Muxray

Inspect tmux panes as JSON: snapshot, diff, and classify Claude/Codex/Copilot agent state across the whole fleet.

muxray turns a live tmux pane into deterministic JSON so you can read what a program is doing without scraping raw terminal bytes. Reach for it when you are supervising one or more interactive CLIs in tmux — especially terminal coding agents (Claude Code, Codex, Copilot) — and you need to know what state a pane is in, what changed, or whose turn it is.

• Output is JSON on stdout; errors are JSON on stderr.
• It runs locally: no network egress. Pane content (which may hold secrets) never leaves the machine. The only exception is the explicit muxray update.
• It is read-only: muxray observes panes. It never sends keystrokes, runs pane content, or mutates sessions. To send input, use the tmux skill.

When to use it

• You manage panes running Claude/Codex/Copilot and need to know which one is running, needs_approval, waiting_for_input, error, or completed.
• You want to detect and show what changed in a pane since a prior capture.
• You want a one-glance fleet view of every pane's program + state.
• You need to block until a pane finishes working before assigning the next task.

When NOT to use it

• To send input / approve a prompt / type a command → use the tmux skill (tmux send-keys). muxray only reads.
• For a one-shot non-interactive command → run it directly in the shell.
• To read a pane the user did not ask about. Only inspect panes relevant to the task; do not sweep unrelated sessions for its own sake.
• For raw scrollback you intend to grep yourself → tmux capture-pane -p is simpler. muxray's value is the classification and diff, not raw bytes.

Safe invocation

Prefer the bundled wrapper for any call whose pane/session target comes from context you do not fully control — it validates the target and restricts you to the fixed read-only subcommands, so no untrusted text reaches a shell:

bash

{baseDir}/scripts/muxray-run.sh status --pane work:1.0
{baseDir}/scripts/muxray-run.sh scan --text

For a fully-supervised, literal target you may call muxray directly. Either way: fixed subcommand, explicit --pane/flags, never an interpolated shell string. muxray takes no shell input and parses no expressions.

Core commands

Command	What you get
muxray list	every tmux session/window/pane, structured
muxray status --pane <t>	the program + state classified for that pane
muxray scan	every pane classified in one call (the fleet view)
muxray watch --pane <t> [--until <states>]	block until the pane settles
muxray snapshot --pane <t> [--out <f>]	capture the pane (stored locally)
muxray diff --pane <t> [--since <f|id>]	what changed vs a previous snapshot
muxray inspect --pane <t>	snapshot + diff + status in one call
muxray doctor	environment check (tmux present, store writable)

--text gives a terse human line instead of JSON. muxray <cmd> -h lists a command's flags.

Pane targets (--pane): session · session:window.pane (work:1.0) · pane id (%3) · session id ($0) · omitted = the current pane when inside tmux. --session <name> is a clearer equivalent for whole-session targeting.

The JSON contract

Every result carries an envelope: schema_version (currently "2"), command, muxray_version, generated_at. Branch on schema_version — if it is not the version you coded against, the shape may have changed.

status and inspect carry a classification:

json

{ "program": "codex", "status": "running", "rule_id": "codex.running",
  "confidence": 0.88, "evidence": "Working (esc to interrupt)" }

• program — claude, codex, copilot, shell (an interactive shell prompt — the harness is not live; reported as idle), or unknown (a pane it does not recognize: editor, pager, transcript, muxray's own output).
• status — one of: idle, running, blocked, waiting_for_input, needs_approval, error, completed, unknown.
• muxray reports state from the current live frame's footer only. A pane that merely mentions an agent in scrolled content is program=unknown — "parse it yourself," not "something failed." A footer that is a shell prompt is program=shell/status=idle: a dropped connection or exited agent is not an agent error.
• Pass --explain to attach a trace of every rule considered — use it to diagnose an unknown.

diff carries changed (bool — both true and false are exit 0; change is not an error), summary, added/removed/context line arrays, hunks, and the previous_snapshot/current_snapshot ids.

