ruflo's cockpit inside Claude Code: the /ruflo command, its pages, and a band above the prompt. It reads ruflo's own files; anything it cannot measure reads n/a. Design notes: ADR-407 (cockpit), ADR-448 (the Room), ADR-446 (plugin mods).

Drive the console from the command line

scripts/drive.sh runs the console through its own model tools (console_open, console_state, console_set, console_run) from a real headless Claude, prints what each tool answered, and can assert on it, so a dev loop or CI job can check that a UI shows something.

RUFLO_E2E_LIVE=1 bash plugins/ruflo-console/scripts/drive.sh \
  --expect 'Waiting for a yes' \
  "$PWD/plugins/ruflo-console" read "Open the Room and quote back what is waiting for a yes."

Output is one CALL <tool> <input> and one RESULT <tool> <text> line per console call (each result cut at 4000 characters), then COST <usd> CALLS <n>, then one EXPECT ok|FAIL /regex/ line per --expect.

Exit	Meaning
0	a console call ran and every `--expect` matched (or the run was skipped, see below)
1	no console tool call ran
2	usage error: bad level or an invalid `--expect` regex
3	at least one `--expect` regex (case-insensitive, repeatable) matched no `RESULT`
4	`CONSOLE_DRIVE_INSTALL` was set and the marketplace add or install failed

What it spends: about $0.10 per run with haiku, hard-capped at $0.40 (--max-budget-usd). Use one to three runs, not a sweep.
Isolated config: the run uses a throwaway CLAUDE_CONFIG_DIR seeded with the level you ask for and confirm mode auto (your login is copied 0600 and shredded on exit), because RUFLO_CONSOLE_CONTROL can only lower the saved setting (ADR-450 T12) and your own saved mode would otherwise decide what a drive does.
Why the level is capped: the level is read, write or manage; full is rejected (exit 2). The run happens in a scratch project (mktemp -d), with Bash, Write and Edit disallowed, so a drive can never spend money or delete anything.
Seeding: CONSOLE_DRIVE_SEED=<dir> copies that directory into the scratch .claude-flow/ first, for example a claims/claims.json to assert the claims view shows it.
Skips cleanly: without RUFLO_E2E_LIVE=1 or a claude binary (on PATH or $CLAUDE_BIN) it prints SKIP and exits 0, so it is safe to call from CI.
Extra plugins: any further arguments are added as --plugin-dir. That loads a plugin but does not install it, so claude plugin configure cannot see it.
Installed plugin (for option rows): CONSOLE_DRIVE_INSTALL=<marketplace-dir>:<plugin>@<marketplace> registers that directory as a marketplace and installs the plugin (user scope) inside the throwaway config only, then runs the drive; the Settings view can then read the plugin's options. Your real ~/.claude/plugins and saved console settings are never touched, and claude plugin update is never run. Prints INSTALLED <id>; a bad value exits 2, a failed install exits 4.

CONSOLE_DRIVE_INSTALL="$PWD:ruflo-mods@ruflo" RUFLO_E2E_LIVE=1 bash plugins/ruflo-console/scripts/drive.sh \
  --expect 'Hide unused agent types' "$PWD/plugins/ruflo-console" read "Open view settings with chip mods, then console_state, and quote the ruflo-mods option rows."

Your project's ADRs (ADR-480)

TOOLS → ADRs manages the Architecture Decision Records of the project you run Claude Code in (not ruflo's own).

Finding them. The console looks for docs/adr, docs/adrs, doc/adr, adr, docs/architecture/decisions, docs/decisions and architecture/decisions, or the folder in Settings → ADR folder. A folder that is a link is never read. With none, initialise ADRs here asks first, then creates the folder and 0001-record-architecture-decisions.md; nothing is overwritten.
Reading them. MADR front matter, Nygard / adr-tools ## Status sections (Superseded by [5. …](0005-….md)), ruflo-style Status: lines, log4brains and plain markdown with no status all parse; a file that does not parse still lists. Filter by status, words and scope; the detail shows links both ways, the decision, and the missions it is attached to; the health strip is the lint.
Writing them. propose takes the next number and the style your ADRs already use (override in Settings: ADR style, ADR file name pattern). Accept, reject, deprecate or supersede from a record's detail. Each shows the exact file and the diff, writes only that on your Yes, and leaves a file that changed since alone.
Missions, loops, swarms. Attach ADRs to the active mission (the page suggests some from the goal's words; you decide). Claude's mission context, each task's instruction and spawned swarm agents (ruflo-swarm 0.3.5) are told the accepted decisions, masked and capped. At verify, changed files are compared with the paths the attached accepted ADRs name: a warning in the mission record. That compares paths only. It does not prove a change follows or breaks a decision, and it never blocks. When a mission is done, draft an ADR from this mission prepares a proposed record from its goal and tasks; the decision is yours to write.
For Claude. /ruflo run adr-propose <title>, adr-accept <n>, adr-supersede <old> <new>, adr-attach <n> and the rest are palette entries, so Claude reaches them through console_run at the write control level (reading the page is read); a change still waits for your Yes on the diff.

