All pages

Docs · AI features

Terminal intel

Availability: macOS. These are terminalPane-tier tools: portable in principle, unimplemented off macOS today.

TRMNLY keeps a structured record of every command that finishes in a pane — the command text, working directory, exit code, duration, and who ran it (you, the agent, the queue, a runbook, a replay). On top of that record sit a provenance overlay, ambient anomaly detection, and a set of introspection tools the agent uses to reason about your terminal instead of re-running things blindly.

Command provenance

Open the command palette and run Toggle Command Provenance to float a panel over the focused pane listing its recent commands: who produced each one, its exit code, and how long ago it ran. It refreshes live as new commands finish — the fastest answer to "did I type that, or did the agent?"

Anomaly detection

TRMNLY keeps per-command statistics (typical duration, failure history, familiar warning lines) and flags a run that deviates from its own history:

  • a command failing for the first time ever
  • a run sharply slower than usual
  • a warning line it has never printed before

Flags surface as a notification and as labels on the command record, so both you and the agent see them.

Agent tools

Tool Use
shell_state Live state of your shell: cwd, last exit code, foreground process, actual environment
command_provenance Recent commands with who ran each (user / agent / queue / runbook / replay), exits, durations
get_command_errors Structured errors parsed from a failed command's output — kind, message, file:line when the toolchain prints one
command_stats Per-command history: typical duration, failure rate, anomaly flags
pipe_panes Move output from one pane into another — e.g. local build errors into an SSH debug session — with regex/head/tail filtering
session_record Capture a pane's command stream (commands, cwds, exits, output tails) into a portable repro bundle
session_replay Re-run a recorded bundle deterministically, one command per prompt, stopping on error
job_output_digest Rolling digest of a background job's log — reads only what's new since the last check, deduped errors/warnings

Error extraction is conservative: it only parses output on non-zero exits (or on demand), so a grep error listing full of matches isn't mistaken for a build failure.

Repro bundles

session_record turns "it broke on my machine" into a file someone — or some agent — can actually run: the ordered commands with their working directories, exit codes, and output tails. session_replay feeds them back through the command queue with stop-on-error, so a divergence points at the first command that behaved differently.