Snapshot & diff: detect and show change

bash

muxray snapshot --pane work:1.0 --out before.json   # capture a baseline
# ... let the program work ...
muxray diff --pane work:1.0 --since before.json      # what changed

--since also accepts a snapshot id, or is omitted/latest to diff against the most recent stored snapshot of that pane (muxray keeps a local store, so you often don't need --out at all). changed: false is deterministic and reproducible across machines — the hash is over cleaned text only.

Interpreting a diff: read summary first, then added (new output the program produced) vs removed (lines that scrolled off or were replaced). hunks counts distinct change regions. A spinner/elapsed-timer line flipping between captures shows up as a tiny diff — treat a 1-line cosmetic change as "still working," not "made progress."

The control loop

Two verbs are the loop — you don't hand-roll poll+sleep+compare:

1. Wait until it's your turn. muxray watch --pane <t> blocks until the pane stops working, then prints the final classification and exits 0.
   • default --until = any settled state (idle, completed, needs_approval, waiting_for_input, blocked, error); it waits through running and transient unknown.
   • narrow it: --until idle,needs_approval.
   • bound it: --timeout 5m exits 5 if it never settles (the last-seen classification is still emitted).
   • then branch on the final status: needs_approval/waiting_for_input → hand off to a human; error → restart/alert; idle/completed → assign the next task (if program=shell, the pane dropped to a shell — relaunch, don't assign to a live agent).
2. See the whole fleet. muxray scan classifies every pane in one call → { "panes": [ { "target": "%3", "session": …, "classification": {…} } … ] }. target is the pane id (%N) — feed it straight back into status/watch/diff. A pane that can't be read is reported unknown with an error class rather than failing the whole scan.

The wrapper {baseDir}/scripts/muxray-watch-diff.sh runs the canonical "baseline → wait until settled → diff → classify" sequence for one pane and prints all three results — use it to summarize a single agent's working turn.

Reading a coding agent's output

To answer "what is this Claude/Codex/Copilot pane doing?":

1. muxray status --pane <t> → the classification is the answer: program names the agent, status names its state, evidence is the footer line that decided it.
2. If you also need what it produced, muxray inspect --pane <t> adds the snapshot and a diff against the last baseline in one call. Read tail_excerpt / the diff added lines for the most recent output.
3. needs_approval / waiting_for_input mean the agent is paused on a human — surface the prompt and stop; do not auto-approve.
4. unknown with a recognizable agent in scrollback usually means the live frame scrolled away or the pane is mid-redraw. Re-status once; if still unknown, fall back to reading tail_excerpt yourself.

Reporting findings to the user

• Lead with the classification, not a wall of terminal text: e.g. "work:1.0 — Codex, needs_approval (asking to run rm -rf build/)."
• For change, summarize the diff.summary + the few added lines that matter; do not paste the whole pane.
• For a fleet, render muxray scan --text (one line per pane) and call out only the panes that need action.

Safety & secrets

• Pane text can contain secrets (tokens, keys, .env echoes). muxray does not redact snapshot.raw/clean, diff lines, or tail_excerpt. You must summarize or redact obvious secrets before showing output to the user or sending it anywhere external. Prefer reporting classification fields over raw pane dumps; pass --no-raw to drop the raw capture from a snapshot.
• muxray performs no network egress except the explicit, opt-in muxray update (downloads a verified release; sends nothing). Do not run update, telemetry, or bundle --include-excerpt as part of an observation loop — they are operator actions, not agent steps.
• Only inspect panes relevant to the user's request.

Exit codes

0 ok (incl. changed:true/false) · 1 internal · 2 usage · 3 tmux/capture · 4 snapshot not found · 5 watch timed out. On failure stderr carries error.class (a stable, branchable id) and error.hint (the next action).

Reference

• references/json-contract.md — compact field/state cheat-sheet.
• examples/inspect-agent.md — a worked end-to-end example.
• muxray usage — the full in-binary calling contract.