v0.5.0

0.5.0 (2026-11-12)

Features

• bench: add competitive benchmark harness with gates (1796661)
• cli: load commands from a generated manifest (67c5881)
• core: detect AI agents from one documented table (758a1a5)
• core: resolve credentials in one place (307e2c3)
• help: explain agent mode with dg help agent-mode (6cb08e4)
• init: resume a rerun in a clone dg init recorded (76ebda9)
• init: setup steps need --yes outside a terminal (8c50ae1)
• listen: add --timeout for prerecorded transcription (d695b92)
• listen: add -o text for the transcript alone (943af31)
• listen: print the server's body for prerecorded -o json (6aa1795)
• listen: save large results to a file in agent sessions (fb7f9ec)
• login: let agents finish device login with --complete (82e37e8)
• speak: add --out-file and deprecate -o/--output for the audio file (8f59531)
• telemetry: add DEEPCTL_TELEMETRY=log and =off (6319d5b)
• telemetry: honor DO_NOT_TRACK and CI (06bd209)

Bug Fixes

• Windows runtime failures in speak, listen, update, init and agent-mode help (6a8e5cd)
• ci: drop unwired workflow_call trigger from competitive-bench.yml (53b720f)
• cli: escape characters the output encoding can't write (d168f59)
• cli: keep warnings and notices off stdout (4364a85)
• cli: key cached root help on the telemetry sink setting (58d8b66)
• cli: treat Ctrl-C while the CLI is loading as cancelled (68d79ea)
• cli: use coded messages for cancel and unexpected errors (89c975b)
• config: count ~/.config on macOS when XDG_CONFIG_HOME names it (403f203)
• config: on macOS, follow ~/.config only where platformdirs does (76d1415)
• config: wait for a locked config file once per run (d290452)
• core: send API requests through one transport (edc5ea7)
• debug: send no stored credentials from debug network (edc5243)
• debug: take the probe's and network's host from the transport (3a6cd9b)
• give Windows agents a runnable next_step from init, login and speak (488520a)
• init: clone only Deepgram templates, with plain names (56dc618)
• init: keep an agent's resume records and logins under the real home (6ae7be3)
• init: keep repository settings from running programs (3bea87c)
• listen: finish live streams cleanly on Ctrl-C and end of input (1230585)
• listen: give each OS its own stdin and microphone help (79b5cb9)
• listen: keep machine output to the final result (b8fa0d7)
• listen: report a failed stdout write as an output error (589b2c9)
• listen: stop live listen before capture when the caption reader is gone (24c5deb)
• listen,speak: name Flux models, not API versions (3373da7)
• login: accept real API keys and ask for --yes instead of cancelling (e4db87a)
• login: check the base URL before dg login --complete saves a key (86e74c8)
• login: save the pending login and init's records on Windows (d4625a0)
• login: show a profile's config-file key in profiles --list (b73e1ce)
• mcp: keep environment proxies away from the authenticated client (de4262e)
• mcp: keep stdout for JSON-RPC from the start of dg mcp (d6e09c2)
• output: drop status chrome from stderr in agent mode (cc2de13)
• plugin: refuse a plugin named like a built-in alias before importing it (dfa89fe)
• preserve agent trust compatibility after rebase (c62445a)
• speak: don't read stdin or write audio to a pipe in agent mode (ab6ceb9)
• speak: don't take a Windows pipe or console for another file (7e2dc8b)
• speak: write --out-file and -o files correctly on Windows (3a3cf14)
• update: detect pipx, uv and npm installs correctly (2c97a87)
• update: never wait for the update check (cb4a70f)
• update: report an unreachable index as an error (3014068)
• update: retry an update check that didn't finish after an hour (906bfcc)

Performance Improvements

• cli: cache plugin discovery and help output (48e5f1b)
• cli: render subcommand help from specs (4667ad3)
• cli: start dg without importing click for --version and mcp (f0ef392)
• core: drop the pre-flight key check and reuse one HTTP client (41988be)
• core: import deepctl_core names lazily (9dc5cb4)
• core: load config without pydantic on the help path (0eb4cee)
• listen: bound live-stream memory and frame latency (3d28b98)
• listen: check URL sources without fetching them (bebdbd0)
• listen: stream prerecorded uploads from disk (728e01b)
• output: write JSON off a terminal without rich (d673225)
• speak: stream audio to pipes and players as it arrives (72c243c)
• telemetry: skip the flush when no report was queued (bcc803c)

Documentation

• compare with 0.4.1 in comments, the release before 0.5.0 (5481f6a)
• fix line wrap and restore workflow_call rationale in bench/README.md (259f07d)
• mcp: describe dg mcp as the documentation search it is (c097bec)
• telemetry: publish the report's field schema (19ea267)
• telemetry: say this version sends no telemetry (f8690b4)

Behavior changes

Upgrading to 0.5.0 changes these:

• -o/--output is deprecated for dg speak's audio file; use --out-file instead (8f59531).
• dg init no longer runs a template's setup commands or starts the app without a person at a terminal: pass --yes (or, outside an AI agent session, --install) to run setup and --start to start the app; otherwise it exits 1 with confirmation_required (8c50ae1).
• dg init clones only Deepgram-owned templates now (56dc618).
• dg listen's prerecorded -o json now prints the server's response body as sent — keys the server omitted no longer appear as null — and -o yaml prints that body instead of the table (set DEEPCTL_LISTEN_JSON=sdk to restore the old JSON shape) (6aa1795).
• In an AI-agent session, a prerecorded dg listen result over 64 KiB is now saved under ~/.deepctl/results/ (or a private temp directory) instead of printed, when the output format was auto-detected and stdout isn't a regular file or /dev/null; stdout gets a summary, and --full-output restores the full result (fb7f9ec).
• Under a hard agent signal, dg speak no longer reads stdin unless the text argument is - (otherwise it exits with E-SPEAK-NO-TEXT), and refuses to write audio to a pipe (E-SPEAK-NO-OUTPUT) (ab6ceb9).
• When no one can answer a prompt, dg login's confirmations now return confirmation_required (exit 1) instead of "Login cancelled" (e4db87a).
• A DEEPGRAM_API_KEY, DEEPGRAM_PROFILE, DEEPGRAM_PROJECT_ID or DEEPGRAM_BASE_URL that still holds an unexpanded ${...} placeholder is now ignored, with a W-ENV-PLACEHOLDER warning; off a terminal, a keyring that doesn't answer within 2 s is skipped for the config file or environment, with a W-KEYRING-TIMEOUT warning. Credential precedence itself is unchanged from 0.4.1 (307e2c3).
• dg debug network no longer attaches ~/.netrc credentials to its probe requests (it never sent your Deepgram key) (edc5243).