Skip to Content
ConceptsThe /nav Pane

The /nav Pane

New in v8.0.0. With the runtime running inside Claude Code it can draw: /nav opens a pane of cards, and a one-line band sits above the prompt. The rule behind both: lead with the next leg, show what surprises you, keep the readouts behind a key.

The /nav pane after a day of work: context, session and reads cards, the judge's tally, the next leg as a button with an ETA, an orange off-route card with park / back / switch, memories, tasks

What you see

 ┌ context ───────────────┐ ┌ session ──────────────────────────┐
 │ 6% ▇░░░░░░░░░          │ │ $0.38   52k · 29% cache           │
 │ compact safe           │ │ 7d $48.15 · 8 commits             │
 │ ⣀⣀⣠⣤⣶⣿                 │ │ ⣀⣀⣤⣶⣿⣿⣷⣄⣀  tokens/min            │
 └────────────────────────┘ └───────────────────────────────────┘
 ┌ judge ────────────────────────────────────────────────────────┐
 │ task · substantial · unclear  → brief shown   ↑ task: jev …   │
 └───────────────────────────────────────────────────────────────┘
 ┌ next ─────────────────────────────────────────────────────────┐
 │ TASK-84   Navigator v8 — mod-first runtime with Python fallback│
 │ [ ● 13/14  Docs: CLAUDE.md "Navigator Runtime" table ]  3 turns│
 │ → then    Dogfood, bump to 8.0.0, tag, release    eta ~25m · 2 legs left
 │ marker    v8-mod-runtime-2026-10-02                           │
 └───────────────────────────────────────────────────────────────┘
 ┌ memories ─────────────────────────────────────────────────────┐
 │ ▸ decision 95%  Complexity model: additive - plan §4 …        │
 │ ▸ pins a memory into your next prompt                         │
 └───────────────────────────────────────────────────────────────┘
 m: marker · c: compact · r: refresh · d: details · j: judge
  • context — fill, one verdict (compact safe / good moment to compact / compact due), and the fill over recent turns as a braille area.
  • session — with the local Prometheus of .agent/grafana/ answering: today’s cost, tokens and cache hit rate, the 7-day cost and commits, tokens/min over two hours. Without it: Claude Code’s own cost and rate-limit window, the phase, the knowledge-graph size. Loopback only, read with curl; dashboard.enabled: false turns it off.
  • judge — only after a judged prompt: the typed judge’s verdict in words, what Navigator did with it (→ brief shown, task mode, loop mode, direct), and the axes where it overrode the keyword rule. j opens the session tally. Since v8.1.0 the tally is followed by the last eight decisions, and y / x confirm or dispute the latest one into ~/.config/navigator/judge-labels.json — the file scripts/judge_label.py and scripts/judge_eval.py --fixture read, so dogfood becomes eval data.
  • next — the destination (the goal of a brief, else the active task), the current leg as a button — n or Enter submits Do the next leg of TASK-84: … as your prompt — turns spent on it, the leg after with an ETA from the pace so far, the newest marker. Legs come from the task doc’s - [ ] checklist, else the numbered steps of its plan section with ### Step n — … ✅ progress headings, else research → impl → verify → complete.
  • off route — only while drifting: two prompts in a row that share no vocabulary with the destination open a warning with park (writes a parked task stub), back, or switch.
  • memories — up to three recalled for the open tasks; ▸ pins one into your next prompt.
  • reads and tasks — behind d: Read calls this session — files read through cat, sed -n, head or grep in Bash count too (use an Agent after three code files in one turn) and the tasks marked in progress — a task doc counts when its **Status** line says 🚧 or “in progress”; a prose mention elsewhere in the doc does not (v8.2.7). Since v8.2.0 the reads card ends with the reject log: 3 rejects over 14:03 stop_completion, in the warning color when today is non-zero.
  • rejects — behind l: the last eight refusals as HH:MM op · reason, read from .agent/.nav-rejects.jsonl after every turn. The file never enters the model’s context; the pane is the only reader besides grep.

The pane follows the session: Prometheus is re-read after every completed turn, and the task list, marker and memories reload after a turn that wrote under .agent/ — through Edit or Write, a Bash heredoc or sed -i, a git mv or commit, or a subagent — and after /clear, /resume or /compact. The next card moves inside the turn: each edit of the active task doc re-reads its checklist. r forces a full reload.

The pane asks for 72 columns and takes what the terminal gives. Below 72 the three top cards stack full-width and the button row wraps, so a narrow split still reads (v8.2.3); the gauge, judge row and next card adapt too (v8.2.4, v8.2.5).

The band

One line above the prompt, three states:

nav · TASK-84 · ● 13/14 Docs: CLAUDE.md "Navigator Runtime" table · 2 legs left
low fuel 72% · compact after this leg · TASK-84 · ● 13/14 …
off route: write sourdough recipe · /nav to park or go back
Hooks as RuntimeTheory of Mind