and subcommands, see CLI reference. For the full env-var and config-file reference, see configuration.