Shell Pilot 1.20.1

September 5, 2026 · Patch

Sessions spent inside a coding agent or any other interactive tool are now named after the tool. If a session that was nothing but pi ended up titled "Working On Update", this is the fix.

The symptom

Launch the pi agent and work inside it. The shell sees a history of pi, pi update, cle, clear, pi — and the session was titled "Working On Update", then left there. It named neither the tool nor the work. The same happened for claude, aider, codex, for REPLs like python, node and psql, and for ollama.

Root cause

Two independent halves, both in SessionTopicAnalyzer.

First, the headline. The analyzer classifies each command's first token into an activity class. pi has no entry in that map, so every command voted generic — and the generic class's headline verb is the literal fallback "Working On".

Second, the subject. Noun extraction skipped only a command's own grammar, via a table of subcommand words keyed per verb (git, npm, kubectl and friends). There is no entry for pi, so the first argument of pi update was taken as the thing the session is about: "Update".

Underneath both: the commands such a tool runs on your behalf happen inside its own process and never reach the shell hooks. The launch line is all the analyzer will ever see, so the tool's own name is the best description of the session that exists.

The new rules

• Tool sessions are named after the tool. When the dominant activity of the recent commands is the generic fallback and the dominant verb has no known activity, the session takes the tool's name instead of a verb. "Working On" is never emitted for one. 31 tools carry a display name — pi, claude, codex, aider, opencode, gemini, copilot, cursor, goose, amp, ollama, llm, python/python3/ipython, node, deno, bun, irb, psql, mysql, sqlite3, redis-cli, mongosh, tmux, htop/top/btop/btm, lazygit, nvtop. Anything else is title-cased from the command name, so foo update reads "Foo".
• A subject is appended only when it is worth reading — a ticket id, a path-like argument, or the model named by ollama run|pull|show and llm -m. Those are kept verbatim rather than title-cased, because llama3 and health_check stop being searchable otherwise.
• Generic subcommand words are never subjects, under any verb. A shared list of 60 words that read as grammar wherever they appear — update, upgrade, install, init, help, version, run, serve, build, test, list, show, chat, config, status, resume, continue, plan, review and the rest — is skipped when looking for what a command is about, on top of the existing per-verb tables. For a verb the analyzer has no grammar for, leading one- and two-letter arguments are skipped too.
• Ordinary shell work is untouched. Verbs the analyzer already knows keep their tables and their titles: brew install jq still reads "Installing Jq", and git status / git log / git diff still read "Reviewing".
• Naming is deterministic. The dominant activity and dominant verb are now picked by highest count with ties going to the most recent occurrence. The old vote table used Dictionary.max(by:), which has no defined order among equal counts, so an evenly mixed session could be titled differently on two identical runs.

Example titles

• pi, pi update, pi → Pi Agent
• claude --continue → Claude Code
• ollama run llama3 → Ollama · llama3
• aider src/main.py → Aider · main
• python3 scripts/health_check.py → Python · health_check
• foo update (a tool with no display name) → Foo

Tests

321 tests in 26 suites pass, up from 313. Eight new cases in SessionTopicAnalyzerTests cover the exact reported history, a Claude Code session, the Ollama model subject, an unrecognized tool, a path subject, a Python script, and regression guards for the git and brew headlines. Analysis stays pure, off the main actor, and well under 5 ms.

Known limits

• The app is not notarized, so a first install still shows a Gatekeeper prompt.
• The commands an agent runs internally are invisible to the shell — they never reach the hooks that feed the analyzer. The tool's name is therefore the best available title for such a session; it cannot describe work it was never shown.
• Everything else in v1.20.0 is unchanged — this release contains no new features.