Toasts (ADR-477)

Settings → Interface and updates → Toasts sets what the ruflo plugins may show over the transcript: all, important (warnings and errors) or off, with a mute chip each for console, swarm, protector and mods. It is kept in the console's store and mirrored to .claude-flow/console/toast-prefs.json, which the other plugins read. Every toast, drawn or not, becomes an event on the Events page (toast <source> <glyph> <text> [off|muted|deduped|…]); the other plugins' digests are read from .claude-flow/console/toasts/<source>.jsonl, masked and capped. The console itself also toasts a mission that finishes (ok) or loses a task (error), and its update notes by level. Design and contract: ADR-477.

Room and Mods

Open it with /ruflo room (menu: Safety → The Room).

Waiting for a yes: the one confirm that is pending, with the seconds left to answer it.
Needs you, Sessions, Preview (ADR-486): approvals, questions, failures and finished work across Claude Code, Codex and Ruflo sessions, including ones started outside Ruflo; the sessions grouped by repository and worktree; and a preview of the one you pick (latest response, current tool, edited files, test result, cost labelled reported, estimated or unavailable). Discovery is local and read-only: no model call, no process, nothing woken (the one write is a four-count summary ruflo-mods can show). The Sessions list starts folded; jumping to a session from the queue opens it. A finished session leaves the queue only after its preview was drawn; a session whose identity is ambiguous goes to an unassigned bucket and raises nothing. Transcript text is never saved, never sent to a model and never included in a text answer. Options sessionWorkspace and sessionPreview (both on).
The feed: events, Claude's console actions and what you said, newest first. Filter by who, search, pause, page back. ⛔ blocked shows only what was refused or failed: Claude's actions the control log marked denied or error, and events that say denied. Press a line to open it; an event that came from a page of its own (swarm, claims, learning, plugins, missions) offers a jump button to that page.
Mods: one line per plugin mod that has written .claude-flow/<name>-mod/status.json. Press a line for its detail: what it guards (the file's own summary), guard, calls, blocked, the class of its last refusal (secret, destructive, path, network, policy, other; the refused text is never kept), modVersion, session start, last write, file age, and a stale marker when the last write was an earlier session.

A status file is data, not instructions: it is size-capped, shape-checked (version: 1 only), and every string is stripped of control and bidi characters and cut to length before it is drawn. summary, modVersion and lastDenied are optional; a mod that does not write them shows "not reported".

Project Anatole in Security & Doctor

An optional Project Anatole section on the Security & Doctor page (key u) lists, runs and edits the ruflo-protector mod (ADR-453). Without the plugin it is one line: claude plugin install ruflo-protector@ruflo.

Reads .claude-flow/protector-mod/status.json, rules.json and the last 200 lines of alerts.jsonl through the bounded, regular-file-only reader (a file over 64 KB is refused; every field is whitelisted and cleaned). It is labelled reported by the mod, unauthenticated: any process can write those files.
Shows the mode, the baseline's maturity ("learning 62%"), open alerts by severity, blocked and degraded; one row per rule (OWASP refs, severity, an off · notify · block chip, hits and acked share over the last 200 alerts, "changed from default"); the open alerts with ack and allow. The Findings meter counts open alerts, labelled as Anatole's.
Runs /protector run and /protector replay into the page's Result panel.
Edits the mode (off, learn, notify, enforce) and per-rule modes through /protector; every change asks first and its confirm row starts with Effect:. enforce is declared an install-class action and reset-baseline a delete-class one, so Claude's console tools always wait for you on both (palette ids anatole-mode, anatole-rule, anatole-ack, anatole-allow, anatole-reset, anatole-run, anatole-replay).