CLI flags

polyglot [options]                 start the interactive TUI
polyglot init                      interactive first-run setup
polyglot share [id|path] [opts]    export a session transcript to a file
polyglot replay [id|path] [opts]   re-run a saved session against the current build
polyglot report generate [opts]    write a local reliability digest (opt-in, no network call)
polyglot report submit <path>      submit a previously-generated digest
polyglot -p "<prompt>" [options]   run one prompt, print the answer, exit
echo "<prompt>" | polyglot -p      read the prompt from stdin

polyglot init

Runs a short interactive wizard - provider (local or Anthropic), model, and base URL for a local server - and writes the answers to ~/.polyglot/settings.json. Existing keys in that file that the wizard doesn’t ask about are preserved.

It runs automatically the first time you start polyglot in an interactive terminal with no configuration, so a fresh install drops you into setup rather than an error. Run it again any time to re-do the global config. Needs a TTY; in a non-interactive shell it prints the equivalent environment variables instead.

polyglot share

Exports a session as a Markdown (default) or standalone HTML file. Secret-looking values are scrubbed by default. export is an alias.

polyglot share                       # most recent session -> ./polyglot-session-<date>.md
polyglot share <id>                  # a specific session by id
polyglot share ./saved.jsonl         # a session file someone sent you
polyglot share <id> --format html --out review.html
polyglot share <id> --full --no-redact

Flag	Description
--out <path>	Output file. Default ./polyglot-session-<date>.<ext>.
--format <md|html>	md (default) or a self-contained html page.
--no-redact	Keep secret-looking values instead of scrubbing them.
--full	Include complete tool-call arguments and result bodies (default: one-line summaries).

Full guide: Sharing sessions.

polyglot replay

Re-runs a saved session against the current build and reports where it diverges from what the session recorded - which tool calls resolve differently, which repairs are no longer needed, whether a since-released fix would have helped. Useful for pinning down “my model broke on this and I don’t know if it’s the model or polyglot”, and for turning a real failure into a regression test.

polyglot replay                      # most recent session, parse-level
polyglot replay <id>                 # a specific session by id
polyglot replay ./saved.jsonl        # a session file, or a scenario capture .json
polyglot replay <id> --execute --seed ./repo-snapshot
polyglot replay <id> --output-format json

By default it’s parse-level: every recorded model completion is re-resolved through today’s tool-call parser and repair pipeline, with no execution and no working directory - fully deterministic. Exit code is 2 if any call that ran before no longer resolves.

Flag	Description
--execute	Re-run the whole agent loop in a temp working directory instead of just re-resolving tool calls. Reports invariants + a tool-call diff.
--seed <dir>	Seed that working directory from <dir> (needed for --execute to be meaningful - read_file calls error otherwise).
--structured / --free-text	Force the tool-call transport instead of detecting it per turn.
--save <name>	Write the session out as a regression fixture (packages/core/src/testing/regressions/<name>.json when run from the polyglot repo).
--output-format <fmt>	text (default) or json.

polyglot report

Aggregates your local audit log into a reliability digest - repair counts by model and by strategy, parse-error rates - and writes it to a Markdown file. Fully local; generate never makes a network call. A separate, explicit submit step is the only way any of it leaves your machine, and only if you run it. Requires audit.enabled: true (there’s nothing to aggregate otherwise).

polyglot report generate                                # last 7 days, aggregate only
polyglot report generate --days 30
polyglot report generate --include-raw-samples           # interactive review screen, see below
polyglot report submit ./polyglot-reliability-digest-2026-01-15.md

By default the digest is aggregate-only: counts, never argument values or file contents - hashArgs’s own default extended to this file. --include-raw-samples additionally flags repairs from the two lowest-confidence strategies (loose_kv, trailing_blob - the ones most likely to be a genuinely new failure mode worth reporting) and opens an interactive screen that auto-redacts each one (the same secret scanner as tool-output scanning) before you see it, then lets you include or exclude each sample individually - nothing is added to the file without your explicit per-sample choice. Needs a real terminal; from a script or CI, pass --i-have-reviewed-and-approve-raw-samples to include every flagged (already-redacted) sample without the per-item screen, or omit --include-raw-samples entirely for the aggregate-only default.

report submit <path> prints the file’s exact contents again and asks for a final confirmation before sending it - a corpus contribution to help improve the tool-call repair heuristics, not telemetry and not your team’s own HQ data (see Data handling if your team runs one).

Flag	Description
--days <n>	generate: how many days back to aggregate. Default 7.
--include-raw-samples	generate: flag low-confidence repairs for interactive review.
--i-have-reviewed-and-approve-raw-samples	generate: include all flagged samples without the interactive screen - for non-interactive runs only.
--out <path>	generate: output file. Default ./polyglot-reliability-digest-<date>.md.

Flags

Flag	Description
-p, --print	Non-interactive: run a single prompt and exit. See Scripting & CI.
--output-format <fmt>	-p mode: text (default) or json.
--allow-all	-p mode: run every tool without prompting.
--permission-mode <mode>	-p mode: manual, auto, or plan for this run.
--no-persist	Ephemeral session - write nothing to ~/.polyglot/. Works in both interactive and -p mode. Can only tighten the persistTranscripts setting, never loosen it.
--resume [id|path]	Resume the most recent session, one by id, or a .jsonl session file by path. Same transcript restoration as /resume.
--probe	openai-compatible only. Ping the endpoint once on startup to detect its real context window and whether it honors structured output, refreshing ~/.polyglot/capabilities.json. For Ollama, the window comes from its own API (/api/ps, /api/show): the loaded model’s window, else num_ctx. Same as probeCapabilities: true for one run.
-v, --version	Print the version and exit.
-h, --help	Print help and exit.

In -p mode the session id is written to stderr (session: <id>) and included in the JSON envelope (session_id), so runs can be chained with --resume.